Ivan Marsic - Rutgers ECE

Ivan Marsic - Rutgers ECE
Last updated: September 10, 2012
Ivan Marsic
Copyright © 2012 by Ivan Marsic. All rights reserved.
Rutgers University, New Brunswick, New Jersey
Permission to reproduce or copy all or parts of this material for non-profit use is granted on the
condition that the author and source are credited. Suggestions and comments are welcomed.
Author’s address:
Rutgers University
Department of Electrical and Computer Engineering
94 Brett Road
Piscataway, New Jersey 08854
[email protected]
Book website: http://www.ece.rutgers.edu/~marsic/books/SE/
Preface
This book reviews important technologies for software development with a particular focus on
Web applications. In reviewing these technologies I put emphasis on underlying principles and
basic concepts, rather than meticulousness and completeness. In design and documentation, if
conflict arises, clarity should be preferred to precision because, as will be described, the key
problem of software development is having a functioning communication between the involved
human parties. My main goal in writing this book has been to make it useful.
The developer should always keep in mind that software is written for people, not for computers.
Computers just run software—a minor point. It is people who understand, maintain, improve, and
use software to solve problems. Solving a problem by an effective abstraction and representation
is a recurring theme of software engineering. The particular technologies evolve or become
obsolete, but the underlying principles and concepts will likely resurface in new technologies.
Audience
This book is designed for upper-division undergraduate and graduate courses in software
engineering. It intended primarily for learning, rather than reference. I also believe that the book’s
focus on core concepts should be appealing to practitioners who are interested in the “whys”
behind the software engineering tools and techniques that are commonly encountered. I assume
that readers will have some familiarity with programming languages and I do not cover any
programming language in particular. Basic knowledge of discrete mathematics and statistics is
desirable for some advanced topics, particularly in Chapters 3 and 4. Most concepts do not
require mathematical sophistication beyond a first undergraduate course.
Approach and Organization
The first part (Chapters 1–5) is intended to accompany a semester-long hands-on team project in
software engineering. In the spirit of agile methods, the project consists of two iterations. The
first iteration focuses on developing some key functions of the proposed software product. It is
also exploratory to help with sizing the effort and setting realistic goals for the second iteration.
In the second iteration the students perform the necessary adjustments, based on what they have
learned in the first iteration. Appendix G provides a worked example of a full software
engineering project.
The second part (Chapters 6–8 and most Appendices) is intended for a semester-long course on
software engineering of Web applications. It also assumes a hands-on student team project. The
focus is on Web applications and communication between clients and servers. Appendix F briefly
surveys user interface design issues because I feel that proper treatment of this topic requires a
book on its own. I tried to make every chapter self-contained, so that entire chapters can be
i
Ivan Marsic

Rutgers University
ii
skipped if necessary. But you will not benefit the most by reading it that way. I tried to avoid
“botanical” approach, telling you in detail what is here and what is there in software engineering,
so you can start from any point and walk it over in any way. Instead, this book takes an
evolutionary approach, where new topics systematically build on previous topics.
The text follows this outline.
Chapter 2 introduces object-oriented software engineering. It is short enough to be covered in few
weeks, yet it provides sufficient knowledge for students to start working on a first version of their
software product. Appendix G complements the material of Chapter 2 by showing a practical
application of the presented concepts. In general, this knowledge may be sufficient for amateur
software development, on relatively small and non-mission-critical projects.
Chapters 3 through 5 offer more detailed coverage of the topics introduced in Chapter 2. They are
intended to provide the foundation for iterative development of high-quality software products.
Chapters 6 – 8 provide advanced topics which can be covered selectively, if time permits, or in a
follow-up course dedicated to software engineering of Web applications.
This is not a programming text, but several appendices are provided as reference material for
special topics that will inevitably arise in many software projects.
Examples, Code, and Solved Problems
I tried to make this book as practical as possible by using realistic examples and working through
their solutions. I usually find it difficult to bridge the gap between an abstract design and coding.
Hence, I include a great deal of code. The code is in the Java programming language, which
brings me to another point.
Different authors favor different languages and students often complain about having to learn yet
another language on not having learned enough languages. I feel that the issue of selecting a
programming language for a software engineering textbook is artificial. Programming language is
a tool and the software engineer should master a “toolbox” of languages so to be able to choose
the tool that best suits the task at hand.
Every chapter (except for Chapters 1 and 9) is accompanied with a set of problems. Solutions to
most problems can be found on the back of this book, starting on page 523.
Design problems are open-ended, without a unique or “correct” solution, so the reader is welcome
to question all the designs offered in this book. I have myself gone through many versions of each
design, and will probably change them again in the future, as I learn more and think more. At the
least, the designs in this book represent a starting point to critique and improve.
Additional information about team projects and online links to related topics can be found at the
book website: http://www.ece.rutgers.edu/~marsic/books/SE/ .
Contents at a Glance
PREFACE
................................................................................................................................ I
CONTENTS AT A GLANCE ..................................................................................................................... III
TABLE OF CONTENTS ........................................................................................................................... V
CHAPTER 1
INTRODUCTION .......................................................................................................... 1
CHAPTER 2
OBJECT-ORIENTED SOFTWARE ENGINEERING .............................................................. 61
CHAPTER 3
MODELING AND SYSTEM SPECIFICATION ................................................................... 170
CHAPTER 4
SOFTWARE MEASUREMENT AND ESTIMATION ............................................................. 217
CHAPTER 5
DESIGN WITH PATTERNS ......................................................................................... 246
CHAPTER 6
XML AND DATA REPRESENTATION ........................................................................... 319
CHAPTER 7
SOFTWARE COMPONENTS ....................................................................................... 361
CHAPTER 8
WEB SERVICES ..................................................................................................... 374
CHAPTER 9
FUTURE TRENDS.................................................................................................... 410
APPENDIX A
JAVA PROGRAMMING ............................................................................................. 417
APPENDIX B
NETWORK PROGRAMMING ...................................................................................... 419
APPENDIX C
HTTP OVERVIEW ................................................................................................... 433
APPENDIX D
DATABASE-DRIVEN WEB APPLICATIONS ................................................................... 442
APPENDIX E
DOCUMENT OBJECT MODEL (DOM) ......................................................................... 443
APPENDIX F
USER INTERFACE PROGRAMMING ............................................................................. 446
APPENDIX G
EXAMPLE PROJECT: TIC-TAC-TOE GAME .................................................................. 449
APPENDIX H
SOLUTIONS TO SELECTED PROBLEMS ....................................................................... 523
iii
Ivan Marsic
REFERENCES

Rutgers University
iv
........................................................................................................................... 596
ACRONYMS AND ABBREVIATIONS ....................................................................................................... 606
INDEX
........................................................................................................................... 608
Software Engineering 
Contents
v
Table of Contents
PREFACE
................................................................................................................................ I
CONTENTS AT A GLANCE ..................................................................................................................... III
TABLE OF CONTENTS ........................................................................................................................... V
CHAPTER 1
INTRODUCTION .......................................................................................................... 1
1.1
What is Software Engineering? ...................................................................................... 2
1.1.1
Why Software Engineering Is Difficult (1) ................................................................................. 7
1.1.2
Book Organization ................................................................................................................ 8
1.2
Software Engineering Lifecycle ...................................................................................... 8
1.2.1
Symbol Language ............................................................................................................... 11
1.2.2
Requirements Analysis and System Specification .................................................................. 13
1.2.3
Object-Oriented Analysis and the Domain Model.................................................................... 15
1.2.4
Object-Oriented Design ....................................................................................................... 17
1.2.5
Project Effort Estimation and Product Quality Measurement .................................................... 20
1.3
Case Studies .............................................................................................................. 25
1.3.1
Case Study 1: From Home Access Control to Adaptive Homes................................................ 26
1.3.2
Case Study 2: Personal Investment Assistant ........................................................................ 30
1.4
The Object Model ....................................................................................................... 39
1.4.1
Controlling Access to Object Elements .................................................................................. 44
1.4.2
Object Responsibilities and Relationships.............................................................................. 47
1.4.3
Reuse and Extension by Inheritance and Composition ............................................................ 48
1.5
Student Team Projects ................................................................................................ 49
1.5.1
Stock Market Investment Fantasy League ............................................................................. 49
1.5.2
Web-based Stock Forecasters ............................................................................................. 52
1.5.3
Remarks about the Projects ................................................................................................. 54
1.6
Summary and Bibliographical Notes ............................................................................. 57
CHAPTER 2
OBJECT-ORIENTED SOFTWARE ENGINEERING .............................................................. 61
2.1
Software Development Methods ................................................................................... 62
2.1.1
Agile Development .............................................................................................................. 63
Ivan Marsic

2.1.2
2.2
Rutgers University
vi
Decisive Methodological Factors .......................................................................................... 65
Requirements Engineering .......................................................................................... 68
2.2.1
Requirements and User Stories ............................................................................................ 70
2.2.2
Requirements Gathering Strategies ...................................................................................... 77
2.2.3
Effort Estimation ................................................................................................................. 78
2.3
Software Architecture .................................................................................................. 80
2.3.1
Problem Architecture ........................................................................................................... 82
2.3.2
Software Architectural Styles ................................................................................................ 86
2.3.3
Recombination of Subsystems ............................................................................................. 87
2.4
Use Case Modeling ..................................................................................................... 88
2.4.1
Actors, Goals, and Sketchy Use Cases ................................................................................. 88
2.4.2
System Boundary and Subsystems....................................................................................... 94
2.4.3
Detailed Use Case Specification ........................................................................................... 96
2.4.4
Security and Risk Management ...........................................................................................107
2.4.5
Why Software Engineering Is Difficult (2) ..............................................................................108
2.5
Analysis: Building the Domain Model .......................................................................... 109
2.5.1
Identifying Concepts ...........................................................................................................110
2.5.2
Concept Associations and Attributes ....................................................................................113
2.5.3
Domain Analysis ................................................................................................................118
2.5.4
Contracts: Preconditions and Postconditions ........................................................................119
2.6
Design: Assigning Responsibilities ............................................................................. 120
2.6.1
Design Principles for Assigning Responsibilities ....................................................................124
2.6.2
Class Diagram ...................................................................................................................130
2.6.3
Why Software Engineering Is Difficult (3) ..............................................................................133
2.7
Test-driven Implementation........................................................................................ 133
2.7.1
Overview of Software Testing ..............................................................................................134
2.7.2
Test Coverage and Code Coverage .....................................................................................136
2.7.3
Practical Aspects of Unit Testing .........................................................................................140
2.7.4
Integration and Security Testing ..........................................................................................143
2.7.5
Test-driven Implementation .................................................................................................146
2.7.6
Refactoring: Improving the Design of Existing Code ..............................................................151
2.8
Summary and Bibliographical Notes ........................................................................... 152
Problems ............................................................................................................................ 156
CHAPTER 3
MODELING AND SYSTEM SPECIFICATION ................................................................... 170
3.1
What is a System? .................................................................................................... 171
3.1.1
World Phenomena and Their Abstractions ............................................................................172
3.1.2
States and State Variables ..................................................................................................176
3.1.3
Events, Signals, and Messages ...........................................................................................181
Software Engineering 
Contents
vii
3.1.4
Context Diagrams and Domains ..........................................................................................183
3.1.5
Systems and System Descriptions .......................................................................................185
3.2
Notations for System Specification ............................................................................. 186
3.2.1
Basic Formalisms for Specifications .....................................................................................186
3.2.2
UML State Machine Diagrams .............................................................................................193
3.2.3
UML Object Constraint Language (OCL) ..............................................................................196
3.2.4
TLA+ Notation ...................................................................................................................201
3.3
Problem Frames ....................................................................................................... 203
3.3.1
Problem Frame Notation .....................................................................................................204
3.3.2
Problem Decomposition into Frames....................................................................................205
3.3.3
Composition of Problem Frames..........................................................................................208
3.3.4
Models ..............................................................................................................................209
3.4
Specifying Goals ....................................................................................................... 210
3.5
Summary and Bibliographical Notes ........................................................................... 211
Problems ............................................................................................................................ 212
CHAPTER 4
SOFTWARE MEASUREMENT AND ESTIMATION ............................................................. 217
4.1
Fundamentals of Measurement Theory ....................................................................... 218
4.1.1
4.2
Measurement Theory .........................................................................................................219
What to Measure? .................................................................................................... 221
4.2.1
Use Case Points ................................................................................................................222
4.2.2
Cyclomatic Complexity .......................................................................................................231
4.3
Measuring Module Cohesion...................................................................................... 233
4.3.1
Internal Cohesion or Syntactic Cohesion ..............................................................................233
4.3.2
Interface-based Cohesion Metrics........................................................................................235
4.3.3
Cohesion Metrics using Disjoint Sets of Elements .................................................................236
4.3.4
Semantic Cohesion ............................................................................................................237
4.4
Coupling .................................................................................................................. 237
4.5
Psychological Complexity .......................................................................................... 238
4.5.1
4.6
4.6.1
4.7
Algorithmic Information Content ...........................................................................................238
Effort Estimation ....................................................................................................... 240
Deriving Project Duration from Use Case Points....................................................................241
Summary and Bibliographical Notes ........................................................................... 242
Problems ............................................................................................................................ 244
CHAPTER 5
DESIGN WITH PATTERNS ......................................................................................... 246
5.1
Indirect Communication: Publisher-Subscriber ............................................................. 247
5.1.1
Applications of Publisher-Subscriber ....................................................................................254
5.1.2
Control Flow ......................................................................................................................255
Ivan Marsic

5.1.3
5.2
Rutgers University
viii
Pub-Sub Pattern Initialization ..............................................................................................257
More Patterns ........................................................................................................... 257
5.2.1
Command .........................................................................................................................258
5.2.2
Decorator ..........................................................................................................................261
5.2.3
State.................................................................................................................................262
5.2.4
Proxy ................................................................................................................................264
5.3
Concurrent Programming .......................................................................................... 271
5.3.1
Threads ............................................................................................................................272
5.3.2
Exclusive Resource Access—Exclusion Synchronization .......................................................274
5.3.3
Cooperation between Threads—Condition Synchronization ...................................................276
5.3.4
Concurrent Programming Example ......................................................................................277
5.4
Broker and Distributed Computing .............................................................................. 283
5.4.1
Broker Pattern ...................................................................................................................286
5.4.2
Java Remote Method Invocation (RMI).................................................................................288
5.5
Information Security .................................................................................................. 295
5.5.1
Symmetric and Public-Key Cryptosystems ............................................................................297
5.5.2
Cryptographic Algorithms ....................................................................................................298
5.5.3
Authentication....................................................................................................................300
5.5.4
Program Security ...............................................................................................................300
5.6
Summary and Bibliographical Notes ........................................................................... 302
Problems ............................................................................................................................ 305
CHAPTER 6
XML AND DATA REPRESENTATION ........................................................................... 319
6.1
Structure of XML Documents ..................................................................................... 322
6.1.1
Syntax ..............................................................................................................................322
6.1.2
Document Type Definition (DTD) .........................................................................................328
6.1.3
Namespaces .....................................................................................................................332
6.1.4
XML Parsers .....................................................................................................................334
6.2
XML Schemas .......................................................................................................... 336
6.2.1
XML Schema Basics ..........................................................................................................337
6.2.2
Models for Structured Content .............................................................................................342
6.2.3
Datatypes..........................................................................................................................345
6.2.4
Reuse ...............................................................................................................................352
6.2.5
RELAX NG Schema Language............................................................................................352
6.3
Indexing and Linking ................................................................................................. 353
6.3.1
XPointer and Xpath ............................................................................................................353
6.3.2
XLink ................................................................................................................................354
6.4
Document Transformation and XSL ............................................................................ 355
6.5
Summary and Bibliographical Notes ........................................................................... 358
Software Engineering 
Contents
ix
Problems ............................................................................................................................ 359
CHAPTER 7
SOFTWARE COMPONENTS ....................................................................................... 361
7.1
Components, Ports, and Events ................................................................................. 362
7.2
JavaBeans: Interaction with Components .................................................................... 363
7.2.1
Property Access.................................................................................................................364
7.2.2
Event Firing .......................................................................................................................364
7.2.3
Custom Methods................................................................................................................365
7.3
Computational Reflection ........................................................................................... 366
7.3.1
Run-Time Type Identification ...............................................................................................367
7.3.2
Reification .........................................................................................................................368
7.3.3
Automatic Component Binding ............................................................................................369
7.4
State Persistence for Transport .................................................................................. 369
7.5
A Component Framework .......................................................................................... 370
7.5.1
Port Interconnections .........................................................................................................370
7.5.2
Levels of Abstraction ..........................................................................................................372
7.6
Summary and Bibliographical Notes ........................................................................... 373
Problems ............................................................................................................................ 373
CHAPTER 8
WEB SERVICES ..................................................................................................... 374
8.1
Service Oriented Architecture .................................................................................... 376
8.2
SOAP Communication Protocol.................................................................................. 377
8.2.1
The SOAP Message Format ...............................................................................................378
8.2.2
The SOAP Section 5 Encoding Rules ..................................................................................383
8.2.3
SOAP Communication Styles ..............................................................................................386
8.2.4
Binding SOAP to a Transport Protocol .................................................................................389
8.3
WSDL for Web Service Description ............................................................................ 390
8.3.1
The WSDL 2.0 Building Blocks ............................................................................................391
8.3.2
Defining a Web Service’s Abstract Interface .........................................................................394
8.3.3
Binding a Web Service Implementation ................................................................................396
8.3.4
Using WSDL to Generate SOAP Binding ..............................................................................397
8.3.5
Non-functional Descriptions and Beyond WSDL ....................................................................398
8.4
UDDI for Service Discovery and Integration ................................................................. 399
8.5
Developing Web Services with Axis ............................................................................ 400
8.5.1
Server-side Development with Axis ......................................................................................400
8.5.2
Client-side Development with Axis .......................................................................................406
8.6
OMG Reusable Asset Specification ............................................................................ 407
8.7
Summary and Bibliographical Notes ........................................................................... 408
Problems ............................................................................................................................ 409
Ivan Marsic

Rutgers University
x
CHAPTER 9
FUTURE TRENDS.................................................................................................... 410
9.1
Aspect-Oriented Programming ................................................................................... 411
9.2
OMG MDA ............................................................................................................... 412
9.3
Autonomic Computing ............................................................................................... 412
9.4
Software-as-a-Service (SaaS) .................................................................................... 413
9.5
End User Software Development................................................................................ 413
9.6
The Business of Software .......................................................................................... 416
9.7
Summary and Bibliographical Notes ........................................................................... 416
APPENDIX A
JAVA PROGRAMMING ............................................................................................. 417
A.1
Introduction to Java Programming .............................................................................. 417
A.2
Bibliographical Notes ................................................................................................ 417
APPENDIX B
NETWORK PROGRAMMING ...................................................................................... 419
B.1
Socket APIs ............................................................................................................. 419
B.2
Example Java Client/Server Application ...................................................................... 424
B.3
Example Client/Server Application in C ....................................................................... 427
B.4
Windows Socket Programming .................................................................................. 430
B.5
Bibliographical Notes ................................................................................................ 432
APPENDIX C
HTTP OVERVIEW ................................................................................................... 433
C.1
HTTP Messages ....................................................................................................... 434
C.2
HTTP Message Headers ........................................................................................... 438
C.3
HTTPS—Secure HTTP ............................................................................................. 441
C.4
Bibliographical Notes ................................................................................................ 441
APPENDIX D
DATABASE-DRIVEN WEB APPLICATIONS ................................................................... 442
APPENDIX E
DOCUMENT OBJECT MODEL (DOM) ......................................................................... 443
E.1
Core DOM Interfaces ................................................................................................ 443
E.2
Bibliographical Notes ................................................................................................ 445
APPENDIX F
USER INTERFACE PROGRAMMING ............................................................................. 446
F.1
Model/View/Controller Design Pattern ......................................................................... 446
F.2
UI Design Recommendations ..................................................................................... 446
F.3
Bibliographical Notes ................................................................................................ 447
APPENDIX G
EXAMPLE PROJECT: TIC-TAC-TOE GAME .................................................................. 449
G.1
Customer Statement of Requirements ........................................................................ 450
G.1.1
Problem Statement ............................................................................................................450
Software Engineering 
G.1.2
G.2
Contents
xi
Glossary of Terms ..............................................................................................................451
System Requirements Engineering ............................................................................. 452
G.2.1
Enumerated Functional Requirements .................................................................................452
G.2.2
Enumerated Nonfunctional Requirements.............................................................................457
G.2.3
On-Screen Appearance Requirements .................................................................................457
G.2.4
Acceptance Tests ..............................................................................................................457
G.3
Functional Requirements Specification ....................................................................... 461
G.3.1
Stakeholders .....................................................................................................................461
G.3.2
Actors and Goals ...............................................................................................................461
G.3.3
Use Cases Casual Description ............................................................................................461
G.3.4
Use Cases Fully-Dressed Description ..................................................................................464
G.3.5
Acceptance Tests for Use Cases .........................................................................................470
G.3.6
System Sequence Diagrams ...............................................................................................473
G.4
User Interface Specification ....................................................................................... 477
G.4.1
Preliminary UI Design .........................................................................................................477
G.4.2
User Effort Estimation.........................................................................................................478
G.5
Domain Analysis ....................................................................................................... 479
G.5.1
Domain Model ...................................................................................................................479
G.5.2
System Operation Contracts ...............................................................................................486
G.5.3
Mathematical Model ...........................................................................................................486
G.6
Design of Interaction Diagrams .................................................................................. 488
G.6.1
First Iteration of Design Sequence Diagrams ........................................................................488
G.6.2
Evaluating and Improving the Design ...................................................................................493
G.7
Class Diagram and Interface Specification .................................................................. 498
G.8
Unit Tests and Coverage ........................................................................................... 499
G.8.1
Deriving the Object States ..................................................................................................499
G.8.2
Events and State Transitions ...............................................................................................505
G.8.3
Unit Tests for States ...........................................................................................................506
G.8.4
Unit Tests for Valid Transitions ............................................................................................510
G.9
Refactoring to Design Patterns ................................................................................... 511
G.9.1
Roadmap for Applying Design Patterns ................................................................................511
G.9.2
Remote Proxy Design Pattern .............................................................................................511
G.9.3
Publish-Subscribe Design Pattern........................................................................................513
G.9.4
Command Design Pattern ...................................................................................................513
G.9.5
Decorator Design Pattern....................................................................................................513
G.9.6
State Design Pattern ..........................................................................................................514
G.9.7
Model-View-Controller (MVC) Design Pattern .......................................................................520
G.10 Concurrency and Multithreading ................................................................................. 521
Ivan Marsic

Rutgers University
xii
APPENDIX H
SOLUTIONS TO SELECTED PROBLEMS ....................................................................... 523
REFERENCES
........................................................................................................................... 596
ACRONYMS AND ABBREVIATIONS ....................................................................................................... 606
INDEX
........................................................................................................................... 608
Chapter 1
Introduction
Contents
“There is nothing new under the sun but there are lots of old
things we don’t know.”
—Ambrose Bierce, The Devil’s Dictionary
Software engineering is a discipline for solving business
problems by designing and developing software-based
systems. As with any engineering activity, a software engineer
starts with problem definition and applies tools of the trade to
obtain a problem solution. However, unlike any other
engineering, software engineering seems to require great
emphasis on methodology or method for managing the
development process, in addition to great skill with tools and
techniques. Experts justify this with the peculiar nature of the
problems solved by software engineering. These “wicked
problems” can be properly defined only after being solved.
This chapter first discusses what software engineering is about
and why it is difficult. Then we give a brief preview of
software development. Next, cases studies are introduced that
will be used throughout the book to illustrate the theoretical
concepts and tools. Software object model forms the
foundation for concepts and techniques of modern software
engineering. Finally, the chapter ends by discussing hands-on
projects designed for student teams.
1
1.1 What is Software Engineering?
1.1.1
1.1.2 Why Software Engineering Is Difficult (1)
1.1.3 Book Organization
1.2 Software Engineering Lifecycle
1.2.1 Symbol Language
1.2.2 Requirements Analysis and System
Specification
1.2.3 Object-Oriented Analysis and the Domain
Model
1.2.4 Object-Oriented Design
1.2.5 Project Effort Estimation and Product
Quality Measurement
1.3 Case Studies
1.3.1 Case Study 1: From Home Access Control
to Adaptive Homes
1.3.2 Case Study 2: Personal Investment
1.4 The Object Model
1.4.1 Controlling Access to Object Elements
1.4.3 Reuse and Extension by Inheritance
and Composition
1.4.2 Object Responsibilities
1.4.4 x
1.5 Student Team Projects
1.5.1 Stock Market Investment Fantasy League
1.5.2 Web-based Stock Forecasters
1.5.3 Remarks about the Projects
1.6 Summary and Bibliographical Notes
Ivan Marsic

Rutgers University
2
1.1 What is Software Engineering?
“To the optimist, the glass is half full. To the pessimist, the glass is half empty. To the engineer, the glass
is twice as big as it needs to be.” —Anonymous
“Computer science is no more about computers than astronomy is about telescopes.” —Edsger W. Dijkstra
The purpose of software engineering is to develop software-based systems that let customers
achieve business goals. The customer may be a hospital manager who needs patient-record
software to be used by secretaries in doctors’ offices; or, a manufacturing manager who needs
software to coordinate multiple parallel production activities that feed into a final assembly stage.
Software engineer must understand the customer’s business needs and design software to help
meet them. This task requires

The ability to quickly learn new and diverse disciplines and business processes

The ability to communicate with domain experts, extract an abstract model of the
problem from a stream of information provided in discipline-specific jargon, and
formulate a solution that makes sense in the context of customer’s business

The ability to design a software system that will realize the proposed solution and
gracefully evolve with the evolving business needs for many years in the future.
Software engineering is often confused with programming. Software engineering is the creative
activity of understanding the business problem, coming up with an idea for solution, and
designing the “blueprints” of the solution. Programming is the craft of implementing the given
blueprints (Figure 1-1). Software engineer’s focus is on understanding the interaction between
the system-to-be and its users and the environment, and designing the software-to-be based on
this understanding. Unlike this, programmer’s focus is on the program code and ensuring that the
code faithfully implements the given design. This is not a one-way process, because sometimes
Chapter 1 
3
Introduction
Customer:
Requires a computer system to achieve some business goals
by user interaction or interaction with the environment
in a specified manner
System-to-be
User
Software-to-be
Environment
Software Engineer’s task:
To understand how the system-to-be needs to interact with
the user or the environment so that customer’s requirement is met
and design the software-to-be
May be the
same person
Programmer’s task:
To implement the software-to-be
designed by the software engineer
Figure 1-1: The role for software engineering.
the designs provided by the “artist” (software engineer) cannot be “carved” in “marble”
(programming infrastructure) as given, and the “craftsman” (programmer) needs to work closely
with the designer to find a workable solution. In an ideal world, both activities would be done by
the same person to ensure the best result; in reality, given their different nature and demands,
software engineering and programming are often done by different people.
Some people say software engineering is about writing loads of documentation. Other people say
software engineering is about writing a running code. It is neither one. Software engineering is
about understanding business problems, inventing solutions, evaluating alternatives, and making
design tradeoffs and choices. It is helpful to document the process (not only the final solution) to
know what alternatives were considered and why particular choices were made. But software
engineering is not about writing documentation. Software engineering is about delivering value
for the customer, and both code and documentation are valuable.
Ivan Marsic

Setting posts
4
Rutgers University
Cutting wood
Nailing
Painting
Figure 1-2: Illustration of complexity on the problem of scheduling construction tasks.
I hope to convey in this text that software is many parts, each of which individually may be easy,
but the problem is that there are too may of them. It is not the difficulty of individual
components; it is the multitude that overwhelms you—you simply lose track of bits and pieces.
Let me illustrate this point on a simple example. Suppose one wants to construct a fence around a
house. The construction involves four tasks: setting posts, cutting wood, painting, and nailing
(Figure 1-2). Setting posts must precede painting and nailing, and cutting must precede nailing.
Suppose that setting posts takes 3 units of time, cutting wood takes 2 units of time, painting takes
5 units of time for uncut wood and 4 units of time otherwise, and nailing takes 2 units of time for
unpainted wood and 3 units of time otherwise. In what order should these tasks be carried out to
complete the project in the shortest possible time?
It is difficult to come up with a correct solution (or, solutions) without writing down possible
options and considering them one by one. It is hard to say why this problem is complicated,
because no individual step seems to be difficult. After all, the most complicated operation
involves adding small integer numbers. Software engineering is full of problems like this: all
individual steps are easy, yet the overall problem may be overwhelming.
Mistakes may occur both in understanding the problem or implementing the solution. The
problem is, for discrete logic, closeness to being correct is not acceptable; one flipped bit can
change the entire sense of a program. Software developers have not yet found adequate methods
to handle such complexity, and this text is mostly dedicated to present the current state of the
knowledge of handling the complexity of software development.
Chapter 1 
5
Introduction
7
4
1
0
8
5
2
6
9
3
ATM machine
Communication link
Bank’s
remote
datacenter
Bank
customer
Figure 1-3: Example: developing software for an Automatic Teller Machine (ATM).
Software engineering relies on our ability to think about space and time, processes, and
interactions between processes and structures. Consider an example of designing a software
system to operate an automatic banking machine, known as Automatic Teller Machine (ATM)
(Figure 1-3). Most of us do not know what is actually going on inside an ATM box; nonetheless,
we could offer a naïve explanation of how ATM machines work. We know that an ATM machine
allows us to deposit or withdraw money, and we can imagine how to split these activities into
simpler activities to be performed by imaginary little “agents” working inside the machine. Figure
1-4 illustrates how one might imagine what should be inside an ATM to make it behave as it
does. We will call the entities inside the system “concepts” because they are imaginary. As seen,
there are two types of concepts: “workers” and “things.”
We know that an ATM machine plays the role of a bank window clerk (teller). The reader may
wonder why we should imagine many virtual agents doing a single teller’s job. Why not simply
imagine a single virtual agent doing the teller’s job?! The reason that this would not help much is
because all we would accomplish is to transform one complicated and inscrutable object (an
ATM machine) into another complicated and inscrutable object (a virtual teller). To understand a
complex thing, one needs to develop ideas about relationships among the parts inside. By
dividing a complicated job into simpler tasks and describing how they interact, we simplify the
problem and make it easier to understand and solve. This is why imagination is critical for
software engineering (as it is for any other problem-solving activity!).
Of course, it is not enough to uncover the static structure of the system-to-be, as is done in Figure
1-4. We also need to describe how the system elements (“workers” and “things”) interact during
the task accomplishment. Figure 1-5 illustrates the working principle (or operational principle) of
the ATM model from Figure 1-4 by a set of step-by-step interactions.
Ivan Marsic

6
Rutgers University
Domain Model
How may I
help you?
Transaction
record
Cash
Bookkeeper
Speakerphone
Safe
Safe keeper
Phone
Window clerk
Datacenter
liaison
Dispenser
Bank’s
remote
datacenter
Customer
Figure 1-4: Imagined static structure of ATM shows internal components and their roles.
PA
Enter
your PIN
C
B
Verify
account
XYZ
D
Verify
this
account
Typing in
PIN number
…
E
How may
I help
you?
Withdraw
$60
Account
valid.
Balance:
$100
XYZ valid.
Balance:
$100
F
G
Release
$60
Dispense
$60
Record
$60 less
Dispensing!
H
Please take
your cash
Figure 1-5: Dynamic interactions of the imagined components during task accomplishment.
Chapter 1 
Introduction
7
rogramming language, like any other formal language, is a set of symbols and rules for
manipulating them. It is when they need to meet the real world that you discover that associations
can be made in different ways and some rules were not specified. A novice all too often sees only
benefits of building a software product and ignores and risks. An expert sees a broader picture
and anticipates the risks. After all, dividing the problem in subproblems and conquering them
piecewise does not guarantee logical rigor and strict consistency between the pieces. Risks
typically include conditions such as, the program can do what is expected of it and then some
more, unexpected capabilities (that may be exploited by bad-intentioned people). Another risk is
that not all environment states are catalogued before commencing the program development.
Depending on how you frame your assumptions, you can come up with a solution. The troubles
arise if the assumptions happen to be inaccurate, wrong, or get altered due to the changing world.
1.1.1
Why Software Engineering Is Difficult (1)
“Software is like entropy. It is difficult to grasp, weighs nothing, and obeys the second law of
thermodynamics; i.e., it always increases.” —Norman R. Augustine
If you are a civil engineer building bridges then all you need to know is about bridges. Unlike
this, if you are developing software you need to know about software domain (because that is
what you are building) and you need to know about the problem domain (because that is what
you are building a solution for). Some problems require extensive periods of dedicated research
(years, decades, or even longer). Obviously, we cannot consider such problem research as part of
software engineering. We will assume that a theoretical solution either exists, or it can be found
in a relatively short time by an informed non-expert.
A further problem is that software is a formal domain, where the inputs and goal states are well
defined. Unlike software, the real world is informal with ill-defined inputs and goal states.
Solving problems in these different domains demands different styles and there is a need to
eventually reconcile these styles. A narrow interpretation of software engineering deals only with
engineering the software itself. This means, given a precise statement of what needs to be
programmed, narrow-scope software engineering is concerned with the design, implementation,
and testing of a program that represents a solution to the stated problem. A broader interpretation
of software engineering includes discovering a solution for a real-world problem. The real-world
problem may have nothing to do with software. For example, the real-world problem may be a
medical problem of patient monitoring, or a financial problem of devising trading strategies. In
broad-scope software engineering there is no precise statement of what needs to be programmed.
Our task amounts to none less than engineering of change in a current business practice.
Software engineering is mainly about modeling the physical world and finding good abstractions.
If you find a representative set of abstractions, the development flows naturally. However, finding
abstractions in a problem domain (also known as “application domain”) involves certain level of
“coarse graining.” This means that our abstractions are unavoidably just approximations—we
cannot describe the problem domain in perfect detail: after all that would require working at the
level of atomic or even subatomic particles. Given that every physical system has very many
parts, the best we can do is to describe it in terms of only some of its variables. Working with
approximations is not necessarily a problem by itself should the world structure be never
changing. But, we live in a changing world: things wear out and break, organizations go bankrupt
or get acquired or restructured, business practices change, government regulations change, fads
Ivan Marsic

8
Rutgers University
and fashions change, and so on. On a fundamental level, one could argue that the second law of
thermodynamics works against software engineers (or anyone else trying to build models of the
world), as well. The second law of thermodynamics states that the universe tends towards
increasing disorder. Whatever order was captured in those comparatively few variables that we
started with, tends to get dispersed, as time goes on, into other variables where it is no longer
counted as order. Our (approximate) abstractions necessarily become invalid with passing time
and we need to start afresh. This requires time and resources which we may not have available.
We will continue discussion of software development difficulties in Sections 2.4.5 and 2.6.3.
Software development still largely depends on heroic effort of select few developers. Product line
and development standardization are still largely missing, but there are efforts in this direction.
Tools and metrics for product development and project management are the key and will be given
considerable attention in this text.
1.1.2
Book Organization
Chapter 2 offers a quick tour of software engineering that is based on software objects, known as
Object-Oriented Software Engineering (OOSE). The main focus is on tools, not methodology, for
solving software engineering problems. Chapter 3 elaborates on techniques for problem
understanding and specification. Chapter 4 describes metrics for measuring the software process
and product quality. Chapter 5 elaborates on techniques for problem solution, but unlike Chapter
2 it focuses on advanced tools for software design. Chapter 6 describes structured data
representation using XML. Chapter 7 presents software components as building blocks for
complex software. Chapter 8 introduces service-oriented architectures and Web services.
I adopt an incremental and iterative refinement approach to presenting the material. For every
new topic, we will scratch the surface and move on, only to revisit later and dig deeper.
The hope with metaphors and analogies is that they will evoke understanding much faster and
allow “cheap” broadening it, based on the existing knowledge.
1.2 Software Engineering Lifecycle
The Feynman Problem-Solving Algorithm:
(i) Write down the problem (ii) think very hard, and (iii) write down the answer.
Any product development can be expected to proceed as an organized process that usually
includes the following phases:

Planning / Specification

Design

Implementation

Evaluation
Chapter 1 
9
Introduction
Requirements
Design
Implementation
Waterfall
method
Testing
Deployment &
Maintenance
Figure 1-6: Waterfall process for software development.
So is with software development. The common software development phases are as follows:
1. Requirements Specification
- Understanding the usage scenarios and deriving the static domain model
2. Design
- Assigning responsibilities to objects and specifying detailed dynamics of their
interactions under different usage scenarios
3. Implementation
- Encoding the design in a programming language
4. Testing
- Individual classes/components (unit testing) and the entire system (integration
testing)
5. Operation and Maintenance
- Running the system; Fixing bugs and adding new features
The lifecycle usually comprises many other activities, some of which precede the above ones,
such as marketing survey to determine the market need for the planned product. This text is
restricted to engineering activities, usually undertaken by the software developer.
The early inspiration for software lifecycle came from other engineering disciplines, where the
above activities usually proceed in a sequential manner (or at least it was thought so). This
method is known as Waterfall Process because developers build monolithic systems in one fell
swoop (Figure 1-6). It requires completing the artifacts of the current phase before proceeding to
the subsequent one. In civil engineering, this approach would translate to: finish all blueprints
neatly before starting construction; finish the construction before testing it for soundness; etc.
There is also psychological attraction of the waterfall model: it is a linear process that leads to a
conclusion by following a defined sequence of steps. However, over the years developers realized
that software development is unlike any other product development in these aspects:
Ivan Marsic

Rutgers University
10

Unlike most other products, software is intangible and hard to visualize. Most people
experience software through what it does: what inputs it takes and what it generates as
outputs

Software is probably the most complex artifact—a large software product consists of so
many bits and pieces as well as their relationships, every single one having an important
role—one flipped bit can change the entire sense of a program

Software is probably the most flexible artifact—it can be easily and radically modified at
any stage of the development process, so it can quickly respond to changes in customer
requirements (or, at least it is so perceived)
Therefore, software development process that follows a linear order of understanding the
problem, designing a solution, implementing and deploying the solution, does not produce best
results. It is easier to understand a complex problem by implementing and evaluating pilot
solutions. These insights led to adopting incremental and iterative (or, evolutionary) development
methods, which are characterized by:
1. Break the big problem down into smaller pieces (increments) and prioritize them.
2. In each iteration progress through the development in more depth.
3. Seek the customer feedback and change course based on improved understanding.
Incremental and iterative process seeks to get to a working instance1 as soon as possible. Having
a working instance available lets the interested parties to have something tangible, to play with,
make inquiries and receive feedback. Through this experimentation (preferably by end users),
unsuspected deficiencies are discovered that drive a new round of development using failures and
the knowledge of things that would not work as a springboard to new approaches. This greatly
facilitates the consensus reaching and building the understanding of all parties of what needs to
be developed and what to expect upon the completion. So, the key of incremental and iterative
methods is to progressively deepen the understanding or “visualization” of the target product, by
both advancing and retracting to earlier activities to rediscover more of its features. A popular
incremental and iterative process is called Unified Process [Jacobson et al., 1999]. Methods that
are even more aggressive in terms of short iterations and heavy customer involvement are
characterized as Agile. The customer is continuously asked to prioritize the remaining work items
and provide feedback about the delivered increments of software.
All lifecycle processes have a goal of incremental refinement of the product design, but different
people hold different beliefs on how this is to be achieved. This has been true in the past and it
continues to be true, and I will occasionally comment on different approaches. Personally, I
enthusiastically subscribe to the incremental and iterative approach, and in that spirit the
exposition in this text progresses in an incremental and iterative manner, by successively
elaborating the software lifecycle phases. For every new topic, we will scratch the surface and
move on, only to revisit later and dig deeper.
A quick review of existing software engineering textbooks reveals that software engineering is
largely about management. Project management requires organizational and managerial skills
1
This is not necessarily a prototype, because “prototype” creates impression of something to be thrown
away after initial experimentation. Conversely, a “working instance” can evolve into the actual product.
Chapter 1 
Introduction
11
such as identifying and organizing the many tasks comprising a project, allocating resources to
complete those tasks, and tracking actual against expected/anticipated resource utilization.
Successful software projects convey a blend of careful objective evaluation, adequate preparation,
continuous observation and assessment of the environment and progress, and adjusting tactics.
It is interesting to compare the issues considered by Brooks [1975] and compare those of the
recent agile methods movement—both put emphasis on communication of the development team
members. My important goal here is, therefore, to present the tools that facilitate communication
among the developers. The key such tools are:

Modular design: Breaking up the system in modules helps to cope with complexity; we
have already seen how the ATM system was made manageable by identifying smaller
tasks and associated “modules” (Figure 1-4). Modules provide building blocks or
“words” of a language when describing complex solutions.

Symbol language: The Unified Modeling Language (UML) is used similar to how the
symbols such as , , , and , are used in mathematics. They abbreviate the exposition
of the material and facilitate the reader’s understanding of the material.

Project and product metrics: Metrics for planning and measuring project progress, and
metrics for measuring the quality of software products provide commonly agreeable tools
for tracking the work quality and progress towards the completion.

Design heuristics: Also known as patterns, they create a design language for naming and
describing the best practices that were proven in many contexts and projects.
Decomposing a problem into simpler ones, so called divide-and-conquer approach, is common
when dealing with complex problems. In software development it is embodied in modularity: The
source code for a module can be written and maintained independently of the source code for
other modules. As with any activity, the value of a structured approach to software development
becomes apparent only when complex problems are tackled.
1.2.1
Symbol Language
“Without images we can neither think nor understand anything.” —Martin Luther (1483-1546)
“There are only 10 types of people in this world. Those who know binary, and those who don’t.”
—Unknown
As part of a design process, it is essential to communicate your ideas. When describing a process
of accomplishing a certain goal, person actually thinks in terms of the abbreviations and symbols
as they describe the “details” of what she is doing, and could not proceed intelligently if she were
not to do so. George Miller found in the 1950s that human short-term memory can store about
seven items at a time [Miller, 1957]. The short-term memory is what we use, for instance, to
remember a telephone number just long enough to look away from the paper on which it is
written to dial the number. It is also known as working memory because in it information is
assumed to be processed when first perceived. It has been likened to the RAM (random access
memory) of a computer. Recall how many times you had to look back in the middle of dialing,
particularly if you are not familiar with the area code, which makes the number a difficult 10
digits! It turns out that the Miller’s hypothesis is valid for any seven “items,” which could be
anything, such as numbers, faces, people, or communities—as we organize information on higher
Ivan Marsic

12
Rutgers University
«interface»
BaseInterface
ClassName
Three common
compartments:
1.
Actor
Classifier name
2.
Attributes
3.
Operations
+ operation()
# attribute_1 : int
# attribute_2 : boolean
# attribute_3 : String
+ operation_1() : void
+ operation_2() : String
+ operation_3(arg1 : int)
Class1Implement
Class2Implement
+ operation()
+ operation()
Stereotype
«» provides
additional info/
annotation/
explanation
Inheritance
relationship:
BaseInterface
is implemented
by two classes
Software Class
Comment
Software Interface Implementation
instance1 : Class1
instance5 : Class2
doSomething()
instance8 : Class3
doSomethingElse()
Interaction Diagram
doSomethingYetElse()
Figure 1-7: Example UML symbols for software concepts.
levels of abstraction, we can still remember seven of whatever it is. This item-level thinking is
called chunking. Symbols can be easier chunked into patterns, which are represented by new
symbols. Using symbols and hierarchical abstraction makes it easier for people to think about
complex systems.
Diagrams and symbols are indispensible to the software engineer. Program code is not the best
way to document a software system, although some agile methodologists have claimed that it is
(more discussion in Section 2.1.1). Code is precise, but it is also riddled with details and
idiosyncrasies of the programming language. Because it is essentially text, is not well-suited for
chunking and abstraction. The visual layout of code can be used to help the reader with chunking
and abstraction, but it is highly subjective with few widely accepted conventions.
Our primary symbol language is UML, but it is not strictly adhered to throughout the text. I will
use other notations or an ad-hoc designed one if I feel that it conveys the message in a more
elegant way. I would prefer to use storyboards and comic-strip sequences to represent that
problem and solution in a comprehensible manner. On the other hand, they are time-consuming
and often ambiguous, so we will settle for the dull but standardized graphics of the UML.
Example UML symbols are shown in Figure 1-7. To become familiar with UML, you can start at
http://www.uml.org, which is the official standard’s website. People usually use different symbols
for different purposes and at different stages of progression. During development there are many
ways to think about your design, and many ways to informally describe it. Any design model or
modeling language has limits to what it can express and no one view of a design tells all. For
example, strict adherence to a standard may be cumbersome for the initial sketches; contrariwise,
documenting the completed design is always recommended in UML simply because so many
people are already familiar with UML.
Chapter 1 
13
Introduction
(a)
1
7
0
8
5
2
9
6
3
System
(ATM machine)
Bank customer
(b)
4
Bank’s remote
datacenter
Window clerk
Datacenter
liaison
Bookkeeper
Safe keeper
Dispenser
Speakerphone
Telephone
Transaction
record
Safe
Cash
Figure 1-8: Gallery of actors (a) and concepts (b) of the system under discussion. The actors
are relatively easy to identify because they are external to the system and visible; conversely,
the concepts are hard to identify because they are internal to the system, hence
invisible/imaginary.
As can be observed throughout this text, the graphic notation is often trivial and can be mastered
relatively quickly. The key is in the skills in creating various models—it can take considerable
amount of time to gain this expertise.
1.2.2
Requirements Analysis and System Specification
We start with the customer statement of work (also known as customer statement of
requirements), if the project is sponsored by a specific customer, or the vision statement, if the
project does not have a sponsor. The statement of work describes what the envisioned system-tobe is about, followed by a list of features/services it will provide or tasks/activities it will support.
Given the statement of work, the first step in the software development process is called
requirements analysis or systems analysis. During this activity the developer attempts to
understand the problem and delimit its scope. The result is an elaborated statement of
Ivan Marsic

A
14
Rutgers University
B
Enter
your PIN
Verify
account
XYZ
C
How may
I help
you?
1
4 2
7 85 6 3
0 9
1
4 2
7 85 6 3
0 9
XYZ valid.
Balance:
$100
Typing in
PIN number
…
D
E
Please take
your cash
7
4
XYZ
withdrew
$60
1
0
Collecting
cash …
Withdraw
$60
2
5 3
8 6
9
Acknowledged
Figure 1-9: Scenario for use case “Withdraw Cash.” Unlike Figure 1-5, this figure only
shows interactions of the actors and the system.
requirements. The goal is to produce the system specification—the document that is an exact
description of what the planned system-to-be is to do. Requirements analysis delimits the system
and specifies the services it offers, identifies the types of users that will interact with the system,
and identifies other systems that interact with ours. For example, the software engineer might ask
the customer to clarify if the ATM machine (Figure 1-3) will support banking for customers of
other banks or only the bank that owns the ATM machine. The system is at first considered a
black box, its services (“push buttons”) are identified, and typical interaction scenarios are
detailed for each service. Requirement analysis includes both fact-finding of how the problem is
solved in the current practice as well as envisioning how the planned system might work.
Recall the ATM example from Figure 1-3. We identified the relevant players in Figure 1-4.
However, this may be too great a leap for a complex system. A more gradual approach is to start
considering how the system-to-be will interact with the external players and defer the analysis of
what happens inside the system until a later time. Figure 1-8(a) shows the players external to the
system (called “actors”). If the ATM machine will support banking for customers of other banks,
then we will need to identify additional actors.
A popular technique for requirements analysis is use case modeling. A set of use cases describes
the elemental tasks a system is to perform and the relation between these tasks and the outside
world. Each use case description represents a dialog between the user and the system, with the
aim of helping the user achieve a business goal. In each dialog, the user initiates actions and the
system responds with reactions. The use cases specify what information must pass the boundary
of the system in the course of a dialog (without considering what happens inside the system).
Because use cases represent recipes for user achieving goals, each use-case name must include a
Chapter 1 
Introduction
15
verb capturing the goal achievement. Given the ATM machine example (Figure 1-3), Figure 1-9
illustrates the flow of events for the use case “Withdraw Cash.”
Use cases are only a beginning of software engineering process. When we elaborate use cases of
a system, it signifies that we know what the system needs to accomplish, not how; therefore, it is
not just “a small matter of system building” (programming) that is left after we specify the use
cases. Requirements analysis is detailed in Sections 2.2 and 2.4.
1.2.3
Object-Oriented Analysis and the Domain Model
“…if one wants to understand any complex thing—be it a brain or an automobile—one needs to develop
good sets of ideas about the relationships among the parts inside. …one must study the parts to know the
whole.” —Marvin Minsky, The Emotion Machine
Use cases consider the system as a black box and help us understand how the system as a whole
interacts with the outside word. The next step is to model the inside of the system. We do this by
building the domain model, which shows what the black box (the system-to-be) encloses. Given a
service description, we can imagine populating the black box with domain concepts that will do
the work. In other words, use cases elaborate the system’s behavioral characteristics (sequence
of stimulus-response steps), while the domain model details the system’s structural
characteristics (system parts and their arrangement) that make it possible for the system to
behave as described by its use cases.
It is useful to consider a metaphor in which software design is seen as creating a virtual
enterprise or an agency. The designer is given an enterprise’s mission description and hiring
budget, with the task of hiring appropriate workers, acquiring things, and making it operational.
The first task is to create a list of positions with a job description for each position. The designer
needs to identify the positions, the roles and responsibilities, and start filling the positions with
the new workers. Recall the ATM machine example from Figure 1-3. We need to identify the
relevant players internal to the system (called “concepts”), as illustrated in Figure 1-8(b).
In the language of requirements analysis, the enterprise is the system to be developed and the
employees are the domain concepts. As you would guess, the key task is to hire the right
employees (identify good concepts, or abstractions). Somewhat less critical is to define their
relationships and each individual’s attributes, which should be done only if they are relevant for
the task the individual is assigned to. I like this metaphor of “hiring workers” because it is in the
spirit of what Richard Feynman considered the essence of programming, which is “getting
something to do something” [Feynman et al., 2000]. It also sets up the stage for the important
task of assigning responsibilities to software objects.
The idea for conducting object-oriented analysis in analogy to setting up an enterprise is inspired
by the works of Fritz Kahn. In the early 20th century, Kahn produced a succession of books
illustrating the inner workings of the human body, using visual metaphors drawn from industrial
society. His illustrations drew a direct functional analogy between human physiology and the
operation of contemporary technologies—assembly lines, internal combustion engines, refineries,
dynamos, telephones, etc. Kahn’s work is aptly referred to as “illustrating the incomprehendable”
and I think it greatly captures the task faced by a software engineer. The interested reader should
search the Web for more information on Kahn.
Ivan Marsic

16
Rutgers University
Solution
modification
How may I
help you?
Transaction
record
(a)
Bookkeeper
Speakerphone
Draftsman
Window clerk
Dispenser
Customer
Solution
modification
How may I
help you?
Transaction
record
(b)
Bookkeeper
Speakerphone
Courier
Window clerk
Dispenser
Remote
bank
Customer
Figure 1-10: Alternative solutions for an ATM system. (Compare to Figure 1-4)
Domain analysis is more than just letting our imagination loose and imagining any model for the
system-to-be. Design problems have unlimited number of alternative solutions. For example,
consider again the design for an ATM system from Figure 1-4. One could imagine countless
alternative solutions, two of which are shown in Figure 1-10. In Figure 1-10(a), we imagine
having a draftsman to draw the banknotes requested by the customer then and there. In Figure
1-10(b), we imagine having a courier run to a nearest bank depository to retrieve the requested
Chapter 1 
17
Introduction
monies. How do we know which solution is best, or even feasible? Implementing and evaluating
all imaginable solutions is impossible, because it takes time and resources. Two factors help
constrain the options and shorten the time to solution:

Knowing an existing solution for the same or similar problem

Analyzing the elements of the external world that the system-to-be will interact with
I created the solution in Figure 1-4 because I have seen how banks with human tellers operate. I
know that solutions in Figure 1-10 would take an unacceptable amount of time for each
withdrawal, and the problem statement does not mention having a stack of blank paper and ink at
disposal for solution in Figure 1-10(a), or having a runner at disposal for solution in Figure
1-10(b). The problem statement only mentions a communication line to a remote datacenter.
There is nothing inherent in any of these solutions that makes some better than others. What
makes some solutions “better” is that they copy existing solutions and take into account what is at
our disposal to solve the problem. The implication is that the analyst needs to consider not only
what needs to be done, but also how it can be done—what are feasible ways of doing it. We need
to know what is at our disposal in the external world: do we have a stack of blank papers, ink, or
a courier to run between the ATM and a depository? If this information is not given, we need to
ask our customer to clarify. For example, the customer may answer that the system-to-be will
have at disposal only a communication line to a remote datacenter. In this case, we demand the
details of the communication protocol and the format of messages that can be exchanged. We
need to know how will the datacenter answer to different messages and what exceptions may
occur. We also need to know about the hardware that accepts the bank cards and disposes
banknotes. How will our software be able to detect that the hardware is jammed?
Our abstractions must be grounded in reality, and the grounding is provided by knowing what is
at the disposal in the external world that the system-to-be can use to function. This is why we
cannot delimit domain analysis to what the black box (software-to-be) will envelop. Rather, we
need to consider entities that are both external and internal to the software-to-be. The external
environment constrains the problem to be solved and by implication constrains the internal design
of the software-to-be. We also need to know what is implementable and what not, either from
own experience, or from that of a person familiar with the problem domain (known as the
“domain expert”).None of our abstractions is realistic, but some are useful and others are not.
Object-oriented analysis is detailed in Section 2.5.
1.2.4
Object-Oriented Design
“Design is not just what it looks like and feels like. Design is how it works.”—Steve Jobs
The act of design involves assigning form and function to parts so to create an esthetical and
functional whole. In software development, the key activity in the design phase is assigning
responsibilities to software objects. A software application can be seen as a set or community of
interacting software objects. Each object embodies one or more roles, a role being defined by a
set of related responsibilities. Roles, i.e., objects, collaborate to carry out their responsibilities.
Our goal is to create a design in which they do it in a most efficient manner. Efficient design
contributes to system performance, but no less important contribution is in making the design
easier to understand by humans.
Ivan Marsic

Rutgers University
18
Design is the creative process of searching how to implement all of the customer’s requirements.
It is a problem-solving activity and, as such, is very much subject to trial and error. Breaking up
the system into modules and designing their interactions can be done in many ways with varying
quality of the results. In the ATM machine example, we came up with one potential solution for
step-by-step interactions, as illustrated Figure 1-5. The key question for the designer is: is this the
best possible way to assign responsibilities and organize the activities of virtual agents? One
could solve the same design problem with a different list of players and different organization of
their step-by-step interactions. As one might imagine, there is no known way for exactly
measuring the optimality of a design. Creativity and judgment are key for good software design.
Knowledge of rules-of-thumb and heuristics are critical in deciding how good a design is.
Luckily, most design work is routine design, where we solve a problem by reusing and adapting
solutions from similar problems.
So, what kinds of designs are out there? Two very popular kinds of software designs are what I
would call Maurits Escher2 and Rube Goldberg3 designs. Both are fun to look at but have little
practical value. Escher designs are impossible to implement in reality. Goldberg designs are
highly-complicated contraptions, which solve the problem, but they are very brittle. If anything
changes in the underlying assumptions, they fail miserably.
A key problem of design is that we cannot know for sure if a design will work unless we
implement it and try it. Therefore, a software engineer who is also a skilled programmer has
advantage in software design, because he knows from experience how exactly to implement the
abstract constructs and what will or will not work. Related to this issue, some agile
methodologists claim that program code is the only faithful representation of program design.
Although it may be faithful, code alone is insufficient to understand software design. One also
needs diagrams to “see the forest for the trees.” Code also usually does not document the design
objectives, alternative designs that were considered, merits of different designs, and the rationale
for the chosen designs.
2
Maurits Cornelis Escher (1898-1972) is one of the world’s most famous graphic artists, known for his socalled impossible structures, such as Ascending and Descending, Relativity, his Transformation Prints,
such as Metamorphosis I, Metamorphosis II and Metamorphosis III, Sky & Water I or Reptiles.
3
Reuben Lucius Goldberg (Rube Goldberg) (1883-1970) was a Pulitzer Prize winning cartoonist, sculptor,
and author. He is best known for his “inventions”—elaborate sets of arms, wheels, gears, handles, cups,
and rods, put in motion by balls, canary cages, pails, boots, bathtubs, paddles, and live animals—that
take simple tasks and make them extraordinarily complicated.
Chapter 1 
19
Introduction
Operator (includes motor and radio control mechanism)
1
2
Remote control transmitter
6
5
3
7
4
Rail with a belt or chain
5
8
6
Pressing of a button on the remote control transmitter (1)
Garage door
authenticates the device & activates the motor in the operator (2).
The motor pulls the chain (or belt) along the rail (3) and winds
the torsion spring (4).
The torsion spring winds the cable on the pulleys (or drums) (5)
on both sides of the door.
The cables lift the door, pushing the different sections of the door
into the horizontal tracks (6)
At the same time, the trolley (or traveler) (7) moves along the rail (3)
and controls how far the door opens (or closes),
Safety reversing sensor
as well as the force the garage door exerts by way of the curved door arm (8)   
Figure 1-11: Top row: A Rube Goldberg machine for garage door opening.
Bottom row: An actual design of a garage door opener.
Consider the garage-door opener designs in Figure 1-11. The top row shows a Rube Goldberg
design and the bottom row shows an actual design. What makes the latter design realistic and
what is lacking in the former design? Some observations:

The Rube Goldberg design uses complex components (the rabbit, the hound, etc.) with
many unpredictable or uncontrollable behaviors; conversely, a realistic design uses
specialized components with precisely controllable functions
Ivan Marsic

20
Rutgers University

The Rube Goldberg design makes unrealistic assumptions, such as that the rabbit will not
move unless frightened by an exploding cap.

The Rube Goldberg design uses unnecessary links in the operational chain.
We will continue discussion of software design when we introduce the object model in Section
1.4. Recurring issues of software design include:

Design quality evaluation: Optimal design may be an unrealistic goal given the
complexity of real-world applications. A more reasonable goal is to find criteria for
comparing two designs and deciding which one is better. The principles for good objectoriented design are introduced in Section2.6 and elaborated in subsequent chapters.

Design for change: Useful software lives for years or decades and must undergo
modifications and extensions to account for the changing world in which it operates.
Chapter 5 describes the techniques for modifiable and extensible design.

Design for reuse: Reusing existing code and designs is economical and allows creating
more sophisticated systems. Chapter 7 considers techniques for building reusable
software components.
Other important design issues include design for security and design for testability.
1.2.5
Project Effort Estimation and Product Quality
Measurement
I will show, on an example of hedge pruning, how project effort estimation and product quality
measurement work hand in hand with incremental and iterative development, particularly in agile
methods. Imagine that you want to earn some extra cash this summer and you respond to an
advertisement by a certain Mr. McMansion to prune the hedges around his property (Figure
1-12). You have never done hedge pruning before, so you will need to learn as you go. The first
task is to negotiate the compensation and completion date. The simplest way is to make a guess
that you can complete the job in two weeks and you ask for a certain hourly wage. Suppose that
Mr. McMansion agrees and happily leaves for vacation. After one week, you realize that you are
much behind the schedule, so to catch up you lower the quality of your work. After two weeks,
the hedges are pruned and Mr. McMansion is back from vacation. He will likely find many
problems with your work and may balk at paying for the work done.
Now suppose that you employ incremental and iterative hedge pruning. You start by dividing the
hedges into smaller sections, because people are better at guessing the relative sizes of object
parts than the absolute size of an entire object. Suppose that you came up with the partitioning
labeled with encircled numbers  to  in Figure 1-12. Think of hedge pruning as traveling along
the hedge at a certain velocity (while pruning it). The velocity represents your work productivity.
To estimate the travel duration, you need to know the length of the path (or, path size). That is
Travel duration 
Path size
Travel velocity
(1.1)
Because you have never pruned hedges, you cannot know your velocity, so the best you can do is
to guess it. You could measure the path size using a tape measure, but you realize there is a
Chapter 1 
21
Introduction








Figure 1-12: Example for project estimation: Formal hedge pruning.
problem. Different sections seem to have varying difficulty of pruning, so your velocity will be
different along different sections. For example, it seems that section  at the corner of Back and
Side Streets (Figure 1-12) will take much more work to prune than section  between the garden
and Main Street. Let us assume you assign “pruning points” to different sections to estimate their
size and complexity. Suppose you use the scale from 1 to 10. Because section  seems to be the
most difficult, so we assign it 10 pruning points. The next two sections in terms of difficulty
appear to be  and , and relative to section  you feel that they are at about 7 pruning points.
Next are sections , , and , and you give them 4 pruning points. Finally, section  gets 3
pruning points and section  gets 2 pruning points. The total for the entire hedge is calculated
simply by adding the individual sizes
N
Total size   points - for - section i 
(1.2)
i 1
Therefore, the total for the entire hedge is 10  27  34  3  2 = 41 pruning points. This
represents your size estimate of the entire hedge. It is very important that this is a relative-size
estimate, because it measures how big individual sections are relative to one another. So, a
section estimated at four pruning points is expected to take twice as long work as a section
estimated at two pruning points.
How accurate is this estimate? Should section  be weighted 3.5
points instead of 3? There are two parts to this question: (a) how
accurate is the relative estimate for each section, and (b) is it
appropriate to simply add up the individual sizes? As for the former
issue, you may wish to break down the hedge sections into smaller parts, because it is easier to do
eyeballing of smaller parts and comparing to one another. Section  is particularly large and it
may be a good idea to split it up to smaller pieces. If you keep subdividing, in the extreme instead
of eyeballing hedge sections you could spend weeks and count all the branches and arrive at a
Ivan Marsic

22
Rutgers University
Estimation accuracy
100%
Estimation cost
Figure 1-13: Exponential cost of estimation. Improving accuracy of estimation beyond a
certain point requires huge cost and effort (known as the law of diminishing returns).
much more accurate estimate. You could even measure density of branches in individual sections,
their length, hardness, etc. Obviously, there is a point beyond which only minor improvement in
estimation accuracy is brought at a huge cost (known as the law of diminishing returns). Many
people agree that the cost-accuracy relationship is exponential (Figure 1-13). It is also interesting
to note that, in the beginning of the curve, we can obtain huge gains in accuracy with modest
effort investment. The key points for size estimation are that (1) the pieces should be fairly small
and (2) they should be of similar size, because it is easier to compare the relative sizes of small
and alike pieces.
As for the latter issue about equation (1.2), the appropriateness of using a linear summation, a key
question is if the work on one section is totally independent on the work on another section. The
independence is equivalent to assuming that every section will be pruned by a different person
and each starts with an equal degree of experience in hedge pruning. I believe there are
confounding factors that can affect the accuracy of the estimate. For example, as you progress,
you will learn about hedge pruning and become more proficient, so your velocity will increase
not because the size of some section became smaller but because you became more proficient. In
Section 2.2.3 I will further discuss the issue of linear superposition in the context of software
project estimation.
All you need now is the velocity estimate, and using equation (1.1) you can give Mr. McMansion
the estimate of how long the entire hedge pruning will take. Say you guess your velocity at 2
pruning points per day. Using equation (1.1) you obtain 41/2  21 working days or 4 weeks. You
tell Mr. McMansion that your initial estimate is 21 days to finish the work. However, you must
make it clear that this is just a guess, not a hard commitment; you cannot make hard
commitments until you do some work and find out what is your actual productivity (or
“velocity”). You also tell Mr. McMansion how you partitioned the work into smaller items
(sections of the hedge) and ask him to prioritize the items, so that you know his preferred
ordering. Say that Mr. McMansion prefers that you start from the back of the house and as a
result you obtain the work backlog list shown in Figure 1-14. He will inspect the first deliverable
after one week, which is the duration of one iteration.
Here comes the power of iterative and incremental work. Given Mr. McMansion’s prioritized
backlog, you pull as many items from the top of the list as will fit into an iteration. Because the
Chapter 1 
23
Introduction
Work backlog
1) Prune Section 8
3.5 days (7pts)
2) Prune Section 7
2 days (4pts)
3) Prune Section 6
1 day (2pts)
4) Prune Section 5
2 days (4pts)
5) Prune Section 4
1.5 days (3p)
6) Prune Section 1
Estimated work duration
Items pulled by the team into an iteration
2 days (4pts)
7) Prune Section 2
3.5 days (7p)
8) Prune Section 3
5 days (10p)
Work items
21 days
1st iteration
2nd iteration
n-th iteration
5 days
List prioritized by the customer
Estimated completion date
Time
Figure 1-14: The key concepts for iterative and incremental project effort estimation.
first two items (sections  and ) add up to 5.5 days, which is roughly one week, i.e., one
iteration, you start by pruning sections  and . Suppose that after the first week, you pruned
have about three quarters of the hedges in sections  and . In other words after the first
iteration you found that your actual velocity is 3/4 of what you originally thought, that is, 1.5
pruning point per day. You estimate a new completion date as follows.
Total number of remaining points = 1/4  11 points remaining from sections  and 
 30 points from all other sections
 33 points
Estimated completion date = 22 days + 5 days already worked = 27 days total
You go to Mr. McMansion and tell him that your new estimate is that it will take you 27 days
total, or 22 more days to complete the work. Although this is still an estimate and may prove
incorrect, you are much more confident about this estimate, because it is based on your own
experience. Note that you do not need to adjust your size estimate of 41 pruning points, because
the relative sizes of hedge sections have not changed! Because of this velocity adjustment, you
need to calculate new work durations for all remaining items in the backlog (Figure 1-14). For
example, the new durations for sections  and  will be 1.3 days and 2.7 days, respectively. As
a result, you will pull into the second iteration the remaining work from the first iteration plus
sections  and . Section  that was originally planned for the second iteration (Figure 1-14)
will be left for the third iteration.
It is important to observe that initially you estimate your velocity, but after the first increment you
use the measured velocity to obtain a more accurate estimate of the project duration. You may
continue measuring your velocity and re-estimating the total effort duration after each increment,
Ivan Marsic

24
Rutgers University
Good Shape
(Low branches get sun)
Snow accumulates
on broad flat tops
Poor Shape
(Low branches
shaded from sun)
Straight lines require
more frequent trimming
Heading back not
recommended as
it alters the natural
shape of the shrub
Peaked and rounded tops
hinder snow accumulation
Remove dead wood
Remove water spouts
and suckers
Rounded forms, which
follow nature’s tendency,
require less trimming
Figure 1-15: Quality metrics for hedge pruning.
but this probably will not be necessary, because after the first few increments you will obtain an
accurate measurement of your pruning velocity. The advantage of incremental work is that you
can quickly gain accurate estimate of the entire effort and will not need to rush it later to complete
on time, while sacrificing product quality.
Speaking of product quality, next we will see how iterative work helps improve product quality.
You may be surprised to find that hedge pruning involves more than simply trimming the shrub.
Some of parameters that characterize the quality of hedge pruning are illustrated in Figure 1-15.
Suppose that after the first iteration (sections  and ), Mr. McMansion can examine the work
and decide if the quality is satisfactory or needs to be adjusted for future iterations.
It is much more likely that Mr. McMansion will be satisfied with your work if he is continuously
consulted then if he simply disappeared to vacation after describing the job requirements.
Regardless of how detailed the requirements description, you will inevitably face unanticipated
situations and your criteria of hedge esthetics may not match those of Mr. McMansion. Everyone
sees things differently, and frequent interactions with your customer will help you better
understand his viewpoint and preferences. Early feedback will allow you to focus on things that
matter most to the customer, rather than facing a disappointment when the work is completed.
This is why it is important that the customer remains engaged throughout the duration of the
project, and participates in all important decisions and inspects the quality of work any time a
visible progress is made.
Chapter 1 
25
Introduction
In summary, we use incremental staging and scheduling strategy to quickly arrive at an effort
estimate and to improve the development process quality. We use the iterative, reworkscheduling strategy to improve the product quality. Of course, for both of these strategies it is
essential to have good metrics. Project and product metrics are described in Chapter 4. We will
also see in Section 2.2.3 how user-story points work similar to hedge-pruning points, and how
they can be used to estimate development effort and plan software releases.
1.3 Case Studies
Two case studies will be used in examples throughout the text to illustrate software development
techniques. In addition, several more projects are designed for student teams later in Section 1.5.
Both case studies (as well as student projects) address relatively complex problems. I favor
complex projects, threading throughout the book, rather than simple, unconnected examples,
because I feel that the former illustrate better the difficulties and merits of the solutions. Both
projects are open-ended and without a clear objective, so that we can consider different features
and better understand the requirements derivation process. My hope is that by seeing software
engineering applied on complex (and realistic) scenarios, the reader will better grasp
compromises that must be made both in terms of accuracy and richness of our abstractions. This
should become particularly evident in Chapter 3, which deals with modeling of the problem
domain and the system that will be developed.
Before we discuss the case studies, I briefly introduce a simple diagrammatic technique for
representing knowledge about problem domains. Concept maps4 are expressed in terms of
concepts and propositions, and are used to represent knowledge, beliefs, feelings, etc. Concepts
are defined as apperceived regularities in objects, events, and ideas, designated by a label, such as
“green,” “high,” “acceleration,” and “confused.” A proposition is a basic unit of meaning or
expression, which is an expression of the relation among concepts. Here are some example
propositions:

Living things are composed of cells

The program was flaky

Ice cube is cold
We can decompose arbitrary sentences into propositions. For example, the sentence
“My friend is coding a new program”
can be written as the following propositions
Proposition Concept
Relation
1.
I
have
I
have
Concept
friend
friend
engages in
coding
4
constructs a
A good introduction about concept maps can be found here: http://en.wikipedia.org/wiki/Concept_map.
CmapTools (http://cmap.ihmc.us/) is free software that facilitates construction of conceptprogram
maps.
is
new
Ivan Marsic

26
Rutgers University
2.
3.
4.
friend
coding
program
engages in
constructs a
is
coding
program
new
How to construct a concept map? A common strategy starts with listing all the concepts that you
can identify in a given problem domain. Next, create the table as above, initially leaving the
“Relation” column empty. Then come up with (or consult a domain expert for) the relations
among pairs of concepts. Note that, unlike the simple case shown in the above table, in general
case some concepts may be related to several other concepts. Finally, drawing the concept map is
easy when the table is completed. We will learn more about propositions and Boolean algebra in
Chapter 3.
Concept maps are designed for capturing static knowledge and relationships, not sequential
procedures. A concept map provides a semiformal way to represent knowledge about a problem
domain. It has reduced ambiguity compared to free-form text, and visual illustration of
relationships between the concepts is easier to understand. I will use concepts maps in describing
the case study problems and they can be a helpful tool is software engineering in general. But
obviously we need other types of diagrammatic representations and our main tool will be UML.
1.3.1
Case Study 1: From Home Access Control to
Adaptive Homes
Figure 1-16 illustrates our case-study system that is used in the rest of the text to illustrate the
software engineering methods. In a basic version, the system offers house access control. The
system could be required to authenticate (“Are you who you claim to be?”) and validate (“Are
you supposed to be entering this building?”) people attempting to enter a building. Along with
controlling the locks, the system may also control other household devices, such as the lighting,
air conditioning, heating, alarms, etc.
As typical of most software engineering projects, a seemingly innocuous problem actually hides
many complexities, which will be revealed as we progress through the development cycle. Figure
1-16 already indicates some of those—for example, houses usually have more than one lock.
Shown are two locks, but there could be additional ones, say for a garage entrance, etc.
Additional features, such as intrusion detection further complicate the system. For example, the
house could provide you with an email report on security status while you are away on vacation.
Police will also attend when they receive notification from a central monitoring station that a
monitored system has been activated. False alarms require at least two officers to check on and
this is a waste of police resources. Many cities now fine residents for excessive false alarms.
Here are some additional features to think about. You could program the system to use timers to
turn lights, televisions and sound systems on and off at different times to give your home a
“lived-in look” when you are away. Install motion-detecting outdoor floodlights around your
home or automatic announcing of visitors with a chime sound. More gadgets include garage door
openers, active badges, and radio-frequency identification (RFID) tags, to detect and track the
tenants. Also, an outside motion sensor may turn on the outdoors light even before the user
unlocks the door. We could dream up all sorts of services; for example, you may want to be able
to open the door for a pizza-deliveryman remotely, as you are watching television, by point-and-
Chapter 1 
27
Introduction
Central
Computer
Alarm bell
System
1
2
3
4
5
X
Y
Light bulb
Lock
Photosensor
Backyard doors:
External &
Internal lock
Switch
Front doors:
External &
Internal lock
Figure 1-16: Our first case-study system provides several functions for controlling the home
access, such as door lock control, lighting control, and intrusion detection and warning.
click remote controls. Moreover, the system may bring up the live video on your TV set from a
surveillance camera at the doors.
Looking at the problem in a broader business context, it is unlikely that all or even the majority of
households targeted as potential customers of this system will be computer-savvy enough to
maintain the system. Hence, in the age of outsourcing, what better idea than to contract a security
company to manage all systems in a given area. This brings a whole new set of problems, because
we need to deal with potentially thousands of distributed systems, and at any moment many new
users may need to be registered or unregistered with the (centralized?) system.
There are problems maintaining a centralized database of people’s access privileges. A key
problem is having a permanent, hard-wired connection to the central computer. This sort of
network is very expensive, mainly due to the cost of human labor involved in network wiring and
maintenance. This is why, even in the most secure settings, a very tiny fraction of locks tend to be
connected. The reader should check for an interesting decentralized solution proposed by a
software company formerly known as CoreStreet (http://www.actividentity.com/). In their proposed
solution, the freshest list of access privileges spreads by “viral propagation” [Economist, 2004].
First Iteration: Home Access Control
Our initial goal is only to support the basic door unlocking and locking functions.
Although at the first sight these actions appear simple, there are difficulties with
both.
Figure 1-16 shows the locks connected by wire-lines to a central personal
computer (PC). This is not necessarily how we want to solve the problem; rather,
the PC just illustrates the problem. We need it to manage the users
(adding/removing valid users) and any other voluminous data entry, which may be
cumbersome from a lock’s keypad—using a regular computer keyboard and
monitor would be much more user friendly. The connections could be wireless,
Ivan Marsic

28
Rutgers University
tenant
enters
wishes
key
can be
upper bound on failed attempts
lock opened
causes
valid key
invalid key
may signal
burglar
launches
can be prevented by enforcing
dictionary attack
Figure 1-17: Concept map representing home access control.
and moreover, the PC may not even reside in the house. In case of an apartment complex, the PC
may be located in the renting office.5
The first choice is about the user identification. Generally, a person can be identified by one of
the following:

What you carry on you (physical key or another gadget)

What you know (password)

Who you are (biometric feature, such as fingerprint, voice, face, or iris)
I start with two constraints set for this specific system: (1) user should not need to carry any
gadgets for identification; and, (2) the identification mechanism should be cheap. The constraint
(1) rules out a door-mounted reader for magnetic strip ID cards or RFID tags—it imposes that the
user should either memorize the key (i.e., “password”) or we should use biometric identification
mechanism(s). The constraint (2) rules out expensive biometric identifiers, such as face
recognition (see, e.g., http://www.identix.com/ and http://passfaces.com/) or voice recognition (see,
e.g., http://www.nuance.com/prodserv/prodverifier.html). There are relatively cheap fingerprint readers
(see, e.g., http://www.biometrics-101.com/) and this is an option, but to avoid being led astray by
technology details, for now we assume that the user memorizes the key. In other words, at present
we do not check the person’s true identity (hence, no authentication)—as long as she knows a
valid key, she will be allowed to enter (i.e., validation only).
For unlocking, a difficulty is with handling the failed attempts
(Figure 1-17). The system must withstand “dictionary attacks” (i.e.,
burglars attempting to discover an identification key by systematic
trial). Yet it must allow the legitimate user to make mistakes.
For locking coupled with light controls, a difficulty is with
detecting the daylight: What with a dark and gloomy day, or if the
photo sensor ends up in a shade. We could instead use the wallclock time, so the light is always turned on between 7:30 P.M. and
7:30 A.M. In this case, the limits should be adjusted for the
seasons, assuming that the clock is automatically adjusted for daylight saving time shift. Note
5
This is an architectural decision (see Section 2.3 about software architecture).
Chapter 1 
29
Introduction
also that we must specify which light should be turned on/off: the one most adjacent to the doors?
The one in the hallway? The kitchen light? … Or, all lights in the house?
Interdependency question: What if the door needs to be locked after the
tenant enters the house—should the light stay on or should different lights
turn on as the tenant moves to different rooms?
bolt
Also, what if the user is not happy with the system’s decision and does opposite of what the
system did, e.g., the user turns off the light when the system turned it on? How do these events
affect the system functioning, i.e., how to avoid that the system becomes “confused” after such an
event?
Figure 1-18 illustrates some of the difficulties in specifying exactly what the user may want from
the system. If all we care about is whether the door is unlocked or locked, identify two possible
states: “unlocked” and “locked.” The system should normally be in the “locked” state and
unlocked only in the event the user supplies a valid key. To lock, the user should press a button
labeled “Lock,” but to accommodate forgetful users, the system should lock automatically
autoLockInterval seconds after being unlocked. If the user needs the door open longer for some
reason, she may specify the holdOpenInterval. As seen, even with only two clearly identified
states, the rules for transitioning between them can become very complex.
I cannot overstate the importance of clearly stating the user’s goals. The goal state can be
articulated as unlocked AND light_on. This state is of necessity temporary, because the door
should be locked once the user enters the house and the user may choose to turn off the hallway
light and turn on the one in the kitchen, so the end state ends up being lockeded AND light_off.
Moreover, this definition of the goal state appears to be utterly incomplete.
Due to the above issues, there are difficulties with unambiguously establishing the action
preconditions. Therefore, the execution of the “algorithm” turns out to be quite complex and
eventually we have to rely only on heuristics. Although each individual activity is simple, the
combination of all is overwhelming and cannot be entirely solved even with an extremely
complex system! Big software systems have too many moving parts to conform to any set of
simple percepts. What appeared a simple problem turns out not to have an algorithmic solution,
and on the other hand we cannot guarantee that the heuristics will always work, which means that
we may end up with an unhappy customer.
Note that we only scratched the surface of what appeared a simple problem, and any of the above
IF validKey AND holdOpenInterval THEN unlock
IF validKey THEN unlock
locked
unlocked
IF pushLockButton THEN lock
IF timeAfterUnlock = max{ autoLockInterval, holdOpenInterval } THEN lock
Figure 1-18: System states and transition rules.
Ivan Marsic

Rutgers University
30
issues can be further elaborated. The designer may be simply unable to explicitly represent or
foresee every detail of the problem. This illustrates the real problem of heuristics: at a certain
point the designer/programmer must stop discerning further details and related issues. But, of
course, this does not mean that they will not arise sooner or later and cause the program to fail.
And we have not mentioned program bugs, which are easy to sneak-in in a complex program.
Anything can happen (and often does).
1.3.2
Case Study 2: Personal Investment Assistant
“The way to make money is to buy stock at a low price, then when the price goes up, sell it.
If the price doesn’t go up, don’t buy it.” —Will Rogers
Financial speculation, ranging from straight gambling and betting to modern trading of financial
securities, has always attracted people. For many, the attraction is in what appears to be a promise
of wealth without much effort; for most, it is in the promise of a steady retirement income as well
as preserving their wealth against worldly uncertainties. Investing in company equities (stocks)
has carried the stigma of speculation through much of history. Only relatively recently stocks
have been treated as reliable investment vehicles (Figure 1-19). Nowadays, more than 50% of the
US households own stocks, either directly, or indirectly through mutual funds, retirement
accounts or other managed assets. There are over 600 securities exchanges around the world.
Many people have experience with financial securities via pension funds, which today are the
largest investor in the stock market. Quite often, these pension funds serve as the “interface” to
the financial markets for individual investors. Since early 1990s the innovations in personal
computers and the Internet made possible for the individual investor to enter the stock markets
without the help from pension funds and brokerage firms. The Internet also made it possible to do
all kinds of researches and comparisons about various companies, in a quick and cheap fashion—
an arena to which brokerage firms and institutional investors had almost exclusive access owing
to their sheer size and money-might.
Computers have, in the eyes of some, further reduced the amount of effort needed for
participation in financial markets, which will be our key motivation for our second case study:
how to increase automation of trading in financial markets for individual investors. Opportunities
for automation range from automating the mechanics of trading to analysis of how wise the
particular trades appear to be and when risky positions should be abandoned.
There are many different financial securities available to investors. Most investment advisers
would suggest hedging the risk of investment loss by maintaining a diversified investment
portfolio. In addition to stocks, the investor should buy less-risky fixed income securities such as
bonds, mutual funds, treasuries bills and notes or simply certificate of deposits. To simplify our
case study, I will ignore such prudent advice and assume that our investor wants to invest in
stocks only.
Chapter 1 
31
Introduction
people who trade
have expectation of
are attracted by
can have
indirect participation
profit returns
hints at
are needed for
direct participation
market’s performance history
is done through
short-term objectives
shows
retirement income
is done through
+10% annual returns over long run
broker
pension fund
negative returns 1/3 of the time
retirement plan
mutual fund
hedge fund
Figure 1-19: Concept map of why people trade and how they do it.
Why People Trade and How Financial Markets Work?
Anyone who trades does so with the expectation of making profits. People take risks to gain
rewards. Naturally, this immediately begets questions about the kind of return the investor
expects to make and the kind of risk he is willing to take. Investors enter into the market with
varied objectives. Broadly, the investor objectives could be classified into short-term-gain and
long-term-gain. The investors are also of varied types. There are institutional investors working
for pension funds or mutual funds, and then there are day-traders and hedge-fund traders who
mainly capitalize on the anomalies or the arbitrages that exist in the markets. Usually the
institutional investors have a “long” outlook while the day-traders and the hedge-funds are more
prone to have a “short” take on the market.
Here I use the terms “trader” and “investor” and synonymous. Some people use these terms to
distinguish market participants with varied objectives and investment styles. Hence, an “investor”
is a person with a long outlook, who invests in the company future by buying shares and holds
onto them expecting to profit in long term. Conversely, a “trader” is a person with a short
outlook, who has no long-term interest in the company but only looks to profit from short-term
price variations and sells the shares at first such opportunity.
As shown in Figure 1-20(a), traders cannot exchange financial securities directly among
themselves. The trader only places orders for trading with his broker and only accredited
financial brokers are allowed to execute transactions. Before the Internet brokers played a more
significant role, often provided investment advice in addition to executing transactions, and
charged significant commission fees. Nowadays, the “discount brokers” mostly provide the
transaction service at a relatively small commission fee.
Ivan Marsic

32
Rutgers University
Trader
Broker
Exchange
Market exchange
Brokers
Bank
Traders/
Investors
(a)
(b)
Figure 1-20: Structure of securities market. (a) Trading transactions can be executed only
via brokers. (b) “Block diagram” of market interactions.
price quote
can be
traded price
indicative price
is set by
can be
last trade
can be
buying stock
selling stock
is executed at
is executed at
ask price
bid price
Figure 1-21: Concept map explaining how quoted stock prices are set.
Mechanics of Trading in Financial Markets
A market provides a forum where people always sell to the highest bidder. For a market to exist
there must be a supply and demand side. As all markets, financial markets operate on a bid-offer
basis: every stock has a quoted bid and a quoted ask (or offer). The concept map in Figure 1-21
summarizes the functioning of stock prices. The trader buys at the current ask and sells at the
current bid. The bid is always lower than the ask. The difference between the bid and the ask is
referred to as the spread. For example, assume there is a price quote of 100/105. That means the
highest price someone is willing to pay to buy is 100 (bid), and the lowest price there is selling
interest at 105 (offer or ask). Remember that there are volumes (number of shares) associated
with each of those rates as well.
Using the bid side for the sake of illustration, assume that the buyer at 100 is willing to purchase
1,000 units. If someone comes in to sell 2,000 units, he would execute the first 1,000 at 100, the
bid rate. That leaves 1,000 units still to be sold. The price the seller can get will depend on the
depth of the market. It may be that there are other willing buyers at 100, enough to cover the
reminder of the order. In an active (or liquid) market this is often the case.
Chapter 1 
33
Introduction
What happens in a thin market, though? In such a situation, there may not be a willing buyer at
100. Let us assume a situation illustrated in the table below where the next best bid is by buyer
B3 at 99 for 500 units. It is followed by B4 at 98 for 100 units, B1 for 300 units at 97, and B2 for
200 at 95. The trader looking to sell those 1,000 remaining units would have to fill part of the
order at 99, more at 98, another bit at 97, and the last 100 at 95. In doing so, that one 2,000 unit
trade lowered the bid down five points (because there would be 100 units left on the bid by B2).
More than likely, the offer rate would move lower in a corresponding fashion.
Trader
Bid
Ask
Num. of shares
Seller S1
market
Buyer B1
Buyer B2
$97
1000
300
Buyer B3
$95
200
Buyer B4
$99
500
$98
100
The above example is somewhat exaggerated but it illustrates the point. In markets with low
volume it is possible for one or more large transactions to have significant impact on prices. This
can happen around holidays and other vacation kinds of periods when fewer traders are active,
and it can happen in the markets that are thinly traded (lack liquidity) in the first place.
When a trader wishes to arrange a trade, he places an order, which is a request for a trade yet to
be executed. An order is an instruction to a broker/dealer to buy, sell, deliver, or receive
securities that commits the issuer of the “order” to the terms specified. An order ticket is a form
detailing the parameters of an Order instruction. Buy or sell orders differ in terms of the time
limit, price limit, discretion of the broker handling the order, and nature of the stock-ownership
position (explained below). Four types of orders are most common and frequently used:
1. Market order: An order from a trader to a broker to buy or sell a stock at the best
available price. The broker should execute the order immediately, but may wait for a
favorable price improvement. A market order to buy 10 shares of Google means buy the
stock at whatever the lowest ask (offer) price is at the time the trade is executed. The
broker could pay more (or less) than the price quoted to the trader, because in the
meantime the market may have shifted (also recall the above example). Market orders are
the quickest but not necessarily the optimal way to buy or sell a security.
2. Limit order: An order to buy or sell at a specific price, or better. The trader using a limit
order specifies the maximum buy price or the minimum sale price at which the
transaction shall be executed. That means when buying it would be at the limit price or
below, while the reverse is true for a sell order. For example, a limit order to sell 100
Google shares at 600 means the trade will be executed at or above 600. A limit order can
only be filled if the stock’s market price reaches the limit price.
3. Stop order: (also referred to as a stop-loss order) A delayed market order to buy or sell a
security when a certain price is reached or passed. A stop order is set at a point above (for
a buy) or below (for a sell) the current price. When the current price reaches or passes
through the specified level, the stop order is converted into an active market order
(defined above in item 1). For example, a sell stop at 105 would be triggered if the
market price touches or falls below 105. A buy stop order is entered at a stop price above
the current market price. Investors generally use a buy stop order to limit a loss or to
protect a profit on a stock that they have sold short. A sell stop order is entered at a stop
Ivan Marsic

34
Rutgers University
owning stock
means
long position
profits through
stock price appreciation
is exited by
selling stock
creates
profit
profit = salePrice – purchasePrice – commissions
owning cash earned by selling borrowed stock
means
short position
profits through
stock price depreciation
is exited by
buying stock
(a)
creates
returning borrowed stock
trader
profit
profit = salePrice – purchasePrice – loanInterest – commissions
expects
share price depreciation
1
broker
2
performs
borrows shares from
are
held by
selling borrowed shares
triggers
creates a
is
executed
short position
at
generates
collects
sale proceeds
commissions
3
interest on loan
bid price
interest on sale proceeds
4
are returned to
repossessed shares
(b)
is exited by
buying shares
yields
is executed at
ask price
Figure 1-22: (a) Concept map of two types of stock-ownership positions: long and short.
(b) Concept map explaining how short position functions.
price below the current market price. Investors generally use a sell stop order to limit a
loss or to protect a profit on a stock that they own.
4. Stop Limit Order: A combination of the stop and limit orders. Unlike the simple stop
order, which is converted into a market order when a certain price is reached, the stop
limit order is converted into a limit order. Hence, the trader can control the price at which
the order can be executed and will get a fill at or better than the limit order price.
For information on other, more advanced order types, the reader should search the Web. There
are two types of security-ownership positions: long and short, see Figure 1-22(a). A long position
represents actual ownership of the security regardless of whether personal funds, financial
leverage (borrowed funds), or both are used in its purchase. Profits are realized if the price of the
security increases.
Chapter 1 
Introduction
35
A short position involves first a sale of the stock, followed by a purchase at, it is hoped, a lower
price, Figure 1-22(b). The trader is “short” (does not own the stock) and begins by borrowing a
stock from the investment broker, who ordinarily holds a substantial number of shares and/or has
access to the desired stock from other investment brokers. The trader then sells the borrowed
stock at the market price. The short position holder owes the shares to the broker; the short
position can be covered by buying back the shares and returning the purchased shares to the
broker to settle the loan of shares. This sequence of steps is labeled by numbers in Figure 1-22(b).
The trader hopes that the stock price will drop and the difference between the sale price and the
purchase price will result in a positive profit.
One can argue that there is no such thing as a “bad market,” there is only the wrong position in
the market. If the trader believes that a particular stock will move upwards, he should establish a
long position. Conversely, if he believes that the stock will slide, he should establish a short
position6. The trader can also hedge his bets by holding simultaneously both long and short
positions on the same stock.
Computerized Support for Individual Investor Trading
We need to consider several choices and constraints for the system-to-be. First, we need to decide
whether the system-to-be will provide brokerage services, or will just provide trading advice.
Online brokerage firms already offer front-end systems for traders, so it will be difficult to insert
our system-to-be between a trader and a broker. Offering our system-to-be as tool for on-a-side
analysis (out of the trading loop) would have limited appeal. The other option is to include
brokerage services, which will introduce significant complexity into the system. An important
constraint on applicability of our system is that real-time price quotations currently are not
available for free. We choose to consider both options in this book. The first five chapters will
consider a case study of a system that includes a trader/broker services. Chapter 8 on Web
services will consider stock analyst services. Both versions are described in Section 1.5.
Knowing how to place a trading order does not qualify one as a trader. It would be equivalent of
saying that one knows how to drive a car just after learning how to use the steering wheel or the
brake. There is much more to driving a car than just using the steering wheel or the brake.
Similarly, there is much more to trading than just executing trades. To continue with the analogy,
we need to have a “road map,” a “travel plan,” and we also need to know how to read the “road
signs,” and so on.
In general, the trader would care to know if a trading opportunity arose and, once he places a
trading order, to track the status of the order. The help of computer technology has always been
sought by traders for number crunching and scenario analysis. The basic desire is to be able to tell
the future based on the knowledge of the past. Some financial economists view price movements
on stock markets as a purely “random walk,” and believe that the past prices cannot tell us
anything useful about future behavior of the price. Others, citing chaos theory, believe that useful
6
This is the idea of the so called inverse funds, see more here: B. Steverman: “Shorting for the 21st century:
Inverse funds allow investors to place bets on predictions of a drop in stocks,” Business Week, no. 4065,
p. 78, December 31, 2007.
Ivan Marsic

36
Rutgers University
Triangles
?
?
50
50
70
45
45
65
40
40
60
35
35
Symmetrical
Ascending
55
Descending
50
30
30
45
25
25
40
20
20
35
15
M
aj
or
t
“Wedge”
?
15
“Flag”
“Pennant”
re
nd
Head
Left
shoulder
Right
shoulder
Neckline
Neckline
Left
shoulder
Head
Right
shoulder
Head and Shoulders Bottom
a
M
Trading
volume
rt
jo
re
nd
Head and Shoulders Top
Figure 1-23: Technical analysis of stock price trends: Some example types of trend patterns.
In all charts the horizontal axis represents time and the vertical axis stock price range. Each
vertical bar portrays the high and low prices of a particular stock for a chosen time unit.
Source: Alan R. Shaw, “Market timing and technical analysis,” in Sumner N. Levine
(Editor), The Financial Analyst’s Handbook, Second Edition, pp. 312-372, Dow Jones-Irwin,
Inc., Homewood, IL, 1988.
regularities can be observed and exploited. Chaos theory states that seemingly random processes
may in fact have been generated by a deterministic function that is not random [Bao, et al., 2004].
Bao, Yukun, Yansheng Lu, Jinlong Zhang. “Forecasting stock prices by SVMs regression,”
Artificial Intelligence: Methodology, Systems, and Applications, vol. 3192, 2004.
A simple approach is to observe prices pricei(t) of a given stock i over a window of time tcurrent 
Window, …, tcurrent  2, tcurrent  1, tcurrent. We could fit a regression line through the observed
points and devise a rule that a positive line slope represents a buying opportunity, negative slope
a need to sell, and zero slope calls for no action. Obviously, it is not most profitable to buy when
the stock already is gaining nor it is to sell when the stock is already sliding. The worst-case
scenario is to buy at a market top or to sell when markets hit bottom. Ideally, we would like to
detect the turning points and buy when the price is just about to start rising or sell when the price
is just about to start falling. Detecting an ongoing trend is relatively easy; detecting an imminent
onset of a new trend is difficult but most desirable.
Chapter 1 
Introduction
37
This is where technical analysis comes into picture. Technical analysts believe that market prices
exhibit identifiable regularities (or patterns or indicators) that are bound to be repeated. Using
technical analysis, various trends could be “unearthed” from the historical prices of a particular
stock and potentially those could be “projected into future” to have some estimation around
where that stock price is heading. Technical analysts believe that graphs give them the ability to
form an opinion about any security without following it in real time. They have come up with
many types of indicators that can be observed in stock-price time series and various
interpretations of the meaning of those indicators. Some chart formations are shown in Figure
1-23. For example, the triangles and flags represent consolidations or corrective moves in market
trends. A flag is a well-defined movement contrary to the main trend. The head-and-shoulder
formations are used as indicators of trend reversals and can be either top or bottom. In the
“bottom” case, for example, a major market low is flanked on both sides (shoulders) by two
higher lows. Cutting across the shoulders is some resistance level called the neckline. (Resistance
represents price levels beyond which the market has failed to advance.) It is important to observe
the trading volume to confirm the price movements. The increasing volume, as you progress
through the pattern from left to right, tells you that more and more traders see the shifting
improvement in the company’s fortunes. A “breakout” (a price advance) in this situation signals
the end of a downtrend and a new direction in the price trend. Technical analysts usually provide
behavioral explanations for the price action and formation of trends and patterns.
However, one may wonder if just looking at a sequence of price numbers can tell us everything
we need to know about the viability of an investment?! Should we not look for actual causes of
price movements? Is the company in bad financial shape? Unable to keep up with competition?
Or, is it growing rapidly? There is ample material available to the investor, both, in electronic and
in print media, for doing a sound research before making the investment decision. This kind of
research is called fundamental analysis, which includes analysis of important characteristics of
the company under review, such as:
1. Market share: What is the market standing of the company under review? How much share of
the market does it hold? How does that compare against the competitors?
2. Innovations: How is the company fairing in terms of innovations? For example in 3M company
no less than 25% of the revenues come from the innovative products of last 5 years. There is even
an index for innovations available for review and comparison (8th Jan’2007 issue of Business
Week could be referred to).
3. Productivity: This relates the input of all the major factors of production – money, materials
and people to the (inflation adjusted) value of total output of goods and services from the outside
4. Liquidity and Cash-flow: A company can run without profits for long years provided it has
enough cash flows, but hardly the reverse is true. A company, if it has a profitable unit, but not
enough cash flows, ends of “putting that on sale” or “spinning that unit out.”
In addition to the above indicators, number crunching is also a useful way to fine-tune the
decision. Various financial numbers are readily available online, such as
- Sales
- EPS: Earning per Share
- P/E – ttm: Trailing 12 months’ ratio of Price per Share to that of Earning per Share
Ivan Marsic

38
Rutgers University
Trader’s goal G1: To profit from investment
Question: How?
Possible answers:
Trader’s goal G1.1: To identify trading opportunity
Trader’s goal G1.3: To track order status
Question: How?
Possible answers:
Trader’s goal G1.2: To ensure timely & reliable transaction
Long-term investor’s goal G1.2.1: To identify growth/value stock
Short-term trader’s goal G1.2.2: To identify arbitrage opportunity (“indicators” in time series)
Figure 1-24: Example of refining the representation of user’s goals.
- P/E – forward: Ratio of Estimated Price per Share for coming 12 months to that of Estimated
Earning of coming 12 months
- ROI: Return on Investment
The key barometer of stock market volatility is the Chicago Board Options Exchange's Volatility
Index, or VIX, which measures the fluctuations of options contracts based on the S&P 100-stock
index.
In fact, one could argue that the single most important decision an investor can make is to get out
of the way of a collapsing market7.
Where the investor is usually found to be handicapped is when she enters into the market with the
objective of short term gains. The stock market, with its inherent volatility offers ample
opportunities to exploit that volatility but what the investor lacks is an appropriate tool to assist in
this “decision-making” process.
The investor would ideally like to “enter” the market after it is open and would “exit” the market
before it is closed, by the end of that day. The investor would seek a particular stock, the price of
which she is convinced would rise by the end of the day, would buy it at a “lower” price and
would sell it at a higher price. If she gets inkling, somehow, that a particular stock is going to go
up, it will be far easier for her to invest in that stock. Usually time is of essence here and this is
where technical analysis comes into picture.
Again, we must clearly state what the user needs: the user’s goals. It is not very helpful to state
that the user’s goal is “to make money.” We must be as specific as possible, which can be
achieved by keeping asking questions “How?” An example of goal refinement is shown in Figure
1-24. Note that in answering how to identify a trading opportunity, we also need to know whether
our trader has a short-term or long-term outlook to investment. In addition, different trader types
may compose differently the same sub-goals (low-level goals) into high-level goals. For example,
the long-term investor would primarily consider the company’s prospects (G1.2.1), but may
employ time-series indicators (G1.2.2) to decide the timing of their investments. Just because one
7
Michael Mandel, “Bubble, bubble, who’s in trouble?” Business Week, p. 34, June 26, 2006.
Chapter 1 
Introduction
39
anticipates that an investment will be held for several years because of its underlying
fundamentals, that does not mean that he should overlook the opportunity for buying at a lower
price (near the bottom of an uptrend).
It is important to understand the larger context of the problem that we are trying to solve. There
are already many people who are trying to forecast financial markets. Companies and
governments spend vast quantities of resources attempting to predict financial markets. We have
to be realistic of what we can achieve with relatively minuscule resources and time period of one
academic semester.
From the universe of possible market data, we have access only to a subset, which is both due to
economic (real-time data are available with paid subscription only) and computational (gathering
and processing large data quantities requires great computing power) reasons. Assuming we will
use freely available data and a modest computing power, the resulting data subset is suitable only
for certain purposes. By implication, this limits our target customer and what he can do with the
software-to-be.
In conclusion, our planned tool is not for a “professional trader.” This tool is not for institutional
investor or large brokerage/financial firm. This tool is for an ordinary single investor who does
not have acumen of financial concepts, yet would like to trade smartly. This tool is for an investor
who does not have too much time to do a thorough research on all aspects of a particular
company, neither does he have understanding and mastery over financial number crunching. It is
unlikely to be used for “frequency trading,” because we lack computing power and domain
knowledge needed for such sophisticated uses.
1.4 The Object Model
“You cannot teach beginners top-down programming, because they don’t know which end is up.”
—C.A.R. Hoare
An object is a software packaging of data and code together into a unit within a running
computer program. Objects can interact by calling other objects for their services. In Figure 1-25,
object Stu calls the object Elmer to find out if 905 and 1988 are coprimes. Two integers are
said to be coprime or relatively prime if they have no common factor other than 1 or,
equivalently, if their greatest common divisor is 1. Elmer performs computation and answers
positively. Objects do not accept arbitrary calls. Instead, acceptable calls are defined as a set of
object “methods.” This fact is indicated by the method areCoprimes() in Figure 1-25. A
method is a function (also known as operation, procedure, or subroutine) associated with an
object so that other objects can call on its services. Every software object supports a limited
number of methods. Example methods for an ATM machine object are illustrated in Figure 1-26.
The set of methods along with the exact format for calling each method (known as the method
“signature”) represents the object’s interface (Figure 1-27). The interface specifies object’s
behavior—what kind of calls it accepts and what it does in response to each call.
Ivan Marsic

40
Rutgers University
Prime factorization:
elmer . areCoprimes(
905, 1988
)
905 = 5  181
1988 = 2  2  7  71
Result:
YES!
Stu
Elmer
Figure 1-25: Client object sends a message to a server object by invoking a method on it.
Server object is the method receiver.
Software objects work together to carry out the tasks required by the program’s business logic. In
object-oriented terminology, objects communicate with each other by sending messages. In the
world of software objects, when an object A calls a method on an object B we say, “A sends a
message to B.” In other words, a client object requests the execution of a method from a server
object by sending it a message. The message is matched up with a method defined by the
software class to which the receiving object belongs. Objects can alternate between a client role
and a server role. An object is in a client role when it is the originator of an object invocation, no
matter whether the objects are located in the same memory space or on different computers. Most
objects play both client and server roles.
In addition to methods, software objects have attributes or properties. An attribute is an item of
data named by an identifier that represents some information about the object. For example, a
person’s attribute is the age, or height, or weight. The attributes contain the information that
differentiates between the various objects. The currently assigned values for object attributes
describe the object’s internal state or its current condition of existence. Everything that a
software object knows (state) and can do (behavior) is expressed by the attributes and the
methods within that object. A class is a collection of objects that share the same set of attributes
and methods (i.e., the interface). Think of a class as a template or blueprint from which objects
are made. When an instance object is created, we say that the objects are instantiated. Each
instance object has a distinct identity and its own copy of attributes and methods. Because objects
are created from classes, you must design a class and write its program code before you can
create an object.
Objects also have special methods called constructors, which are called at the creation of an
object to “construct” the values of object’s data members (attributes). A constructor prepares the
new object for use, often accepting parameters which the constructor uses to set the attributes.
Unlike other methods, a constructor never has a return value. A constructor should put an object
in its initial, valid, safe state, by initializing the attributes with meaningful values. Calling a
constructor is different from calling other methods because the caller needs to know what values
are appropriate to pass as parameters for initialization.
Chapter 1 
41
Introduction
Object:
ATM machine
1
4 2
7 85 6 3
0 9
method-1:
Accept card
method-3:
Take selection
method-2:
Read code
1234
5678
1
12345
2
4
7
8
0
5
3
6
9
Figure 1-26: Acceptable calls are defined by object “methods,” as shown here by example
methods for an ATM machine object.
Interface
attributes
method-1
method-2
method-3
Figure 1-27: Software object interface is a set of object’s methods with the format for
calling each method.
Ivan Marsic

42
Rutgers University
Traditional approach to program development, known as procedural approach, is process
oriented in that the solution is represented as a sequence of steps to be followed when the
program is executed. The processor receives certain input data and first does this, then that, and
so on, until the result is outputted. The object-oriented approach starts by breaking up the whole
program into software objects with specialized roles and creating a division of labor. Objectoriented programming then, is describing what messages get exchanged between the objects in
the system. This contrast is illustrated on the safe home access system case study (Section 1.3.1).
Example 1.1
Procedural approach versus Object-oriented approach
The process-based or procedural approach represents solution as a sequence of steps to be followed
when the program is executed, Figure 1-28(a). It is a global view of the problem as seen by the single
agent advancing in a stepwise fashion towards the solution. The step-by-step approach is easier to
understand when the whole problem is relatively simple and there are few alternative choices along the
path. The problem with this approach is when the number of steps and alternatives becomes
overwhelming.
Object-oriented (OO) approach adopts a local view of the problem. Each object specializes only in a
relatively small subproblem and performs its task upon receiving a message from another object,
Figure 1-28(b). Unlike a single agent travelling over the entire process, we can think of OO approach
as organizing many tiny agents into a “bucket brigade,” each carrying its task when called upon, Figure
1-28(c). When an object completes its task, it sends a message to another object saying “that does it for
me; over to you—here’s what I did; now it’s your turn!” Here are pseudo-Java code snippets for two
objects, KeyChecker and LockCtrl:
Listing 1-1: Object-oriented code for classes KeyChecker (left) and LockCtrl (right).
public class KeyChecker {
protected LockCtrl lock_;
protected java.util.Hashtable
validKeys_;
...
public class LockCtrl {
protected boolean
locked_ = true; // start locked
protected LightCtrl switch_;
...
/** Constructor */
public KeyChecker(
LockCtrl lc, ...) {
lock_ = lc;
...
}
/** Constructor */
public LockCtrl(
LightCtrl sw, ...) {
switch_ = sw;
...
}
/** This method waits for and
* validates the user-supplied key
*/
public keyEntered(
String key
) {
if (
validKeys.containsKey(key)
) {
lock_.unlock(id);
}
} else {
// deny access
// & sound alarm bell?
}
/** This method sets the lock state
* and hands over control to the switch
*/
public unlock() {
... operate the physical lock device
locked_ = false;
switch_.turnOn();
}
public lock(boolean light) {
... operate the physical lock device
locked_ = true;
if (light) {
switch_.turnOff();
}
Chapter 1 
43
Introduction
}
}
}
}
Two important observations:
Object roles/responsibilities are focused (each object is focused on one task, as its name says); later,
we will see that there are more responsibilities, like calling other objects
Object’s level of abstraction must be carefully chosen: here, we chose key checker and its method
keyEntered(), instead of specifying the method of key entry (type in code vs. acquire biometric
identifier), and LockCtrl does not specify how exactly the lock device functions. Too low level
specifies such details (which could be specified in a derived class), or too high abstraction level just
says control-the-access().
The key developer skill in object-oriented software development is performing the division of labor for
software objects. Preferably, each object should have only one clearly defined task (or, responsibility)
and that is relatively easy to achieve. The main difficulty in assigning responsibilities arises when an
object needs to communicate with other objects in accomplishing a task.
When something goes wrong, you want to know where to look or whom to single out. This is
particularly important for a complex system, with many functions and interactions. Object-oriented
approach is helpful because the responsibilities tend to be known. However, the responsibilities must
be assigned adequately in the first place. That is why assigning responsibilities to software objects is
probably the most important skill in software development. Some responsibilities are obvious. For
example, in Figure 1-28 it is natural to assign the control of the light switch to the LightCtrl object.
However, assigning the responsibility of communicating messages is harder. For example, who should
send the message to the LightCtrl object to turn the switch on? In Figure 1-28, LockCtrl is charged
with this responsibility. Another logical choice is KeyChecker, perhaps even more suitable, because it
is the KeyChecker who ascertains the validity of a key and knows whether or not unlocking and
lighting actions should be initiated. More details about assigning responsibilities are presented in
Section 2.6.
The concept of objects allows us to divide software into smaller pieces to make it manageable.
The divide-and-conquer approach goes under different names: reductionism, modularity, and
structuralism. The “object orientation” is along the lines of the reductionism paradigm: “the
(b)
Valid
key
?
(a)
Key
Key
Checker
Checker
Lock
Lock
Ctrl
Ctrl
turnOn()
Light
Light
Ctrl
Ctrl
No
Yes
Unlock the
lock
unlock()
unlock()
(c)
Key
Key
Checker
Checker
turnOn()
Lock
Lock
Ctrl
Ctrl
Light
Light
Ctrl
Ctrl
Turn the
light on
Figure 1-28: Comparison of process-oriented (procedural) and object-oriented methods on
the safe home access case study. (a) A flowchart for a procedural solution; (b) An objectoriented solution. (c) An object can be thought of as a person with expertise and
responsibilities.
Ivan Marsic

Rutgers University
44
tendency to or principle of analysing complex things into simple constituents; the view that a
system can be fully understood in terms of its isolated parts, or an idea in terms of simple
concepts” [Concise Oxford Dictionary, 8th Ed., 1991]. If your car does not work, the mechanic
looks for a problem in one of the parts—a dead battery, a broken fan belt, or a damaged fuel
pump. A design is modular when each activity of the system is performed by exactly one unit,
and when inputs and outputs of each unit are well defined. Reductionism is the idea that the best
way to understand any complicated thing is to investigate the nature and workings of each of its
parts. This approach is how humans solve problems, and it comprises the very basis of science.
SIDEBAR 1.1: Object Orientation
 Object orientation is a worldview that emerged in response to real-world problems faced by
software developers. Although it has had many successes and is achieving wide adoption, as
with any other worldview, you may question its soundness in the changing landscape of
software development. OO stipulates that data and processing be packaged together, data being
encapsulated and unreachable for external manipulation other than through object’s methods. It
may be likened to disposable cameras where film roll (data) is encapsulated within the camera
mechanics (processing), or early digital gadgets with a built-in memory. People have not really
liked this model, and most devices now come with a replaceable memory card. This would
speak against the data hiding and for separation of data and processing. As we will see in
Chapter 8, web services are challenging the object-oriented worldview in this sense.
There are three important aspects of object orientation that will be covered next:

Controlling access to object elements, known as encapsulation

Object responsibilities and relationships

Reuse and extension by inheritance and composition
1.4.1
Controlling Access to Object Elements
Modular software design provides means for breaking software into meaningful components,
Figure 1-29. However, modules are only loose groupings of subprograms and data. Because there
is no strict ownership of data, subprograms can infringe on each other’s data and make it difficult
to track who did what and when. Object oriented approach goes a step further by emphasizing
state encapsulation, which means hiding the object state, so that it can be observed or modified
only via object’s methods. This approach enables better control over interactions among the
modules of an application. Traditional software modules, unlike software objects, are more
“porous;” encapsulation helps prevent “leaking” of the object state and responsibilities.
In object-orientation, object data are more than just program data—they are object’s attributes,
representing its individual characteristics or properties. When we design a class, we decide what
internal state it has and how that state is to appear on the outside (to other objects). The internal
state is held in the attributes, also known as class instance variables. UML notation for software
class is shown in Figure 1-30. Many programming languages allow making the internal state
directly accessible through a variable manipulation, which is a bad practice. Instead, the access to
object’s data should be controlled. The external state should be exposed through method calls,
Chapter 1 
(a)
45
Introduction
Subprograms
(behavior)
Data
(state)
Software Module 1
Methods
(behavior)
Software Module 2
Software Module 3
Attributes
/data
(state)
(b)
Software Object 1
Software Object 2
Software Object 3
Figure 1-29: Software modules (a) vs. software objects (b).
called getters and setters, to get or set the instance variables. Getters and setters are sometimes
called accessor and mutator methods, respectively. For example, for the class LightController in
Figure 1-31 the getter and setter methods for the attribute lightIntensity are
getLightIntensity() and setLightIntensity(), respectively. Getter and setter
methods are considered part of object’s interface. In this way, the interface exposes object’s
behavior, as well as its attributes via getters and setters.
Access to object attributes and methods is controlled using access designations, also known as
visibility of attributes and methods. When an object attribute or method is defined as public,
other objects can directly access it. When an attribute or method is defined as private, only
that specific object can access it (not even the descendant objects that inherit from this class).
Another access modifier, protected, allows access by related objects, as described in the next
section. The UML symbols for access designations in class diagrams are as follows (Figure 1-30):
+ for public, global visibility; # for protected visibility; and, − for private within-theclass-only visibility.
We separate object design into three parts: its public interface, the terms and conditions of use
(contracts), and the private details of how it conducts its business (known as implementation).
The services presented to a client object comprise the interface. The interface is the fundamental
means of communication between objects. Any behavior that an object provides must be invoked
by a message sent using one of the provided interface methods. The interface should precisely
describe how client objects of the class interact with the class. Only the methods that are
designated as public comprise the class interface (“+” symbol in UML class diagrams). For
example, in Figure 1-31 the class HouseholdDeviceController has three public methods that
constitute its interface. The private method sendCommandToUSBport() is not part of the
Ivan Marsic

46
Rutgers University
ClassName
private attributes
protected attribute
public operations
protected operation
– attribute_1 : int
– attribute_2 : boolean
# attribute_3 : String
Three compartments:
1.
Classifier name
+ operation_1() : void
+ operation_2() : String
# operation_3(arg1 : int)
2.
Attributes
3.
Operations
Figure 1-30: UML notation for software class.
interface. Note that interfaces do not normally include attributes—only methods. If a client needs
to access an attribute, it should use the getter and setter methods.
Encapsulation is fundamental to object orientation. Encapsulation is the process of packaging
your program, dividing its classes into the public interface and the private implementation. The
basic question is, what in a class (which elements) should be exposed and what should be hidden.
This question pertains equally to attributes and behavior. (Recall that attributes should never be
exposed directly, but instead by using getter and setter methods.) Encapsulation hides everything
that is not necessary for other classes to know about. By localizing attributes and behaviors and
preventing logically unconnected functions from manipulating object elements, we ensure that a
change in a class will not cause a rippling effect around the system. This property makes for
easier maintaining, testing, and extending the classes.
Object orientation continues with the black-box approach of focusing on interface. In Section
1.2.2, the whole system was considered as a black box, and here we focus on the micro-level of
individual objects. When specifying an interface, we are only interested in what an object does,
not how it does it. The “how” part is considered in implementation. Class implementation is the
program code that specifies how the class conducts its business, i.e., performs the computation.
Normally, the client object does not care how the computation is performed as long as it produces
the correct answer. Thus, the implementation can change and it will not affect the client’s code.
For example, in Figure 1-25, object Stu does not care that the object Elmer answers if numbers
are coprimes. Instead, it may use any other object that provides the method areCoprimes() as
part of its interface.
anymathematician . areCoprimes(
905, 1988
)
[ some correct
computation ]
Result:
YES!
Stu
Any Mathematician
Chapter 1 
Introduction
47
Contracts can specify different terms and conditions of object. Contract may apply at design time
or at run time. Programming languages such as Java and C# have two language constructs for
specifying design-time contracts.
Run time contracts specify the conditions under which an object methods can be called upon
(conditions-of-use guarantees), and what outcome methods achieve when they are finished
(aftereffect guarantees).
It must be stressed that the interchangeable objects must be identical in every way—as far as the
client object’s perceptions go.
1.4.2
Object Responsibilities and Relationships
The key characteristic of object-orientation is the concept of responsibility that an object has
towards other objects. Careful assignment of responsibilities to objects makes possible the
division of labor, so that each object is focused on its specialty. Other characteristics of object
orientation, such as polymorphism, encapsulation, etc., are characteristics local to the object
itself. Responsibilities characterize the whole system design. To understand how, you need to read
Chapters 2, 4, and 5. Because objects work together, as with any organization you would expect
that the entities have defined roles and responsibilities. The process of determining what the
object should know (state) and what it should do (behavior) is known as assigning the
responsibilities. What are object’s responsibilities? The key object responsibilities are:
1. Knowing something (memorization of data or object attributes)
2. Doing something on its own (computation programmed in a “method”)
3. Calling methods of other objects (communication by sending messages)
We will additionally distinguish a special type of doing/computation responsibilities:
2.a) Business rules for implementing business policies and procedures
Business rules are important to distinguish because, unlike algorithms for data processing and
calculating functions, they require knowledge of customer’s business context and they often
change. We will also distinguish communication responsibilities:
3.a) Calling constructor methods; this is special because the caller must know the
appropriate parameters for initialization of the new object.
Assigning responsibilities essentially means deciding what methods an object gets and who
invokes those methods. Large part of this book deals with assigning object responsibilities,
particularly Section 2.6 and Chapter 5.
The basic types of class relationships are inheritance, where a class inherits elements of a base
class, and composition, where a class contains a reference to another class. These relationships
can be further refined as:
Base Class
Container

Is-a relationship (hollow triangle symbol ∆ in UML diagrams): A class “inherits” from
another class, known as base class, or parent class, or superclass

Has-a relationship: A class “contains” another class
Ivan Marsic

Rutgers University
48
-
Composition relationship (filled diamond symbol ♦ in UML diagrams): The
contained item is an integral part of the containing item, such as a leg in a desk
-
Aggregation relationship (hollow diamond symbol ◊): The contained item is an
element of a collection but it can also exist on its own, such as a desk in an office

Uses-a relationship (arrow symbol ↓ in UML diagrams): A class “uses” another class

Creates relationship: A class “creates” another class (calls a constructor method)
Has-a and Uses-a relationships can be seen as types of composition.
1.4.3
Reuse and Extension by Inheritance and
Composition
One of the most powerful characteristics of object-orientation is code reuse. Procedural
programming provides code reuse to a certain degree—you can write a procedure and then reuse
it many times. However, object-oriented programming goes an important step further, allowing
you to define relationships between classes that facilitate not only code reuse, but also better
overall design, by organizing classes and factoring in commonalities of various classes.
Two important types of relationships in the object model enable reuse and extension: inheritance
and composition. Inheritance relations are static—they are defined at the compile time and cannot
change for the object’s lifetime. Composition is dynamic, it is defined at run time, during the
participating objects’ lifetimes, and it can change.
When a message is sent to an object, the object must have a method defined to respond to that
message. The object may have its own method defined as part of its interface, or it may inherit a
method from its parent class. In an inheritance hierarchy, all subclasses inherit the interfaces from
their superclass. However, because each subclass is a separate entity, each might require a
separate response to the same message. For example, in Figure 1-31 subclasses Lock Controller
and Light Controller inherit the three public methods that constitute the interface of the superclass
Household Device Controller. The private method is private to the superclass and not available to
the derived subclasses. Light Controller overrides the method activate() that it inherits from
its superclass, because it needs to adjust the light intensity after turning on the light. The method
deactivate() is adopted unmodified. On the other hand, Lock Controller overrides both
methods activate() and deactivate() because it requires additional behavior. For
example, in addition to disarming the lock, Lock Controller’s method deactivate() needs to
start the timer that counts down how long time the lock has remained unlocked, so it can be
automatically locked. The method activate() needs to clear the timer, in addition to arming
the lock. This property that the same method behaves differently on different subclasses of the
same class is called polymorphism.
Inheritance applies if several objects have some responsibilities in common. The key idea is to
place the generic algorithms in a base class and inherit them into different detailed contexts of
derived classes. With inheritance, we can program by difference. Inheritance is a strong
relationship, in that the derivatives are inextricably bound to their base classes. Methods from the
base class can be used only in its own hierarchy and cannot be reused in other hierarchies.
Chapter 1 
49
Introduction
HouseholdDeviceController
– deviceStatus : boolean
+
+
+
–
activate( )
deactivate( )
isActivated( ) : boolean
sendCommandToUSBport(cmd : string)
LockController
Inheritance
relationship:
Base class
is extended
by two classes
LightController
– autoLockInterval : long
– lightIntensity : int
+
+
–
–
+ activate( )
+ getLightIntensity(value : int)
+ setLightIntensity( ) : int
activate( )
deactivate( )
startAutolockTimer( )
performAutoLock( ) : boolean
Figure 1-31: Example of object inheritance.
1.5 Student Team Projects
“Knowledge must come through action; you can have no test which is not fanciful, save by trial.”
—Sophocles
“I have been impressed with the urgency of doing. Knowing is not enough; we must apply.
Being willing is not enough; we must do.” —Leonardo da Vinci
The book website, given in Preface, describes several student team projects. These projects are
selected so each can be accomplished by a team of undergraduate students in the course of one
semester. At the same time, the basic version can be extended so to be suitable for graduate
courses in software engineering and some of the projects can be extended even to graduate theses.
Here I describe only two projects and more projects along with additional information about the
projects described here is available at the book’s website, given in Preface.
Each project requires the student to learn one or more technologies specific for that project. In
addition, all student teams should obtain a UML diagramming tool.
1.5.1
Stock Market Investment Fantasy League
This project is fashioned after major sports fantasy leagues, but in the stock investment domain.
You are to build a website which will allow investor players to make virtual investments in realworld stocks using fantasy money. The system and its context are illustrated in Figure 1-32. Each
Ivan Marsic

50
Rutgers University
Stock Market Investment Fantasy League System
Servers
and data
storage
Web
clients
System
admin
Players
Advertisers
Payment
system
Stock
reporting
website
(e.g., PayPal)
(e.g., Google/Yahoo! Finance)
Real-world
stock
exchanges
Figure 1-32: Stock market fantasy league system and the context within which it operates.
player has a personal account with fantasy money in it. Initially, the player is given a fixed
amount of startup funds. The player uses these funds to virtually buy the stocks. The system then
tracks the actual stock movement on real-world exchanges and periodically adjusts the value of
players’ investments. The actual stock prices are retrieved from a third-party source, such as
Yahoo! Finance, that monitors stock exchanges and maintains up-to-date stock prices. Given a
stock in a player’s portfolio, if the corresponding actual stock loses value on a real-world stock
exchange, the player’s virtual investment loses value equally. Likewise, if the corresponding
actual stock gains value, the player’s virtual investment grows in the same way.
The player can sell the existing stocks or buy new ones at any time. This system does not provide
any investment advice. When player sells a stock, his/her account is credited with fantasy money
in the amount that corresponds to the current stock price on a stock exchange. A small
commission fee is charged on all trading transactions (deducted from the player’s account).
Your business model calls for advertisement revenues to support financially your website.
Advertisers who wish to display their products on your website can sign-up at any time and create
their account. They can upload/cancel advertisements, check balance due, and make payments
(via a third party, e.g., a credit card company or PayPal.com). Every time a player navigates to a
new window (within this website), the system randomly selects an advertisement and displays the
advertisement banner in the window. At the same time, a small advertisement fee is charged on
the advertiser’s account. A more ambitious version of the system would fetch an advertisement
dynamically from the advertiser’s website, just prior to displaying it.
To motivate the players, we consider two mechanisms. One is to remunerate the best players, to
increase the incentive to win. For example, once a month you will award 10 % of advertisement
profits to the player of the month. The remuneration is conducted via a third party, such as
PayPal.com. In addition, the system may support learning by analyzing successful traders and
extracting information about their trading strategies. The simplest service may be in the form
stock buying recommendations: “players who bought this stock also bought these five others.”
More complex strategy analysis may be devised.
Chapter 1 
51
Introduction
User Functions
Web Client for Players
Web Client for Advertisers
• Registration
• Account/Portfolio management
• Trading execution & history
• Stock browsing/searching
• Viewing market prices & history
• Account management
• Banner uploading and removal
• Banner placement selection
System Administration
Functions
• User management
• Selection of players for awards
• Dashboard for monitoring
the league activities
Backend Operations
Player Management
• Account balance
• Trading support
• Transactions archiving
• Portfolio valuation
• Periodic reporting (email)
Real-World Market
Observation & Analysis
• Retrieval of stock prices
- On-demand vs. periodic
• Analysis
- Technical & fundamental
•?
Advertiser Management
• Account balance
• Uploading new banners
• Banner placement selection
League Management
• Player ranking
• Awards disbursement control
• Trading performance analysis
• Coaching of underperformers
Figure 1-33: Logical grouping of required functions for Stock Market Fantasy League.
Statement of Requirements
Figure 1-33 shows logical grouping of functions requested from our system-to-be.
Player portfolio consists of positions—individual stocks owned by the player. Each position
should include company name, ticker symbol, the number of shares owned by this player, and
date and price when purchased. Player should be able to specify stocks to be tracked without
owning any of those stocks. Player should also be able to specify buy- and sell thresholds for
various stocks; the system should alert (via email) the player if the current price exceeds any of
these thresholds.
Stock prices should be retrieved periodically to valuate the portfolios and at the moment when the
user wishes to trade. Because price retrieval can be highly resource demanding, the developer
should consider smart strategies for retrieval. For example, cache management strategies could be
employed to prioritize the stocks based on the number of players that own it, the total number of
shares owned, etc.
Additional Information
I would strongly encourage the reader to look at Section 1.3.2 for an overview of financial
investment. Additional information about this project can be found at the book website, given in
Preface.
http://finance.yahoo.com/
Ivan Marsic

Rutgers University
52
http://www.marketwatch.com/
See also Problem 2.29 and Problem 2.32 at the end of Chapter 2, the solutions of which can be
found at the back of the text.
1.5.2
Web-based Stock Forecasters
“Business prophets tell what is going to happen, business profits tell what has happened.” —Anonymous
There are many tools available to investors but none of them removes entirely the element of
chance from investment decisions. Large trading organizations can employ sophisticated
computer systems and armies of analysts. Our goal is to help the individual investor make better
investment decisions. Our system will use the Delphi method,8 which is a systematic interactive
forecasting method for obtaining consensus expectation from a panel of independent experts.
The goal of this project is to have multiple student teams implement Web services (Chapter 8) for
stock-prediction. Each Web service (WS) will track different stocks and, when queried, issue a
forecast about the price movement for a given stock. The client module acts as a “facilitator”
which gathers information from multiple Web services (“independent experts”) and combines
their answers into a single recommendation. If different Web services offer conflicting answers,
the client may repeat the process of querying and combining the answers until it converges
towards the “correct” answer.
There are three aspects of this project that we need to decide on:
 What kind of information should be considered by each forecaster? (e.g., stock prices, trading
volumes, fundamental indicators, general economic indicators, latest news, etc. Stock prices
and trading volumes are fast-changing so must be sampled frequently and the fundamental
and general-economy indicators are slow-moving so could be sampled at a low frequency.)
 Who is the target customer? Organization or individual, their time horizon (day trader vs. longterm investor)
 How the application will be architected? The user will run a client program which will poll the
WS-forecasters and present their predictions. Should the client be entirely Web-based vs.
locally-run application? A Web-based application would be downloaded over the Web every
time the user runs the client; it could be developed using AJAX or a similar technology.
As a start, here are some suggested answers:
 Our target customers are individuals who are trading moderately frequently (up to several times
per week), but not very frequently (several times per day).
 The following data should be gathered and stored locally. Given a list of about 50–100
companies, record their quoted prices and volumes at the maximum available sampling
8
An introductory description is available here: http://en.wikipedia.org/wiki/Delphi_method . An in-depth review
is available here: http://web.njit.edu/~turoff/Papers/delphi3.html (M. Turoff and S. R. Hiltz: “Computer Based
Delphi Processes,” in M. Adler and E. Ziglio (Editors), Gazing Into the Oracle: The Delphi Method and
Its Application to Social Policy and Public Health, London, UK: Kingsley Publishers, 1995.)
Chapter 1 
53
Introduction
density (check http://finance.yahoo.com/); also record some broad market indices, such as DJIA
or S&P500.
 The gathered data should be used for developing the prediction model, which can be a simple
regression-curve fitting, artificial neural network, or some other statistical method. The model
should consider both the individual company’s data as well as the broad market data. Once
ready for use, the prediction model should be activated to look for trends and patterns in stock
prices as they are collected in real time.
Potential services that will be provided by the forecaster service include:

Given a stock x, suggest an action, such as “buy,” “sell,” “hold,” or “sit-out;” we will
assume that the forecaster provides recommendation for one stock at a time

Recommend a stock to buy, from all stocks that are being tracked, or from all in a given
industry/sector
A key step in specifying the forecaster service is to determine its Web service interface: what will
go in and what will come out of your planned Web service? Below I list all the possible
parameters that I could think of, which the client and the service could exchange. The
development team should use their judgment to decide what is reasonable and realistic for their
own team to achieve within the course of an academic semester, and select only some of these
parameters for their Web service interface.
Parameters sent by the facilitator to a forecaster (from the client to a Web service) in the
inquiry include:

Stock(s) to consider: individual (specified by ticker symbol), select-one-for-sector (sector
specified by a standard category), any (select the best candidate)

Trade to consider: buy, sell, hold, sit-out

Time horizon for the investment: integer number

Funds available: integer number for the capital amount/range

Current portfolio (if any) or current position for the specified symbol
OR Position to consider: long, short, any
Some of these parameters may not be necessary, particularly in the first instantiation of the
system. Also, there are privacy issues, particularly with the last two items above, that must be
taken into account. The forecaster Web-services are run by third parties and the trader may not
wish to disclose such information to third parties.
Results returned by a forecaster to the facilitator (for a single stock per inquiry):

Selected stock (if the inquiry requested selection from “sector” or “any”)

Prediction: price trend or numeric value at time t in the future

Recommended action and position: buy, sell, hold, sit-out, go-short

Recommended horizon for the recommended action: time duration

Recommendation about placing a protective sell or buy Stop Order.

Confidence level (how confident is the forecaster about the prediction): range 0 – 100 %
Ivan Marsic

Rutgers University
54
The performance of each prediction service should be evaluated as follows. Once activated, each
predicted price value should be stored in a local database. At a future time when the actual value
becomes known, it should be recorded along with the previously predicted value. A large number
of samples should be collected, say over the period of tens of days. We use absolute mean error
and average relative error as indices for performance evaluation. The average relative error is
defined as i yi  yˆ i i yi , where yi and ŷi are the actual and predicted prices at time i,


respectively.
Statement of Requirements
Extensions
Risk analysis to analyze “what if” scenarios.
Additional information about this project can be found at the book website, given in Preface.
1.5.3
Remarks about the Projects
My criteria in the selection of these projects was that they are sufficiently complex so to urge the
students to enrich their essential skills (creativity, teamwork, communication) and professional
skills (administration, leadership, decision, and management abilities when facing risk or
uncertainty). In addition, they expose the students to at least one discipline or problem domain in
addition to software engineering, as demanded by a labor market of growing complexity, change,
and interdisciplinarity.
The reader should observe that each project requires some knowledge of the problem domain.
Each of the domains has myriads of details and selecting the few that are relevant requires a
major effort. Creating a good model of any domain requires skills and expertise and this is
characteristic of almost all software engineering projects—in addition to software development
skills, you always must learn something else in order to build a software product.
The above projects are somewhat deceptive insofar as the reader may get impression that all
software engineering projects are well defined and the discovery of what needs to be developed is
done by someone else so the developer’s job is just software development. Unfortunately, that is
rarely the case. In most cases the customer has a very vague idea of what they would like to be
developed and the discovery process requires a major effort. That was the case for all of the
above projects—it took me a great deal of fieldwork and help from many people to arrive at the
project descriptions presented above. In the worst case you may not even know who will be your
customer, as is the case for traffic monitoring (described at the book website, given in Preface)
and the investment fantasy league (Section 1.5.1). In such cases, you need to invent your own
customers—you need to identify who might benefit from your product and try and interest them
in participating in the development.
Chapter 1 
Introduction
55
Frederick Brooks, a pioneer of software
engineering, wrote that “the hardest
single part of building a software system
is deciding precisely what to build”
[Brooks, 1995: p. 199]. By this token,
the hardest work on these projects is
already done. The reader should not feel
short-changed,
though,
because
difficulties
in
deriving
system
requirements will be illustrated.
Example 1.2
RFID tags in retail
The following example illustrates of
how a typical idea for a software
engineering project might evolve. The
management of a grocery supermarket
(our customer) contacted us with an idea
for a more effective product promotion.
Their plan is to use a computer system
to track and influence people’s buying
habits. A set of logical rules would
define the conditions for generating
promotional offers for customers, based
on the products the customer has already
chosen. For example, if customer
removed a product A from a shelf, then
she may be offered a discount coupon
on product B. Alternatively, the
customer may be asked if she may also
need product C. This last feature serves
as a reminder, rather than for offering
discount coupons. For example, if a
customer removes a soda bottle from a
shelf, she may be prompted to buy
potato chips, as well.
To implement this idea, the store will
use Radio Frequency Identification
(RFID) tags on all store items. Each tag carries a 96-bit EPC (Electronic Product Code). The RFID tag
readers will be installed on each shelf on the sales floor, as well as in the cashier registers at the sales
point. When a tag is removed from the region of a reader’s coverage, the reader will notify the
computer system that the given tag disappeared from its coverage area. In turn, the system will apply
the logical rules and show a promotional offer on a nearest display. We assume that each shelf will
have an “offers display” that will show promotional offers or reminders related to the last item that was
removed from this shelf.
As we consider the details of the idea, we realize that the system will not be able to identify individual
customers and tailor promotional offers based on the customer identity. In addition to privacy
concerns, identifying individual customers is a difficult technological problem and the store
management ruled out potential solutions as too expensive. We do not care as much to know who the
customer is; rather, we want to know the historic information about other items that this customer
placed in her cart previously during the current shopping episode to customize the offer. Otherwise, the
current offer must be based exclusively on the currently removed item and not on prior shopping
Ivan Marsic

Rutgers University
56
history. Next, we come up with an idea of installing RFID tag readers in the shopping carts, so we can
track the current items in each shopping cart. However, the supermarket management decides against
this approach, because of a high price of the readers and concerns about their robustness to weather
and handling or vandalism.
As a result, we conclude that logical IF-THEN-ELSE rules for deciding about special offers will take
as input only a single product identity, based on the RFID tag of the item the customer has just
removed from the shelf. The discount coupon will be a “virtual coupon,” which means that the
customer is told about the discounted product, and the discount amount will be processed at the
cashier’s register during the checkout. The display will persist for a specified amount of time and then
automatically vanish. The next question is whether each display will be dedicated to a single product
or shared among several adjacently shelved products? If the display will be shared, we have a problem
if other items associated with this display are removed (nearly) simultaneously. How do we show
multiple offers, and how to target each to the appropriate customer? A simple but difficult question is,
when the displayed coupon should vanish? What if the next customer arrives and sees it before it
vanishes? Perhaps there is nothing bad with that, but now we realize that we have a difficulty targeting
the coupons. In addition, because the system does not know what is in the customer’s cart, it may be
that the customer already took the product that the system is suggesting. After doing some market
research, we determine that small displays are relatively cheap and an individual display can be
assigned to each product. We give up targeting customers, and just show a virtual coupon as specified
by the logical rules.
Given that the store already operates in the same way with physical, paper-based coupons, the question
is if it is worth to install electronic displays or use RFID tags? Is there any advantage of upgrading the
current system? If the RFID system input is used, then the coupon will appear when an item is
removed. We realize that this makes no sense and just show the product coupon all the time, same as
with paper-based coupons. An advantage of electronic displays is that they preclude having the store
staff go around and place new coupons or remove expired ones.
We started with the idea of introducing RFID tags and ended up with a solution that renders them
useless. An argument can be made that tags can be used to track product popularity and generate
promotional offers based on the current demand or lack thereof. A variation of this project, with a
different goal, will be considered in Problem 2.15 at the end of Chapter 2.
There are several lessons to be learned about software engineering from the above example:

One cannot propose a solution without a deep understanding of the problem domain and
working closely with the customer

Requirements change dynamically because of new insights that were not obvious initially

Final solution may be quite different from the initial idea.
T
he project descriptions presented earlier in this chapter are relatively precise and include
more information than what is usually known as the customer statement of work, which is an
expression, from a potential customer, of what they require of a new software system. I expressed
the requirements more precisely to make them suitable for one-semester (undergraduate) student
projects. Our focus here will be on what could be called “core software engineering.”
On the other hand, the methods commonly found in software engineering textbooks would not
help you to arrive at the above descriptions. Software engineering usually takes from here—it
assumes a defined problem and focuses on finding a solution. Having defined a problem sets the
Chapter 1 
Introduction
57
constraints within which to seek for the solution. If you want to broaden the problem or reframe
it, you must go back and do some fieldwork. Suppose you doubt my understanding of financial
markets or ability to extract the key aspects of the security trading process (Section 1.3.2) and you
want to redefine the problem statement. For that, software engineering methods (to be described
in Chapter 2) are not very useful. Rather, you need to employ ethnography or, as an engineer you
may prefer Jackson’s “problem frames” [Jackson 2001], see Chapter 3. Do I need to mention that
you better become informed about the subject domain? For example, in the case of the financial
assistant, the subject domain is finance.
1.6 Summary and Bibliographical Notes
Because software is pure invention, it does not have physical reality to keep it in check. That is,
we can build more and more complex systems, and pretend that they simply need a little added
debugging. Simple models are important that let us understand the main issues. The search for
simplicity is the search for a structure within which the complex becomes transparent. It is
important to constantly simplify the structure. Detail must be abstracted away and the underlying
structure exposed.
Although this text is meant as an introduction to software engineering, I focus on critical thinking
rather than prescriptions about structured development process. Software development can by no
means be successfully mastered from a single source of instruction. I expect that the reader is
already familiar with programming, algorithms, and basic computer architecture. The reader may
also wish to start with an introductory book on software engineering, such as [Larman, 2005;
Sommerville, 2004]. The Unified Modeling Language (UML) is used extensively in the
diagrams, and the reader unfamiliar with UML should consult a text such as [Fowler, 2004]. I
also assume solid knowledge of the Java programming language. I do offer a brief introductionto/refresher-of the Java programming language in Appendix A, but the reader lacking in Java
knowledge should consult an excellent source by Eckel [2003].
The problem of scheduling construction tasks (Section 1.1) is described in [Goodaire &
Parmenter, 2006], in Section 11.5, p. 361. One solution involves first setting posts, then cutting,
then nailing, and finally painting. This sequence is shown in Figure 1-2 and completes the job in
11 units of time. There is a second solution that also completes the job in 11 units of time: first
cut, then set posts, then nail, and finally paint.
Although I emphasized that complex software systems defy simple models, there is an interesting
view advocated by Stephen Wolfram in his NKS (New Kind of Science):
http://www.wolframscience.com/ , whereby some systems that appear extremely complex can be
captured by very simple models.
In a way, software development parallels the problem-solving strategies in the field of artificial
intelligence or means-ends analysis. First we need to determine what are our goals (“ends”); next,
represent the current state; then, consider how (“means” to employ) to minimize the difference
between the current state and the goal state. As with any design, software design can be seen as a
Ivan Marsic

Rutgers University
58
difference-reduction activity, formulated in terms of a symbolic description of differences.
Finally, in autonomic computing, the goals are represented explicitly in the program that
implements the system.
There are many reasons why some systems succeed (e.g., the Web, the Internet, personal
computer) and others fail, including:

They meet a real need

They were first of their kind

They coevolved as part of package with other successful technologies and were more
convenient or cheaper (think MS Word versus WordPerfect)

Because of their technical excellence
Engineering excellence alone is not guarantee for success but a clear lack of it is a guarantee for
failure.
There are many excellent and/or curious websites related to software engineering, such as:
Teaching Software Engineering – Lessons from MIT, by Hal Abelson and Philip Greenspun:
http://philip.greenspun.com/teaching/teaching-software-engineering
Software Architecture – by Dewayne E. Perry: http://www.ece.utexas.edu/~perry/work/swa/
Software Engineering Academic Genealogy – by Tao Xie:
http://www.csc.ncsu.edu/faculty/xie/sefamily.htm
Section 1.2.1:
Symbol Language
Most people agree that symbols are useful, even if some authors invent their own favorite
symbols. UML is the most widely accepted graphical notation for software design, although it is
sometimes criticized for not being consistent. Even in mathematics, the ultimate language of
symbols, there are controversies about symbols even for such established subjects as calculus (cf.,
Newton’s vs. Leibnitz’s symbols for calculus), lest to bring up more recent subjects. To sum up,
you can invent your own symbols if you feel it absolutely necessary, but before using them,
explain their meaning/semantics and ensure that it is always easy to look-up the meanings of your
symbols. UML is not ideal, but it is the best currently available and most widely adopted.
Arguably, symbol language has a greater importance than just being a way of describing one’s
designs. Every language comes with a theory behind it, and every theory comes with a language.
Symbol language (and its theory) helps you articulate your thoughts. Einstein knew about the
general relativity theory for a long time, but only when he employed tensors was he able to
articulate the theory (http://en.wikipedia.org/wiki/History_of_general_relativity).
Section 1.2.3:
Object-Oriented Analysis and the Domain Model
I feel that this is a more gradual and intuitive approach than some existing approaches to domain
analysis. However, I want to emphasize that it is hard to sort out software engineering approaches
into right or wrong ones—the developer should settle on the approach that produces best results
Chapter 1 
Introduction
59
for him or her. On the downside of this freedom of choice, choices ranging from the dumbest to
the smartest options can be defended on the basis of a number of situation-dependent
considerations.
Some authors consider object-oriented analysis (OOA) to be primarily the analysis of the existing
practice and object-oriented design (OOD) to be concerned with designing a new solution (the
system-to-be).
Modular design was first introduced by David Parnas in 1960s.
A brief history of object orientation [from Technomanifestos] and of UML, how it came together
from 3 amigos. A nice introduction to programming is available in [Boden, 1977, Ch. 1],
including the insightful parallels with knitting which demonstrates surprising complexity.
Also, from [Petzold] about ALGOL, LISP, PL/I.
Objects: http://java.sun.com/docs/books/tutorial/java/concepts/object.html
N. Wirth, “Good ideas, through the looking glass,” IEEE Computer, vol. 39, no. 1, pp. 28-39,
January 2006.
H. van Vliet, “Reflections on software engineering education,” IEEE Software, vol. 23, no. 3, pp.
55-61, May-June 2006.
[Ince, 1988] provides a popular account of the state-of-the-art of software engineering in mid
1980s. It is worth reading if only for the insight that not much has changed in the last 20 years.
The jargon is certainly different and the scale of the programs is significantly larger, but the
issues remain the same and the solutions are very similar. Then, the central issues were reuse,
end-user programming, promises and perils of formal methods, harnessing the power of hobbyist
programmers (today known as open source), and prototyping and unit testing (today’s equivalent:
agile methods).
Section 1.3.1: Case Study 1: From Home Access Control to
Adaptive Homes
The Case Study #1 Project (Section 1.3.1) – Literature about the home access problem domain:
A path to the future may lead this project to an “adaptive house” [Mozer, 2004]. See also:
Intel: Home sensors could monitor seniors, aid diagnosis (ComputerWorld)
http://www.computerworld.com/networkingtopics/networking/story/0,10801,98801,00.html
Another place to look is: University of Florida’s Gator Tech Smart House [Helal et al., 2005],
online at: http://www.harris.cise.ufl.edu/gt.htm
For the reader who would like to know more about home access control, a comprehensive, 1400pages two-volume set [Tobias, 2000] discusses all aspects of locks, protective devices, and the
methods used to overcome them. For those who like to tinker with electronic gadgets, a great
companion is [O’Sullivan & T. Igoe, 2004].
Biometrics:
Wired START: “Keystroke biometrics: That doesn’t even look like my typing,” Wired, p. 42,
June 2005. Online at: http://www.wired.com/wired/archive/13.06/start.html?pg=9
Ivan Marsic

Rutgers University
60
Researchers snoop on keyboard sounds; Computer eavesdropping yields 96 percent accuracy rate.
Doug Tygar, a Berkeley computer science professor and the study's principal investigator
http://www.cnn.com/2005/TECH/internet/09/21/keyboard.sniffing.ap/index.html
Keystroke Biometric Password; Wednesday, March 28, 2007 2:27 PM/EST
BioPassword purchased the rights to keystroke biometric technology held by the Stanford
Research Institute. On March 26, 2007, the company announced BioPassword Enterprise Edition
3.0 now with optional knowledge-based authentication factors, integration with Citrix Access
Gateway Advanced Edition, OWA (Microsoft Outlook Web Access) and Windows XP embedded
thin clients.
http://blogs.eweek.com/permit_deny/content001/seen_and_heard/keystroke_biometric_password.
html?kc=EWPRDEMNL040407EOAD
See also [Chellappa et al., 2006] for a recent review on the state-of-the-art in biometrics.
Section 1.4:
The Object Model
The concept of information hiding originates from David Parnas [1972].
D. Coppit, “Implementing large projects in software engineering courses,” Computer Science
Education, vol. 16, no. 1, pp. 53-73, March 2006. Publisher: Routledge, part of the Taylor &
Francis Group
J. S. Prichard, L. A. Bizo, and R. J. Stratford, “The educational impact of team-skills training:
Preparing students to work in groups,” British Journal of Educational Psychology, vol. 76, no. 1,
pp. 119-140, March 2006.
(downloaded: NIH/Randall/MATERIALS2/)
M. Murray and B. Lonne, “An innovative use of the web to build graduate team skills,” Teaching
in Higher Education, vol. 11, no. 1, pp. 63-77, January 2006. Publisher: Routledge, part of the
Taylor & Francis Group
Chapter 2
Object-Oriented Software Engineering
Contents
“When a portrait painter sets out to create a likeness, he relies
above all upon the face and the expression of the eyes, and
pays less attention to the other parts of the body. In the same
way, it is my intention to dwell upon those actions which
illuminate the workings of the soul.” —Plutarch
This chapter describes concepts and techniques for objectoriented software development. The first chapter introduced
the stages of software engineering lifecycle (Section 1.2).
Now, the tools and techniques for each stage are gradually
detailed and will be elaborated in later chapters.
We start with the methodology and project management
issues, which is a first concern faced with large-scale product
development. Next we review elements of requirements
engineering: how system requirements are gathered, analyzed,
and documented. Real-world projects rarely follow exclusive
“bottom-up” approach, from requirements through objects to
program code. Instead, high-level factors commonly
considered under “software architecture” influence the system
design in a top-down manner. The rest of this chapter takes a
bottom-up approach, with top-down forces shaping our design
choices.
A popular approach to requirements engineering is use case
modeling, which elaborates usage scenarios of the system-tobe. A similar approach, common in agile methods, centers on
user stories. Requirements engineering is followed by domain
modeling, where we model the problem domain with the main
emphasis on modeling the internal elements (“objects”) of our
system-to-be. Following analysis, the design stage specifies
how objects interact to produce desired behaviors of the
system-to-be. This chapter concludes with the techniques for
software implementation and testing. While studying this
chapter, the reader may find it useful to check Appendix G
and see how the concepts are applied in an example project.
61
2.1 Software Development Methods
2.1.1 Agile Development
2.1.2 Decisive Methodological Factors
2.2 Requirements Engineering
2.2.1 Requirements and User Stories
2.2.2 Requirements Gathering Strategies
2.2.3 Effort Estimation
2.3 Software Architecture
2.3.1 Problem Architecture
2.3.2 Software Architectural Styles
2.3.3 Recombination of Subsystems
2.4 Use Case Modeling
2.4.1
2.4.2
2.4.3
2.4.4
2.4.5
Actors, Goals, and Sketchy Use Cases
System Boundary and Subsystems
Detailed Use Case Specification
Security and Risk Management
Why Software Engineering Is Difficult (2)
2.5 Analysis: Building the Domain Model
2.5.1
2.5.2
2.5.3
2.5.4
Identifying Concepts
Concept Associations and Attributes
Domain Analysis
Contracts: Preconditions and
Postconditions
2.6 Design: Assigning Responsibilities
2.6.1 Design Principles for Assigning
Responsibilities
2.6.2 Class Diagram
2---3
2.6.3 Why Software Engineering Is Difficult (3)
2.7 Test-driven Implementation
2.7.1
2.7.2
2.7.3
2.7.4
2.7.5
2.7.6
Overview of Software Testing
Test Coverage and Code Coverage
Practical Aspects of Unit Testing
Integration and Security Testing
Test-driven Implementation
Refactoring: Improving the Design of
Existing Code
2.8 Summary and Bibliographical Notes
Problems
Ivan Marsic

Rutgers University
62
2.1 Software Development Methods
“Plan, v.t. To bother about the best method of accomplishing an accidental result.”
—Ambrose Bierce, The Devil’s Dictionary
The goal of software methodologists is to understand how high quality software can be developed
efficiently. The hope is that new insights will emerge about effective product development, so
both students and experts might benefit from learning and applying methodology. Ideally, the
developer would adhere to the prescribed steps and a successful project would result—regardless
of the developer’s knowledge and expertise. Methodology development often works by observing
how expert developers work and deriving an abstract model of the development process. In
reality, life cycle methods are often not followed; when they are, it is usually because of
employer’s policy in place. Why is it so, if following a method should be a recipe for success?
There are several reasons why methodologies are ignored or resisted in practice. One reason is
that methodology is usually derived from past experience. But, what worked for one person may
not work for another. Both developers and projects have different characteristics and it is difficult
to generalize across either one. Software development is so complex that it is impossible to create
precise instructions for every scenario. In addition, method development takes relatively long
time to recognize and extract “best practices.” By the time a method is mature, the technologies it
is based on may become outdated. The method may simply be inappropriate for the new and
emerging technologies and market conditions.
A development method usually lays out a prescriptive process by mandating a sequence of
development tasks. Some methods devise very elaborate processes with a rigid, documentationheavy methodology. The idea is that even if key people leave the project or organization, the
project should go on as scheduled because everything is properly documented. This approach is
known as “Big Design Up Front” (BDUF). However, experience teaches us that it is impossible
to consider all potential scenarios just by thinking. And, regardless of how well the system is
documented, if key people leave, the project suffers. It is much more sensible to develop initial
versions of the system-to-be from a partial understanding of the problem, let users play with such
a prototype, and then redesign and develop a new iteration based on the gained understanding.
One difficulty with product development is that when thinking about a development plan,
engineer usually thinks in terms of methodology: what to do first, what next, etc. Naturally, first
comes discovery (studying the problem domain and finding out how the problem is solved now
and proposing how it can be solved better with the to-be-developed technology); then comes
development (designing and implementing the system); lastly, the system is deployed and
evaluated. This sequential thinking naturally leads to the “waterfall model” (Section 1.2) and
heavy documentation.
The customer does not see it that way. The customer would rather see some rudimentary
functionality soon, and then refinement and extension.
Recent methods, known as agile, attempt to deemphasize process-driven documentation and
detailed specifications. They also consider the number and experience of the people on the
development team.
Four major software development methodologies can be classified as:
Chapter 2 
Object-Oriented Software Engineering

Structured analysis and design (SAD), developed in late 1960s and 1970s

Object-oriented analysis and design (OOAD), developed in 1980s and 1990s

Agile software development (ASD), developed in late 1990s and 2000s

Aspect-oriented software development (AOSD), developed in 2000s
63
The structured analysis and design (SAD) methodology emerged in the 1970s and introduced
functional decomposition and data-flow analysis as key modeling tools.
The object-oriented analysis and design (OOAD) methodology emerged in the late 1980s and was
widely adopted by the mid 1990s. It introduced use cases and the Unified Modeling Language
(UML) as key modeling tools.
The ideas of agile software development (ASD) emerged at the end of 1990s and rapidly gained
popularity in the software industry as a “lightweight” way to develop software. Agile
development is reviewed in Section 2.1.1.
The aspect-oriented software development (AOSD) methodology emerged in the late 1990s. It is
not a replacement for any of the other methodologies. Rather, it helps deal with scattered
crosscutting concerns. Functional features of a software system could be divided into two
categories: (1) core features that provide basic functionality and allow the end-user to achieve
specific business goals; and, (2) supplementary features that provide support for entitlements,
connectivity, concurrency, system interface, etc. Many of the complementary features can be
scattered across the application and tangled with core features, which is why they are called
crosscutting concerns. By “tangled” I mean that these crosscutting concerns are invoked in the
context of core features and are part of the affected core functionality. Aspect-oriented software
development helps deal with crosscutting concerns in a systematic manner.
2.1.1
Agile Development
“People forget how fast you did a job—but they remember how well you did it.” —An advertising executive
“Why do we never have time to do it right, but always have time to do it over?” —Anonymous
Agility is both a development philosophy and a collection of concepts embedded into
development methodologies. An agile approach to development is essentially a results-focused
method that iteratively manages changes and risks. It also actively engages customers in
providing feedback on successive implementations, in effect making them part of the
development team. Unlike process-driven documentation, it promotes outcome-driven
documentation. The emphasis of agile practices is on traveling lightweight, producing only those
artifacts (documentation) that are absolutely necessary. The philosophy of the agile approach is
formulated by the Manifesto for Agile Software Development (http://agilemanifesto.org/).
Agile development evangelists recommend that the development should be incremental and
iterative, with quick turnover, and light on documentation. They are believers in perfection being
the enemy of innovation. Agile methods are not meant to entirely replace methodologies such as
structured analysis and design, or object-oriented analysis and design. Rather, agile methods are
often focused on how to run the development process (“project management”), perhaps using the
tools for software development inherited from other methods, but in a different way. A popular
Ivan Marsic

Rutgers University
64
agile-development tool is user stories, which are intended to represent the system requirements,
estimate effort and plan software releases (Section 2.2.3).
SIDEBAR 2.1: Agile vs. Sloppy
 I have had students complain that demanding readability, consistency, and completeness in
project reports runs against the spirit of agile development. Some software engineering
textbooks insist on showing snapshots of hand drawn UML diagrams, as opposed to neat
diagrams created electronically, to emphasize the evanescent nature of designs and the need for
dynamic and untidy artifacts. This may work for closely knit teams of professionals, working
in adjacent offices exclusively on their project. But, I found it not to be conducive for the
purpose of grading student reports: it is very difficult to discern sloppiness from agility and
assign grades fairly. Communication, after all, is the key ingredient of teamwork, and
communication is not improved if readability, consistency, and completeness of project reports
are compromised. I take it that agility means: reduce the amount of documentation but not at
the expense of the communicative value of project artifacts. Brevity is a virtue, but we also
know that redundancy is the most effective way to protect the message from noise effects. (Of
course, you need to know the right type of redundancy!)
Agile methodologists seem not to have much faith in visual representations, so one can find few
if any graphics and diagrams in agile software development books. Some authors take the agile
principles to the extreme and I would caution against this. I have seen claims that working code is
the best documentation of a software product. I can believe that there are people for whom
program code is the most comprehensible document, but I believe that most people would
disagree. Most people would find easiest to understand carefully designed diagrams with
accompanying narrative in a plain natural language. Of course, the tradeoff is that writing proper
documentation takes time, and it is difficult to maintain the documentation consistent with the
code as the project progresses.
Even greater problem is that the code documents only the result of developer’s design decisions,
but not the reasoning behind those decisions. Code is a solution to a problem. It is neither a
description of the problem, nor of the process by which the problem was solved. Much of the
rationale behind the solution is irretrievably lost or hidden in the heads of the people who chose
it, if they are still around. After a period of time, even the person who made a design decision
may have difficulty explaining it if the reasons for the choice are not explicitly documented.
However, although documentation is highly desirable it is also costly and difficult to maintain in
synchrony with the code as the lifecycle progresses. Outdated documentation may be source of
confusion. It is said that the code is the only unambiguous source of information. Such overgeneralizations are not helpful. It is like saying that the building itself is the only unambiguous
source of information and one need not be bothered with blueprints. You may not have blueprints
for your home or even not know where to find them, but blueprints for large public buildings are
carefully maintained as they better be. After all, it is unethical to leave a customer with working
code, but without any documentation. There is a spectrum of software projects, so there should be
a matching spectrum of documentation approaches, ranging from full documentation, through
partial and outdated one, to no documentation. I believe that even outdated documentation is
better than no documentation. Outdated documents may provide insight into the thinking and
evolution that went into the software development. On most projects, documentation should be
created with the understanding that it will not always be up to date with the code, resulting in
Chapter 2 
Object-Oriented Software Engineering
65
“stale” parts. A discrepancy usually arises in subsequent iterations, so we may need to prioritize
and decide what to keep updated and what to mark as stale.
There are other issues with maintaining adequate documentation. The developer may even not be
aware of some choices that he made, because they appear to be “common sense.” Other decisions
may result from company’s policies that are documented separately and may be changed
independently of the program documentation. It is useful to consider again the exponential curve
in Figure 1-13, which can be modified for documentation instead of estimation. Again, a
relatively small effort yields significant gains in documentation accuracy. However, after a
certain point the law of diminishing returns triggers and any further improvement comes at a
great cost. It is practically impossible to achieve perfect documentation.
SIDEBAR 2.2: How Much Diagramming?

I often hear inquiries and complaints that the amount of diagramming in this book is
excessive. This book is primarily intended for students learning software engineering and
therefore it insists on tidiness and comprehensiveness for instructive purposes. If I were doing
real projects, I would not diagram and document every detail, but only the most difficult and
important parts. Unfortunately, we often discover what is “difficult and important” only long
after the project is completed or after a problem arises. Experience teaches us that the more
effort you invest in advance, the more you will be thankful for it later. The developer will need
to use their experience and judgment as well as contextual constraints (budget, schedule, etc.)
to decide how much diagramming is appropriate.
Many books and software professionals place great emphasis on the management software
engineering projects. In other words, it is not about the engineering per se but it is more about
how you go about engineering software, in particular, knowing what are the appropriate steps to
take and how you put them together. Management is surely important, particularly because most
software projects are done by teams, but it should not be idolized at the detriment of product
quality. This book focuses on techniques for developing quality software.
2.1.2
Decisive Methodological Factors
Software quality can be greatly improved by paying attention to factors such as traceability,
testing, measurement, and security.
Traceability
Software development process starts with an initial artifact, such as customer statement of work,
and ends with source code. As the development progresses, being able to trace the links among
successive artifacts is key. If you do not make explicit how an entity in the current phase evolved
from a previous-phase entity, then it is unclear what was the purpose of doing all that previous
work. Lack of traceability renders the past creations irrelevant and we might as well have started
with this phase. It makes it difficult for testers to show that the system complies with its
requirements and maintainers to assess the impact of a change. Therefore, it is essential that a
precise link is made from use cases back to requirements, from design diagrams back to use
cases, and from source code back to design diagrams. Traceability refers to the property of a
software artifact, such as a use case or a class, of being traceable to the original requirement or
Ivan Marsic

66
Rutgers University
rationale that motivated its existence. Traceability must be maintained across the lifecycle.
Maintaining traceability involves recording, structuring, linking, grouping, and maintaining
dependencies between requirements and other software artifacts. We will see how traceability
works on examples in this chapter.
Requirements
Use Cases
Concepts/Objects
Source Code
UC-1
CO-1
Code-1
CO-2
Code-2
CO-3
Code-3
UC-M
CO-S
Code-W
UC-N
CO-T
Code-X
Use Cases
(Section 2.3)
OOA/OOD
(Sections 2.4 & 2.5)
Req-1
UC-2
Req-K
Requirements
Engineering
(Section 2.2)
Implementation
(Section (2.7)
Testing
The key idea of Test-Driven Development (TDD) is that every step in the development process
must start with a plan of how to verify that the result meets some goal. The developer should not
create a software artifact (such as a system requirement, a UML diagram, or source code) unless
he has a plan of how it will be tested. For example, a requirement is not well-specified if an
automated computer program cannot be written to test it for compliance. Such a requirement is
vague, subjective, or contradictory and should be reworked.
The testing process is not simply confined to coding. Testing the system design with
walkthroughs and other design review techniques is very helpful. Agile TDD methodology
prescribes to make progress just enough to pass a test and avoid detailed analysis. When a
problem is discovered, fix it. This approach may not be universally appropriate, e.g., for mission
critical applications. Therein, when a problem is discovered, it might have led to a major human
or economic loss. Discovering that you missed something only when system failed in actual use
may prove very costly. Instead, a thorough analysis is needed in advance of implementation.
However, the philosophy of thinking while creating a software artifact about how it will be tested
and designing for testability applies more broadly than agile TDD.
Software defects (or, bugs) are typically not found by looking at source code. Rather, defects are
found by mistreating software and observing how it fails, by reverse engineering it (approach
used by people who want to exploit its security vulnerabilities), and by a user simply going about
his business until discovering that a program has done something like delete all of the previous
hour’s work. Test plans and test results are important software artifacts and should be preserved
along with the rest of software documentation. More about testing in Section 2.7.
Chapter 2 
Object-Oriented Software Engineering
67
Agile TDD claims to improve the code, and detect design brittleness and lack of focus. It may
well do that, but that is not the main purpose of testing, which is to test the correctness, not
quality of software. Even a Rube-Goldberg design can pass tests under the right circumstances.
And we cannot ever check all circumstances for complex software systems. Therefore, it would
be helpful to know if our system works correctly (testing) and if it is of high quality, not a RubeGoldberg machine. This is why we need software measurement.
Measurement
While testing is universally practiced and TDD widely adopted, metrics and measurement are
relatively rarely used, particularly for assessing software product quality. Agile methods have
emphasized using metrics for project estimation, to track progress and plan the future iterations
and deliverables. Software product metrics are intended to assess program quality, not its
correctness (which is assessed by testing and verification). Metrics do not uncover errors; they
uncover poor design.
More about software measurement in Chapter 4.
Security
Most computers, telephones, and other computing devices are nowadays connected to the public
Internet. Publicly accessible Web applications and services can be abused and twisted to
nefarious ends. Even if the computer does not contain any “sensitive” information, its computing
and communication resources may be abused to send out spam and malware as part of a
distributed botnet. Such hijacked systems provide a “safe” means of distribution of illicit goods or
services on someone else’s server without that person’s knowledge. Because of ubiquitous
connectivity, anyone’s security problems impact everyone else, with only rare exceptions.
There are two kinds of technology-based security threats in software systems. One arises because
of bad software, where the attacker exploits software defects. The other arises because of network
interconnectedness, when the attacker exploits other infected systems to poison the traffic to or
from targeted computers. Hence, even if software is designed with security features to prevent
unauthorized use of system resources, it may be denied data or services from other computers.
Attackers rely on exploitable software defects as well as continuing to develop their own
infrastructure. An experienced developer must understand both the principles of software design
and the principles of network security. Otherwise, he will be prone to making naïve mistakes
when assessing the security benefits of a particular approach to software development. This book
focuses on better software design and does not cover network security.
The Security Development Lifecycle (SDL), promoted by Microsoft and other software
organizations, combines the existing approaches to software development with security-focused
activities throughout the development lifecycle. Security risk management focuses on minimizing
design flaws (architectural and design-level problems) and code bugs (simple implementation
errors in program code). Identifying security flaws is more difficult than looking for bugs,
because it requires deep understanding of the business context and software architecture and
design. We work to avoid design flaws while building secure software systems. Techniques
include risk analysis, abuse cases (trying to misuse the system while thinking like an attacker),
and code quality auditing.
Ivan Marsic

Rutgers University
68
Functional security features should not be confused with software security. Software security is
about developing high quality, problem-free software. Functional security features include
cryptography, key distribution, firewalls, default security configuration, privilege separation
architecture, and patch quality and response time. Poorly designed software is prone to security
threats regardless of built-in security functionality. Security functionality design is detailed in
Section 5.5.
2.2 Requirements Engineering
“The hardest single part of building a software system is deciding what to build. No part of the work so
cripples the resulting system if done wrong. No other part is more difficult to rectify later.”—Fred Brooks
“You start coding. I’ll go find out what they want.” —Computer analyst to programmer
Requirements engineering helps software engineers understand the problem they are to solve. It
involves activities that lead to understanding the business context, what the customer wants, how
end-users will interact with the software, and what the business impact will be. Requirements
engineering starts with the problem definition: customer statement of work (also known as
customer statement of requirements). This is an informal description of what the customers think
they need from a software system to do for them. The problem could be identified by
management personnel, through market research, by ingenious observation, or some other means.
The statement of work captures the perceived needs and, because it is opinion-based, it usually
evolves over time, with changing market conditions or better understanding of the problem.
Defining the requirements for the system-to-be includes both fact-finding about how the problem
is solved in the current practice as well as envisioning how the planned system might work. The
final outcome of requirements engineering is a requirements specification document.
The key task of requirements engineering is formulating a well-defined problem to solve. A welldefined problem includes

A set of criteria (“requirements”) according to which proposed solutions either definitely
solve the problem or fail to solve it

The description of the resources and components at disposal to solve the problem.
Requirements engineering involves different stakeholders in defining the problem and specifying
the solution. A stakeholder is an individual, team, or organization with interests in, or concerns
related to, the system-to-be. Generally, the system-to-be has several types of stakeholders:
customers, end users, business analysts, systems architects and developers, testing and quality
assurance engineers, project managers, the future maintenance organization, owners of other
systems that will interact with the system-to-be, etc. The stakeholders all have a stake, but the
stakes may differ. End users will be interested in the requested functionality. Architects and
developers will be interested in how to effectively implement this functionality. Customers will
be interested in costs and timelines. Often compromises and tradeoffs need to be made to satisfy
different stakeholders.
Chapter 2 
69
Object-Oriented Software Engineering
Aspect-Oriented
Requirements
Object-Oriented
Analysis & Design
Requirements
gathering
Requirements
analysis
Requirements
specification
Structured
Analysis & Design
Agile Development
User Stories
Figure 2-1: Requirements process in different methodologies.
Although different methodologies provide different techniques for requirements engineering, all
of them follow the same requirements process: requirements gathering, requirements analysis,
and requirements specification (Figure 2-1). The process starts with customer’s requirements or
surveying the potential market and ends with a specification document that details how the
system-to-be will behave. This is simply a logical ordering of requirements engineering activities,
regardless of the methodology that is used. Of course, the logical order does not imply that each
step must be perfectly completed before the next is taken.
Requirements gathering (also known as “requirements elicitation”) helps the developer
understand the business context. The customer needs to define what is required: what is to be
accomplished, how the system will fit into the needs of the business, and how the system will be
used on a day-to-day basis. This turns out to be very hard to achieve, as discussed in Section
2.2.2. The statement of work is rarely precise and complete enough for the development team to
start working on the software product.
Requirements analysis involves refining of and reasoning about the requirements received from
the customer during requirements gathering. Analysis is driven by the creation and elaboration of
user scenarios that describe how the end-user will interact with the system. Negotiation with the
customer will be needed to determine the priorities, what is essential, and what is realistic. A
popular tool is the use cases (Section 2.4). It is important to ensure that the developer’s
understanding of the problem coincides with the customer’s understanding of the problem.
Requirements specification represents the problem statement in a semiformal or formal manner to
ensure clarity, consistency, and completeness. It describes the function and quality of the
software-to-be and the constraints that will govern its development. A specification can be a
written document, a set of graphical models, a formal mathematical model, a collection of usage
scenarios (or, “use cases”), a prototype, or any combination of these. The developers could use
UML or another symbol language for this purpose.
Ivan Marsic

Rutgers University
70
As mentioned, logical ordering of the development lifecycle does not imply that we must achieve
perfection in one stage before we progress to the next one. Quite opposite, the best results are
achieved by incremental and iterative attention to different stages of the requirements engineering
process. This is an important lesson of the agile development philosophy. Traditional prescriptive
processes are characterized by their heavy emphasis on getting all the requirements right and
written early in the project. Agile projects, on the other hand, acknowledge that it is impossible to
identify all the requirements in one pass. Agile software development introduced a light way to
model requirements in the form of user stories, which are intended to capture customer needs,
and are used to estimate effort and plan releases. User stories are described in Section 2.2.3.
Section 2.3.1 introduces different problem types and indicates that different tools for
requirements engineering work best with different types of problems. In addition to problem
types, the effectiveness of requirements tools depends on the intended stakeholders. Different
requirements documents may be needed for different stakeholders. For example, the requirements
may be documented using customer’s terminology so that customers unfamiliar with software
engineering jargon may review and approve the specification of the system-to-be. A
complementary document may be prepared for developers and testing engineers in a semi-formal
or formal language to avoid ambiguities of natural languages.
2.2.1
Requirements and User Stories
“The best performance improvement is the transition from the nonworking state to the working state.”
—John Ousterhout
The statement of requirements is intended to precisely state the capabilities of the system that the
customer needs developed. Software system requirements are usually written in the form of
statements “The system shall …” or “The system should …” The “shall” form is used for features
that must be implemented and the “should” form for desirable but not mandatory features. IEEE
has published a set of guidelines on how to write software requirements. This document is known
as IEEE Standard 830.
Statement of Requirements, Case Study 1: Secure Home Access
Table 2-1 enumerates initial requirements for the home access control system extracted
from the problem description in Section 1.3.1. Each requirement is assigned a unique
identifier. The middle column shows the priority weight (PW) of each requirement, with a
greater number indicating a higher priority. The priority weight may be assigned by the
customer or derived from the urgency-to-deliver the requested capabilities to the customer.
The range of priority weights is decided arbitrarily, in our example it is 1–5. It is preferable
to have a small range (10 or less), because the priorities are assigned subjectively and it is
difficult to discern finely-grained priorities. Larger projects with numerous requirements
may need larger range of priorities.
An important issue is the granularity of requirements. Some of the requirements in Table
2-1 are relatively complex or compound requirements. Test-Driven Development (TDD)
stipulates writing requirements so that they are individually testable. In a software lifecycle,
requirements eventually result in source code, which is then Verified and Validated by running a
test set that exercises each requirement individually (Section 2.7.1). In the end, a report is created
Chapter 2 
71
Object-Oriented Software Engineering
Table 2-1: Requirements for the first case study, safe home access system (see Section 1.3.1).
Identifier Priority
Requirement
REQ1
5
The system shall keep the door locked at all times, unless commanded otherwise by
authorized user. When the lock is disarmed, a countdown shall be initiated at the
end of which the lock shall be automatically armed (if still disarmed).
REQ2
2
The system shall lock the door when commanded by pressing a dedicated button.
REQ3
5
The system shall, given a valid key code, unlock the door and activate other devices.
REQ4
4
The system should allow mistakes while entering the key code. However, to resist
“dictionary attacks,” the number of allowed failed attempts shall be small, say three,
after which the system will block and the alarm bell shall be sounded.
REQ5
2
The system shall maintain a history log of all attempted accesses for later review.
REQ6
2
The system should allow adding new authorized persons at runtime or removing
existing ones.
REQ7
2
The system shall allow configuring the preferences for device activation when the
user provides a valid key code, as well as when a burglary attempt is detected.
REQ8
1
The system should allow searching the history log by specifying one or more of
these parameters: the time frame, the actor role, the door location, or the event type
(unlock, lock, power failure, etc.). This function shall be available over the Web by
pointing a browser to a specified URL.
REQ9
1
The system should allow filing inquiries about “suspicious” accesses. This function
shall be available over the Web.
that says what requirements passed and what requirements failed. For this purpose, no
requirement should be written such that there are several “tests” or things to verify
simultaneously. If there is a compound requirement that failed, it may not be clear what part of
the requirement has failed. For example, if we were to test requirement REQ1 in Table 2-1, and
the door was found unlocked when it should have been locked, the entire requirement would fail
Verification. It would be impossible to tell from the report if the system accidentally disarmed the
lock, or the autolock feature failed. Therefore, when we group several “elemental” requirements
which apply to one functional unit into one compound requirement, we have a problem of not
being able to individually test requirements in this group. By splitting up REQ1 we obtain:
REQ1a: The system shall keep the doors locked at all times, unless commanded otherwise by
authorized user.
REQ1b: When the lock is disarmed, a countdown shall be initiated at the end of which the lock
shall be automatically armed (if still disarmed).
However, requirements fragmentation accommodates only the testing needs. Other considerations
may favor compounding of “elemental” requirements which apply to one functional unit. A
problem with elemental requirements is that none of them describes a stand-alone, meaningful
unit of functionality—only together they make sense. From customer’s viewpoint, good
requirements should describe the smallest possible meaningful units of functionality.
Although the choice of requirements granularity is subject to judgment and experience and there
is no clear metrics, the best approach is to organize one’s requirements hierarchically.
Ivan Marsic

Rutgers University
72
Note that Table 2-1 contains two types of requirement prioritization. There is an implicit priority
in “shall” vs. “should” wording, as well as explicit Priority Weight column. We need to ensure
that they are consistent. In principle, all features that must be implemented (“shall” type) should
be of higher priority then any feature that is not mandatory. Any inconsistency between the
prioritizations must be resolved with the customer. To avoid potential inconsistencies and
ambiguities, agile methods adopt a work backlog (Figure 1-14) that simply lists the work items in
the order in which they should be done.
Following the Test-Driven Development paradigm, we write tests for the requirements during the
requirements analysis. These tests are known as user acceptance tests (UATs) and they are
specified by the customer (Section 2.7.1). The system-to-be will be created to fulfill the
customer’s vision, so the customer decides that a requirement has been correctly implemented
and therefore the implementation is “accepted.” Acceptance tests capture the customer’s
assumptions about how the functionality specified with the requirement will work, under what
circumstances it may behave differently, and what could go wrong. The customer can work with
a programmer or tester to write the actual test cases. A test case is a particular choice of input
values to be used in testing a program and expected output values. A test is a finite collection of
test cases. For example, for the requirement REQ3, the customer may suggest these test cases:

Test with the valid key of a current tenant on his or her apartment (pass)

Test with the valid key of a current tenant on someone else’s apartment (fail)

Test with an invalid key on any apartment (fail)

Test with the key of a removed tenant on his or her previous apartment (fail)

Test with the valid key of a just-added tenant on his or her apartment (pass)
These test cases provide only a coarse description of how a requirement will be tested. It is
insufficient to specify only input data and expected outcomes for testing functions that involve
multi-step interaction. Use case acceptance tests in Section 2.4.3 will provide step-by-step
description of acceptance tests.
The table includes the requirement REQ7 that allows the user to configure the preferences for
activating various household devices in response to different events. The preferences would be set
up using a user interface (sketched in Figure 2-2). This is not to advocate user interface design at
this early stage of project development. However, the developer should use all reasonable means
to try and understand the customer’s needs as early as possible. Drawing sketches of user
interfaces is a useful tool for eliciting what the customer needs and how he would like to interact
with the system.
Table 2-1 contains only a few requirements that appear to be clear at the outset of the project.
Some of the requirements are somewhat imprecise and will be enhanced later, as we learn more
about the problem and about the tools used in solving it. Other requirements may be discovered
or the existing ones altered as the development lifecycle iteratively progresses. Refining and
modifying the initial requirements is the goal of requirements analysis.
Chapter 2 
73
Object-Oriented Software Engineering
Device Preferences
File Configure Help
Activate for valid key
Activate for burglary attempt
Lights
AirAir-conditioning
Alarm bell
Send SMS
Music
Heating
Police
…
Apply
Close
Figure 2-2: Envisioning the preference configuration for the control of household devices.
Statement of Requirements, Case Study 2: Investment Assistant
Here we extract initial requirements for the personal investment assistant system based on the
description given in Section 1.3.2. The requirements are shown in Table 2-2.
The statement of requirements is only a digest, and the reader should keep in mind that it must be
accompanied with a detailed description of customer’s business practices and rules, such as the
market functioning described earlier.
The stock trading ticket in REQ2 is a form containing the client’s instructions to the broker or
dealer. A stock trading ticket contains four parts: the client’s information, the security
information, the order information and any special instructions. The ticket specifies the action
(buy/sell), the order type (market/limit/stop), the symbol of the stock to trade, the number of
shares, and additional parameters in case of limit and stop orders. If the action is to buy, the
system shall check that the investor has sufficient funds in his/her account.
The order management window lists working, filled, cancelled, and parked orders, as well as
exceptions and all orders. In the working window, an order can be cancelled, replaced, and
designed as “go to market” for immediate execution, as well as be chained for order-cancelsorder status.
Similar to Table 2-1, Table 2-2 contains only a few requirements that appear to be clear at the
outset of the project. Other requirements may be discovered or the existing ones enhanced or
altered as the development lifecycle progresses.
Ivan Marsic

74
Rutgers University
Table 2-2: Requirements for the second case study, investment assistant (see Section 1.3.2).
Identifier PW
Requirement
REQ1
5
The system shall support registering new investors by providing a real-world email,
which shall be external to our website. Required information shall include a unique login
ID and a password that conforms to guidelines, as well as investor’s first and last name
and other demographic information. Upon successful registration, the system shall set up
an account with a zero balance for the investor.
REQ2
5
The system shall support placing orders by filling out a form known as “order ticket,”
which contains the client’s information, the stock information, the order information, and
any special instructions. The ticket shall be emailed to the client and enqueued for
execution when the specified conditions are satisfied.
REQ3
5
The system shall periodically review the enqueued orders and for each order ticket in the
queue take one of the following actions:
(i) If the order type is Market Order, the system shall execute the trade instantly;
(ii) Else, if the order conditions are matched, convert it to a Market Order at the current
stock price;
(iii) Else, if the order has expired or been cancelled, remove it from the queue, declare it a
Failed Order and archive as such;
(iv) Else, leave the order untouched.
If either of actions (i), (ii), or (iii) is executed, the system shall archive the transaction and
notify the trader by sending a “brokerage trade confirmation.”
REQ4
2
The system shall allow the trader to manage his or her pending orders, for example to
view the status of each order or modify the order, where applicable.
REQ5
2
The system shall continuously gather the time-series of market data (stock prices, trading
volumes, etc.) for a set of companies or sectors (the list to be decided).
REQ6
3
The system shall process the market data for two types of information:
(i) on-demand user inquiries about technical indicators and company fundamentals (both
to be decided), comparisons, future predictions, risk analysis, etc.
(ii) in-vigilance watch for trading opportunities or imminent collapses and notify the
trader when such events are detected
REQ6
3
The system shall record the history of user’s actions for later review.
User Stories
Agile development methods have promoted “user stories” as an alternative to traditional
requirements. A user story is a brief description of a piece of system functionality as viewed by a
user. It represents something a user would be likely to do in a single sitting at the computer
terminal. User stories are written in a free-form, with no mandatory syntax, but generally they are
fitting the form:
user-role + capability + business-value
Here is an example of a user story for our case study of secure home access:
As a tenant, I can unlock the doors to enter my apartment.
user-role
capability
business-value
Chapter 2 
75
Object-Oriented Software Engineering
Table 2-3: User stories for the first case study, safe home access. (Compare to Table 2-1.)
The last column shows the estimated effort size for each story (described in Section 2.2.3).
Identifier
User Story
Size
ST-1
As an authorized person (tenant or landlord), I can keep the doors locked at
all times.
4
points
ST-2
As an authorized person (tenant or landlord), I can lock the doors on demand.
3 pts
ST-3
The lock should be automatically locked after a defined period of time.
6 pts
ST-4
As an authorized person (tenant or landlord), I can unlock the doors.
(Test: Allow a small number of mistakes, say three.)
9
points
ST-5
As a landlord, I can at runtime manage authorized persons.
10 pts
ST-6
As an authorized person (tenant or landlord), I can view past accesses.
6 pts
ST-7
As a tenant, I can configure the preferences for activation of various devices.
6 pts
ST-8
As a tenant, I can file complaint about “suspicious” accesses.
6 pts
The business-value part is often omitted to maintain the clarity and conciseness of user stories.
Table 2-3 shows the user stories for our first case study of home access control (Section 1.3.1). If
we compare these stories to the requirements derived earlier (Table 2-1), we will find that stories
ST-1 and ST-2 roughly correspond to requirement REQ1, story ST-3 corresponds to REQ2, story
ST-4 corresponds to REQ3 and REQ4, and story ST-6 corresponds to REQ8, etc. Note, however,
that unlike the IEEE-830 statements “The system shall …,” user stories put the user at the center.
Types of Requirements
System requirements make explicit the characteristics of the system-to-be. Requirements are
usually divided into functional and non-functional. Functional requirements determine the
system’s expected behavior and the effects it should produce in the problem domain. These
requirements generally represent the main product features.
Non-functional requirements describe some quality characteristic that the system-to-be shall
exhibit. They are also known as “quality” or “emergent” requirements, or the “-ilities” of the
system-to-be. An example non-functional requirement is: Maintain a persistent data backup, for
the cases of power outages.
The term FURPS+ refers to the non-functional system properties:

Functionality lists additional functional requirements that might be considered, such as
security, which refers to ensuring data integrity and authorized access to information

Usability refers to the ease of use, esthetics, consistency, and documentation—a system
that is difficult and confusing to use will likely fail to accomplish its intended purpose

Reliability specifies the expected frequency of system failure under certain operating
conditions, as well as recoverability, predictability, accuracy, and mean time to failure

Performance details the computing speed, efficiency, resource consumption, throughput,
and response time
Ivan Marsic


Rutgers University
76
Supportability characterizes testability, adaptability, maintainability, compatibility,
configurability, installability, scalability, and localizability
For example, in terms of usability of our safe home access case study, we may assume a lowbudget customer, so the system will be installed and configured by the developer, instead of
“plug-and-play” operation.
All requirements must be written so that they are testable in that it should be obvious how to
write acceptance tests that would demonstrate that the product meets the requirement. We have
seen earlier example of acceptance tests for functional requirements in Table 2-1. Non-functional
requirements are more susceptible for vague formulations. For example, we often hear that a
system should be “easy to use.” It is difficult to design tests to verify such a claim. There is little
value in writing requirements that are not testable.
For example, for our case study of safe home access system, we envisioned three types of
computing devices. Users will use these devices in different contexts and for different tasks, so
we can expect that they have different usability requirements. We should consider the time
constraints of user type and produce order-of-magnitude time limits for computer interaction
required to accomplish a certain activity. For example, the user interacting with the door device
expects that the number of keystrokes, clicks, or touches will be minimized for quick task
completion. The property manager interacting with the desktop computer is less concerned with
efficiency and more with rich features to review the data and examine trends. Similarly, the
reliability requirements for different devices are likely to be different. The door device must be
highly reliable (e.g., system failure rate of 4 in a year or less), while the desktop application can
tolerate much lower reliability level.
Although at first it may appear easy, the distinction between functional and non-functional
requirements is often difficult to make. More often than not, these requirements are intertwined
and satisfying a non-functional requirement usually necessitates modifications in the system
function. For example, if performance objectives cannot be met, some functional features may
need to be left out.
The reader should be cautioned against regarding non-functional requirements as secondary to
functional requirements. The satisfaction of non-functional requirements must be as thoroughly
and rigorously ensured as that of functional requirements. In either case, satisfaction of a
requirement results in visible properties of the system-to-be, which means they will affect
customer or user satisfaction with the product.
I
n most cases, not all requirements can be realized because of budgetary or time constraints.
Therefore, it is necessary to prioritize the requirements. We have seen examples of assigning
priority weights to requirements in Table 2-1 and Table 2-2, where the weights were guessed by
the customer. A systematic method for prioritizing software product requirements is the costvalue approach. The basic idea is to determine for each candidate requirement its cost of
implementing and how much value the requirement would have. It is critical that the customer is
involved in requirements prioritization, assisted by tools that help highlight the tradeoffs.
Requirements prioritization is not helpful if all or most requirements are assigned high priority.
We distinguish four types of requirements:
Chapter 2 
77
Object-Oriented Software Engineering
1. Essential: have to be realized to make the system acceptable to the customer.
2. Desirable: highly desirable, but not mandatory requirements
3. Optional: might be realized if time and resources permit
4. Future: will not be realized in the current version of the system-to-be, but should be
recorded for consideration in future versions
The priority of requirements determines the order in which they will be implemented.
2.2.2
Requirements Gathering Strategies
“Everything true is based on need.” —George Bernard Shaw
“Well, as the new Hummer H2 ads observe, ‘need’ is a highly subjective word.” —Peter Coffee (in 2003)
If the developer is lucky, the customer will arrive with a clear statement of work that needs to be
done (“customer statement of requirements”). In reality, this rarely happens. Requirements for the
system-to-be should be devised based on observing the current practice and interviewing the
stakeholders, such as end users, managers, etc. To put it simply, you can’t fix it if you don’t know
what’s broken. Structured interviews help in understanding what stakeholders do, how they might
interact with the planned system, and the difficulties they are facing with the existing technology.
Agile methodologists recommend that the customers or users stay continuously involved
throughout the project duration, instead of only providing the requirements initially and
disappearing until the system is completed. (The reader may wish to check again Section 1.2.5
about the benefits of continuous customer involvement.)
How to precisely specify what system needs to do is a problem, but sometimes it is even more
difficult is to get the customer to say what he or she expects from the system. Gathering domain
knowledge by interviews is difficult because domain experts use terminology and jargon specific
to their domain that is unfamiliar and hard for an outsider to grasp. While listening to a domain
expert talk, a software engineer may find herself thinking “These all are words that I know, but
together they mean nothing to me.” Some things may be so fundamental or seem too obvious to a
person doing them habitually, that he thinks those are not worth mentioning.
In addition, it is often difficult for the user to imagine the work with a yet-to-be-built system.
People can relatively easily offer suggestions on how to improve the work practices in small
ways, but very rarely can they think of great leaps, such as, to change their way of doing business
on the Internet before it was around, or to change their way of writing from pen-and-paper when
word processors were not around. So, they often cannot tell you what they need or expect from
the system. What often happens is that the customer is paralyzed by not knowing what technology
could do and the developer is stuck by not knowing what the customer needs to have. Of great
help in such situation is having a working instance, a prototype, or performing a so called
Wizard-of-Oz experiment with a mock-up system.
See also Ch. 2 of “Wicked Problems”—problems that cannot be fully defined.
A popular technique for functional requirements engineering is the use case modeling, which is
described in Section 2.4.
Ivan Marsic

Rutgers University
78
W
e should keep in mind that we are trying to achieve several goals in requirements
engineering. As Figure 2-1 illustrates, we are trying to understand the problem in the
context of current practice (requirements gathering), then envision, elaborate, and negotiate
potential solutions (requirements analysis), and finally write down an engineering description of
what needs to be developed (requirements specification). Different tools have been proposed for
requirements engineering. As one would expect, none of these tools works best for all tasks of
requirements engineering and for all types of problems. Some tools work great for requirements
gathering, but may not be suitable for requirements analysis or specification. For example, user
stories (Section 2.2.1) work well in requirements gathering and analysis, but may be less suitable
for specification. Other tools work well on all three tasks, but not for all problem types. For
example, use case modeling (Section 2.4) works well on all three tasks, but only for certain
problem types. Further details are provided in the sections that follow. More tools that are better
suited for different problem types will be described in Chapter 3.
2.2.3
Effort Estimation
Requirements and user stories can be used to estimate effort and plan software
releases. The estimation process works very similarly to the example described
in Section 1.2.5. Similar to “hedge pruning points” described in Section 1.2.5,
to measure the relative size of the user stories we assign user-story points to
each user story. My preliminary estimates of the relative sizes of the user
stories on the scale 1–10 are shown in the rightmost column of Table 2-3.
I have to admit that, as I am making these estimates, I do not have much confidence in them. I am
very familiar with the home access case study and went many times over the solutions in
subsequent chapters. While making the estimates in Table 2-3, I am trying to make a holistic
guess, which requires a great deal of subjectivity. It is impossible to hold all those experiences in
one’s head at once and combine them in a systematic manner. The resulting estimates simply
reflect a general feeling about the size of each user story. This may be sufficient to start the
project, but I prefer using more structured methods for software size estimation. One such method
is based on use case points, described later in Chapter 4. However, more structured methods
come at a cost—they require time to derive the design details. I recommend that the reader should
always be mindful about which part of the exponential curve in Figure 1-13 he is operating on.
The desired accuracy of the estimate is acceptable only if the effort to achieve it (or, cost) is
acceptable, as well.
To apply equation (1.1) and estimate the effort (duration) needed to develop the system, we also
need to know the development team’s velocity. In physics, velocity is defined as the distance an
object travels during a unit of time. If in software project estimation size is measured in story
points, then the development team’s velocity is defined as the number of user-story points that the
team can complete per single iteration (the unit of time). That is, the velocity represent’s the
team’s productivity.
In software projects linear sum of sizes for individual user stories is rarely appropriate because of
reuse or shared code. Some functionality will be shared by several stories, so adding up sizes for
individual stories when estimated independently is not appropriate. Let me illustrate on an
analogy. Consider you are charged to build highways from city A to cities B and C (Figure 2-3).
Chapter 2 
79
Object-Oriented Software Engineering
B
A
(b)
C
City C
City B
B
A
City A
(a)
(c)
C
Figure 2-3: Combining the part sizes illustrated on a highway building example (a). Cities
may be connected independently (b), or parts of the product may be “reused” (c).
You eyeball a geographic map of the area and you estimate that the highway A–C will be twice
longer than the highway A–B. So, you estimate the size for the entire effort as 1s + 2s = 3s,
where s is a scaling constant. However, upon more careful inspection you realize that parts of
highways to cities B and C can be shared (reused), as illustrated in Figure 2-3(c). If you choose
this option, you cannot estimate the total size just by adding the individual sizes (AB  AC).
Instead, you need to consider them together. The total effort will be considerably smaller.
Reuse is common in software objects (consider how ubiquitous subroutines and libraries are!).
Therefore, my concern is that simply adding the story sizes introduces a gross inaccuracy in the
overall effort estimation. Recall the exponential relationship of cost and accuracy (Figure 1-13).
The reader would be mistaken to assume that reuse always means less work. Considering again
the highway analogy, the solution in Figure 2-3(c) may require more effort or cost than the one in
Figure 2-3(b). The infrastructure-sharing solution in Figure 2-3(c) requires building highway
interchanges and erecting traffic signs. You may wonder, why should anyone bother with reuse if
it increases the effort? The reason may be to preserve resources (conserve the land and protect
nature), or to make it easier to connect all three cities, or for esthetic reasons, etc. Reducing the
developer’s effort is not always the most important criterion for choosing problem solutions. The
customer who is sponsoring the project decides about the priorities.
Agile methodologists recommend avoiding dependencies between user stories.
High dependencies between stories make story size estimation difficult. (Note that
the assumption is as follows. The individual story sizes are still combined in a linear
sum, but the dependencies are tackled by adjusting the individual size estimations.)
When dependencies are detected, the developer can try these ways around it:

Combine the dependent user stories into one larger but independent story

Find a different way of splitting the stories
An expert developer might easily do this. But then, the expert might as well get an accurate effort
estimate by pure guessing. The problem is with beginner developers, who need the most a
Ivan Marsic

80
Rutgers University
Product line (or product family)
highest abstraction level
System or product
Subsystems/Modules
Packages
Classes/Objects
Methods
lowest level
Figure 2-4: Hierarchy of software system scope levels. At the highest scope level is a
product line—a family of products.
systematic way of estimating the project effort. The beginner may find it difficult to detect and
tackle dependencies between user stories.
2.3 Software Architecture
“Conceptual integrity is the most important consideration in system design.”
—Fred Brooks, The Mythical Man-Month
A simplest manifestation of a system-level design is the familiar “block diagram,” which shows
the subsystems or modules (as rectangular boxes) and their relations (lines connecting the boxes).
However, software architecture is much more than decomposing the system into subsystems.
Software architecture is a set of high-level decisions made during the development and
evolution of a software system. A decision is “architectural” if, given the current level of system
scope (Figure 2-4), the decision must be made by considering the current scope level. Such
decision could not be made from a more narrowly-scoped, local perspective.
Figure 2-5
Chapter 2 
81
Object-Oriented Software Engineering
product line
architecture decisions
product or system
architecture decisions
systemic impact
Product line scope
Product/system A scope
Product B scope
Subsystem scope
Class scope
local impact
Figure 2-5: Architectural decisions are made at certain scope levels and cannot be made at
lower hierarchical levels.
Architectural decisions should focus on high impact, high priority areas that are in strong
alignment with the business strategy. We already discussed some architectural decisions for our
case study system for safe home access in Section 1.3.1 (footnote 5). It might have looked as a
simple decision with a self-evident choice to have a central computer and embedded computers at
each door.
Some key questions that we are faced with include:
Q1: How to decompose the system (into parts)?
Q2: How the parts relate to one another?
Q3: How to document the system’s software architecture?
One way to start is by considering an abstraction hierarchy of different parts of the system (Figure
2-4). Such diagrams show only the parts of the system and their inclusion hierarchy. They do not
convey the dependencies in terms of mutual service uses: which part uses the services of what
other parts?
A good path to designing software architecture (i.e., solution architecture) starts by considering
the problem architecture (Section 2.3.1). That is, we start with the requirements (i.e., the problem
statement), which define how the system will interact with its environment.
Objects through their relationships form confederations, which are composed of potentially many
objects and often have complex behavior. The synergy of the cooperative efforts among the
members creates a new, higher-level conceptual entity.
Organizations are partitioned into departments—design, manufacturing, human resources,
marketing, etc. Of course, partitioning makes sense for certain size of the organization;
partitioning a small organization into departments and divisions does not make much sense.
Similarly, software systems should be partitioned into subsystems or modules where each
subsystem performs a set of logically related functions.
Ivan Marsic

82
Rutgers University
Safe Home Access System
Subsystem
for device
control
Subsystem
for
administration
Subsystem
for remote
data access
Decision on system
decomposition
On embedded
computer
On office
desktop
On tenant’s
smartphone
Decision on software-tohardware mapping
Figure 2-6: Architectural decisions for safe home access system.
Keypad and
Embedded processor
Light bulb
RS-232
Interface cable
Switch
Photosensor
Computer
Alarm bell
Figure 2-7: Hardware components for the system implementation.
Figure 2-6
Assume we have an embedded processor with a keypad, wired to other hardware components of
the system, as shown in Figure 2-7. The embedded processor accepts “commands” from the
computer via a RS-232 serial port and simply passes them on the corresponding device. The
many intricacies of serial communication are omitted and the interested reader is directed to the
bibliography review at the end of this chapter. The embedded processor may in an advanced
design become a full-featured computer, communicating with the main computer via a local area
network (LAN).
System architects may decompose an application into subsystems early in design. But subsystems
can be also discovered later, as the complexity of the system unfolds.
2.3.1
Problem Architecture
The most powerful ways of dealing with complex problems include recognizing and exploiting
regularities (or, patterns), and dividing the problem into smaller subproblems and solving each
individually (known as divide-and-conquer approach). When faced with a difficult software
engineering problem, it helps to recognize if it resembles to known typical problems. If it does,
we employ known solutions.
Chapter 2 
83
Object-Oriented Software Engineering
(a)
(b)
Figure 2-8: Contrasting decomposition types: (a) projection; (b) partition.
Problem can be decomposed in different ways, such as “projection” vs. “partition” (Figure 2-8).
There are significant differences between them. Partition isolates the parts from one another—it
simplifies by removing the relationships. Projection just simplifies the representation (by
removing some dimensions), while preserving the relationships between the parts. It allows any
kind of overlap between the elements of one subproblem and the elements of another. We favor
problem projection for its relationship-preserving trait.
For example, consider our first case study of safe home access. Figure 2-9 shows the elements of
the problem domain and how they relate to the system-to-be. There are eleven sub-domains of the
problem domain. The key sub-domains are the tenant (1), landlord (2), and the lock (3). Some
sub-domains are people or physical objects and some sub-domains are digital artifacts, such as
the list of valid keys (4), tenant accounts (10), and log of accesses (11). The system-to-be is
shown as composed of subsystems (shown as smaller boxes inside the system’s box) that
implement different requirements from Table 2-1. As seen, the concerns of different requirements
overlap and the system-to-be cannot be partitioned neatly into isolated subsystems. Initially, we
consider different requirements as subproblems of the entire problem and describe the subsystems
that solve different subproblems. Then we consider how the subsystems are integrated and how
they interact to satisfy all the requirements.
We start by identifying some typical elementary problems encountered by software engineers.
This classification of problems is empirical, not deduced by logical reasoning. Of course, there is
no proof that it is complete, unique, non-overlapping, etc.
There are three key players in software engineering problems: the user who uses the system to
achieve a goal, the software system (to be developed, i.e., the system-to-be), and the
Ivan Marsic

84
Rutgers University
REQ3
REQ1, REQ2,
REQ3, REQ4
PROBLEM DOMAIN
(5) Device
preferences
(6) Photosensor
(3) Key
(7) Light
Subsystem-2
(1) Tenant
(4) List of
valid keys
Subsystem-1
(3) Lock
REQ4
Subsystem-3
(8) Alarm bell
(11) Log of
accesses
Software-to-be
(2) Landlord
Subsystem-4
(9) Desktop computer
(10) Tenant
accounts
REQ5, REQ7,
REQ8, REQ9
Figure 2-9: Components of the problem domain for safe home access. Requirements from
Table 2-1 specify what the software-to-be should accomplish in the problem domain. We
can decompose the software-to-be into subsystems related to the requirements satisfaction.
environment—the rest of the world that may include other systems, considered as “black boxes”
because we either do not know or do not care about their structure. Figure 2-10 illustrates some
typical elementary software engineering problems. In problems of type 1.a) the user feeds the
system with a document and the system transforms the input document to an output document.
An example is a compiler that transforms source code written in a computer language (the source
language) into another computer language (the target language, often having a binary form known
as “object code”). Another example is a PDF writer, which takes a Web page or a word-processor
document and generates a PDF document.
In problems of type 1.b) the system helps the user edit and maintain a richly structured body of
information (Figure 2-10). The information must typically be manipulated in many different
ways. The data is long-lived and its integrity is important. Example applications include wordprocessing, graphics authoring, or relational database systems.
In problems of type 2 the system is programmed to control the environment (Figure 2-10, second
row). The system continuously observes the environment and reacts to predefined events. For
example, a thermostat monitors the room temperature and regulates it by switching heating or
cooling devices on or off to maintain the temperature near a desired setpoint value.
Chapter 2 
85
Object-Oriented Software Engineering
1.a) System transforms input document to output document
1. User works with computer system
(environment irrelevant/ignored)
IN doc
System
OUT doc
1.b) User edits information stored in a repository
User
System
System
Repository
User
2. Computer system controls the environment
(user not involved)
System
Environment
3. Computer system intermediates between
the user and the environment
3.a) System observes the environment and displays information
User
User
System
Environment
System
Environment
3.b) System controls the environment as commanded by the user
User
System
Environment
Figure 2-10: Some of the typical elementary problems encountered in software engineering.
In problems of type 3.a) the system monitors the environment and displays the information for the
user. The display may be continuous or filtered to notify the user only of predefined events. For
example, a patient-monitoring system measures physiological signals and displays them
continuously on a computer screen. Additionally, the system may be programmed to look for
trends, or sudden changes, or anomalous values and alert the clinician (user) by audio signals.
In problems of type 3.b) the system helps the user control the environment. The system receives
and executes the user’s commands. An example is controlling industrial processes. In our first
case study of safe home access (Section 1.3.1), the user commands the system to disarm the door
lock (and possibly activate other household devices).
Complex software engineering problems may combine several elementary problems from Figure
2-10. Consider our case study of safe home access (Figure 2-9). We already mentioned that it
includes the type 3.b) problem of commanding the system to disarm the lock. The requirements
(Table 2-1) also include managing the database of current tenant accounts (REQ5), which is a
problem of type 1.b). The system should also monitor if the door is unlocked for an extended
period of time and lock it automatically (REQ1), which is a problem of type 2.
Ivan Marsic

86
Rutgers University
1.a) Transformation:
Feeding
subsystem
Transformation
subsystem
Receiving
subsystem
Data
editor
1.b) Simple workpieces:
User
Data repository
2. Required behavior:
Controlling
subsystem
Controlled
subsystem
3.a) Information display:
Monitoring
subsystem
Monitored
subsystem
Controlling
subsystem
Controlled
subsystem
Display
3.b) Commanded behavior:
Operator
Figure 2-11: Problem architectures of typical software engineering problems.
To deal with complex problems that involve several subproblems, we apply the divide-andconquer approach. We decompose the problem into simpler problems, design computer
subsystems to solve each subproblem individually, and then compose the subsystems into an
integrated system that solves the original complex problem.
Figure 2-11 illustrates the elementary “building bricks” that correspond to different subproblem
types in Figure 2-10. We continue the discussion of problem decomposition and subsystem
specification in Section 2.4.2. More details will be provided later, in Section 3.3, when we
introduce problem frames.
2.3.2
Software Architectural Styles
So far the development process was presented as a systematic derivation of a software design
from system requirements. Although this process is iterative, every iteration presumably starts
with (possibly revised) requirements and progresses towards an implementation. However, in
reality such “bottom-up” design approaches at the local level of objects are insufficient to achieve
optimal designs, particularly for large systems. There are many contextual constraints and
influences other than requirements that determine the software architecture. For example, the
development team may prefer certain designs based on their expertise; their actual progress
compared to the plan; currently prevailing practices; available assets, such as lack of expertise in
certain areas, such as databases or visualization; hardware and networking constraints; etc. Most
problems do not start completely new development, but rather reuse existing designs, software
packages, libraries, etc. For example, many contemporary systems are based on Web architecture,
using a browser to access a database or Web services (see Appendix D). Complementary to
Chapter 2 
87
Object-Oriented Software Engineering
User Interaction
User Interface Layer
User Authentication
Archiving
Management of
Sensors and Devices
Communication w.
Police Station
Domain Layer
(Application Logic)
Technical Services
Layer
Figure 2-12: Software packages for the case study system. The system has a layered
architecture, with the three layers as indicated.
bottom-up approach are system-level (macro-level), global design approaches which help us to
“see the forest for the trees.” These “top-down” approaches decompose the system into logical
units or follow some global organizational patterns.
Program Flow Control
One can also set up “daemons” that spend their lifetime on the lookout for a certain type of event,
and do what they have to do whenever a happening of that type occurs. A more flexible IFTHEN-ELSE is to say, “If this happens, use that method to choose an appropriate procedure from
this list of procedures,” where the contents of the list in question can vary as the program runs.
IF-THEN-ELSE partitions the set of all possible situations in two or more cases. The partition
may turn out to be too rigid, there may be some exception cases that were not anticipated and
now need to be accounted for. Possibly even by the user! A key issue is, How to let the user to
“rewire” the paths/flows within the program if a need arises?
The program code that implements software classes and subsystems is usually organized into
software packages. Each package contains a set of logically related classes (Figure 2-12).
2.3.3
Recombination of Subsystems
After decomposition, different subsystems are usually developed and tested independently. At
some point, all subsystems need to be recombined and integrated into the whole system-to-be.
The recombination (or composition) problem is unsolved and very tricky. Key issues:

Cross-platform compatibility, particularly trust and privilege issues

Concurrent data access in multithreaded systems
The key problem of recombination of subsystems or frames into the system-to-be is the diversity
of infrastructures and platforms used for development. Modern software applications are rarely
written as a single monolithic program. Instead, they are built on top of complex middleware
frameworks such as .NET and Java technology, using multiple programming languages, and run
Ivan Marsic

Rutgers University
88
on several computers with different operating systems. Developers rely on outside libraries,
frameworks, COTS (Commercial-Off-The-Shelf) components, etc. The subsystems are usually
distributed over different computers. This diversity of platforms introduces many unknowns that
are hard or impossible to control by the developer.
Even most secure components can be assembled into an unsecure mess.
2.4 Use Case Modeling
A use case is a description of how a user will use the planned system to accomplish business
goals. As any description, it can be sketchy or it can be very detailed. Both versions (and many
degrees of detail in between) have important uses in requirements engineering. It is natural to
start with summary descriptions of use cases and gradually progress towards detailed descriptions
that thoroughly specify the planned system.
Use cases were already introduced in Section 1.2.2 and the current section presents details of use
case modeling. We start with summary descriptions of use cases and end with detailed
descriptions that represent the specification of the planned system.
2.4.1
Actors, Goals, and Sketchy Use Cases
In system development, we are mainly concerned with the actors that interact directly with the
system-to-be, including end users and other systems. However, all stakeholders have certain goals
for the system-to-be and occasionally it may be appropriate to list those goals. The consideration
of system requirements starts with identifying the actors for the system-to-be.
Actors and Their Goals
An actor is any entity (human, physical object, or another system) external to the system-to-be
that interacts with the system-to-be. Actors have their responsibilities and seek the system’s
assistance in managing those responsibilities. In our case-study example of secure home access,
resident’s responsibilities are to maintain the home secured and in proper order, as well as seek
comfortable living. The property manager’s responsibilities include keeping track of current and
departed residents. Maintenance personnel’s responsibilities include checks and repairs. There are
also some physical devices depicted in Figure 1-16 that are not part of the system-to-be but
interact with it. They also count as actors for our system, as will be seen later.
To carry out its responsibilities, an actor sets goals, which are time and context-dependent. For
example, a resident leaving the apartment for work has a goal of locking the door; when coming
back, the resident’s goal is to open the door and enter the apartment.
To achieve its goals, an actor performs some actions. An action is the triggering of an interaction
with the system-to-be. While preparing a response to the actor’s action, the system-to-be may
need assistance from external entities other than the actor who initiated the process. Recall how in
Chapter 2 
89
Object-Oriented Software Engineering
Figure 1-9 the system-to-be (ATM machine) needed assistance from a remote datacenter to
successfully complete the use case “Withdraw Cash.” This is why we will distinguish initiating
actors and participating actors. If a participating actor delivers, then the initiating actor is closer
to reaching the goal. All actors should have defined responsibilities. The system-to-be itself is an
actor and its responsibility is to assist the (initiating) actors in achieving their goals. In this
process, system-to-be may seek help from other systems or (participating) actors.
To this point we have identified the following actors:

Tenant is the home occupant

Landlord is the property owner or manager

Device is a physical device to be controlled by the system-to-be, such as lock-mechanism
and light-switch, that are controlled by our system (see Figure 1-16)

Other potential actors: Maintenance, Police, etc. (some will be introduced later)
When deciding about introducing new actors, the key question is: “Does the system provide
different service(s) to the new actor?” It is important to keep in mind that an actor is associated
with a role rather than with a person. Hence, a single actor should be created per role, but a
person can have multiple roles, which means that a single person can appear as different actors.
Also, different persons may play the same actor role, perhaps at different times.
In addition, our system may receive assistance from other systems in the course of fulfilling the
actor’s goal. In this case, the other systems will become different actors if they offer different
type of service to the system-to-be. Examples will be seen later.
Table 2-4 summarizes preliminary use cases for our case-study example.
Table 2-4: Actors, goals, and the associated use cases for the home access control system.
Actor
Actor’s Goal (what the actor intends to accomplish)
Landlord
To disarm the lock and enter, and get space lighted up.
Unlock (UC-1)
Landlord
To lock the door & shut the lights (sometimes?).
Lock (UC-2)
Landlord
To create a new user account and allow access to home.
AddUser (UC-3)
Landlord
To retire an existing user account and disable access.
RemoveUser (UC-4)
Tenant
To find out who accessed the home in a given interval of InspectAccessHistory
time and potentially file complaints.
(UC-5)
Tenant
To disarm the lock and enter, and get space lighted up.
Unlock (UC-1)
Tenant
To lock the door & shut the lights (sometimes?).
Lock (UC-2)
Tenant
To configure the device activation preferences.
SetDevicePrefs (UC-6)
LockDevice
To control the physical lock mechanism.
UC-1, UC-2
LightSwitch
To control the lightbulb.
UC-1, UC-2
[to be
identified]
To auto-lock the door if it is left unlocked for a given AutoLock (UC-2)
interval of time.
Use Case Name
Because the Tenant and Landlord actors have different responsibilities and goals, they will utilize
different use cases and thus they should be seen differently by the system. The new actor can use
more (or less, subset or different) use cases than the existing actor(s), as seen in Table 2-4. We
Ivan Marsic

90
Rutgers University
could distinguish the Maintenance actor who can do everything as the Landlord, except to
manage users. If we want to include the Maintenance but the same use cases apply for this actor
as for the Tenant, this means that there is no reason to distinguish them—we just must come up
with an actor name that covers both.
Note that the last row contains a yet-to-be-identified actor, whose goal is to
automatically arm the lock after a certain period of time expires, to account for
forgetful persons. Obviously, this is not a person’s goal, but neither is it the
system’s goal because system-to-be does nothing on its own—it must receive an
external stimulus to take action. We will see later how this can be solved.
An actor can be a person or another system which interacts with our system-to-be.
There are two main categories of actors, defined relative to a particular use case:
1
2
3
4
5
X
Y
1. Initiating actor (also called primary actor or simply user): initiates the use
case to realize a goal, which depends on the actor’s responsibilities and the
current context
2. Participating actor (also called secondary actor): participates in the use
case but does not initiate it; there are two subcategories:
(a) Supporting actor: helps the system-to-be to complete the use
case—that is, our system-to-be initiates the supporting actor
Resident
(b) Offstage actor: passively participates in the use case, i.e.,
neither initiates nor helps complete the use case, but may be
notified about some aspect of it
Landlord
Actors may be defined in generalization hierarchies, in which an abstract actor Tenant
description is shared and augmented by one or more specific actor descriptions. Actor generalization.
Table 2-4 implies that a software system is developed with a purpose/responsibility—this purpose
is assisting its users (actors) to achieve their goals. Use cases are usage scenarios and therefore
there must be an actor intentionally using this system. The issue of developer’s intentions vs.
possible usage scenarios is an important one and can be tricky to resolve. There is a tacit but
important assumption made by individual developers and large organizations alike, and that is
that they are able to control the types of applications in which their products will ultimately be
used. Even a very focused tool is designed not without potential to do other things—a clever user
may come up with unintended uses, whether serendipitously or intentionally.
Summary Use Cases
A use case is a usage scenario for an external entity, known as actor, and the system-to-be. A use
case represents an activity that an actor can perform on the system and what the system does in
response. It describes what happens when an actor disturbs our system from its “stationary state”
as the system goes through its motions until it reaches a new stationary state. It is important to
keep in mind that the system is reactive, not proactive; that is, if left undisturbed, the system
would remain forever in the equilibrium state.
Table 2-4 names the preliminary use cases for our case-study example. The reader may observe
that the summary use cases are similar to user stories (Table 2-3). Like user stories, summary use
Chapter 2 
Object-Oriented Software Engineering
91
cases do not describe details of the business process. They just identify the user’s role (actor type)
and the capability that the system-to-be will provide (to assist in achieving the actor’s goals).
The same technique for effort estimation that works for user stories (Section 2.2.3) can be applied
to summary use cases. We can use again user story points and the development velocity to
estimate the project duration by applying equation (1.1), given in Section 1.2.5. Later, in Section
4.2.2, we will describe use case points for software size measurement and effort estimation.
However, use case points cannot be applied on summary use cases, because they require detailed
use case descriptions. Detailed use case descriptions require time and effort to obtain, so they will
become available only at a later stage in the project lifecycle (see Section 2.4.3).
Casual Description of Use Cases
SIDEBAR 2.3: The Task-Artifact Cycle

Use case analysis as well as task analysis (Kirwan & Ainsworth, 1992) emerged in the
tradition of mechanization of work and the division of labor pioneered by F. W. Taylor
(Kanigel, 2005), which assumes that detailed procedure can be defined for every task. So far
we aimed to define the use cases in a technology-independent manner—the system usage
procedure should be independent of the device/artifact that currently implements the use case.
However, this is not easy or perhaps even possible to achieve, because the user activities will
depend on what steps are assumed to have been automated. For example, the details of the use
case UC-1 (Unlock) depend on the user identification technology. If a face-recognition
technology were available that automatically recognizes authorized from unauthorized users,
then UC-1 becomes trivial and requires no explicit user activities.
Consider a simple example of developing a digital wristwatch. For a
regular watch, the owner needs to manually adjust time or date when
traveling to a different time zone or at the end of a month. Therefore, we
need to define the use cases for these activities. On the other hand, if the
watch is programmed to know about the owner’s current GPS location,
time zones, calendar, leap years, daylight-saving schedule, etc., and it
has high quality hardware, then it needs no buttons at all. Some of this
information could be automatically updated if the watch is wirelessly
connected to a remote server. This watch would always show the correct
time because everything is automated. It has no use cases that are
initiated by the human owner, unlike a manually-operated watch. The
interested reader should consult (Vicente, 1999: p. 71, 100-106) for
further critique of how tasks affect artifacts and vice versa.

Ivan Marsic
92
Rutgers University
system
boundary
First tier use cases
actor
Second tier use cases
«inc
«initiate»
«in
it
Tenant
i at
»
lude
UC7: AuthenticateUser
UC1: Unlock
e»
«participate»
»
p at e
rtici
«pa
«particip
ate»
LockDevice
ate»
«particip
te»
i t ia
n
i
«
«initiate»
UC2: Lock
«pa
rt
icip
a
LightSwitch
te»
«initiate + participate»
Landlord
communication
use case
Timer
Figure 2-13: UML use case diagram for the device-control subsystem of the home access
system. Compare with Table 2-4.
Use Case Diagram
Figure 2-13 sums up the actors, use cases, and their relationships in a so-called use case
diagram. There are two use-case categories distinguished: “first-” vs. “second tier.” The “first
tier” use cases represent meaningful services provided by the system to an actor. The “second
tier” use cases represent elaborations or sub-services of the main services. In a sense, they are
equivalent of subroutines in programs because they capture some repetitive activity and can be
reused in multiple locations. The figure also shows the relationship of use cases in different tiers.
The two stereotypical- or cliché types of relationship are:

«extend» – optional extensions of the main case

«include» – required subtasks of the main case
The developer can introduce new stereotypes or clichés for representing the use case
relationships. Note also that the labels on communication lines («initiate» and «participate») are
often omitted to avoid cluttering.
The AuthenticateUser use case is not a good candidate for first tier use cases, because it does not
represent a meaningful stand-alone goal for an initiating actor. It is, however, useful to show it
explicitly as a second tier use case, particularly because it reveals which use cases require user
authentication. For example, one could argue that Lock does not need authentication, because
performing it without authentication does not represent a security threat. Similarly, Disable
should not require authentication because that would defeat the purpose of this case. It is of note
that these design decisions, such as which use case does or does not require authentication, may
need further consideration and justification. The reader should not take these lightly, because
each one of them can have serious consequences, and the reader is well advised to try to come up
with scenarios where the above design decisions may not be appropriate.
SIDEBAR 2.4: Is Login a Use Case?
Chapter 2 
93
Object-Oriented Software Engineering
Account Management Subsystem
«initiate»
«ini
tiat
UC3: AddUser
« in
clu
de
»
e»
Landlord
at
ip
tr ic
a
«p
e»
te»
«initia
«inclu
de
UC4: RemoveUser
»
UC8: Login
de»
clu
«in
UC5: InspectAccessHistory
«initia
te»
«
cl
in
e»
ud
UC6: SetDevicePrefs
Tenant
Figure 2-14: Use cases for the account-management subsystem of the home access system.
 A novice developer frequently identifies user login as a use case. On the other hand, expert
developers argue that login is not a use case. Recall that use case is motivated by user’s goal;
The user initiates interaction with the system to achieve a certain goal. You are not logging in
for the sake of logging in—you are logging in to do some work, and this work is your use case.
BAD:
GOOD:
Login
AddUser
AddUser
«inc
lude
»
Login
Landlord
SetDevicePrefs
Landlord
SetDevicePrefs
«inc
»
lude
The reader should not mistake the use case diagram for use cases. The diagram serves only to
capture an overview of the system services in a concise visual form. It summarizes the system
features and their relationships, without detailing how each feature should operate. Unlike this,
use cases are text stories that detail exactly what happens when an actor attempts to obtain a
service from the system. A helpful analogy is a book’s index vs. contents: a table-of-contents or
index is certainly useful, but the actual content represents the book’s main value. Similarly, a use
case diagram provides a useful overview index, but you need the actual use cases (contents) to
understand what the system does or is supposed to do.
Figure 2-14 shows the use cases for the second subsystem of the safe home access system, which
supports various account management activities. The diagrams in Figure 2-13 and Figure 2-14
form the use case diagram of the entire system.
Ivan Marsic 
94
Rutgers University
Figure 2-15 shows additional relationships among use cases that can be used to improve the
informativeness of use case diagrams. For example, use cases that share common functionality
can be abstracted in a more general, “base” use case (Figure 2-15(a)). If a user’s goal has several
subgoals, some of which are optional, we can indicate this information in a use case diagram
using the «extend» stereotype. For example, we may design a use case to allow the user to
manage his account. As part of account management, optional activities that may or may not take
place are the inspection of access history and configuring the device-activation preferences
(Figure 2-15(b)).
2.4.2
System Boundary and Subsystems
Determining the System Boundary
ManageUsers
»
tend
«ex
UC5: InspectAccessHistory
ManageAccount
UC3: AddUser
UC4: RemoveUser
(a)
«ex
tend
»
UC6: SetDevicePrefs
(b)
Figure 2-15: More relationships among use cases: (a) Use case generalization; (b) Optional
use cases, denoted with the «extend» stereotype.
Unfortunately, there are no firm guidelines of delineating the boundary of the system under
development. Drawing the system boundary is a matter of choice. However, once the boundary is
drawn, the interactions for all the actors must be shown in use cases in which they interact with
the system-to-be.
Chapter 2 
95
Object-Oriented Software Engineering
Case (a):
Local face recognition
Local
computer
Face
image
Apartment building
Security
camera
Case (b):
Remote face recognition
Network
Network
FaceReco, Ltd.
Figure 2-16: Alternative cases of face recognition the for secure home access system.
Consider a variation of the home access control system which will be used for an apartment
building, or a community of apartment buildings, rather than a single-family home. The
management demands user identification based on face recognition, instead of alphanumeric
password keys. Roughly speaking, a face recognition system works by taking an image of a
person’s face (“mug shot”), compares it with the known faces, and outputs a Boolean result:
“authorized” or “unauthorized” user. Here are two variations (see Figure 2-16):
(a) You procure face recognition software, install it on your local computer, and link it up
with a standard relational/SQL database for memorizing the faces of legitimate users.
(b) After a preliminary study, you find that maintaining the database of legitimate faces,
along with training the recognition system on new faces and unlearning the faces of
departed residents, are overly complex and costly. You decide that the face recognition
processing should be outsourced to a specialized security company, FaceReco, Ltd. This
company specializes in user authentication, but they do not provide any applicationspecific services. Thus, you still need to develop the rest of the access control system.
The first task is to identify the actors, so the issue is: Are the new tools (face recognition software
and relational database) new actors or they are part of the system and should not be distinguished
from the system? In case (a), they are not worth distinguishing, so they are part of the planned
system. Although each of these is a complex software system developed by a large organization,
as far as we (the developer) are concerned, they are just modules that provide data-storage and
user-authentication. Most importantly, they are under our control, so there is nothing special
about them as opposed to any other module of the planned system.
Therefore, for case (a), everything remains the same as in the original design. The use case
diagram is shown in Figure 2-13 and Figure 2-14.
Ivan Marsic 
96
Rutgers University
UC2: Lock
«p
Tenant
art
ic
UC1: Unlock
ipa
te»
te»
«initia
UC7: AuthenticateUser
«in
clu
de
»
UC4: RemoveUser
«pa
rtic
«participate»
UC3: AddUser
«initia
te»
Landlord
LockDevice
«inc
lude
»
«inclu
de
ate»
«particip
ipa
te»
FaceReco, Ltd.
»
UC8: Login
Figure 2-17: Part of the modified use case diagram for that includes a new actor: FaceReco.
See text for details.
For case (b), a part of the new use case diagram is shown in Figure 2-17 (the rest remains the
same as in Figure 2-13 and Figure 2-14). Now we need to distinguish a new actor, the FaceReco
Company which provides authentication services. There is a need-to-know that they are part of
the process of fulfilling some goal(s) of initiating actors.
Subsystems and Software Architecture
Figure 2-18 shows the traceability matrix that maps the system requirements to use cases. Its
purpose is to check that all requirements are covered by the use cases and none of the use cases is
created without a reason (i.e., without a requirement from which it was derived). If a use case is
derived from a requirement, then the corresponding entry in the matrix is checked. The Max PW
(priority weight) row shows the maximum priority of any checked requirement in the column
above. The bottom row shows the Total PW of each use case obtained by summing up the
priorities of the checked requirements in the column above. The Max PW and Total PW values
are used to schedule the work on implementing the sue cases. The highest-priority use cases will
be elaborated, implemented, and delivered the first.
2.4.3
Detailed Use Case Specification
A detailed use case description represents a use case of the system as a sequence of interactions
between external entities (actors) and the system-to-be. Detailed use cases are usually written as
usage scenarios or scripts, listing a specific sequence of actions and interactions between the
actors and the system. For use case scenarios, we will use a stepwise, “recipe-like” description. A
scenario describes in a step-by-step manner activities that an actor does and how the system
responds. A scenario is also called a use case instance, in the sense that it represents only one of
Chapter 2 
97
Object-Oriented Software Engineering
Req’t PW UC1 UC2 UC3 UC4 UC5 UC6 UC7 UC8
REQ1 5
X
X
REQ2 2
REQ3 5
X
REQ4 4
X
X
REQ5 2
X
X
X
X
X
REQ6 1
X
X
X
X
REQ7 2
REQ8 1
X
X
REQ9 1
X
X
Max PW
5
2
2
2
1
5
2
1
Total PW
15
3
2
2
3
9
2
3
Figure 2-18: Requirements-to-use-cases traceability matrix for the safe home access case
study. Priority weight (PW) given in Table 2-1. (Traceability continued in Figure 2-28.)
several possible courses of action for a given use case. Use cases specify what information must
pass the boundary of a system when a user or another system interacts with it.
We usually first elaborate the “normal” scenario, also called main success scenario, which
assumes that everything goes perfect. Because everything flows straightforward, this scenario
usually does not include any conditions or branching—it flows linearly. It is just a causal
sequence of action/reaction or stimulus/response pairs. Figure 2-19 shows the use case schema.9
Alternate scenarios or extensions in a use case can result from:

Inappropriate data entry, such as the actor making a wrong menu-item choice (other than
the one he/she originally intended), or the actor supplies an invalid identification.

System’s inability to respond as desired, which can be a temporary condition or the
information necessary to formulate the response may never become available.
For each of the alternate cases we must create an event flow that describes what exactly happens
in such a case and lists the participating actors. Alternate scenarios are even more important than
the main success scenario, because they often deal with security issues.
Although we do not know what new uses the user will invent for the system, purposeful
development is what governs the system design. For example, attempt to burglarize a home may
be a self-contained and meaningful goal for certain types of system users, but this is not what we
are designing the system for—this is not a legal use case; rather, this must be anticipated and
treated as an exception to a legal use case. We will consider such “abuse cases” as part of security
and risk management (Section 2.4.4).
9
The UML standard does not specify a use case schema, so the format for use cases varies across different
textbooks. Additional fields may be used to show other important information, such as non-functional
requirements associated with the use case.
Ivan Marsic 
Rutgers University
Use Case UC-#:
Name / Identifier [verb phrase]
Related Require’ts:
List of the requirements that are addressed by this use case
Initiating Actor:
Actor who initiates interaction with the system to accomplish a goal
Actor’s Goal:
Informal description of the initiating actor’s goal
98
Participating Actors: Actors that will help achieve the goal or need to know about the outcome
Preconditions:
What is assumed about the state of the system before the interaction starts
Postconditions:
What are the results after the goal is achieved or abandoned; i.e., what
must be true about the system at the time the execution of this use case is
completed
Flow of Events for Main Success Scenario:
1. The initiating actor delivers an action or stimulus to the system (the arrow indicates

the direction of interaction, to- or from the system)
2. The system’s reaction or response to the stimulus; the system can also send a message

to a participating actor, if any
3. …

Flow of Events for Extensions (Alternate Scenarios):
What could go wrong? List the exceptions to the routine and describe how they are handled
 1a. For example, actor enters invalid data
 2a. For example, power outage, network failure, or requested data unavailable
…
The arrows on the left indicate the direction of communication:  Actor’s action;  System’s reaction
Figure 2-19: A general schema for UML use cases.
A note in passing, the reader should observe that use cases are not specific to the object-oriented
approach to software engineering. In fact, they are decidedly process-oriented rather than objectoriented, for they focus on the description of activities. As illustrated in Figure 1-9(a), at this
stage we are not seeing any objects; we see the system as a “black box” and focus on the
interaction protocol between the actor(s) and the black box. Only at the stage of building the
domain model we encounter objects, which populate the black box.
Detailed Use Cases
Detailed use cases elaborate the summary use cases (Table 2-4). For example, for the use case
Unlock, the main success scenario in an abbreviated form may look something like this:
Use Case UC-1:
Unlock
Related Requirem’ts:
REQ1, REQ3, REQ4, and REQ5 stated in Table 2-1
Initiating Actor:
Any of: Tenant, Landlord
Actor’s Goal:
To disarm the lock and enter, and get space lighted up automatically.
Participating Actors:
LockDevice, LightSwitch, Timer
Preconditions:
• The set of valid keys stored in the system database is non-empty.
• The system displays the menu of available functions; at the door
Chapter 2 
Object-Oriented Software Engineering
99
keypad the menu choices are “Lock” and “Unlock.”
Postconditions:
The auto-lock timer has started countdown from autoLockInterval.
Flow of Events for Main Success Scenario:
 1. Tenant/Landlord arrives at the door and selects the menu item “Unlock”
2. include::AuthenticateUser (UC-7)
 3. System (a) signals to the Tenant/Landlord the lock status, e.g., “disarmed,” (b) signals
to LockDevice to disarm the lock, and (c) signals to LightSwitch to turn the light on
 4. System signals to the Timer to start the auto-lock timer countdown
 5. Tenant/Landlord opens the door, enters the home [and shuts the door and locks]
In step 5 above, the activity of locking the
door is in brackets, because this is covered
under the use case Lock, and does not concern
this use case. Of course, we want to ensure
that this indeed happens, which is the role of
an auto-lock timer, as explained later. An
extension scenario for the above use case may
specify how the system-to-be will behave
should the door be unlocked manually, using a
physical key.
Although extensions or alternate scenarios are
not listed in the description of UC-1, for each
of the steps in the main success scenario we must consider what could go wrong. For example,

In Step 1, the actor may make a wrong menu selection

Exceptions during the actor authentication are considered related to UC-7

In Step 5, the actor may be held outside for a while, e.g., greeting a neighbor
For instance, to address the exceptions in Step 5, we may consider installing an infrared beam in
the doorway that detects when the person crosses it. Example alternate scenarios are given next
for the AuthenticateUser and Lock use cases.
In step 2 of UC-1, I reuse a “subroutine” use case, AuthenticateUser, by keyword “include,”
because I anticipate this will occur in other use cases, as well. Here is the main scenario for
AuthenticateUser as well as the exceptions, in case something goes wrong:
Use Case UC-7:
AuthenticateUser (sub-use case)
Related Requirements:
REQ3, REQ4 stated in Table 2-1
Initiating Actor:
Any of: Tenant, Landlord
Actor’s Goal:
To be positively identified by the system (at the door interface).
Participating Actors:
AlarmBell, Police
Preconditions:
• The set of valid keys stored in the system database is non-empty.
• The counter of authentication attempts equals zero.
Postconditions:
None worth mentioning.
Flow of Events for Main Success Scenario:
Ivan Marsic 



100
Rutgers University
1. System prompts the actor for identification, e.g., alphanumeric key
2. Tenant/Landlord supplies a valid identification key
3. System (a) verifies that the key is valid, and (b) signals to the actor the key validity
Flow of Events for Extensions (Alternate Scenarios):
2a. Tenant/Landlord enters an invalid identification key
1. System (a) detects error, (b) marks a failed attempt, and (c) signals to the actor

1a. System (a) detects that the count of failed attempts exceeds the maximum

allowed number, (b) signals to sound AlarmBell, and (c) notifies the Police actor
of a possible break-in
2. Tenant/Landlord supplies a valid identification key

3. Same as in Step 3 above
When writing a usage scenario, you should focus on what is essential to achieve
the initiating actor’s goal and avoid the details of how this actually happens. Focus
on the “what” and leave out the “how” for the subsequent stages of the
development lifecycle. For example, in Step 2 of the use case AuthenticateUser, I
just state that the user should provide identification; I do not detail whether this is
done by typing on a keypad, by scanning an RFID tag, or by some biometric
technology.
Latch
bolt
Strike
plate
Dead
bolt
At the time of writing detailed use cases, we also write the corresponding user acceptance tests. In
the context of use cases, a user acceptance test case is a detailed procedure that fully tests a use
case or one of its flows of events. Recall that use cases are part of requirements engineering, and
the customer should help with specifying the acceptance tests. The focus is on what the user does,
not what the system does. This means that the test cases must be designed around the actual tasks
that the user will need to perform. Use-case-based acceptance tests are similar to acceptance tests
described in Section 2.2.1. As mentioned, testing functions that involve multi-step interaction
requires more than just specifying the input data and expected outcomes. Here we are able to
provide detailed steps for pass and fail conditions, because by now we have elaborated step-bystep scenarios for use cases. Here is an example test case for testing the use case UC-1.
Test-case Identifier:
TC-1.01
Use Case Tested:
UC-1, main success scenario, and UC-7
Pass/fail Criteria:
The test passes if the user enters a key that is contained in the database,
with less than a maximum allowed number of unsuccessful attempts
Input Data:
Numeric keycode, door identifier
Test Procedure:
Expected Result:
Step 1. Type in an incorrect keycode and a System beeps to indicate failure;
valid door identifier
records unsuccessful attempt in the database;
prompts the user to try again
Step 2. Type in the correct keycode and door System flashes a green light to indicate success;
identifier
records successful access in the database;
disarms the lock device
An acceptance test needs to convince the customer that the system works as expected.
Chapter 2 
Object-Oriented Software Engineering
101
We continue the elaboration of use cases with the main success scenario for the Lock use case:
Use Case UC-2:
Lock
Related Requirements:
REQ1, REQ2, and REQ5 stated in Table 2-1
Initiating Actor:
Any of: Tenant, Landlord, or Timer
Actor’s Goal:
To lock the door & get the lights shut automatically (?)
Participating Actors:
LockDevice, LightSwitch, Timer
Preconditions:
The system always displays the menu of available functions.
Postconditions:
The door is closed and lock armed & the auto-lock timer is reset.
Flow of Events for Main Success Scenario:
 1. Tenant/Landlord selects the menu item “Lock”
 2. System (a) signals affirmation, e.g., “lock armed,” (b) signals to LockDevice to arm
the lock (if not already armed), (c) signal to Timer to reset the auto-lock counter, and
(d) signals to LightSwitch to turn the light off (?)
Flow of Events for Extensions (Alternate Scenarios):
2a. System senses that the door is not closed, so the lock cannot be armed
 1. System (a) signals a warning that the door is open, and (b) signal to Timer to start the
alarm counter
 2. Tenant/Landlord closes the door
 3. System (a) senses the closure, (b) signals affirmation to the Tenant/Landlord, (c)
signals to LockDevice to arm the lock, (d) signal to Timer to reset the auto-lock
counter, and (e) signal to Timer to reset the alarm counter
Note that in this case, the auto-lock timer appears as both the initiating and participating actor for
this use case. (This is also indicated in the use case diagram in Figure 2-13.) This is because if the
timeout time expires before the timer is reset, Timer automatically initiates the Lock use case, so
it is an initiating actor. Alternatively, if the user locks the door before the timeout expires, the
timer will be reset, so it is an offstage actor, as well.
I also assume that a single Timer system can handle multiple concurrent requests. In the Lock use
case, the timer may be counting down the time since the lock has been disarmed. At the same
time, the system may sense that the door is not closed, so it may start the alarm timer. If the door
is not shut within a given interval, the system activates the AlarmBell actor and may notify the
Police actor.
You may wonder why not just say that the system will somehow handle the auto-lock
functionality rather than going into the details of how it works. Technically, the timer is part of
the system-to-be, so why should it be declared an external actor?! Recall that the system is always
passive—it reacts to an external stimulus but does nothing on its own initiative. Thus, to get the
system perform auto-lock, somebody or something must trigger it to do so. This is the
responsibility of Timer. Timer is an external stimulus source relative to the software under
development, although it will be part of the end hardware-plus-software system.
Next follows the description of the ManageUsers use case:
Use Case UC-3:
AddUser
Related Requirements:
REQ6 stated in Table 2-1
Ivan Marsic 
Rutgers University
102
Initiating Actor:
Landlord
Actor’s Goal:
To register new or remove departed residents at runtime.
Participating Actors:
Tenant
Preconditions:
None worth mentioning. (But note that this use case is only available
on the main computer and not at the door keypad.)
Postconditions:
The modified data is stored into the database.
Flow of Events for Main Success Scenario:
1. Landlord selects the menu item “ManageUsers”

2. Landlord identification: Include Login (UC-8)
3. System (a) displays the options of activities available to the Landlord (including “Add

User” and “Remove User”), and (b) prompts the Landlord to make selection
4. Landlord selects the activity, such as “Add User,” and enters the new data

5. System (a) stores the new data on a persistent storage, and (b) signals completion

Flow of Events for Extensions (Alternate Scenarios):
4a. Selected activity entails adding new users: Include AddUser (UC-3)
4b. Selected activity entails removing users: Include RemoveUser (UC-4)
The Tenant is a supporting actor for this use case, because the tenant will input his identification
(password or a biometric print) during the registration process. Note that in UC-3 we include the
subordinate use case Login (UC-8), which is not the same as AuthenticateUser, numbered UC-7.
The reason is that UC-7 is designed to authenticate persons at the entrance(s). Conversely, user
management is always done from the central computer, so we need to design an entirely different
use case. The detailed description of the use case AddUser will be given in Problem 2.19 and
RemoveUser is similar to it.
In Table 2-4 we introduced UC-5: Inspect Access History, which roughly addresses REQ8 and
REQ9 in Table 2-1. I will keep it as a single use case, although it is relatively complex and the
reader may wish to split it into two simpler use cases. Here is the description of use case UC-5:
Use Case UC-5:
Inspect Access History
Related Requirements: REQ8 and REQ9 stated in Table 2-1
Initiating Actor:
Any of: Tenant, Landlord
Actor’s Goal:
To examine the access history for a particular door.
Participating Actors:
Database, Landlord
Preconditions:
Tenant/Landlord is currently logged in the system and is shown a
hyperlink “View Access History.”
Postconditions:
None.
Flow of Events for Main Success Scenario:
 1. Tenant/Landlord clicks the hyperlink “View Access History”
 2. System prompts for the search criteria (e.g., time frame, door location, actor role, event
type, etc.) or “Show all”
 3. Tenant/Landlord specifies the search criteria and submits
 4. System prepares a database query that best matches the actor’s search criteria and
Chapter 2 

5.
6.


7.
8.

Object-Oriented Software Engineering
103
retrieves the records from the Database
Database returns the matching records
System (a) additionally filters the retrieved records to match the actor’s search criteria;
(b) renders the remaining records for display; and (c) shows the result for
Tenant/Landlord’s consideration
Tenant/Landlord browses, selects “interesting” records (if any), and requests further
investigation (with an accompanying complaint description)
System (a) displays only the selected records and confirms the request; (b) archives the
request in the Database and assigns it a tracking number; (c) notifies Landlord about
the request; and (d) informs Tenant/Landlord about the tracking number
The following example illustrates deriving a use case in a different domain. The main point of
this example is that use cases serve as a vehicle to understand the business context and identify
the business rules that need to be implemented by the system-to-be.
Example 2.1
Restaurant Automation, terminating a worker employment
Consider the restaurant automation project described on the book website (given in Preface). One of
the requirements states that the restaurant manager will be able to manage the employment status:
REQ8: The manager shall be able to manage employee records for newly hired employees, job
reassignments and promotions, and employment termination.
Based on the requirement, a student team derived a use case for employment termination, as shown:
Use Case UC-10:
Terminate Employee
(FIRST VERSION)
REQ8
Related Requirem’ts:
Manager
Initiating Actor:
To fire or layoff an employee.
Actor’s Goal:
Participating Actors:
Preconditions:
Postconditions:
 System successfully updated Employee List.
Failed End Condition:
 Employee List failed to update
Flow of Events for Main Success Scenario:
1. Manager selects Delete option next to employee’s name

2. System asks Manager to confirm that the selected employee should be deleted from list

3. Manager confirms action to delete the employee

4. (a) System removes the employee from Employee List; (b) updates Employee List

Flow of Events for Extensions (Alternate Scenarios):
3a. Manager selects Cancel option
1. System does not delete the employee

So then, dismissing an employee is as simple as deleting a list item! I pointed out that in real world
nothing works so simple. We are not talking about some arbitrary database entries that can be edited as
someone pleases. These entries have certain meaning and business significance and there must be
some rules on how they can be edited. This is why the developer must talk to the customer to learn the
business rules and local laws. Even a student team doing an academic project (and not having a real
customer) should visit a local restaurant and learn how it operates. As a minimum, they could do some
Web research. For example, employee rights in the state of New York are available here:
http://www.ag.ny.gov/bureaus/labor/rights.html. The manager must ensure that the employee received any
remaining pay; that the employee returned all company belongings, such as a personal computer, or
whatever; the manager may also need to provide some justification for the termination; etc. As a result,
it is helpful to refine our requirement:
Ivan Marsic 
104
Rutgers University
REQ8: The manager shall be able to manage employee records for newly hired employees, job
reassignments and promotions, and employment termination in accord with local laws.
Back to the drawing board, and the second version looked like this:
Use Case UC-10:
Related Requirem’ts:
Initiating Actor:
Actor’s Goal:
Participating Actors:
Preconditions:
Terminate Employee
REQ8
Manager
To fire or layoff an employee.
(SECOND VERSION)
 Removed employee is not currently logged into the system.
 Removed employee is has already clocked out.
 Removed employee has returned all company belongings.
Postconditions:
 System successfully updated Employee List.
Failed End Condition:
 Employee List failed to update.
Flow of Events for Main Success Scenario:
1. Manager selects Delete option next to employee’s name

2. System asks Manager to confirm that the selected employee should be deleted from list

3. Manager confirms action to delete the employee

4. (a) System confirms the employee has been paid; (b) removes the employee from

Employee List; (c) updates Employee List
Flow of Events for Extensions (Alternate Scenarios):
3a. Manager selects Cancel option
1. System does not delete the employee

4a. System alerts Manager that the employee has not been paid
1. System does not remove the employee from employee roster and aborts this use case

2. System opens Payroll <<include>> ManagePayroll (UC-13)

I thought this is an amazing trick: whatever you find difficult to solve in your use case, just state it in
the preconditions, so that the initiating actor ensures that the system can do its work smoothly! The
user serves the system instead the system serving the user. Compared to the first version, almost
nothing was changed except that all the hard issues are now resolved as preconditions. Also, in step
4(a), it is not clear how can System confirm that the employee has been paid? And if the employee has
not been paid, the alternative scenario throws the user into another use case, ManagePayroll (UC-13),
where he will again just update some database record. However, updating a database record does not
ensure that the employee has actually received his payment!
In the age of automation, we should strive to have computer systems do more work and human users
do less work. A professionally prepared use case for restaurant employment termination should look
something like this:
Use Case UC-10:
Related Requirem’ts:
Initiating Actor:
Actor’s Goal:
Participating Actors:
Preconditions:
Postconditions:
Terminate Employee
REQ8
Manager
To fire or layoff an employee.
(THIRD VERSION)
Employee’s record is moved to a table of past employees for auditing
purposes.
Failed End Condition:
Flow of Events for Main Success Scenario:
1. Manager enters the employee’s name or identifier

2. System displays the employee’s current record

3. Manager reviews the record and requests termination of employment

4. System asks the manager to select the reason for termination, such as layoff, firing, etc.

and the date when the termination will take effect
Chapter 2 


Object-Oriented Software Engineering
105
5.
6.
Manager selects the reason for termination and the effective date
(a) System checks that in case of firing the decision is justified with past written warnings
documenting poor performance, irresponsibility or breaches of policy.
(b) System informs the Manager about the benefits the employee will receive upon
departure, such as severance pay or unused vacation days, and asks the Manager to
confirm after informing the employee (in person or by email)
7. Manager confirms that the employee has been informed.

8. (a) System makes a record of any outstanding wages and the date by which they should

be mailed to the employee as required by the local laws. A new record is created in a
table of pending actions.
(b) System checks if the employee is currently logged in into the company’s computer
system; if yes, it automatically logs off and blocks the employee’s account
(c) System checks the employee record and informs the Manager the employee before
leaving should return any restaurant-owned uniforms, keys or property that was issued to
the employee
9.
Manager
confirms that the employee has returned all company belongings

10.
System
moves
the employee record to a table of past employees, informs the Manager,

and queries the Manager if he or she wishes to post a classifieds advertisement for a new
employee
 11. Manager declines the offer and quits this use case
Flow of Events for Extensions (Alternate Scenarios):
6a. System determines that the decision firing has not been justified with past written warnings
1. System informs the Manager that because of a lack of justification, the company may be

liable to legal action, and asks the Manager to decide whether to continue or cancel
2. Manager selects one option and the System continues from Step 6(b) in Main Scenario

9a. Manager selects the items currently with the employee
1. System (a) asks the Manager whether to continue or suspend the process until the

employee has returned all company belongings; (b) saves the current unfinished
transaction in a work-in-progress table and sets a period to send reminders for completion
Note that preconditions are not indicated because I could not think of any condition that, if not met,
would make impossible the execution of this use case. Similarly, no postconditions are indicated. One
may think that an appropriate precondition is that this employee should exist in the database. However,
it is conceivable that in Step 2 of the main success scenario the system cannot find any record of this
employee, in which case this should be handled as an alternate scenario.
An additional feature to consider may be that the system initiates a classifieds advertisement to fill the
vacant position created by terminating this employee. It is great to invent new features, but the
developer must make it clear that adding new features will likely postpone the delivery date and
increase project costs. Only our customer can make such decisions.
System Sequence Diagrams
A system sequence diagram represents in a visual form a usage scenario that an actor experiences
while trying to obtain a service from the system. In a way, they summarize textual description of
the use case scenarios. As noted, a use case may have different scenarios, the main success
scenario and the alternate ones. A system sequence diagram may represent one of those scenarios
in entirety, or a subpart of a scenario. Figure 2-20 shows two examples for the Unlock use case.
Ivan Marsic 
106
Rutgers University
: System
User
«initiating actor»
select function(“unlock")
LockDevice
LightSwitch
«supporting actor» «supporting actor»
Timer
«offstage actor»
prompt for the key
(a)
enter key
verify key
signal: valid key, lock open
open the lock
turn on the light
start ("duration“)
: System
User
«initiating actor»
select function(“unlock")
loop
(b)
AlarmBell
Police
«supporting actor»
«offstage actor»
prompt for the key
enter key
verify key
signal: invalid key
prompt to try again
sound alarm
notify intrusion
Figure 2-20: UML System sequence diagrams for the Unlock use case: (a) main success
scenario; (b) burglary attempt scenario. The diagram in (b) shows UML “loop” interaction
frame, which delineates a fragment that may execute multiple times.
The key purpose of system sequence diagrams (as is the case with use cases) is to represent what
information must pass the system boundary and in what sequence. Therefore, a system sequence
diagram can contain only a single box, named System, in the top row, which is our system-to-be.
All other entities must be actors. At this stage of the development cycle, the system is still
considered atomic and must not be decomposed into its constituent parts.
It may appear that we are “peering inside” the black box when stating what system does internally
during the actor  system exchanges. But, note that we are specifying the “what” that the black
box does, not “how” this gets done.
Activity Diagrams
Chapter 2 
107
Object-Oriented Software Engineering
Use Cases for Requirements Engineering
Use cases are a popular tool for gathering requirements and specifying system behavior.
However, I do not want to leave the reader with an illusion that use cases are the ultimate solution
for requirements analysis. As any other tool, they have both virtues and shortcomings. I hope that
the reader experienced some virtues from the preceding presentation. On the shortcomings side,
the suitability of use cases for gathering requirements may be questioned because you start
writing the use cases given the requirements. Also, use cases are not equally suitable for all
problems. Considering the projects defined in Section 1.5 and the book website (given in
Preface), use cases seem to be suitable for the restaurant automation project. In general, the use
case approach is best suited for the reactive and interactive systems. However, they do not
adequately capture activities for systems that are heavily algorithm-driven, such as the virtual
biology lab and financial markets simulation (both described at the book website, given in
Preface), or data-intensive, such as databases or large data collections. Some alternative or
complementary techniques are described in Chapter 3.
2.4.4
Security and Risk Management
A business process risk is the chance of something happening that will have an impact on the
process objectives, such as financial or reputational damage, and is measured in terms of
likelihood and consequence. Identifying and preempting the serious risks that will be faced by the
system is important for any software system, not only for the ones that work in critical settings,
such as hospitals, etc. Potential risks depend on the environment in which the system is to be
used. The root causes and potential consequences of the risks should be analyzed, and reduction
or elimination strategies devised. Some risks are intolerable while others may be acceptable and
should be reduced only if economically viable. Between these extremes are risks that have to be
tolerated only because their reduction is impractical or grossly expensive. In such cases the
probability of an accident arising because of the hazard should be made as low as reasonably
practical (ALARP).
Risk Identification
Lock left disarmed (when it should be armed)
Lock does not disarm (faulty mechanism)
Risk Type
Intolerable
ALARP
Risk Reduction Strategy
Auto-lock after autoLockInterval
Allow physical key use as alternative
To address the intolerable risk, we can design an automatic locking system which observes the
lock state and auto-locks it after autoLockInterval seconds elapses. The auto-locking system
could be made stand-alone, so its failure probability is independent of the main system. For
example, it could run in a different runtime environment, such as separate Java virtual machines,
or even on a separate hardware and energy source.
The probability of a risk occurrence is usually computed based on historical data.
a risk condition capturing the situation upon which the risk of a given fault may occur. a fault in a
business process is an undesired state of a process instance which may lead to a process failure
(e.g. the violation of a policy may lead to a process instance being interrupted). Identifying a fault
in a process requires determining the condition upon which the fault occurs. If a risk is detected
Ivan Marsic 
108
Rutgers University
during requirements analysis, remedial actions should be taken to rectify the use case design and
prevent an undesired state of the business process (fault for short), from occurring.
Risk Identification phase, where risk analysis is carried out to identify risks in the process model
to be designed. Traditional risk analysis methods such as IEC 61025 Fault Tree Analysis (FTA)
and Root Cause Analysis can be employed in this phase. The output of this phase is a set of risks,
each expressed as a risk condition.
M. Soldal Lund, B. Solhaug, and K. Stolen. Model-Driven Risk Analysis. Springer, 2011.
Risk analysis involves more than just considering “what could go wrong” in different steps of
use-case scenarios. It is possible that each step is executed correctly, but system as a whole fails.
Such scenarios represent misuse cases. For example, an important requirement for our safehome-access system is to prevent dictionary attacks (REQ4 in Table 2-1). As described later in
Section 2.5.2, we need to count the unsuccessful attempts, but also need to reset the counter if the
user leaves before providing a valid key or reaching the maximum allowed number of
unsuccessful attempts. To detect such situations, the system may run a timer for
maxAttemptPeriod duration and then reset the counter of unsuccessful attempts.
Assume that an intruder somehow learned the maximum of allowed unsuccessful
attempts and maxAttemptPeriod. The intruder can try a dictionary attack with the
following misuse case:
invalid-key, invalid, …  maxNumOfAttempts ; wait maxAttemptPeriod ; invalid, invalid, …
To ensure fault tolerance, a stand-alone system should monitor the state-variable values and
prohibit the values out of the safe range, e.g., by overwriting the illegal value. Ideally, a different
backup system should be designed for each state variable. This mechanism can work even if it is
unknown which part of the main program is faulty and causes the illegal value to occur.
A positive aspect of a stand-alone, one-task-only system is its simplicity and lack of
dependencies, inherent in the main system, which makes it resilient; a negative aspect is that the
lack of dependencies makes it myopic, not much aware of its environment and unable to respond
in sophisticated manner.
2.4.5
Why Software Engineering Is Difficult (2)
“It’s really hard to design products by focus groups. A lot of times, people don’t know what they want until
you show it to them.” —Steve Jobs, BusinessWeek, May 25 1998
A key reason, probably the most important one, is that
we usually know only approximately what we are to
do. But, a general understanding of the problem is not
enough for success. We must know exactly what to do
because programming does not admit vagueness—it is
a very explicit and precise activity.
Disciplined development
Idea
Product
SOFTWARE ENGINEERING
History shows that projects succeed more often when requirements are well managed.
Requirements provide the basis for agreement with the users and the customer, and they serve as
the foundation for the work of the development team. Software defects often result from
Chapter 2 
109
Object-Oriented Software Engineering
(a)
(b)
Domain Model
System
Concept 3
Concept 1
Concept n
1
7
4
0
8
5
2
9
6
3
Actor
(Bank
customer)
(AT
Actor
Concept 2
Actor
Mm
a ch
ine
)
(Remote
datacenter)
Actor
Figure 2-21: (a) Use case model sees the system as a “black box.” (b) Domain model peers
inside the box to uncover the constituent entities and their (static) relations that make the
black box behave as described by its use cases.
misunderstanding the requirements, not only because of inadequate developmental skills. This
means that requirements provide a vital linkage to ensure that teams deliver systems that solve
real business problems. You need to ensure that you “do the right thing the right way.”
When faced with a difficult decision, it is a good idea to ask the customer for help. After all, the
customer can judge best what solution will work best for him and he will easier accept
compromises if they were involved in making them. However, this is not always simple. Consider
the projects described at the book website (given in Preface). Asking the customer works fine in
the restaurant automation project. Even in the virtual biology lab, we can interview a biology
course instructor to help with clarifying the important aspects of cell biology. However, who is
your customer in the cases of vehicle traffic monitoring and stock investment fantasy league
(Section 1.5.1)? As discussed in the description of the traffic-monitoring project, we are not even
sure whom the system should be targeted to.
More about requirements engineering and system specification can be found in Chapter 3.
2.5 Analysis: Building the Domain Model
“I am never content until I have constructed a mechanical model of the subject I am studying.
If I succeed in making one, I understand; otherwise I do not.” —Lord Kelvin (Sir William Thomson)
Use cases looked at the system’s environment (actors) and the system’s external behavior. Now
we turn to consider the inside of the system. This shift of focus is contrasted in Figure 2-21. In
Section 1.2.3 I likened object-oriented analysis to setting up an enterprise. The analysis phase is
concerned with the “what” aspect—identifying what workers need to be hired and what things
acquired. Design (Section 2.6) deals with the “how” aspect—how these workers interact with
each other and with the things at their workplace to accomplish their share in the process of
fulfilling a service request. Of course, as any manager would tell you, it is difficult to make a
Ivan Marsic 
110
Rutgers University
Step 1: Identifying the boundary concepts
Boundary concepts
Step 2: Identifying the internal concepts
Concept 3
Concept 3
Concept 1
Concept 1
Actor C
Actor A
Actor C
Concept 6
Actor A
Concept 5
Concept 2
Actor B
Concept 4
Concept 2
Actor D
Internal
concepts
Concept 4
Actor B
Figure 2-22: A useful strategy for building a domain model is to start with the “boundary”
concepts that interact directly with the actors (Step 1), and then identify the internal
concepts (Step 2).
clear boundary between the “what” and the “how.” We should not be purists about this—the
distinction between Analysis and Design is primarily to indicate where the emphasis should be
during each stage of development.
We already encountered concepts and their relations in Section 1.3 when describing concept
maps as a diagrammatic representation of knowledge about problem domains. Domain model
described here is similar to a concept map, although somewhat more complex, as will become
apparent soon.
2.5.1
Identifying Concepts
Back to our setting-up-an-enterprise approach, we need to hire workers with appropriate expertise
and acquire things they will work with. To announce the openings under a classifieds section, we
start by listing the positions or, better, responsibilities, for which we are hiring. We identify the
responsibilities by examining the use case scenarios and system sequence diagrams. For example,
we need a worker to verify whether or not the key entered by the user is valid, so we title this
position KeyChecker. We also need a worker to know (memorize, or keep track of) the collection
of valid keys, so we advertise an opening for KeyStorage. Further, to operate the lock and the
light/switch, we come up with LockOperator and LightOperator positions, respectively. Note that
concept name is always a noun phrase.
In building the domain model, a useful strategy is to start from the “periphery” (or “boundary”) of
the system, as illustrated in Figure 2-22. That is, we start by assigning concepts that handle
interactions between the organization and the outside world, that is, between the actors and the
system. Each actor interacts with at least one boundary object. The boundary object collects the
information from the actor and translates it into a form that can be used by “internal” objects. As
well, the boundary object may translate the information in the other direction, from “internal”
objects to a format suitable for an actor.
Organizations are often fronted by a point-of-contact person. A common pattern is to have a
specialized worker to take orders from the clients and orchestrate the workings of the workers
Actor D
Chapter 2 
Object-Oriented Software Engineering
111
inside the system. This type of object is known as Controller. For a complex system, each use
case or a logical group of use cases may be assigned a different Controller object.
When identifying positions, remember that no task is too small—if it needs to be done, it must be
mentioned explicitly and somebody should be given the task responsibility. Table 2-5 lists the
responsibilities and the worker titles (concept names) to whom the responsibilities are assigned.
In this case, it happens that a single responsibility is assigned to a single worker, but this is not
necessarily the case. Complex responsibilities may be assigned to multiple workers and vice versa
a single worker may be assigned multiple simple responsibilities. Further discussion of this issue
is available in the solution of Problem 2.29 at the end of this chapter.
Table 2-5: Responsibility descriptions for the home access case study used to identify the
concepts for the domain model. Types “D” and “K” denote doing vs. knowing
responsibilities, respectively.
Responsibility Description
Typ Concept Name
Coordinate actions of all concepts associated with a use case, a logical D Controller
grouping of use cases, or the entire system and delegate the work to other
concepts.
Container for user’s authentication data, such as pass-code, timestamp, K Key
door identification, etc.
Verify whether or not the key-code entered by the user is valid.
D KeyChecker
Container for the collection of valid keys associated with doors and users. K KeyStorage
Operate the lock device to armed/disarmed positions.
D LockOperator
Operate the light switch to turn the light on/off.
D LightOperator
Operate the alarm bell to signal possible break-ins.
D AlarmOperator
Block the input to deny more attempts if too many unsuccessful attempts. D Controller
Log all interactions with the system in persistent storage.
D Logger
Based on Table 2-5 we draw a draft domain model for our case-study #1 in Figure 2-23. During
analysis, objects are used only to represent possible system state; no effort is made to describe
how they behave. It is the task of design (Section 2.6) to determine how the behavior of the
system is to be realized in terms of the behavior of objects. For this reason, objects at analysis
time have no methods/operations (as seen in Figure 2-23).
UML does not have designated symbols for domain concepts, so it is usual to adopt the symbols
that are used for software classes. I added a smiley face or a document symbol to distinguish
“worker” vs. “thing” concepts. Workers get assigned mainly doing responsibilities, while things
get assigned mainly knowing responsibilities. This labeling serves only as a “scaffolding,” to aid
the analyst in the process of identifying concepts. The distinction may not always be clear cut,
because some concepts may combine both knowing- and doing types of responsibilities. In such
cases, the concepts should be left unlabeled. This is the case for KeycodeEntry and StatusDisplay
in Figure 2-23. Like a real-world scaffolding, which is discarded once construction is completed,
this scaffolding is also temporary in nature.
Another useful kind of scaffolding is classifying concepts into the following three categories:
«boundary», «control», and «entity». This is also shown in Figure 2-23. At first, Key may be
considered a «boundary» because keys are exchanged between the actors and the system. On the
other hand, keys are also stored in KeyStorage. This particular concept corresponds to neither one
of those, because it contains other information, such as timestamp and the door identifier. Only
Ivan Marsic 
112
Rutgers University
Symbolizes
“worker”-type
concept.
Symbolizes
“thing”-type
concept.
«entity»
KeyChecker
«boundary»
KeycodeEntry
«boundary»
StatusDisplay
«entity»
KeyStorage
«entity»
Key
«control»
Controller
LockDevice
«boundary»
HouseholdDeviceOperator
Resident
LightSwitch
Figure 2-23: Partial domain model for the case study #1, home access control.
«boundary»
HouseholdDeviceOperator
«boundary»
LockOperator
«boundary»
LightOperator
«boundary»
MusicPlayerOperator
«boundary»
AlarmOperator
Figure 2-24: Generalization of the concept HouseholdDeviceOperator (Figure 2-23) as a
conceptual superclass obtained by identifying commonality among the concepts that operate
different household devices.
pass-codes identifying the actors are exchanged between the actors and the system (concept:
KeycodeEntry) and this information is transferred to the Key concept.
Figure 2-23 shows a single concept for operating household devices. This concept is obtained by
abstracting common properties of different device-operating concepts in Table 2-5. We show
such generalization diagrammatically as in Figure 2-24. Currently, the single concept appears
sufficient and, given that we prefer parsimonious designs, we leave Figure 2-23 unmodified.
Later, more detailed consideration will reveal the need for distinguishing different device
operators (see Figure 2-25(b)).
Responsibilities for use case UC-5: Inspect Access History can be derived based on the detailed
description of UC-5 (Section 2.4.3). We can gather the doing (D) and knowing (K)
responsibilities as given in Table 2-6.
Table 2-6: Responsibility descriptions for UC-5: Inspect Access History of the home access
case study.
Responsibility Description
Rs1. Coordinate actions of concepts associated with this use case and
delegate the work to other concepts.
Rs2. Form specifying the search parameters for database log retrieval
Type Concept Name
D
Controller
K
Search Request
Chapter 2 
113
Object-Oriented Software Engineering
(from UC-5, Step 2).
Rs3. Render the retrieved records into an HTML document for sending
to actor’s Web browser for display.
Rs4. HTML document that shows the actor the current context, what
actions can be done, and outcomes of the previous actions.
Rs5. Prepare a database query that best matches the actor’s search
criteria and retrieve the records from the database (from UC-5, Step 4).
Rs6. Filter the retrieved records to match the actor’s search criteria (from
UC-5, Step 6).
Rs7. List of “interesting” records for further investigation, complaint
description, and the tracking number.
Rs8. Archive the request in the database and assign it a tracking number
(from UC-5, Step 8).
Rs9. Notify Landlord about the request (from UC-5, Step 8).
D
Page Maker
K
Interface Page
D
Database
Connection
D
Postprocessor
K
Investigation
Request
D
Archiver
D
Notifier
Note that upon careful examination we may conclude that responsibility Rs6 is relatively simple
and it should be assigned to the Page Maker (Postprocessor concept would be rejected). Similarly,
responsibilities Rs8 and Rs9 may be deemed relatively simple and assigned to a single concept
Archiver (Notifier concept would be rejected).
Let us assume that we reject Postprocessor and keep Notifier because it may need to send followup notifications.
The partial domain model corresponding to the subsystem that implements UC-5 is shown later in
Figure 2-26, completed with attributes and associations.
I
t is worth noting at this point how an artifact from one phase directly feeds into the subsequent
phase. We have use case scenarios feed into the system sequence diagrams, which in turn feed
into the domain model. This traceability property is critical for a good development method
(process), because the design elaboration progresses systematically, without great leaps that are
difficult to grasp and/or follow.
Domain model is similar to a concept map (described in Section 1.3)—it also represents concepts
and their relations, here called associations—but domain model is a bit more complex. It can
indicate the concept’s stereotype as well as its attributes (described in the next section).
Note that we construct a single domain model for the whole system. The domain model is
obtained by examining different use case scenarios, but they all end up contributing concepts to
the single domain model.
2.5.2
Concept Associations and Attributes
Associations
Associations (describe who needs to work together and why, not how they work together).
Associations for use case UC-5: Inspect Access History can be derived based on the detailed
description of UC-5 (Section 2.4.3).
Table 2-7: Identifying associations for use case UC-5: Inspect Access History.
Concept pair
Association description
Association name
Ivan Marsic 
Controller  Page
Maker
Page Maker 
Database Connection
Page Maker 
Interface Page
Controller 
Database Connection
Controller 
Archiver
114
Rutgers University
Controller passes requests to Page Maker and
receives back pages prepared for displaying
Database Connection passes the retrieved data to
Page Maker to render them for display
Page Maker prepares the Interface Page
Controller passes search requests to Database
Connection
Controller passes a list of “interesting” records and
complaint description to Archiver, which assigns the
tracking number and creates Investigation Request
Archiver 
Archiver generates Investigation Request
Investigation Request
Archiver  Database Archiver requests Database Connection to store
investigation requests into the database
Connection
Archiver requests Notifier to notify Landlord about
Archiver  Notifier
investigation requests
conveys requests
provides data
prepares
conveys requests
conveys requests
generates
requests save
requests notify
Figure 2-25(a) (completed from Figure 2-23) and Figure 2-26 also show the associations between
the concepts, represented as lines connecting the concepts. Each line also has the name of the
association and sometimes an optional “reading direction arrow” is shown as ►. The labels on
the association links do not signify the function calls; you could think of these as just indicating
that there is some collaboration anticipated between the linked concepts. It is as if to know
whether person X and person Y collaborate, so they can be seated in adjacent cubicles/offices.
Similarly, if objects are associated, they logically belong to the same “package.”
The reader should keep in mind that it is more important to identify the domain concepts than
their associations (and attributes, described next). Every concept that the designer can discover
should be mentioned. Conversely, for an association (or attribute), in order to be shown it should
pass the “does it need to be mentioned?” test. If the association in question is obvious, it should
be omitted from the domain model. For example, in Figure 2-25(a), the association Controller–
obtains–Key is fairly redundant. Several other associations could as well be omitted, because the
reader can easily infer them, and this should be done particularly in schematics that are about to
become cluttered. Remember, clarity should be preferred to accurateness, and, if the designer is
not sure, some of these can be mentioned in the text accompanying the schematic, rather than
drawn in the schematic itself.
Chapter 2 
115
Object-Oriented Software Engineering
retrieves valid keys
«entity»
KeyChecker
(a)
«entity»
KeyStorage
conveys requests
verifies
«boundary»
KeycodeEntry
«entity»
Key
obtains
userIdentityCode
timestamp
doorLocation
Association
name
«boundary»
StatusDisplay
LockDevice
«control»
Controller
«boundary»
HouseholdDeviceOperator
conveys requests
numOfAttempts
maxNumOfAttempts
deviceStatuses
“Reading direction arrow.”
Has no meaning; it only helps reading
the association label, and is often left out.
Resident
(b)
Attributes
LightSwitch
«boundary»
HouseholdDeviceOperator
deviceStatus
«boundary»
LockOperator
autoLockInterval
holdOpenInterval
acceptingInterval
«boundary»
LightOperator
«boundary»
MusicPlayerOperator
asks-illumination
«boundary»
IlluminationDetector
provides-playschedule
«entity»
PlayList
«boundary»
AlarmOperator
notifies-members
«entity»
NotificationList
Figure 2-25: (a) Domain model from Figure 2-23 completed with attributes and
associations. (b) Concepts derived from HouseholdDeviceOperator in Figure 2-24
Attributes
The domain model may also include concept attributes, as is for example shown in Figure 2-25.
Example attributes are deviceStatuses (with valid values “activated” and “stopped”)
that record the current state of the physical devices operated by HouseholdDeviceOperator.
Careful consideration reveals that a single household-device-operator concept is not sufficient.
Although all physical devices share a common attribute (deviceStatus), they also have
specific needs (Figure 2-25(b)). The lock device needs to be armed after an auto-lock interval, so
the corresponding concept needs an extra attribute autoLockInterval. We also discussed allowing
the user to explicitly set the interval to hold open the lock open, which requires another attribute.
The LightOperator needs to check the room illumination before activating the light bulb, so it is
associated with the illumination detector concept. The MusicPlayerOperator needs the playlist of
tracks, so it is associated with the Playlist concept. Even the deviceStatus attribute may have
Ivan Marsic 
116
Rutgers University
«control»
Controller
«entity»
Archiver
conveys-requests
«boundary»
SearchRequest
generates
currentTrackNum
recordsList
complaintDescr
trackingNum
receives
userID
searchParams
«entity»
InvestigationRequest
posts
prepares
«boundary»
InterfacePage
Resident
conveys requests
uses
«entity»
PageMaker
«boundary»
Notifier
saves-data-to
contactInfo
Landlord
conveys-requests
provides-data
«boundary»
DatabaseConnection
Database
Figure 2-26: The domain model for UC-5: Inspect Access History of home access control.
different values for different devices, such as “disarmed” and “armed” for LockOperator;
“unlit” and “lit” for LightOperator; etc., which are more descriptive than the generic ones
“activated” and “stopped”. Although device operators differ from one another, they also
share common properties, so it is useful to indicate in the domain model diagram that they are
related through the common base concept HouseholdDeviceOperator (Figure 2-25(b)).
Table 2-8 shows how the subset of attributes related to use case UC-5: Inspect Access History is
systematically derived based on the detailed description of UC-5 (Section 2.4.3).
Table 2-8: Attributes for use case UC-5: Inspect Access History.
Concept
Attributes
user’s
identity
Search
Request
search
parameters
search
Postprocessor
parameters
records list
complaint
Investigation
description
Request
tracking
number
current
Archiver
tracking
number
contact
Notifier
information
Attribute Description
Used to determine the actor’s credentials, which in turn specify
what kind of data this actor is authorized to view.
Time frame, actor role, door location, event type (unlock, lock,
power failure, etc.).
Copied from search request; needed to Filter the retrieved records
to match the actor’s search criteria.
List of “interesting” records selected for further investigation.
Describes the actor’s suspicions about the selected access records.
Allows tracking of the investigation status.
Needed to assign a tracking number to complaints and requests.
Contact information of the Landlord who accepts complaints and
requests for further investigation.
One more attribute that could be considered is for the Page Maker to store the data received from
Database Connection. Recall that earlier we merged the Postprocessor concept with Page Maker,
which now also has the responsibility to filter the retrieved records to match the actor’s search
criteria.
Chapter 2 
Object-Oriented Software Engineering
117
In Figure 2-25(a), another possible candidate attribute is numOfKeys in the KeyStorage.
However, this is a kind of trivial attribute not worth mentioning, because it is self-evident that the
storage should know how many keys it holds inside.
An attribute numOfAttempts counts the number of failed attempts for the user before sounding
the alarm bell, to tolerate inadvertent errors when entering the key code. In addition, there should
be defined the maxNumOfAttempts constant. At issue here is which concept should possess
these attributes? Because a correct key is needed to identify the user, the system cannot track a
user over time if it does not know user’s identity (a chicken and egg problem!). One option is to
introduce real-world constraints, such as temporal continuity, which can be stated as follows. It is
unlikely that a different user would attempt to open the doors within a very short period of time.
Thus, all attempts within, say, two minutes can be ascribed to the same user10. For this we need to
introduce an additional attribute maxAttemptPeriod or, alternatively, we can specify the
maximum interval between two consecutive attempts, maxInterAttemptInterval.
The knowledge or expertise required for the attempts-counting worker comprises the knowledge
of elapsed time and the validity of the last key typed in within a time window. The
responsibilities of this worker are:
1. If numOfAttempts  maxNumOfAttempts, sound the alarm bell and reset
numOfAttempts = 0
2. Reset numOfAttempts = 0 after a specified amount of time (if the user discontinues
the attempts before reaching maxNumOfAttempts)
3. Reset numOfAttempts = 0 after a valid key is presented
A likely candidate concept to contain these attributes is the KeyChecker, because it is the first to
know the validity of a presented key. On the other hand, if we introduce the AlarmOperator
concept (Figure 2-24), then one may argue that AlarmOperator should contain all the knowledge
about the conditions for activating the alarm bell. However, we should remember that, once the
threshold of allowed attempts is exceeded, the system should activate the alarm, but also deny
further attempts (recall the detailed description of UC-7: AuthenticateUser in Section 2.4.3). In
Table 2-5, the responsibility for blocking the input to deny more attempts was assigned to the
Controller. Therefore, we decide that the best concept to place the attributes related to counting
unsuccessful attempts is the Controller, as shown in Figure 2-25(a).
10
Of course, this assumption is only an approximation. We already considered a misuse case in Section
2.4.4. We could imagine, for instance, the following scenario. An intruder makes numOfAttempts =
maxNumOfAttempts - 1 failed attempts at opening the lock. At this time, a tenant arrives and the
intruder sneaks away unnoticed. If the tenant makes a mistake on the first attempt, the alarm will be
activated, and the tenant might assume that the system is malfunctioning. Whether the developer should
try to address this scenario depends on the expected damages, as well as the time and budget constraints.
Ivan Marsic 
118
Rutgers University
Professional user
Speaks Esperanto
Conformity-driven
Expected commands:
Ekzemplero, Alglui,
Redakti, Viŝi
Potential user
Speaks Klingon
Curiosity-driven
Expected commands:
nIH, Suq, naw‘, Degh
CURRENT USER
C
Accidental user
Does not speak
Chance-driven
Expected commands:
<meaningless - ignore>
Figure 2-27: In domain analysis, we look at the external world from inside out and specify
only what the system-to-be needs to know to function as required.
As already said, during the analysis we should make every effort to stay with what needs to be
done and avoid considering how things are done. Unfortunately, as seen, this is not always
possible. The requirement of tolerating inadvertent typing errors (“unsuccessful attempts”) has
led us into considerable design detail and, worse, we are now committed to a specific technical
solution of the problem.
2.5.3
Domain Analysis
“Scientists should always state the opinions upon which their facts are based.”—Author unknown
Figure 2-27
When developing a software-based system, we are modeling the user and environment in the
software system. The model incorporates internal structures representing the problem domain.
These structures include data representing entities and relations in the domain, and a program
prescribing how these data may be manipulated. Modeling necessarily involves simplification and
abstraction. The purpose of simplification is to make the development manageable: The system
should solve one problem, not all problems.
As discussed in Section 1.2.3, the analyst needs to consider not only what needs to be done, but
also how it can be done—what are feasible ways of doing it. We cannot limit the domain analysis
to the inside of the system-to-be, because all useful systems operate by interacting with their
environment. We need to know what is at our disposal in the external world …
Chapter 2 
119
Object-Oriented Software Engineering
X
X
UC2
3
X
X
UC3
2
X
X
X
X
UC4
2
UC5
3
X
X
X
X
X
X
X
X
UC6
9
X
X
X
X
UC7
2
UC8
3
X
X
X
X
X
X
InvestigationRequest
X
Notifier
DatabaseConnection
PageMaker
X
Archiver
InterfacePage
SearchRequest
Key
Controller-SS2
KeycodeEntry
15
KeyChecker
StatusDisplay
UC1
KeyStorage
Use
PW
Case
Controller-SS1
HouseholdDeviceOperator
Domain Concepts
X
X
X
X
X
X
X
X
X
Figure 2-28: Use-cases-to-domain-model traceability matrix for the safe home access case
study. (PW = priority weight) (Continued from Figure 2-18 and continues in Figure 2-36.)
T
he traceability matrix of the domain model is shown in Figure 2-28. This matrix traces the
domain concepts to the use cases from which they were derived. This mapping continues the
development lifecycle traceability from Figure 2-18. Note that we have two Controller concepts,
Controller-SS1 for the first subsystem that controls the devices (Figure 2-25(a)) and ControllerSS2 for the second subsystem that supports desktop interaction with the system (Figure 2-26).
A more detailed traceability may be maintained for critical projects including risk analysis
traceability (Section 2.4.4) that traces potential hazards to their specific cause; identified
mitigations to the potential hazards; and specific causes of software-related hazards to their
location in the software.
2.5.4
Contracts: Preconditions and Postconditions
Contracts express any important conditions about the attributes in the domain model. In addition
to attributes, contract may include facts about forming or breaking relations between concepts,
and the time-points at which instances of concepts are created or destroyed. You can think of a
software contract as equivalent to a rental contract, which spells out the condition of an item prior
to renting, and will spell out its condition subsequent to renting. For example, for the operations
Unlock and Lock, the possible contracts are:
Operation
Unlock
Ivan Marsic 
Rutgers University
Preconditions
120
• set of valid keys known to the system is not empty
• numOfAttempts  maxNumOfAttempts
• numOfAttempts = 0, for the first attempt of the current user
Postconditions
• numOfAttempts = 0, if the entered Key  Valid keys
• current instance of the Key object is archived and destroyed
The system should be fair, so each user should be allowed the same number of retries
(maxNumOfAttempts). Thus, the precondition about the system for a new user is that
numOfAttempts starts at zero value. (I already discussed the issue of detecting a new user and
left it open, but let us ignore it for now.) The postcondition for the system is that, after the current
user ends the interaction with the system, numOfAttempts is reset to zero.
Operation
Lock
Preconditions
None (that is, none worth mentioning)
Postconditions
• lockStatus = “armed”, and
The operation postconditions specify the guarantees of what the system will do, given that the
actor fulfilled the preconditions for this operation. The postconditions must specify the outcomes
for worst-case vs. average-case vs. best-case scenarios, if such are possible.
2.6 Design: Assigning Responsibilities
“A designer knows he has achieved perfection not when there is nothing left to add, but when there is
nothing left to take away.” —Antoine De Saint-Exupéry
“… with proper design, the features come cheaply. This approach is arduous, but continues to succeed.”
—Dennis Ritchie
Analysis (Section 2.5) dealt with what is needed for our system; we determined how the customer
interacts with the system to obtain services and what workers (concepts) need to be acquired to
make this possible. Analysis is in a way the acquisition phase of the enterprise establishment.
Design (this section), on the other hand, deals with organization, how the elements of the system
work and interact. Therefore, design is mainly focused on the dynamics of the system. Unlike
analysis, where we deal with abstract concepts, here we deal with concrete software objects.
OB
JEC
T-O
RIE
NT
ED
In the postconditions for Lock, we explicitly state that lightStatus remains unchanged
because this issue may need further design attention before fully solved. For example, we may
want to somehow detect the last person leaving the home and turn off the light behind them.
DE
SIG
N
• lightStatus remains unchanged (see text for discussion)
Chapter 2 
121
Object-Oriented Software Engineering
Design
Sequence Diagram
System Sequence Diagram
Controller
: KeyStorage
: LockCtrl
checkKey()
ystem
: System
User
«initiating actor»
select function(“unlock")
: Checker
sk := getNext()
Timer
«offstage actor»
prompt for the key
alt
enter key
val != null
setOpen(true)
verify key
signal: valid key, lock open
open the lock,
turn on the light
[else]
val == null : setLit(true)
start ("duration“)
Figure 2-29: Designing object interactions: from system sequence diagrams to interaction
diagrams. The magnifier glass symbolizes looking at interactions inside the system.
We already encountered system sequence diagrams in Section 2.4.3. As Figure 2-29 illustrates, in
the design phase we are zooming-in inside the system and specifying how its software objects
interact to produce the behaviors observed by the actors. One way to think about the design
problem is illustrated in Figure 2-30. Imagine that you draw a map showing the actors and objects
as “stations” to be visited in the course of executing a use case scenario. The goal of design is to
“connect the dots/stations” in a way that is in some sense “optimal.” Initially, we know that the
path starts with the initiating actor, because the purpose of the system is to assist the initiating
actor in achieving a goal. The path also ends with the initiating actor after the system returns the
computation results. In between, the path should visit the participating actors. So, we know the
entry and exit point(s), and we know the computing responsibilities of concepts (Section 2.5).
Objects need to be called (by sending messages) to fulfill their computing (or, “doing”)
responsibility, and we need to decide how to “connect the dots.” That is, we need is to assign the
messaging responsibilities—who calls each “worker” object, and for “thing” objects we need to
decide who creates them, who uses them, and who updates them. Software designer’s key activity
is assigning responsibilities to the software objects acquired in domain analysis (Section 2.5).
Initially, we start designing the object interactions using the concepts from the domain model. As
we progress and elaborate our design and get a better understanding of what can be implemented
and how (having in mind the capabilities and peculiarities of our chosen programming language),
we will need to substitute some concepts with one or more actual classes. It is important to trace
the evolution from the abstract domain model to specific classes (see Section 2.6.2).
Ivan Marsic 
Resident
122
Rutgers University
:InterfacePage
:SearchRequest
:Controller
:PageMaker
:DatabaseConn
:Archiver
:Notifier
:InvestigRequest
Database
Landlord
Figure 2-30: The design problem seen as “connecting the dots” on the “map” of
participating objects.
Consider, for example, use case UC-5: Inspect Access History for which the doing (D) and
knowing (K) responsibilities are given in Table 2-6 (Section 2.5.1). Suppose that we want to
design the interactions only for Steps 4 – 6 of use case UC-5. Start at the point when the system
receives the search criteria from the actor and stop at the point when an HTML page is prepared
and sent to actor’s browser for viewing. Dynamic object interactions can be represented using
UML sequence diagrams (Sidebar 2.5).
Figure 2-31(a) shows the dilemma of responsibility assignment for the example of use case UC-5.
First, we observe that among the objects in Table 2-6 Archiver, Notifier, and Investigation
Request do not participate in Steps 4–6 of UC-5. Hence, we need to consider only Database
Connection and Page Maker. (Controller participates in every interaction with the initiating
actor.) Second, because this is a Web-based solution, the design will need to be adjusted for the
Web context. For example, Interface Page will not be a class, but an HTML document (with no
class specified). The Search-Request will be sent from the browser to the server as plain text
embedded in an HTTP message.
List of the responsibilities to be assigned (illustrated in Figure 2-31(a)):
R1. Call Database Connection (to fulfill Rs5, defined in Table 2-6 as: retrieve the records
from the database that match the search criteria)
R2. Call Page Maker (to fulfill Rs3, defined in Table 2-6 as: render the retrieved records into
an HTML document)
There is also the responsibility (R3) to check if the list of records retrieved from the database is
empty (because there are no records that match the given search criteria). Based on the outcome,
a different page will be shown to the actor.
Chapter 2 
?
123
Object-Oriented Software Engineering
: DatabaseConn
?
: PageMaker
accessList := retrieve(params : string)
: Controller
: Checker
: DeviceCtrl
checkKey()
R1.
interfacePage :=
render(accessList : string)
?
activate( "lock" )
R2.
(a)
(b)
Figure 2-31: Example of assigning responsibilities. (a) Which objects should be assigned
responsibilities R1 and R2? (b) Once the Key Checker decides the key is valid, the
DeviceCtrl should be notified to unlock the lock. Whose responsibility should this be?
SIDEBAR 2.5: Interaction Diagrams
 Interaction diagrams display protocols—permitted dynamic relations among objects in the
course of a given activity. This sidebar highlights the main points and the reader should check
the details in a UML reference. You read a UML sequence diagram from the top down:

At the top, each box represents an object, which may be named or not. If an object is
named, the name is shown in the box to the left of the colon. The class to which the
object belongs is shown to the right of the colon.

Each timeline (dashed vertical line) describes the world from the vantage point of the
object depicted at the top of the timeline. As a convention, time proceeds downward,
although in a concurrent program the activities at the same level in the diagram do not
necessarily occur at the same time (see Section 5.3).

Thin elongated boxes on a timeline represent the activities of the particular object (the
boxes/bars are optional and can be omitted)

Links (solid horizontal lines with arrows) between the timelines indicate the followedby relation (not necessarily the immediately-followed-by relation). The link is
annotated with a message being sent from one object to another or to itself.

Normally, all “messages” are method calls and, as such, must return. The return action
is denoted by a dashed horizontal link at the bottom of an activity box, oriented
opposite of the message arrow. Although this link is often omitted if the method has no
return value, the call returns nonetheless. Some novices just keep drawing message
arrows in one direction and forget that these must return at some point and the caller
cannot proceed (send new messages) before the callee returns.
Another example is shown in Figure 2-31(a) for use case UC-1: Unlock. Here the dilemma is,
who should invoke the method activate("lock") on the DeviceCtrl to disarm the lock once
the key validity is established? One option is the Checker because it is the first to acquire the
information about the key validity. (Note that the KeyChecker is abbreviated to Checker to save
space in the diagram.) Another option is the Controller, because the Controller would need to
know this information anyway—to signal to the user the outcome of the key validation. An
advantage of the latter choice is that it maintains the Checker focused on its specialty (key
Ivan Marsic 
Rutgers University
124
checking) and avoids assigning other responsibilities to it. Recall that in Figure 2-25 the
Controller has an association with the HouseholdDeviceOperator named “conveysRequests.”
Domain model concept associations provide only a useful hint for assigning communication
responsibilities in the design, but more is needed for making design decisions. Before I present
solutions to problems in Figure 2-31, I first describe some criteria that guide our design decisions.
Our goal is to derive a “good” design or, ideally, an optimal design. Unfortunately, at present
software engineering discipline is unable to precisely specify the quantitative criteria for
evaluating designs. Some criteria are commonly accepted, but there is no systematic framework.
For example, good software designs are characterized with:

Short communication chains between the objects

Balanced workload across the objects

Low degree of connectivity (associations) among the objects
While optimizing these parameters we must ensure that messages are sent in the correct order and
other important constraints are satisfied. As already stated, there are no automated methods for
software design; software engineers rely on design heuristics. The design heuristics used to
achieve “optimal” designs can be roughly divided as:
1. Bottom-up (inductive) approaches that are applying design principles and design patterns
locally at the level of software objects (micro-level design). Design principles are
described in the next section and design patterns are presented in Chapter 5.
2. Top-down (deductive) approaches that are applying architectural styles globally, at the
system level, in decomposing the system into subsystems (macro-level design). Software
architectures are reviewed in Section 2.3.
Software engineer normally combines both approaches opportunistically. While doing design
optimization, it is also important to enforce the contracts (Section 2.5.4) and other constraints,
such as non-functional requirements. Object constraint specification is reviewed in Section 3.2.3.
2.6.1
Design Principles for Assigning Responsibilities
A popular approach to micro-level design is known as responsibility-driven design (RDD). We
know the types of responsibilities that objects can have (Section 1.4.2):

Type 1 responsibility (knowing): Memorizing data or references, such as data values,
data collections, or references to other objects, represented as a property

Type 2 responsibility (doing): Performing computations, such as data processing, control
of physical devices, etc., represented as a method

Type 3 responsibility (communicating): Communicating with other objects, represented
as message sending (method invocation)
Hence, we need to decide what properties and methods belong to what object and what messages
are sent by objects. We have already performed responsibility assigning in the analysis phase
(Section 2.5). There, we “hired workers” to perform certain tasks, which in effect covers the first
two types of responsibility: assigning attributes, associations, and methods for performing
Chapter 2 
125
Object-Oriented Software Engineering
computations. In the design stage of software lifecycle, we are dealing mainly with the third
responsibility type: sending messages to (invoking methods on) other objects.
Low cohesion
High cohesion
Tight coupling
Loose coupling
Important design principles at the local, objects level include:

Expert Doer Principle: that who knows should do the task

High Cohesion Principle: do not take on too many responsibilities of Type 2
(computation)

Low Coupling Principle: do not take on too many responsibilities of Type 3
(communication)
Expert Doer Principle helps shorten the communication chains between the objects. It essentially
states that, when assigning a responsibility for message sending, select the object which first
Ivan Marsic 
Rutgers University
126
learns the information needed to send the message. High Cohesion Principle helps in balancing
the workload across the objects and keeping them focused. Object’s functional cohesion is
inversely proportional to the number of computing responsibilities assigned to it. Low Coupling
Principle helps to reduce the number of associations among the objects. Object’s coupling is
directly proportional to the number of different messages the object sends to other objects.
Consider how to employ these design principles to the example of Figure 2-31(b). For example,
because the Checker is the first to acquire the information about the key validity, by the Expert
Doer Principle it is considered a good candidate to send a message to the DeviceCtrl to disarm the
lock. However, the High Cohesion Principle favors maintaining the Checker functionally focused
on its specialty (key checking) and opposes assigning other responsibilities to it. Ideally, High
Cohesion allows a single non-trivial responsibility per object. Suppose we let High Cohesion
override Expert Doer. A reasonable compromise is to assign the responsibility of notifying the
DeviceCtrl to the Controller. Note that this solution violates the Low Coupling Principle, because
Controller acquires relatively large number of associations. We will revisit this issue later.
As seen, design principles are not always in agreement with each other. Enforcing any particular
design principle to the extreme would lead to absurd designs. Often, the designer is faced with
conflicting demands and must use judgment and experience to select a compromise solution that
he feels is “optimal” in the current context. Another problem is that cohesion and coupling are
defined only qualitatively: “do not take on too many responsibilities.” Chapter 4 describes
attempts to quantify the cohesion and coupling.
Because precise rules are lacking and so much depends on the developer’s judgment, it is critical
to record all the decisions and reasoning behind them. It is essential to document the alternative
solutions that were considered in the design process, identify all the tradeoffs encountered, and
explain why the alternatives were abandoned. The process may be summarized as follows:
1. Identify the responsibilities; domain modeling (Section 2.5) provides a starting point;
some will be missed at first and identified in subsequent iterations
2. For each responsibility, identify the alternative assignments; if the choice appears to be
unique then move to the next responsibility
3. Consider the merits and tradeoffs of each alternative by applying the design principles;
select what you consider the “optimal” choice
4. Document the process by which you arrived to each responsibility assignment.
Some responsibility assignments will be straightforward and only few may require extensive
deliberation. The developer will use his experience and judgment to decide.
Example of Assigning Responsibilities
Let us go back to the problem of assigning responsibilities for UC-1 and UC-5 of the safe home
access case study, presented in Figure 2-31. We first consider use case UC-5: Inspect Access
History, and design the interactions only for its Steps 4 – 6. We first identify and describe
alternative options for assigning responsibilities identified above with Figure 2-31(a).
Assigning responsibility R1 for retrieving records from the Database Connection is relatively
straightforward. The object making the call must know the query parameters; this information is
Chapter 2 
127
Object-Oriented Software Engineering
Resident
«html»
interfacePage :
specify
query
request
: Controller
: PageMaker
: DatabaseConnection
Database
get( queryRequest : string )
accessList := retrieve(params : string)
retrieve records
result
interfacePage := render(accessList : string)
alt
accessList != NULL
page :=
renderList()
[else]
result
displayed
page :=
warning()
«post page»
Figure 2-32: Sequence diagram for part of use case UC-5: Inspect Access History.
first given to the Controller, so by Expert Doer design principle, the Controller should be
assigned responsibility R1.
As for R2 (rendering the retrieved list), there are alternative options. The object making the call
must know the access list as retrieved from the Database. The feasible alternatives are:
1. Database Connection is the first to get hold of the access list records
2. Controller will probably be posting the Interface Page rendered by Page Maker, so it
would be convenient if Controller receives directly the return value from Page Maker
Finally, responsibility R3 to check if the list of records is empty could be assigned to:
1. The object that will be assigned responsibility R2, which can call different methods on
Page Maker
2. Page Maker, which will simply generate a different page for different list contents.
Next, let us employ the design principles, such as Expert Doer, High Cohesion, or Low Coupling
to decide on which object should be given which responsibility.
Consider first assigning responsibility R2. By the Expert Doer principle, the Database Connection
should make the call. However, this choice would lower the cohesion and increase coupling of
the Database Connection, which would need to pass the retrieved list to the rendering object
(Page Maker); in addition, it would need to know what to do with the rendered page returned by
the Page Maker. If we assign R2 to the Database Connection then Database Connection should
probably return the rendered page that it obtains from Page Maker, rather than the retrieved list
(as shown in Figure 2-31(a)). In other words, the method signature should be modified.
Alternatively, the Controller generally has the responsibility to delegate tasks, so the High
Cohesion design principle favors assigning R2 to the Controller. Both options (Database
Connection vs. Controller) contribute the same amount of coupling.
Therefore, we have a conflict among the design principles: Expert Doer favors assigning R2 to
the Database Connection while High Cohesion favors assigning R2 to the Controller). In this
Ivan Marsic 
Rutgers University
128
case, one may argue that maintaining high cohesion is more valuable than satisfying Expert Doer.
Database Connection already has a relatively complex responsibility and adding new
responsibilities will only make things worse. Therefore, we opt for assigning R2 to the Controller.
Responsibility R3 should be assigned to Page Maker because this choice yields highest cohesion.
Figure 2-32 shows the resulting UML sequence diagram. Note that in Figure 2-32 the system is
not ensuring that only authorized users access the database. This omission will be corrected later
in Section 5.2.4 where it is used as an example for the Protection Proxy design pattern.
N
ext consider the use case UC-1: Unlock. Table 2-9 lists the communication (message
sending / method invocation) responsibilities for the system function “enter key” (shown in
the system sequence diagram in Figure 2-20).
Table 2-9: Communicating responsibilities identified for the system function “enter key.”
Compare to Table 2-5.
Responsibility Description
Send message to Key Checker to validate the key entered by the user.
Send message to DeviceCtrl to disarm the lock device.
Send message to DeviceCtrl to switch the light bulb on.
Send message to PhotoObserver to report whether daylight is sensed.
Send message to DeviceCtrl to sound the alarm bell.
Based on the responsibilities in Table 2-9, Figure 2-33 shows an example design for the system
function “enter key.” The Controller object orchestrates all the processing logic related to this
system function. The rationale for this choice was discussed earlier, related to Figure 2-31(b). We
also have the Logger to maintain the history log of accesses.
Note that there is a data-processing rule (also known as “business rule” because it specifies the
business policy for dealing with a given situation) hidden in our design:
IF key  ValidKeys THEN disarm lock and turn lights on
ELSE
increment failed-attempts-counter
IF failed-attempts-counter equals maximum number allowed
THEN block further attempts and raise alarm
By implementing this rule, the object possesses the knowledge of conditions under which a
method can or cannot be invoked. Hence, the question is which object is responsible to know this
rule? The needs-to-know responsibility has implications for the future upgrades or modifications.
Changes to the business rules require changes in the code of the corresponding objects.
(Techniques for anticipating and dealing with change are described in Chapter 5.)
Apparently, we have built an undue complexity into the Controller while striving to preserve high
degree of specialization (i.e., cohesion) for all other objects. This implies low cohesion in the
design; poor cohesion is equivalent to low degree of specialization of (some) objects.
Chapter 2 
: Controller
enterKey()
129
Object-Oriented Software Engineering
k : Key
: Checker
: KeyStorage
: DeviceCtrl
: PhotoObsrv
: Logger
«create»
val := checkKey( k )
loop
[for all stored keys]
sk := getNext()
compare(k, sk)
logTransaction( k, val )
«destroy»
alt
val == true
activate( "lock" )
dl := isDaylight()
opt
[else]
alt
dl == false
activate( "bulb" )
numOfAttempts++
numOfAttempts == maxNumOfAttempts
denyMoreAttempts()
activate( "alarm" )
[else]
prompt: "try again"
Figure 2-33: Sequence diagram for the system function “enter key” (Figure 2-20). Several
UML interaction frames are shown, such as “loop,” “alt” (alternative fragments, of which
only the one with a condition true will execute), and “opt” (optional, the fragment executes
if the condition is true).
The reader should not think that the design in Figure 2-33 is the only one possible. Example
variations are shown in Figure 2-34. In variation (a) the Checker sets the key validity as a flag in
the Key object, rather than reporting it as the method call return value. The Key is now passed on
and DeviceCtrl obtains the key validity flag and decides what to do. The result: business logic is
moved from the Controller into the object that operates the devices. Such solution where the
correct functioning of the system depends on a flag in the Key object is fragile—data can become
corrupted as it is moves around the system. A more elegant solution is presented in Chapter 5,
where we will see how Publish/Subscribe design pattern protects critical decisions by
implementing them as operations, rather than arguments of operations. It is harder to make a
mistake of calling a wrong operation, than to pass a wrong argument value.
Ivan Marsic 
130
Rutgers University
: Controller
enterKey()
k : Key
: Checker
: KeyStorage
: DeviceCtrl
: PhotoObsrv
c
: Logger
k := create()
val := checkKey(k)
loop
: DeviceCtrl
: PhotoSObs
[for all stored keys]
sk := getNext()
compare()
activate( "light" )
dl := isDaylight()
logTransaction(k, val)
«destroy»
alt
val == true
activate(“lock”)
opt
dl := isDaylight()
a
: Controller
opt
k : Key
[else]
prompt:
"try again"
dl == false
: Checker
numOfTrials++
dl == false
setLit(true)
activate(“bulb”)
: KeyStorage
: DeviceCtrl
k := create()
opt
checkKey(k)
numOfTrials == maxNumOfTrials
activate(“alarm”)
b
loop
sk := getNext()
: DeviceCtrl
: PhotoSObs
setValid(ok)
checkIfDaylightAndIfNotThenSetLit()
dl := isDaylight()
controlLock(k)
ok := isValid()
opt
ok == true
setOpen(true)
The caller
could be
Controller or
Checker
opt
dl == false
setLit(true)
Figure 2-34: Variations on the design for the use case “Unlock,” shown in Figure 2-33.
Although the variation in Figure 2-34(b) is exaggerated, I have seen similar designs. It not only
assigns an awkward method name, checkIfDaylightAndIfNotThenSetLit(), but
worse, it imparts the knowledge encoded in the name onto the caller. Anyone examining this
diagram can infer that the caller rigidly controls the callee’s work. The caller is tightly coupled to
the callee because it knows the business logic of the callee. A better solution is in Figure 2-34(c).
Note that the graphical user interface (GUI) design is missing, but that is acceptable because the
GUI can be designed independently of the system’s business logic.
2.6.2
Class Diagram
Class diagram is created simply by reading the class names and their operations off of the
interaction diagrams. The class diagram of our case-study system is shown in Figure 2-35. Note
the similarities and differences with the domain model (Figure 2-25). Unlike domain models, the
class diagram notation is standardized by UML.
Because class diagram gathers class operations and attributes in one place, it is easier to size up
the relative complexity of classes in the system. The number of operations in a class correlates
with the amount of responsibility handled by the class. Good object-oriented designs distribute
expertise and workload among many cooperating objects. If you observe that some classes have
considerably greater number of operations than others, you should examine the possibility that
there may be undiscovered class(es) or misplaced responsibilities. Look carefully at operation
names and ask yourself questions such as: Is this something I would expect this class to do? Or, Is
there a less obvious class that has not been defined?
Chapter 2 
131
Object-Oriented Software Engineering
PhotoSObsrv
Key
– code_ : string
– timestamp_ : long
– doorLocation_ : string
+ isDaylight() : boolean
1
sensor
1..*
KeyStorage
1
+ getNext() : Key
validKeys
1
Controller
# numOfAttemps_ : long
# maxNumOfAttempts_ : long
+ enterKey(k : Key)
– denyMoreAttempts()
1
1
1
logger
Logger
+ logTransaction(k : Key)
devCtrl
KeyChecker
checker + checkKey(k : Key) : boolean
– compare(k : Key, sk : Key) : boolean
DeviceCtrl
# devStatuses_ : Vector
+ activate(dev : string) : boolean
+ deactivate(dev :string) : boolean
+ getStatus(dev : string) : Object
Figure 2-35: Class diagram for the home access software-to-be. Compare to Figure 2-25.
Based on review of the class diagram, we may need to go back and revise (or, refactor, see
Section 2.7.6) the domain model and interaction diagrams. For example, one may see that the
Controller has significantly more connections than other classes (Figure 2-35), which will be
addressed in Chapter 5. This approach is characteristic of iterative development methodology.
W
e also continue maintaining the traceability between the software artifacts. Figure 2-36
traces how software classes evolved from the domain model (Section 2.5). The class
diagram in Figure 2-35 is partial, so we include the classes from Figure 2-32. We see that some
concepts have not been (yet) implemented as classes. Generally, it should be possible to trace all
concepts from the domain model to the class diagram. Some concepts will be mapped directly to
individual classes (although the concept’s name may be changed in the process); others may be
split into several classes. Concepts are derived from the system requirements, and they cannot
disappear without a reason. There are two reasons for a concept to be missing in the traceability
matrix: (i) the concept was derived from a low-priority requirement and the implementation of
this functionality has been deferred for later; or (ii) the corresponding requirement was dropped.
On the other hand, all classes must be traceable back to domain concepts. In iterative and
incremental development, the domain model is not derived completely up front. Rather, the
analysis in Section 2.5 only represents a first iteration. During the design, we may realize that the
domain model is incomplete and we need additional concepts to implement the requested
functionality. In this case, we go back and modify our domain model.
Some classes may not have a directly corresponding abstract concept, because they are introduced
for reasons specific to the programming language in which the system is implemented. Both
missing concepts and emerged (non-traceable) classes must be documented, with the reason for
their disappearance or emergence explained. Tracing elements from the requirements
specification to the corresponding elements in the design specification is part of design
verification and validation.
Ivan Marsic 
132
Rutgers University
Controller-SS1
StatusDisplay
Key
KeyStorage
KeyChecker
HouseholdDeviceOperator
IlluminationDetector
Controller-SS2
SearchRequest
InterfacePage
PageMaker
Archiver
DatabaseConnection
Notifier
InvestigationRequest
DatabaseConnection
PageMaker
«html» interfacePage
SearchRequest
Controller-SS2
Logger
PhotoSObsrv
DeviceCtrl
KeyChecker
KeyStorage
Key
Domain Concepts
Controller-SS1
Software Classes
X
X
X
X
X
X
X
X
X
X
X
Figure 2-36: Domain-model-to-class-diagram traceability matrix for the safe home access
case study. (Continued from Figure 2-28.)
Class Relationships
Class diagram both describes classes and shows the relationships among them. We already
discussed object relationships in Section 1.4. In our particular case, Figure 2-35, there is an
aggregation relationship between KeyStorage and Key; all other relationships happen to be of the
“uses” type. The reader should also recall the access designations that signify the visibility of
class attributes and operations to other classes: + for public, global visibility; # for protected
visibility within the class and its descendant classes; and, − for private within-the-class-only
visibility (not even for its descendants).
Class diagram is static, unlike interaction diagrams, which are dynamic.
Generic Object Roles
As a result of having specific responsibilities, the members of object community usually develop
some stereotype roles.

Structurer

Bridge
Chapter 2 
133
Object-Oriented Software Engineering
SS11
AA
BB
BB
SS22
PP
(a)
AA
(b)
SSNN
(c)
Figure 2-37: Example object communication patterns. (a) One-to-one direct messages. (b)
One-to-many untargeted messages. (c) Via a shared data element.
Write
test
Write
code
Run
test
Verify &
validate
Figure 2-38: Test-driven implementation.
Note that objects almost never play an exclusive role; several roles are usually imparted to
different degree in each object.
Object Communication Patterns
Communication pattern is a message-sending relation imposed on a set of objects. As with any
relation, it can be one-to-one or one-to-many and it can be deterministic or random (Section
3.1.1). Some of these patterns are illustrated in Figure 2-37.
Object-oriented design, particularly design patterns, is further elaborated in Chapter 5.
2.6.3
Why Software Engineering Is Difficult (3)
Another key cause is the lack of analytical methods for software design. Software engineers are
aiming at optimal designs, but quantitative criteria for optimal software design are largely
unknown. Optimality criteria appear to be mainly based upon judgment and experience.
2.7 Test-driven Implementation
“The good news about computers is that they do what you tell them to do. The bad news is that they do
what you tell them to do.” —Ted Nelson
Ivan Marsic 
134
Rutgers University
Given a feature selected for implementation, test-driven implementation works by writing the
code for tests, writing the code that implements the feature, running the tests, and finally
verifying and validating the test results (Figure 2-38). If the results meet the expectations, we
move onto the next feature; otherwise, we need to debug the code, identify and fix the problem,
and test again.
2.7.1
Overview of Software Testing
“Testing shows the presence, not the absence of bugs.” —Edsger W. Dijkstra
Testing is often viewed as executing a program to see if it produces the correct output for a given
input. This implies testing the end-product, the software itself, which in turn means that testing
activities are postponed until late in the lifecycle. This is wrong because experience has shown
that errors introduced during the early stages of software lifecycle are the costliest and most
difficult to discover. A more general definition is that testing is the process of finding faults in
software artifacts, such as UML diagrams or code. A fault, also called “defect” or “bug,” is an
erroneous hardware or software element of a system that can cause the system to fail, i.e., to
behave in a way that is not desired or even harmful. We say that the system experienced failure
because of an inbuilt fault.
Any software artifact can be tested, including requirements specification, domain model, and
design specification. Testing activities should be started as early as possible. An extreme form of
this approach is test-driven development (TDD), one of the practices of Extreme Programming
(XP), in which development starts with writing tests. The form and rigor of testing should be
adapted to the nature of the artifact that is being tested. Testing of design sketches will be
approached differently than testing a software code.
Testing works by probing a program with different combinations of inputs to detect faults.
Therefore, testing shows only the presence of faults, not their absence. Showing the absence of
faults requires exhaustively trying all possible combinations of inputs (or following all possible
paths through the program). The number of possible combinations generally grows exponentially
with software size. However, it is not only about inadvertent bugs—a bad-intended programmer
might have introduced purposeful malicious features for personal gain or revenge, which are
activated only by a very complex input sequence. Therefore, it is impossible to test that a program
will work correctly for all imaginable input sequences. An alternative to the brute force approach
of testing is to prove the correctness of the software by reasoning (or, theorem proving).
Unfortunately, proving correctness generally cannot be automated and requires human effort. In
addition, it can be applied only in the projects where the requirements are specified in a formal
(mathematical) language. We will discuss this topic further in Chapter 3.
A key tradeoff of testing is between testing as many possible cases as possible while keeping the
economic costs limited. Our goal is to find faults as cheaply and quickly as possible. Ideally, we
would design a single “right” test case to expose each fault and run it. In practice, we have to run
many “unsuccessful” test cases that do not expose any faults. Some strategies that help keep costs
down include (i) complementing testing with other methods, such as design/code review,
reasoning, or static analysis; (ii) exploiting automation to increase coverage and frequency of
testing; and (iii) testing early in the lifecycle and often. Automatic checking of test results is
Chapter 2 
Component
code
Component
code
135
Object-Oriented Software Engineering
Unit
test
Te
Unit
test
st
ed
co
m
po
ne
nt
Integrated
modules
Integration
test
System
test
System
in use
Ensures that all
components work
together
Component
code
Unit
test
Ensure that each
component works
as specified
Function
test
Verifies that functional
requirements are satisfied
Quality
test
Verifies non-functional
requirements
Acceptance
test
Customer verifies
all requirements
Installation
test
Testing in user
environment
Figure 2-39:Logical organization of software tests.
preferred to keep the costs low, but may not always be feasible. For example, how to check the
display content of a graphical user interface?
Testing is usually guided by the hierarchical structure of the system (software architecture,
Section 2.3) as designed in the analysis and design phases (Figure 2-39). We may start by testing
individual components, which is known as unit testing. These components are incrementally
integrated into a system. Testing the composition of the system components is known as
integration testing. System testing ensures that the whole system complies with the functional
and non-functional requirements. The customer performs acceptance testing of the whole
system. (Acceptance tests and examples are described in Sections 2.2 and 2.4, when describing
requirements engineering.) As always, the logical organization does not imply that testing steps
should be ordered in time as shown in Figure 2-39. Instead, the development lifecycle evolves
incrementally and iteratively, and corresponding cycles will occur in testing as well.
Unit testing finds differences between the object design model and its corresponding
implementation. There are several benefits of focusing on individual components. One is the
common advantage of the divide-and-conquer approach—it reduces the complexity of the
problem and allows us to deal with smaller parts of the system separately. Second, unit testing
makes it easier to locate and correct faults because only few components are involved in the
process. Lastly, unit testing supports division of labor, so several team members can test different
components in parallel. Practical issues with unit testing are described in Section 2.7.3.
Regression testing seeks to expose new errors, or “regressions,” in existing functionality after
changes have been made to the system. A new test is added for every discovered fault, and tests
are run after every change to the code. Regression testing helps to populate test suite with good
test cases, because every regression test is added after it uncovered a fault in one version of the
code. Regression testing protects against reversions that reintroduce faults. Because the fault that
resulted in adding a regression test already happened, it may be an easy error to make again.
Ivan Marsic 
136
Rutgers University
Another useful distinction between testing approaches is what document or artifact is used for
designing the test cases. Black box testing refers to analyzing a running program by probing it
with various inputs. It involves choosing test data only from the specification, without looking at
the implementation. This testing approach is commonly used by customers, for example for
acceptance testing. White box testing chooses test data with knowledge of the implementation,
such as knowledge of the system architecture, used algorithms, or program code. This testing
approach assumes that the code implements all parts of the specification, although possibly with
bugs (programming errors). If the code omitted a part of the specification, then the white box test
cases derived from the code will have incomplete coverage of the specification. White box tests
should not depend on specific details of the implementation, which would prevent their
reusability as the system implementation evolves.
2.7.2
Test Coverage and Code Coverage
Because exhaustive testing often is not practically achievable, a key issue is to know when we
have done enough testing. Test coverage measures the degree to which the specification or code
of a software program has been exercised by tests. In this section we interested in a narrower
notion of code coverage, which measures the degree to which the source code of a program has
been tested. There are a number of code coverage criteria, including equivalence testing,
boundary testing, control-flow testing, and state-based testing.
To select the test inputs, one may make an arbitrary choice of what one “feels” should be
appropriate input values. A better approach is to select the inputs randomly by using a random
number generator. Yet another option is choosing the inputs systematically, by partitioning large
input space into a few representatives. Arbitrary choice usually works the worst; random choice
works well in many scenarios; systematic choice is the preferred approach.
Equivalence Testing
Equivalence testing is a black-box testing method that divides the space of all possible inputs into
equivalence groups such that the program “behaves the same” on each group. The goal is to
reduce the total number of test cases by selecting representative input values from each
equivalence group. The assumption is that the system will behave similarly for all inputs from an
equivalence group, so it suffices to test with only a single element of each group. Equivalence
testing has two steps: (i) partitioning the values of input parameters into equivalence groups and
(ii) choosing the test input values.
The trouble with this approach is that it is just as hard to find the equivalence classes of inputs as
it is to prove correctness. Therefore, we use heuristics (rules of thumb that are generally useful
but do not guarantee correctness) to select a set of test cases. We are essentially guessing based
on experience and domain knowledge, and hoping that at least one of the selected test cases
belongs to each of the true (unknown) equivalence classes.
Partitioning the values of input parameters into equivalence classes may be performed according
to the following heuristics:

For an input parameter specified over a range of values, partition the value space into one
valid and two invalid equivalence classes. For example, if the allowed input values are
integers between 0 and 100, the valid equivalence class
valid equivalence class
0
100
invalid equivalence classes
Chapter 2 
Object-Oriented Software Engineering
137
contains integers between 0 and 100, one invalid equivalence class contains all negative
integers, and the other invalid equivalence class contains all integers greater than 100.

For an input parameter specified with a single value, partition the value space into one
valid and two invalid equivalence classes. For example, if the allowed value is a real
number 1.4142, the valid equivalence class contains a single element {1.4142}, one
invalid equivalence class contains all real number smaller than 1.4142, and the other
invalid equivalence class contains all real number greater than 1.4142.

For an input parameter specified with a set of values, partition the value space into one
valid and one invalid equivalence class. For example, if the allowed value is any element
of the set {1, 2, 4, 8, 16}, the valid equivalence class contains the elements {1, 2, 4, 8,
16}, and the invalid equivalence class contains all other elements.

For an input parameter specified as a Boolean value, partition the value space into one
valid and one invalid equivalence class (one for TRUE and the other for FALSE).
Equivalence classes defined for an input parameter must satisfy the following criteria:
1. Coverage: Every possible input value belongs to an equivalence class.
2. Disjointedness: No input value belongs to more than one equivalence class.
3. Representation: If an operation is invoked with one element of an equivalence class as an
input parameter and returns a particular result, then it must return the same result if any
other element of the class is used as input.
If an operation has more than one input parameter, we must define new equivalence classes for
combinations of the input parameters (known as Cartesian product or cross product, see Section
3.2.1).
For example, consider testing the Key Checker’s operation checkKey(k : Key) :
boolean. As shown in Figure 2-35, the class Key has three string attributes: code,
timestamp, and doorLocation. The operation checkKey() as implemented in Listing
2-4 does not use timestamp, so its value is irrelevant. However, we need to test that the output
of checkKey() does not depend on the value of timestamp. The other two attributes, code
and doorLocation, are specified with a set of values for each. Suppose that the system is
installed in an apartment building with the apartments numbered as {196, 198, 200, 202, 204,
206, 208, 210}. Assume that the attribute doorLocation takes the value of the associated
apartment number. On the other hand, the tenants may have chosen their four-digit access codes
as {9415, 7717, 8290, …, 4592}. Although a code value “9415” and doorLocation value
“198” are each valid separately, their combination is invalid, because the code value for the
tenant in apartment 198 is “7717.”
Therefore, we must create a cross product of code and doorLocation values and partition
this value space into valid and invalid equivalence classes. For the pairs of test input values
chosen from the valid equivalence class, the operation checkKey() should return the Boolean
value TRUE. Conversely, for the pairs of test input values from invalid equivalence classes it
should return FALSE.
When ensuring test coverage, we should consider not only the current snapshot, but also historic
snapshots as well. For example, when testing the Key Checker’s operation checkKey(), the
Ivan Marsic 
138
Rutgers University
previously-valid keys of former tenants of a given apartment belong to an invalid equivalence
class, although in the past they belonged to the valid equivalence class. We need to include the
corresponding test cases, particularly during integration testing (Section 2.7.4).
Boundary Testing
Boundary testing is a special case of equivalence testing that focuses on the boundary values of
input parameters. After partitioning the input domain into equivalence classes, we test the
program using input values not only “inside” the classes, but also at their boundaries. Rather than
selecting any element from an equivalence class, boundary testing selects elements from the
“edges” of the equivalence class, or “outliers,” such as zero, min/max values, empty set, empty
string, and null. Another frequent “edge” fault results from the confusion between > and >=. The
assumption behind this kind of testing is that developers often overlook special cases at the
boundary of equivalence classes.
For example, if an input parameter is specified over a range of values from a to b, then test cases
should be designed with values a and b as well as just above and just below a and b.
Control Flow Testing
Statement coverage selects a test set such that every elementary statement in the program is
executed at least once by some test case in the test set.
Edge coverage selects a test set such that every edge (branch) of the control flow is traversed at
least once by some test case. We construct the control graph of a program so that statements
become the graph edges, and the nodes connected by an edge represent entry and exit to/from the
statement. A sequence of edges (without branches) should be collapsed into a single edge.
a;
a; b;
if a then b;
if a then b else c;
a
a
a
while a do b;
not a
a
a
b
not a
b
c
b
b
Condition coverage (also known as predicate coverage) selects a test set such that every
condition (Boolean statement) takes TRUE and FALSE outcomes at least once in some test case.
Path coverage determines the number of distinct paths through the program that must be
traversed (travelled over) at least once to verify the correctness. This strategy does not account for
loop iterations or recursive calls. Cyclomatic complexity metric (Section 4.2.2) provides a simple
way of determining the number of independent paths.
State-based Testing
State-based testing defines a set of abstract states that a software unit can take and tests the
unit’s behavior by comparing its actual states to the expected states. This approach has become
popular with object-oriented systems. The state of an object is defined as a constraint on the
values of object’s attributes. Because the methods use the attributes in computing the object’s
behavior, the behavior depends on the object state.
Chapter 2 
139
Object-Oriented Software Engineering
event
guard condition
invalid-key [numOfAttemps  maxNumOfAttempts] /
signal-failure
action
invalid-key /
signal-failure
Locked
state
Accepting
invalid-key
[numOfAttemps  maxNumOfAttempts] /
sound-alarm
valid-key /
signal-success
transition
valid-key /
signal-success
Blocked
Unlocked
Figure 2-40: UML state diagram for the Controller class in Figure 2-35. The notation for
UML state diagrams is introduced in Section 3.2.
The first step in using state-based testing is to derive the state diagram for the tested unit. We start
by defining the states. Next, we define the possible transitions between states and determine what
triggers a transition from one state to another. For a software class, a state transition is usually
triggered when a method is invoked. Then we choose test values for each individual state.
The second step is to initialize the unit and run the test. The test driver exercises the unit by
calling methods on it, as described in Section 2.7.3. When the driver has finished exercising the
unit, assuming no errors have yet occurred, the test then proceeds to compare the actual state of
the unit with its expected state. If the unit reached the expected state, the unit is considered
correct regardless of how it got to that state.
Assume that we are to test the Controller class of our safe home access case study (the class
diagram shown in Figure 2-35). The process of deriving the state diagrams and UML state
diagram notation are described in Chapter 3. A key responsibility of the Controller is to prevent
the dictionary attacks by keeping track of unsuccessful attempts because of an invalid key.
Normally, we assume that the door is locked (as required by REQ1 in Table 2-1). The user
unlocks the door by providing a valid key. If the user provided an invalid key, the Controller will
allow up to maxNumOfAttempts unsuccessful attempts, after which it should block and sound
alarm. Therefore, we identify the following elements of the state diagram (Figure 2-40):

Four states { Locked, Unlocked, Accepting, Blocked }

Two events { valid-key, invalid-key }

Five valid transitions { LockedUnlocked, LockedAccepting, AcceptingAccepting,
AcceptingUnlocked, AcceptingBlocked }
A test set consists of scenarios that exercise the object along a given path through the state
diagram. In general the number of state diagram elements is
all-events, all-states  all-transitions  all-paths
Ivan Marsic 
Rutgers University
140
Because the number of possible paths in the state diagram is generally infinite, it is not practical
to test each possible path. Instead, we ensure the following coverage conditions:

Cover all identified states at least once (each state is part of at least one test case)

Cover all valid transitions at least once

Trigger all invalid transitions at least once
Testing all valid transitions implies (subsumes) all-events coverage, all-states coverage, and allactions coverage. This is considered a minimum acceptable strategy for responsible testing of a
state diagram. Note that all-transitions testing is not exhaustive, because exhaustive testing
requires that every path over the state machine is exercised at least once, which is usually
impossible or at least unpractical.
2.7.3
Practical Aspects of Unit Testing
Executing tests on single components (or “units”) or a composition of components requires that
the tested thing be isolated from the rest of the system. Otherwise we will not be able to localize
the problem uncovered by the test. But system parts are usually interrelated and cannot work
without one another. To substitute for missing parts of the system, we use test drivers and test
stubs. A test driver simulates the part of the system that invokes operations on the tested
component. A test stub is a minimal implementation that simulates the components which are
called by the tested component. The thing to be tested is also known as the fixture.
A stub is a trivial implementation of an interface that exists for the purpose of performing a unit
test. For example, a stub may be hard-coded to return a fixed value, without any computation. By
using stubs, you can test the interfaces without writing any real code. The implementation is
really not necessary to verify that the interfaces are working properly (from the client’s
perspective—recall that interfaces are meant for the client object, Section 1.4). The driver and
stub are also known as mock objects, because they pretend to be the objects they are simulating.
Each testing method follows this cycle:
1. Create the thing to be tested (fixture), the test driver, and the test stub(s)
2. Have the test driver invoke an operation on the fixture
3. Evaluate that the results are as expected
More specifically, a unit test case comprises three steps performed by the test driver:
1. Setup objects: create an object to be tested and any objects it depends on, and set them up
2. Act on the tested object
3. Verify that the outcome is as expected
Suppose you want to test the Key Checker class of the safe-home-access case study that we
designed in Section 2.6. Figure 2-41(a) shows the relevant excerpt sequence diagram extracted
from Figure 2-33. Class Checker is the tested component and we need to implement a test
driver to substitute Controller and test stubs to substitute KeyStorage and Key classes.
Chapter 2 
141
Object-Oriented Software Engineering
Test driver
: Controller
k : Key
: Checker
: KeyStorage
testDriver :
start()
enterKey()
loop [for all stored keys]
sk := getNext()
compare()
: Checker
Test stubs
k : Key
: KeyStorage
k := create()
k := create()
val := checkKey(k)
Tested component
result :=
checkKey(k) loop [for all stored keys]
sk := getNext()
compare()
display
result
(a)
(b)
Figure 2-41: Testing the Key Checker’s operation checkKey() (use case Unlock).
(a)Relevant part of the sequence diagram excerpted from Figure 2-33. (b) Test stubs and
drivers for testing the Key Checker.
As shown in Figure 2-41(b), the test driver passes the test inputs to the tested component and
displays the results. In JUnit testing framework for Java, the result verification is done using the
assert*() methods that define the expected state and raise errors if the actual state differs.
The test driver can be any object type, not necessarily an instance of the Controller class.
Unlike this, the test stubs must be of the same class as the components they are simulating. They
must provide the same operation APIs, with the same return value types. The implementation of
test stubs is a nontrivial task and, therefore, there is a tradeoff between implementing accurate test
stubs and using the actual components. That is, if KeyStorage and Key class implementations
are available, we could use them when testing the Key Checker class.
Listing 2-1: Example test case for the Key Checker class.
public class CheckerTest {
// test case to check that invalid key is rejected
@Test public void
checkKey_anyState_invalidKeyRejected() {
// 1. set up
Checker checker = new Checker( /* constructor params */ );
// 2. act
Key invalidTestKey = new Key( /* setup with invalid code */ );
boolean result = checker.checkKey(invalidTestKey);
// 3. verify
assertEqual(result, false);
}
}
We use the following notation for methods that represent test cases (see Listing 2-1):
Ivan Marsic 
142
Rutgers University
1. Set up
methodName_startingState_expectedResult
2. Act
3. Verify
where methodName is the name of the method (i.e., event) we are testing on the tested object;
startingState are the conditions under which the tested method is invoked; and, expectedResult is
what we expect the tested method to produce under the specified condition. In our example, we
are testing Checker’s method checkKey(). The Checker object does not have any
attributes, so it is always in an initial state. The expected result is that checkKey() will reject
an
invalid
key.
Thus
the
test
case
method
name
checkKey_anyState_invalidKeyRejected().
Testing objects with different states is a bit more complex, because we must bring the object to
the tested state and in the end verify that the object remains in an expected state. Consider the
Controller object and its state diagram shown in Figure 2-40. One test case needs to verify
that when Controller receives maxNumOfAttempts invalid keys, it correctly transitions to
the Blocked state.
Listing 2-2: Example test case for the Controller class.
public class ControllerTest {
// test case to check that the state Blocked is visited
@Test public void
enterKey_accepting_toBlocked() {
// 1. set up: bring the object to the starting state
Controller cntrl = new Controller( /* constructor params */ );
// bring Controller to the Accepting state, just before it blocks
Key invalidTestKey = new Key( /* setup with invalid code */ );
for (i=0; i < cntrl.getMaxNumOfAttempts(); i++) {
cntrl.enterKey(invalidTestKey);
}
assertEqual(
// check that the starting state is set up
cntrl.getNumOfAttempts(), cntrl.getMaxNumOfAttempts() – 1
);
// 2. act
cntrl.enterKey(invalidTestKey);
// 3. verify
assertEqual(
// the resulting state must be "Blocked"
cntrl.getNumOfAttempts(), cntrl.getMaxNumOfAttempts()
);
assertEqual(cntrl.isBlocked(), true);
}
}
Chapter 2 
Object-Oriented Software Engineering
143
It is left to the reader to design the remaining test cases and ensure the coverage conditions
(Section 2.7.2).
A key challenge of unit testing is to sufficiently isolate the units so that each unit can be tested
individually. Otherwise, you end up with a “unit” test that is really more like an integration test.
The most important technique to help achieve this isolation is to program to interfaces instead of
concrete classes.
2.7.4
Integration and Security Testing
In traditional methods, testing takes place relatively late in the development lifecycle and follows
the logical order Figure 2-39. Unit testing is followed by integration testing, which in turn is
followed by system testing. Integration testing works in a step-by-step fashion by linking together
individual components (“units”) and testing the correctness of the combined component.
Components are combined in a horizontal fashion and integration processes in different direction,
depending on the horizontal integration testing strategy.
In agile methods, testing is incorporated throughout the development cycle. Components are
combined in a vertical fashion to implement an end-to-end functionality. Each vertical slice
corresponds to a user story (Section 2.2.3) and user stories are implemented and tested in parallel.
Horizontal Integration Testing Strategies
There are various ways to start by combining the tested units. The simplest, known as “big bang”
integration approach, tries linking all components at once and testing the combination.
Bottom-up integration starts by combining the units at the lowest level of hierarchy. The
“hierarchy” is formed by starting with the units that have no dependencies to other units. For
example, in the class diagram of Figure 2-35, classes PhotoSObsrv, Logger, and
DeviceCtrl do not have navigability arrow pointing to any other class—therefore, these three
classes form the bottommost level of the system hierarchy (Figure 2-42(a)). In bottom-up
integration testing, the bottommost units (“leaf units”) are tested first by unit testing (Figure
2-42(b)). Next, the units that have navigability to the bottommost units are tested in combination
with the leaf units. The integration proceeds up the hierarchy until the topmost level is tested.
There is no need to develop test stubs: The bottommost units do not depend on any other units;
for all other units, the units on which the currently tested unit depends on are already tested. We
do need to develop test drivers for bottom-up testing, although these can be relatively simple.
Note that in real-world systems unit hierarchy may not necessarily form a “tree” structure, but
rather may include cycles making it difficult to decide the exact level of a unit.
Top-down integration starts by testing the units at the highest level of hierarchy that no other
unit depends on (Figure 2-42(c)). In this approach, we never need to develop test drivers, but we
do need test stubs.
Ivan Marsic 
144
Rutgers University
Controller
Level-4
Level-3
KeyChecker
Level-2
KeyStorage
Level-1
Logger
PhotoSObsrv
DeviceCtrl
Test
Logger
Key
(a)
Test Controller &
KeyChecker & KeyStorage &
Key & Logger & PhotoSObsrv
& DeviceCtrl
Test
PhotoSObsrv
Test
DeviceCtrl
Test KeyChecker
& KeyStorage &
Key
(b)
Test Key &
KeyStorage
Test
Controller
Test
Controller &
KeyChecker
Test Controller &
KeyChecker &
KeyStorage & Key
Test Controller &
KeyChecker & KeyStorage &
Key & Logger & PhotoSObsrv
& DeviceCtrl
(c)
Figure 2-42: Integration testing strategies for the system from Figure 2-35. (a) Units
hierarchy; (b) Bottom-up integration testing; (c) Top-down integration testing.
Sandwich integration approach combines top-down and bottom-up by starting from both ends
and incrementally using components of the middle level in both directions. The middle level is
known as the target level. In sandwich testing, usually there is need to write stubs for testing the
top components, because the actual components from the target level can be used. Similarly, the
actual target-level components are used as drivers for bottom-up testing of low-level components.
In our example system hierarchy of Figure 2-42(a), the target layer contains only one component:
Key Checker. We start by top-down testing of the Controller using the Checker. In parallel, we
perform bottom-up testing of the Key Storage again by using the Checker. Finally, we test all
components together.
There are advantages and drawbacks of each integration strategy. Bottom-up integration is
suitable when the system has many low-level components, such as utility libraries. Moving up the
hierarchy makes it easier to find the component-interface faults: if a higher-level component
violates the assumption made by a lower-level component, it is easier to find where the problem
is. A drawback is that the topmost component (which is usually the most important, such as user
interface), is tested last—if a fault is detected, it may lead to a major redesign of the system.
Chapter 2 
User
story-1
145
Object-Oriented Software Engineering
User
story-2
User
story-N
Write a failing
acceptance test
Write a
failing
unit test
Make the
test pass
Refactor
(a)
(b)
Figure 2-43: Vertical integration in agile methods develops functional vertical slices (user
stories) in parallel (a). Each story is developed in a cycle that integrates unit tests in the
inner feedback loop and the acceptance test in the outer feedback loop (b).
Top-down integration has the advantage of starting with the topmost component (usually the user
interface, which means possibility of early end-user involvement). The test cases can be derived
directly from the requirements. Its disadvantage is that developing test stubs is time consuming
and error prone.
The advantages of sandwich testing include no need to write stubs or drivers and the ability of
early testing of the user interface and thus early involvement of end users. A drawback is that
sandwich testing does not thoroughly test the units of the target (middle) level before integration.
This problem can be remedied by the modified sandwich testing that tests the lower, middle, and
upper levels individually before combining them in incremental tests with one another.
Vertical Integration Testing Strategies
Agile methods use the vertical integration approach to develop the user stories in parallel (Figure
2-43(a)). Each story is developed in a feedback loop (Figure 2-43(b)), where the developers use
unit tests in the inner loop and the customer runs the acceptance test in the outer loop. Each cycle
starts with the customer/user writing the acceptance test that will test a particular user story.
Based on the acceptance test, the developer writes the unit tests and develops only the code that is
relevant, i.e., needed to pass the unit tests. The unit tests are run on daily basis, soon after the
code is written, and the code is committed to the code base only after it passes the unit tests. The
acceptance test is run at the end of each cycle (order of weeks or months).
The advantage of vertical integration is that it yields a working deliverable quickly. A potential
drawback is that because each subsystem (vertical slice—user story) is developed independently,
the system may lack uniformity and “grand design.” Therefore, the system may need a major
redesign late in the development cycle.
Ivan Marsic 
146
Rutgers University
Security Testing
Functional testing is testing for “positives”—that the required features and functions are correctly
implemented. However, a majority of security defects and vulnerabilities are not directly related
to security functionality, such as encryption or privilege management. Instead, security issues
involve often unexpected but intentional misuses of the system discovered by an attacker.
Therefore, we also need to test for “negatives,” such as abuse cases, to determine how the system
behaves under attack. Security tests are often driven by known attack patterns.
2.7.5
Test-driven Implementation
“Real programmers don’t comment their code. If it was hard to write, it should be hard to understand.”
—Unknown
This section shows how the designed system might be implemented. (The reader may wish to
review the Java programming refresher in Appendix A before proceeding.) One thing that
programmers often neglect is that the code must be elegant and readable. This is not for the sake
of the computer which will run the code, but for the sake of humans who will read, maintain, and
improve on the original code. I believe that writing good comments is at least as difficult as
writing good code. It may be even more important, because comments describe the developer’s
intention, while the code expresses only what the developer did. The code that lacks aesthetics
and features poor writing style in comments is likely to be a poor quality code.11 In addition to
comments, languages such as Java and C# provide special syntax for writing the documentation
for classes and methods. Javadoc is a tool for generating API documentation in HTML format
from documentation comments in source code. Sandcastle is the equivalent tool for C#.
The hardware architecture of our system-to-be is described in Section [@@@] (Figure 2-7).
The following code uses threads for concurrent program execution. The reader not familiar with
threads should consult Section 5.3.
The key purpose of the main class is to get hold of the external information: the table of valid
keys and a connection to the embedded processor that controls the devices. Following is an
implementation for the main system class.
Listing
2-3:
Implementation
Java
code
of
the
main
HomeAccessControlSystem, of the case-study home-access system.
import
import
import
import
import
11
class,
called
java.io.IOException;
java.io.InputStream;
java.util.TooManyListenersException;
javax.comm.CommPortIdentifier;
javax.comm.NoSuchPortException;
On a related note, writing user messages is as important. The reader may find that the following funny
story is applicable to software products way beyond Microsoft’s: “There was once a young man who
wanted to become a great writer and to write stuff that millions of people would read and react to on an
emotional level, cry, howl in pain and anger, so now he works for Microsoft, writing error messages.”
[ Source: A Prairie Home Companion, February 3, 2007. Online at:
http://prairiehome.publicradio.org/programs/2007/02/03/scripts/showjokes.shtml ]
Chapter 2 
import
import
import
import
Object-Oriented Software Engineering
147
javax.comm.SerialPort;
javax.comm.SerialPortEvent;
javax.comm.SerialPortEventListener;
javax.comm.UnsupportedCommOperationException;
public class HomeAccessControlSystem extends Thread
implements SerialPortEventListener {
protected Controller ctrler_; // entry point to the domain logic
protected InputStream inputStream_; // from the serial port
protected StringBuffer key_ = new StringBuffer(); // user key code
public static final long keyCodeLen_ = 4; // key code of 4 chars
public HomeAccessControlSystem(
KeyStorage ks, SerialPort ctrlPort
) {
try {
inputStream_ = ctrlPort.getInputStream();
} catch (IOException e) { e.printStackTrace(); }
LockCtrl lkc = new LockCtrl(ctrlPort);
LightCtrl lic = new LightCtrl(ctrlPort);
PhotoObsrv sns = new PhotoObsrv(ctrlPort);
AlarmCtrl ac = new AlarmCtrl(ctrlPort);
ctrler_ =
new Controller(new KeyChecker(ks), lkc, lic, sns, ac);
try {
ctrlPort.addEventListener(this);
} catch (TooManyListenersException e) {
e.printStackTrace(); // limited to one listener per port
}
start(); // start the thread
}
/** The first argument is the handle (filename, IP address, ...)
* of the database of valid keys.
* The second arg is optional and, if present, names
* the serial port. */
public static void main(String[] args) {
KeyStorage ks = new KeyStorage(args[1]);
SerialPort ctrlPort;
String portName = "COM1";
if (args.length > 1) portName = args[1];
try {
// initialize
CommPortIdentifier cpi =
CommPortIdentifier.getPortIdentifier(portName);
if (cpi.getPortType() == CommPortIdentifier.PORT_SERIAL) {
ctrlPort = (SerialPort) cpi.open();
// start the thread for reading from serial port
new HomeAccessControlSystem(ks, ctrlPort);
} catch (NoSuchPortException e) {
System.err.println("Usage: ... ... port_name");
}
Ivan Marsic 
Rutgers University
148
try {
ctrlPort.setSerialPortParams(
9600, SerialPort.DATABITS_8, SerialPort.STOPBITS_1,
SerialPort.PARITY_NONE
);
} catch (UnsupportedCommOperationException e) {
e.printStackTrace();
}
}
/** Thread method; does nothing, just waits to be interrupted
* by input from the serial port. */
public void run() {
while (true) { // alternate between sleep/awake periods
try { Thread.sleep(100); }
catch (InterruptedException e) { /* do nothing */ }
}
}
/** Serial port event handler
* Assume that the characters are sent one by one, as typed in. */
public void serialEvent(SerialPortEvent evt) {
if (evt.getEventType() == SerialPortEvent.DATA_AVAILABLE) {
byte[] readBuffer = new byte[5]; // 5 chars, just in case
try {
while (inputStream_.available() > 0) {
int numBytes = inputStream_.read(readBuffer);
// could check if "numBytes" == 1 ...
}
} catch (IOException e) { e.printStackTrace(); }
// append the new char to the user key
key_.append(new String(readBuffer));
if (key_.length() >= keyCodeLen_) { // got the whole key?
// pass on to the Controller
ctrler_.enterKey(key_.toString());
// get a fresh buffer for a new user key
key_ = new StringBuffer();
}
}
}
}
The class HomeAccessControlSystem is a thread that runs forever and accepts the input
from the serial port. This is necessary to keep the program alive, because the main thread just sets
up everything and then terminates, while the new thread continues to live. Threads are described
in Section 5.3.
Next shown is an example implementation of the core system, as designed in Figure 2-33. The
coding of the system is directly driven by the interaction diagrams.
Listing 2-4: Implementation Java code of the classes Controller, KeyChecker, and
Chapter 2 
Object-Oriented Software Engineering
149
LockCtrl.
public class Controller {
protected KeyChecker checker_;
protected LockCtrl lockCtrl_;
protected LightCtrl lightCtrl_;
protected PhotoObsrv sensor_;
protected AlarmCtrl alarmCtrl_;
public static final long maxNumOfAttempts_ = 3;
public static final long attemptPeriod_ = 600000; // msec [=10min]
protected long numOfAttempts_ = 0;
public Controller(
KeyChecker kc, LockCtrl lkc, LightCtrl lic,
PhotoObsrv sns, AlarmCtrl ac
) {
checker_ = kc;
lockCtrl_ = lkc; alarmCtrl_ = ac;
lightCtrl_ = lic; sensor_ = sns;
}
public enterKey(String key_code) {
Key user_key = new Key(key_code)
if (checker_.checkKey(user_key)) {
lockCtrl_.setArmed(false);
if (!sensor_.isDaylight()) { lightCtrl_.setLit(true); }
numOfAttempts_ = 0;
} else {
// we need to check the attempt period as well, but ...
if (++numOfAttempts_ >= maxNumOfAttempts_) {
alarmCtrl_.soundAlarm();
numOfAttempts_ = 0; // reset for the next user
}
}
}
}
import java.util.Iterator;
public class KeyChecker {
protected KeyStorage validKeys_;
public KeyChecker(KeyStorage ks) { validKeys_ = ks; }
public boolean checkKey(Key user_key) {
for (Iterator e = validKeys_.iterator(); e.hasNext(); ) {
if (compare((Key)e.next(), user_key) { return true; }
}
return false;
}
protected boolean compare(Key key1, Key key2) {
}
}
import javax.comm.SerialPort;
Ivan Marsic 
Rutgers University
150
public class LockCtrl {
protected boolean armed_ = true;
public LockCtrl(SerialPort ctrlPort) {
}
}
In Listing 2-4 I assume that KeyStorage is implemented as a list, java.util.ArrayList. If
the keys are simple objects, e.g., numbers, then another option is to use a hash table,
java.util.HashMap. Given a key, KeyStorage returns a value of a valid key. If the return
value is null, the key is invalid. The keys must be stored in a persistent storage, such as
relational database or a plain file and loaded into the KeyStorage at the system startup time,
which is not shown in Listing 2-4.
The reader who followed carefully the stepwise progression from the requirements from the code
may observe that, regardless of the programming language, the code contains many details that
usually obscure the high-level design choices and abstractions. Due to the need for being precise
about every detail and unavoidable language-specific idiosyncrasies, it is difficult to understand
and reason about software structure from code only. I hope that at this point the reader
appreciates the usefulness of traceable stepwise progression and diagrammatic representations.
Chapter 2 
2.7.6
Object-Oriented Software Engineering
151
Refactoring: Improving the Design of Existing
Code
A refactoring of existing code is a transformation that improves its design while preserving its
behavior. Refactoring changes the internal structure of software to make it easier to understand
and cheaper to modify that does not change its observable behavior. The process of refactoring
involves removing duplication, simplifying complex logic, and clarifying unclear code. Examples
of refactoring include small changes, such as changing a variable name, as well as large changes,
such as unifying two class hierarchies.
Refactoring applies sequences of low-level design transformations to the code. Each
transformation improves the code by a small increment, in a simple way, by consolidating ideas,
removing redundancies, and clarifying ambiguities. A major improvement is achieved gradually,
step by step. The emphasis is on tiny refinements, because they are easy to understand and track,
and each refinement produces a narrowly focused change in the code. Because only small and
localized block of the code is affected, it is less likely that a refinement will introduce defects.
Agile methods recommend test-driven development (TDD) and continuous refactoring. They go
together because refactoring (changing the code) requires testing to ensure that no damage was
done.
Using Polymorphism Instead of Conditional Logic
An important feature of programming languages is the conditional. This is a statement that causes
another statement to execute only if a particular condition is true. One can use simple “sentences”
to advise the computer, “Do these fifteen things one after the other; if by then you still haven’t
achieved such-and-such, start all over again at Step 5.” Equally, one can readily symbolize a
complex conditional command such as: “If at that particular point of runtime, this happens, then
do so-and-so; but if that happens, then do such-and-such; if anything else happens, whatever it is,
then do thus-and-so.” Using the language constructs such as IF-THEN-ELSE, DO-WHILE, or
SWITCH, the occasion for action is precisely specified. The problem with conditionals is that
they make code difficult to understand and prone to errors.
Polymorphism allows avoiding explicit conditionals when you have objects whose behavior
varies depending on their types. As a result you find that switch statements that switch on type
codes or if-then-else statements that switch on type strings are much less common in an objectoriented program. Polymorphism gives you many advantages. The biggest gain occurs when this
same set of conditions appears in many places in the program. If you want to add a new type, you
have to find and update all the conditionals. But with subclasses you just create a new subclass
and provide the appropriate methods. Clients of the class do not need to know about the
subclasses, which reduces the dependencies in your system and makes it easier to update.
some conditionals are needed, like checks for boundary conditions, but when you keep working
with similar variables, but apply different operations to them based on condition, that is the
perfect place for polymorphism and reducing the code complexity. Now there are usually two
types of conditionals you can’t replace with Polymorphism. Those are comparatives (>, <) (or
working with primitives, usually), and boundary cases, sometimes. And those two are language
Ivan Marsic 
152
Rutgers University
Behavior
: Communicator
Lock
Tenant
: Checker
: KeyStorage
: LockCtrl
import javax.com
checkKey()
: System
User
import java.io.I
sk := getNext()
«primary actor»
selectFunction(“unlock")
alt
val != null
import java.io.I
setOpen(true)
import java.util
prompt for the key
Unlock
[else]
val == null : setLit(true)
open the lock,
turn on the light
enterKey()
Landlord
addElement()
public class Hom
signal: valid key, lock open
implemen
System Description
Use Cases
protected Co
Interaction
Diagrams
System Sequence
Diagrams
protected In
protected St
public stati
public HomeA
KeyStora
) {
KeyChecker
Key
# code_ : long
+ checkKey() : boolean
try {
1..*
+ getCode() : long
obtains
Key
1
inpu
checker
} catch
conveysRequests
Structure
Communicator
Controller
es
rifi
ve
# numOfTrials_ : long
# maxNumOfTrials_ : long
1
PhotoSObsrv
sensor + isDaylight() : boolean
LockCtrl
+ enterKey(k : Key)
LightCtr
KeyChecker
notifiesKeyValidity
LockOperator
numOfTrials
maxNumOfTrials
lockStatus
lockCtrl
1
1
LockCtrl
# open_ : boolean
+ isOpen() : boolean
+ setOpen(v : boolean)
PhotoObs
alarmCtrl
AlarmCtrl
+ soundAlarm()
Domain Model
Class Diagram
Implementation
Program
Figure 2-44: Summary of a single iteration of the software development lifecycle. The
activity alternates between elaborating the system’s behavior vs. structure. Only selected
steps and artifacts are shown.
specific as well, as in Java only. Some other languages allow you to pass closures around, which
obfuscate the need for conditionals.
2.8 Summary and Bibliographical Notes
“Good judgment comes from experience, and experience comes from bad judgment.”
—Frederick P. Brooks
This chapter presents incremental and iterative approach to software design and gradually
introduces software engineering techniques using a running case study. Key phases of the process
are summarized in Figure 2-44. (Note that package diagram, which is a structural description, is
not shown for the lack of space.) To ensure meaningful correspondence between the successive
software artifacts, we maintain traceability matrices across the development lifecycle. The
traceability matrix links requirements, design specifications, hazards, and validation. Traceability
among these activities and documents is essential.
Figure 2-44 shows only the logical order in which activities take place and does not imply that
software lifecycle should progress in one direction as in the waterfall method. In practice there is
significant intertwining and backtracking between the steps and Figure 2-44 shows only one
iteration of the process. The sequential presentation of the material does not imply how the actual
Chapter 2 
Object-Oriented Software Engineering
153
development is carried out. Teaching works from a known material and follows logical ordering,
but practice needs to face unknown problem and the best ordering is known only after the fact.
A general understanding of the problem domain does not guarantee project success; you need a
very detailed understanding of what is expected from the system. A detailed understanding is best
developed incrementally and iteratively.
Key points:

Object orientation allows creation of software in solution objects which are directly
correlated to the objects (physical objects or abstract concepts) in the problem to be
solved. The key advantage of the object-oriented approach is in the localization of
responsibilities—if the system does not work as intended, it is easier to pinpoint the
culprit in an object-oriented system.

The development must progress systematically, so that the artifacts created in the
previous phase are always being carried over into the next phase, where they serve as the
foundation to build upon.

The traceability matrix acts as a map, providing the links necessary for determining
where information is located. It demonstrates the relationship between design inputs and
design outputs, ensures that design is based on predecessor, established requirements, and
helps ensure that design specifications are appropriately verified and the requirements are
appropriately validated. The traceability matrix supports bidirectional traceability,
“forwards” from the requirements to the code and “backwards” in the opposite direction.

Use case modeling is an accepted and widespread technique to gather and represent the
business processes and requirements. Use cases describe the scenarios of how the system
under discussion can be used to help the users accomplish their goals. Use cases represent
precisely the way the software system interacts with its environment and what
information must pass the system boundary in the course of interaction. Use case steps
are written in an easy-to-understand structured narrative using the vocabulary of the
domain. This is engaging for the end users, who can easily follow and validate the use
cases, and the accessibility encourages users to be actively involved in defining the
requirements.

The analysis models are input to the design process, which produces another set of
models describing how the system is structured and how the system’s behavior is realized
in terms of that structure. The structure is represented as a set of classes (class diagram),
and the desired behavior is characterized by patterns of messages flowing between
instances of these classes (interaction diagrams).

Finally, the classes and methods identified during design are implemented in an objectoriented programming language. This completes a single iteration. After experimenting
with the preliminary implementation, the developer iterates back and reexamines the
requirements. The process is repeated until a satisfactory solution is developed.
The reader should be aware of the capabilities and limitations of software engineering methods.
The techniques presented in this chapter help you to find a solution once you have the problem
properly framed and defined, as is the case with example projects in Section 1.5. Requirements
analysis can help in many cases with framing the problem, but you should also consider
Ivan Marsic 
Rutgers University
154
ethnography methods, participatory design, and other investigative techniques beyond software
engineering.
A short and informative introduction to UML is provided by [Fowler, 2004]. The fact that I adopt
UML is not an endorsement, but merely recognition that many designers presently use it and
probably it is the best methodology currently available. The reader should not feel obliged to
follow it rigidly, particularly if he/she feels that the concept can be better illustrated or message
conveyed by other methods.
Section 2.1: Software Development Methods
[MacCormack, 2001; Larman & Basili, 2003; Ogawa & Piller, 2006]
Section 2.2: Requirements Engineering
IEEE Standard 830 was last revised in 1998 [IEEE 1998]. The IEEE recommendations cover
such topics as how to organize requirements specifications document, the role of prototyping, and
the characteristics of good requirements.
The cost-value approach for requirement prioritization was created by Karlsson and Ryan [1997].
A great introduction to user stories is [Cohn, 2004]. It describes how user stories can be used to
plan, manage, and test software development projects. It is also a very readable introduction to
agile methodology.
More powerful requirements engineering techniques, such as Jackson’s “problem frames”
[Jackson, 2001], are described in the next chapter.
Section 2.4: Use Case Modeling
An excellent source on methodology for writing use cases is [Cockburn, 2001].
System sequence diagrams were introduced by [Coleman et al., 1994; Malan et al., 1996] as part
of their Fusion Method.
Section 2.5: Analysis: Building the Domain Model
The approach to domain model construction presented in Section 2.5 is different from, e.g., the
approach in [Larman, 2005]. Larman’s approach can be summarized as making an inventory of
the problem domain concepts. Things, terminology, and abstract concepts already in use in the
problem domain are catalogued and incorporated in the domain model diagram. A more inclusive
and complex model of the business is called Business Object Model (BOM) and it is also part of
the Unified Process.
An entrepreneurial reader may wish to apply some of the analysis patterns described by Fowler
[1997] during the analysis stage. However, the main focus at this stage should be to come up with
any idea of how to solve the problem, rather than finding an optimal solution. Optimizing should
be the focus of subsequent iterations, after a working version of the system is implemented.
Chapter 2 
Object-Oriented Software Engineering
155
Section 2.6: Design: Assigning Responsibilities
Design with responsibilities (Responsibility-Driven Design):
[Wirfs-Brock & McKean, 2003; Larman, 2005]
Coupling and cohesion as characteristics of software design quality introduced in [Constantine et
al., 1974; Yourdon & Constantine, 1979]. More on coupling and cohesion in Chapter 4.
See also: http://c2.com/cgi/wiki?CouplingAndCohesion
J. F. Maranzano, S. A. Rozsypal, G. H. Zimmerman, G. W. Warnken, P. E. Wirth, and D. M.
Weiss, “Architecture reviews: Practice and experience,” IEEE Software, vol. 22, no. 2, pp. 34-43,
March-April 2005.
Design should give correct solution but should also be elegant (or optimal). Product design is
usually open-ended because it generally has no unique solution, but some designs are “better”
than others, although all may be “correct.” Better quality matters because software is a living
thing—customer will come back for more features or modified features because of different user
types or growing business. This is usually called maintenance phase of the software lifecycle and
experience shows that it represents the dominant costs of a software product over its entire
lifecycle. Initial design is just a start for a good product and only a failed product will end with a
single release.
Class diagrams do not allow describing the ordering of the constituent parts of an aggregation.
The ordering is important in some applications, such as XML Schema (Chapter 6). We could use
the stereotype «ordered» on the “Has-a” relationship, although this approach lacks the advantage
of graphical symbols. More importantly, «ordered» relationship just says the collection is
ordered, but does not allow showing each element individually to specify where it is in the order,
relative to other elements.
Section 2.3: Software Architecture
Section 2.7: Test-driven Implementation
[Raskin, 2005] [Malan & Halland, 2004] [Ostrand et al., 2004]
Useful information on Java programming is available at:
http://www.developer.com/ (Gamelan) and http://www.javaworld.com/ (magazine)
For serial port communication in Java, I found useful information here (last visited 18 January
2006):
http://www.lvr.com/serport.htm
http://www.cs.tufts.edu/~jacob/150tui/lecture/01_Handyboard.html
http://show.docjava.com:8086/book/cgij/exportToHTML/serialPorts/SimpleRead.java.html
Also informative is Wikibooks: Serial Data Communications, at:
http://en.wikibooks.org/wiki/Programming:Serial_Data_Communications
Ivan Marsic 
Rutgers University
156
http://en.wikibooks.org/wiki/Serial_communications_bookshelf
A key book on refactoring is [Fowler, 2000]. The refactoring literature tends to focus on specific,
small-scale design problems. Design patterns focus on larger-scale design problems and provide
targets for refactorings. Design patterns will be described in Chapter 5.
A number of studies have suggested that code review reduces bug rates in released software.
Some studies also show a correlation between low bug rates and open source development
processes. It is not clear why it should be so.
The most popular unit testing framework is the xUnit family (for many languages), available at
http://www.junit.org. For Java, the popular version is JUnit, which is integrated into most of the
popular IDEs, such as Eclipse (http://www.eclipse.org). The xUnit family, including JUnit, was
started by Kent Beck (creator of eXtreme Programming) and Eric Gamma (one of the Gang-ofFour design pattern authors (see Chapter 5), and the chief architect of Eclipse. A popular free
open source tool to automatically rebuild the application and run all unit tests is CruiseControl
(http://cruisecontrol.sourceforge.net).
Testing aims to determine program’s correctness—whether it performs computations correctly, as
expected. However, a program may perform correctly but be poorly designed, very difficult to
understand and modify. To evaluate program quality, we use software metrics (Chapter 4).
Problems
“To learn is no easy matter and to apply what one has learned is even harder.”
—Chairman Mao Tse-Tung
Problem 2.1
Consider the following nonfunctional requirements and determine which of them can be verified
and which cannot. Write acceptance tests for each requirement or explain why it is not testable.
(a) “The user interface must be user-friendly and easy to use.”
(b) “The number of mouse clicks the user needs to perform when navigating to any window
of the system’s user interface must be less than 10.”
(c) “The user interface of the new system must be simple enough so that any user can use it
with a minimum training.”
(d) “The maximum latency from the moment the user clicks a hyperlink in a web page until
the rendering of the new web page starts is 1 second over a broadband connection.”
(e) “In case of failure, the system must be easy to recover and must suffer minimum loss of
important data.”
Chapter 2 
Object-Oriented Software Engineering
157
Problem 2.2
Problem 2.3
You are hired to develop an automatic patient monitoring system for a
home-bound patient. The system is required to read out the patient’s heart
rate and blood pressure and compare them against specified safe ranges. The
system also has activity sensors to detect when the patient is exercising and
adjust the safe ranges. In case an abnormality is detected, the system must
alert a remote hospital. (Note that the measurements cannot be taken
continuously, since heart rate is measured over a period of time, say 1
minute, and it takes time to inflate the blood-pressure cuff.) The system must
also (i) check that the analog devices for measuring the patient’s vital signs
are working correctly and report failures to the hospital; and, (ii) alert the owner when the battery
power is running low.
Enumerate and describe the requirements for the system-to-be.
Problem 2.4
Problem 2.5
Problem 2.6
Problem 2.7
Problem 2.8
Consider an online auction site, such as eBay.com, with selling, bidding,
and buying services. Assume that you are a buyer, you have placed a bid
for an item, and you just received a notification that the bidding process is
closed and you won it. Write a single use case that represents the
subsequent process of purchasing the item with a credit card. Assume the
business model where the funds are immediately transferred to the seller’s account, without
waiting for the buyer to confirm the receipt of the goods. Also, only the seller is charged selling
fees. Start from the point where you are already logged in the system and consider only what
happens during a single sitting at the computer terminal. (Unless otherwise specified, use cases
Ivan Marsic 
158
Rutgers University
Remote
Receiver
External
Light
Motion
Detector
Manual
Opener
Switch
Electric
Eye
Motor
Motion detection perimeter
Remote
Transmitter
Figure 2-45: Depiction of the problem domain for Problem 2.10.
are normally considered only for the activities that span a single sitting.) List also some alternate
scenarios.
Problem 2.9
Consider the online auction site described in Problem 2.8. Suppose that by observation you
determine that the generic Buyer and Seller roles can be further differentiated into more
specialized roles:

Occasional Buyer, Frequent Buyer, and Collector

Small Seller, Frequent Seller, and Corporate Seller
Identify the use cases for both situations: generic Buyers and Sellers vs. differentiated Buyers and
Sellers. Discuss the similarities and differences. Draw the use case diagrams for both situations.
Problem 2.10
You are hired to develop a software system for motion detection and garage door control.
The system should turn the garage door lights on automatically when it detects motion within a
given perimeter.
The garage door opener should be possible to control either by a remote radio transmitter or by a
manual button switch. The opener should include the following safety feature. An “electric eye”
sensor, which projects invisible infrared light beams, should be used to detect if someone or
something passes under the garage door while it closes. If the beam is obstructed while the door is
going down, the door should not close—the system should automatically stop and reverse the
door movement.
The relevant hardware parts of the system are as follows (see Figure 2-45):
 motion detector
 external light bulb
 motor for moving the garage door
 “electric eye” sensor
 remote control radio transmitter and receiver
 manual opener button switch
Chapter 2 
159
Object-Oriented Software Engineering
«initiate»
TurnLightOn
AnimateObject
te»
itia
n
i
«
TurnLightOff
«initiate»
«initia
te»
StealOpenerCode
te»
itia
RemoteOpen
Thief
«in
Figure 2-46: A fragment of a possible use case diagram for Problem 2.11.
Assume that all the hardware components are available and you only need to develop a software
system that controls the hardware components.
(a) Identify the actors for the system and their goals
(b) Derive only the use cases relevant to the system objective and write brief or casual text
description of each
(c) Draw the use case diagram for the system
(d) For the use case that deals with the remote-controlled garage door opening, write a fully
dressed description
(e) Draw the system sequence diagram(s) for the use case selected in (d)
(f) Draw the domain model with concepts, associations, and attributes
[Note: derive the domain model using only the information that is available so far—do
not elaborate the other use cases]
(g) Show the operation contracts for the operations of the use case selected in (d)
Problem 2.11
For the system described in Problem 2.10, consider the following security issue. If the remote
control supplied with the garage door opener uses a fixed code, a thief may park near your house
and steal your code with a code grabber device. The thief can then duplicate the signal code and
open your garage at will. A solution is to use so called rolling security codes instead of a fixed
code. Rolling code systems automatically change the code each time you operate your garage
door.
(f) Given the automatic external light control, triggered by motion detection, and the above
security issue with fixed signaling codes, a possible use case diagram is as depicted in
Figure 2-46. Are any of the shown use cases legitimate? Explain clearly your answer.
(g) For the use case that deals with the remote-controlled garage door closing, write a fully
dressed description.
(h) Draw the system sequence diagram(s) for the use case selected in (b).
(i) Draw the domain model with concepts, associations, and attributes .
[Note: derive the domain model using only the information that is available so far—do
not elaborate the other use cases.]
Ivan Marsic 
160
Rutgers University
(j) Show the operation contracts for the operations of the use case selected in (b).
Problem 2.12
Derive the basic use cases for the restaurant automation system (described at the book website,
given in Preface). Draw the use case diagram.
Problem 2.13
Identify the actors and derive the use cases for the vehicular traffic information system (described
at the book website, given in Preface). Draw the use case diagram. Also, draw the system
sequence diagram for the use case that deals with data collection.
Problem 2.14
Consider the automatic patient monitoring system described in Problem 2.3. Identify the
actors and their goals. Briefly, in one sentence, describe each use case but do not elaborate
them. Draw the use case diagram.
Problem 2.15
Consider a grocery supermarket planning to computerize their inventory management. This
problem is similar to one described in Example 1.2 (Section 1.5.3), but has a different goal. The
items on shelves will be marked with Radio Frequency Identification (RFID) tags and a set of
RFID reader-devices will be installed for monitoring the movements of the tagged items. Each
tag carries a 96-bit EPC (Electronic Product Code) with a Global Trade Identification number,
which is an international standard. The RFID readers are installed on each shelf on the sales floor.
1. Request
2. Response
Tag
Reader
Tag
RFID System:
Tag
Tag
The RFID system consists of two types of components (see figure above): (1) RFID tag or
transponder, and (2) RFID reader or transceiver. RFID tags are passive (no power source), and
use the power induced by the magnetic field of the RFID reader. An RFID reader consists of an
antenna, transceiver and decoder, which sends periodic signals to inquire about any tag in
vicinity. On receiving any signal from a tag it passes on that information to the data processor.
You are tasked to develop a software system for inventory management. The envisioned system
will detect which items will soon be depleted from the shelves, as well as when shelves run out of
Chapter 2 
161
Object-Oriented Software Engineering
stock and notify the store management. The manager will be
able to assign a store associate to replenish the shelf, and the
manager will be notified when the task is completed.
RFID tag
Based on the initial ideas for the desired functions of the
software system, the following requirements are derived:
REQ1. The system shall continuously monitor the tagged
items on the shelves. Every time an item is removed,
this event is recorded in the system database by
recording the current item count from the RFID reader.
The system should also be able to handle the cases
when the customer takes an item, puts it in her
shopping cart, continues shopping, and then changes
her mind, comes back and returns the item to the shelf.
REQ2. The system shall keep track when stock is running low
on shelves. It shall detect a “low-stock” state for a
product when the product’s item count falls below a
given threshold while still greater than zero.
RFID reader
REQ3. The system shall detect an “out-of-stock” state for a
product when the shelf becomes empty and the
product’s item count reaches zero.
Main computer
REQ4. The system shall notify the store manager when a
“low-stock” or “out-of-stock” state is detected, so the
shelves will be replenished. The notification will be
sent by electronic mail, and the manager will be able to
read it on his mobile phone.
REQ5. The store manager shall be able to assign a store
associate with a task to replenish a particular shelf with
a specific product. The store associate shall be notified
by electronic mail about the details of the assigned
task.
Store manager
REQ6. While the store associate puts items on the shelf, the RFID system shall automatically
detect the newly restocked items by reading out their EPC. The system should support the
option that customers remove items at the same time while the store associate is
replenishing this shelf.
REQ7. The store associate shall be able to explicitly inform the system when the replenishment
task is completed. The number of restocked items will be stored in the database record.
The item count obtained automatically (REQ5) may be displayed to the store associate
for verification. After the store associate confirms that the shelf is replenished, the task
status will be changed to “completed,” and a notification event will be generated for the
store manager.
To keep the hardware and development costs low, we make the following assumptions:
Ivan Marsic 
Rutgers University
162
A1. You will develop only the software that runs on the main computer and not that for the
peripheral RFID devices. Assume that the software running the RFID readers will be purchased
together with the hardware devices.
A2. The tag EPC is unique for a product category, which means that the system cannot
distinguish different items of the same product. Therefore, the database will store only the total
count of a given product type. No item-specific information will be stored.
A3. Assume that the RFID system works perfectly which, of course, is not true in reality. As of
this writing (2011) on an average 20% of the tags do not function properly. Accurate read rates on
some items can be very low, because of physical limitations like reading through liquid or metals
still exist or interference by other wireless sources that can disrupt the tag transmissions.
A4. Assume that the item removal event is a clean break, which again, may not be true. For
example, if the user is vacillating between buying and not buying, the system may repeatedly
count the item as removed or added and lose track of correct count. Also, the user may return an
item and take another one of the same kind because she likes the latter more than the former. (A
solution may be periodically to scan all tags with the same EPC, and adjust incorrect counts in the
database.)
A5. Regarding REQ1, each RFID reader will be able to detect correctly when more than one item
of the same type is removed simultaneously. If a customer changed her mind and returned an item
(REQ1), we assume that she will return it to the correct shelf, rather than any shelf.
A6. The communication network and the computing system will be able to handle correctly large
volume of events. Potentially, there will be many simultaneous or nearly simultaneous RFID
events, because there is a large number of products on the shelves and there may be a great
number of customers currently in the store, interacting with the items. We assume that the great
number of events will not “clog” the computer network or the processors.
Do the following:
(a) Write all the summary use cases that can be derived from the requirements REQ1–REQ7.
For each use case, indicate the related requirements. Note that one use case may be
related to several requirements and vice versa, one requirement may be related to several
use cases.
(b) Draw the use case diagram for the use cases described in item (a).
(c) Discuss additional requirements and use cases that could be added to this system.
Problem 2.16
Consider again the Grocery Inventory Management system described in Problem 2.15. Focus
only on the summary use cases that deal with depleted stock detection, related to the requirements
REQ1–REQ4. Write the detailed specification for these use cases only.
Problem 2.17
Chapter 2 
Object-Oriented Software Engineering
163
Problem 2.18
Problem 2.19
Consider a variation of the home access control system which will do user identification based on
face recognition, as described in Section 2.4.2. Write the detailed use case descriptions of use
cases UC3: AddUser and UC4: RemoveUser for both cases given in Figure 2-16, that is locally
implemented face recognition (Case (a)) and remotely provided face recognition (Case (b)).
Problem 2.20
Consider an automatic bank machine, known as Automatic Teller Machine (ATM), and a
customer who wishes to withdraw some cash from his or her banking account. Draw a UML
activity diagram to represent this use case.
Problem 2.21
Derive the domain model with concepts, associations, and attributes for the virtual mitosis lab
(described at the book website, given in Preface).
Note: You may wonder how is it that you are asked to construct the domain model without first
having the use cases derived. The reason is, because the use cases for the mitosis lab are very
simple, this is left as an exercise for the reader.
Problem 2.22
Explain the relationship between use cases and domain model objects and illustrate by example.
Problem 2.23
Problem 2.24
Problem 2.25
Problem 2.26
Problem 2.27
Ivan Marsic 
Rutgers University
164
Problem 2.28
Problem 2.29
An example use case for the system presented in Section 1.5.1 is given as follows. (Although the
advertisement procedure is not shown to preserve clarity, you should assume that it applies where
appropriate, as described in Section 1.5.1.)
Use Case UC-x:
BuyStocks
Player
[full name: investor player]
Initiating Actor:
To buy stocks, get them added to his portfolio automatically
Actor’s Goal:
[e.g., Yahoo! Finance]
Participating Actors: StockReportingWebsite
Player is currently logged in the system and is shown a hyperlink “Buy
Preconditions:
stocks.”
System has informed the player of the purchase outcome. The logs and
Postconditions:
the player’s portfolio are updated.
Flow of Events for Main Success Scenario:
 1. Player clicks the hyperlink “Buy stocks”
 2. System prompts for the filtering criteria (e.g., based on company names, industry
sector, price range, etc.) or “Show all”
3.
Player
specifies the filtering criteria and submits

 4. System contacts StockReportingWebsite and requests the current stock prices for
companies that meet the filtering criteria
 5. StockReportingWebsite responds with HTML document containing the stock prices
 6. From the received HTML document, System extracts, formats, and displays the stock
prices for Player’s consideration; the display also shows the player’s account balance
that is available for trading
 7. Player browses and selects the stock symbols, number of shares, and places the order
_
8. System (a) updates the player’s portfolio; (b) adjusts the player’s account balance,
including a commission fee charge; (c) archives the transaction in a database; and (d)

informs Player of the successful transaction and shows his new portfolio standing

Note that in Step 8 above only virtual trading takes place because this is fantasy stock trading.
Derive (a part of) the domain model for the system-to-be based on the use case BuyStocks.
(a) Write a definition for each concept in your domain model.
(b) Write a definition for each attribute and association in your domain model.
(c) Draw the domain model.
(d) Indicate the types of concepts, such as «boundary», «control», or «entity».
Chapter 2 
: GUI
Customer
enterCard()
165
Object-Oriented Software Engineering
: Controller
: CustomerID
: IDChecker
: CustomerIDStore
: AcctInfo
: AcctManager
: CashDispenserCtrl
askPIN()
enterPIN()
askAmt()
enterAmt()
Figure 2-47: Sequence diagram for the ATM machine of Problem 2.30 (see text for
explanation). GUI = Graphical user interface.
Problem 2.30
Suppose you are designing an ATM machine (also see Problem 2.20). Consider the use case
“Withdraw Cash” and finish the sequence diagram shown in Figure 2-47. The
CustomerID object contains all the information received from the current customer.
IDChecker compares the entered ID with all the stored IDs contained in
CustomerIDStorage. AcctInfo mainly contains information about the current account
balance. AcctManager performs operations on the AcctInfo, such as subtracting the
withdrawn amount and ensuring that the remainder is greater than or equal to zero.
Lastly, CashDispenserCtrl control the physical device that dispenses cash.
One could argued that AcctInfo and AcctManager should be combined into a single
object Account, which encapsulates both account data and the methods that operate
on the data. The account data is most likely read from a database, and the container
object is created at that time. Discuss the pros and cons for both possibilities.
7
4
1
0
2
5 3
8 6
9
Indicate any design principles that you employ in the sequence diagram.
Problem 2.31
You are to develop an online auction site, with selling, bidding, and buying services. The buying
service should allow the users to find an item, bid for it and/or buy it, and pay for it. The use case
diagram for the system may look as follows:
Online Auction Site
ListItem
»
«in
itia
«p te»
ar
tic
ipa
te
»
Seller
itia
«in
»
te
«participate»
BidForItem
ViewBids
CloseAuction
BuyItem
RateTransaction
Creditor
«include»
«init
iate
«initiate
FindItem
»
«initiate»
»
ate
icip»
t
r
a
«p tiate
i
«in
Buyer
«in
itia
te
»
e»
«include»
itiat
«in
?
Shipping
Agency
Ivan Marsic 
166
Rutgers University
ItemInfo
1
– name : String
– startPrice : float
– reserved : boolean
+
+
+
+
+
+
getName() : String
getStartPrice() : float
getSeller() : SellerInfo
getBidsList() : BidsList
setReserved(ok : boolean)
isReserved() : boolean
ItemsCatalog
*
+
+
+
+
Controller
add(item: ItemInfo) : int
remove(idx : int)
getNext(): ItemInfo
hasMore() : boolean
bids
1
+
+
+
+
+
+
+
listItem(item: ItemInfo)
findItem(name : String)
bidForItem(name : String)
viewBids(itemName : String)
closeAuction(itmNam : String)
buyItem(name : String)
payForItem(price: float)
BidsList
seller
1
+
+
+
+
SellerInfo
– name : String
– address : String
add(bid: Bid) : int
remove(idx : int)
getNext(): Bid
hasMore() : boolean
BuyerInfo
– name : String
– address : String
+ getName() : String
+ getAddress() : String
+ getName() : String
+ getAddress() : String
*
1
– amount : float
Payment
+ getBuyer() : BuyerInfo
… Etc.
1
bidder
+ getBidder() : BuyerInfo
+ getAmount() : float
– amount : float
item
1
Bid
seller
buyer
Figure 2-48: A possible class diagram for the online auction site of Problem 2.31.
We assume a simple system to which extra features may be added, such as auction expiration
date on items. Other features may involve the shipment agency to allow tracking the shipment
status.
A possible class diagram for the system is shown in Figure 2-48. Assume that ItemInfo is marked
as “reserved” when the Seller accepts the highest bid and closes the auction on that item only.
Before closing, Seller might want to review how active the bidding is, to decide whether to wait
for some more time before closing the bid. That particular ItemInfo is removed from
ItemsCatalog once the payment is processed.
In the use case CloseAuction, the Seller reviews the existing bids for a given item, selects the
highest and notifies the Buyer associated with the highest bid about the decision (this is why
«participate» link between the use case CloseAuction and Buyer). Assume that there are more
than one bids posted for the selected item.
Complete the interaction diagram shown below for this use case. Do not include processing the
payment (for this use case see Problem 2.8). (Note: You may introduce new classes or modify the
existing classes in Figure 2-48 if you feel it necessary for solving the problem.)
Chapter 2 
Object-Oriented Software Engineering
167
: Controller
Seller
Buyer
viewBids(itemName)
?
closeAuction(itemName)
Problem 2.32
Consider the use case BuyStocks presented in Problem 2.29. The goal is to draw the UML
sequence diagram only for Step 6 in this use case. Start at the point when the system receives the
HTML document from the StockReportingWebsite and stop at the point when an HTML page is
prepared and sent to player’s browser for viewing.
(a) List the responsibilities that need to be assigned to software objects.
(b) Assign the responsibilities from the list in (a) to objects. Explicitly mention any design
principles that you are using in your design, such as Expert Doer, High Cohesion, or Low
Coupling. Provide arguments as to why the particular principle applies.
(c) Draw the UML sequence diagram.
Problem 2.33
Problem 2.34
In the patient-monitoring scenario of Problem 2.3 and Problem 2.14, assume that the hospital
personnel who gets notified about patient status is not office-bound but can be moving around the
hospital. Also, all notifications must be archived in a hospital database for a possible future
auditing. Draw a UML deployment diagram representing the hardware/software mapping of this
system.
Problem 2.35
Consider the automatic patient monitoring system described in Problem 2.3 and analyzed in
Problem 2.14. Focus on the patient device only and ignore any software that might be
running in the remote hospital. Suppose you are provided with an initial software design as
follows.
Ivan Marsic 
168
Rutgers University
The domain model consists of the following concepts and their responsibilities:
Responsibility
Read out the patient’s blood pressure from a sensor
Read out the patient’s heart rate from a sensor
Compare the vital signs to the safe ranges and detect if the vitals are outside
Hold description of the safe ranges for patient vital signs; measurements
outside these ranges indicate elevated risk to the patient; should be
automatically adjusted for patient’s activity
Accept user input for constraints on safe ranges
Read the patient’s activity indicators
Recognize the type of person’s activity
Hold description of a given type of person’s activity
Send an alert to a remote hospital
Hold information sent to the hospital about abnormal vitals or faulty sensors
Run diagnostic tests on analog sensors
Interpret the results of diagnostic tests on analog sensors
Hold description of a type of sensor failure
Read the remaining batter power
Send an alert to the patient
Hold information sent to the patient about low battery
Coordinate activity and delegate work to other concepts
Concept
Blood Pressure Reader
Heart Rate Reader
Abnormality Detector
Vitals Safe Ranges
Safe Range Entry
Activity Observer
Activity Classifier
Activity Model
Hospital Alerter
Hospital Alert
Sensor Diagnostic
Failure Detector
Sensor Failure Mode
Battery Checker
Patient Alerter
Patient Alert
Controller
A sketchy UML sequence diagram is designed using the given concepts as in Figure 2-49. Note
that this diagram is incomplete: the part for checking the batter power is not shown for the lack of
space. However, it should be clear from the given part how the missing part should look like.
Recall that the period lengths for observations made by our system are related as:
BP Reader & HR Reader < Sensor Diagnostic < Activity Observer < Battery Checker
In other words, vital signs are recorded frequently and battery is checked least frequently. These
relationships also indicate the priority or relative importance of the observations. However, the
initial design takes a simplified approach and assumes a single timer that periodically wakes up
the system to visit all different sensors, and acquire and process their data. You may but do not
need to stick with this simplified design in your solution.
Using the design principles from Section 2.6 or any other principles that you are aware of, solve:
(a) Check if the design in Figure 2-49 already uses some design principles and, if so, explain
your claim.
- If you believe that the given design or some parts of it are sufficiently good then
explain how the application of any interventions would make the design worse.
- Be specific and avoid generic or hypothetical explanations of why some designs
are better than others. Use concrete examples and UML diagrams or pseudo-code
to illustrate your point and refer to specific qualities of software design.
(b) Carefully examine the sketchy design in Figure 2-49 and identify as many opportunities
as you can to improve it by applying design principles.
- If you apply a principle, first argue why the existing design may be problematic.
- Provide as much details as possible about how the principle will be implemented
and how the new design will work (draw UML sequence diagrams or write
pseudo-code).
Chapter 2 
Object-Oriented Software Engineering
: Controller
: VitalSignReader
wakeup
Blood pressr
Heart rate
169
: VSafeRanges
: SensDiagnostc
: FailureDetectr
vital := readVitalSign( )
: AbnormalDetect
: HospitalAlerter
: ActivityObserv
: ActivityClassif
ranges := getValues()
abnormal := isOutOfRange(vital)
opt
check if in/out
[ abnormal == TRUE ] send( Hospital Alert Abnormal Vitals )
faulty := isFaulty()
result :=
run tests
isFailed(result)
opt
[ faulty == TRUE ]
send( Hospital Alert Sensor Failure )
exrcs := isExercising()
data :=
read sensor
classify(data)
opt
[ exrcs == TRUE ]
adjust( exercise-mode )
Figure 2-49: A sketchy UML sequence diagram for patient monitoring in Problem 2.35.
-
Explain how the principle that you introduced improved the original design (i.e.,
what are the expected benefits compared to the original design).
Feel free to introduce new concepts, substitute the given concepts with different ones, or modify
their responsibilities. You may also discard existing concepts if you find them redundant. In
addition, you may change how acquisition of different sensory data is initiated. However, when
you do so, explain the motivation for your actions.
Problem 2.36
Chapter 3
Modeling and System Specification
Contents
“The beginning is the most important part of the work.” —Plato
The term “system specification” is used both for the process
of deriving the properties of the software system as well as for
the document that describes those properties. As the system is
developed, its properties will change during different stages of
its lifecycle, and so it may be unclear which specification is
being referred to. To avoid ambiguity we adopt a common
meaning: The system specification states what should be valid
(true) about the system at the time when the system is
delivered to the customer. Specifying system means stating
what we desire to achieve, not how we plan to accomplish it or
what has been achieved at an intermediate stage. The focus of
this chapter is on describing the system function, not its form.
Chapter 5 will focus on the form, how to build the system.
3.1 What is a System?
3.1.1
3.1.2
3.1.3
3.1.4
3.1.5
3.2 Notations for System Specification
3.2.1
3.2.2
3.2.3
3.2.4
Understanding the problem and determining what
needs to be specified

Selecting notation(s) to use for the specification

Verifying that
requirements
the
specification
meets
Basic Formalisms for Specifications
UML State Machine Diagrams
UML Object Constraint Language (OCL)
TLA+ Notation
3.3 Problem Frames
3.3.1 Problem Frame Notation
3.3.2 Problem Decomposition into Frames
3.3.3 Composition of Problem Frames
3.3.4
3.4 Specifying Goals
3.4.1
3.4.2
3.4.3
3.4.4
There are several aspects of specifying the system under
development, including:

World Phenomena and Their Abstractions
States and State Variables
Events, Signals, and Messages
Context Diagrams and Domains
Systems and System Descriptions
3.5
3.5.1
3.5.2
3.5.3
3.6 Summary and Bibliographical Notes
the
Problems
Of course, this is not a linear sequence of activities. Rather, as we achieve better understanding of
the problem, we may wish to switch to a different notation; also, the verification activity may
uncover some weaknesses in understanding the problem and trigger an additional study of the
problem at hand.
We have already encountered one popular notation for specification, that is, the UML standard.
We will continue using UML and learn some more about it as well as about some other notations.
Most developers agree that a single type of system model is not enough to specify any non-trivial
system. You usually need several different models, told in different “languages” for different
stakeholders. The end user has certain requirements about the system, such as that the system
allows him to do his job easier. The business manager may be more concerned about the policies,
170
Chapter 3 
Modeling and System Specification
171
rules, and processes supported by the system. Some stakeholders will care about engineering
design’s details, and others will not. Therefore, it is advisable to develop the system specification
as viewed from several different angles, using different notations.
My primary concern here is the developer’s perspective. We need to specify what are the
resting/equilibrium states and the anticipated perturbations. How does the system appear in an
equilibrium state? How does it react to a perturbation and what sequence of steps it goes through
to reach a new equilibrium? We already saw that use cases deal with such issues, to a certain
extent, although informally. Here, I will review some more precise approaches. This does not
necessarily imply formal methods. Some notations are better suited for particular types of
problems. Our goal is to work with a certain degree of precision that is amenable to some form of
analysis.
The system specification should be derived from the requirements. The specification should
accurately describe the system behavior necessary to satisfy the requirements. Most developers
would argue that the hardest part of software task is arriving at a complete and consistent
specification, and much of the essence of building a program is in fact the debugging its
specification—figuring out what exactly needs to be done. The developer might have
misunderstood the customer’s needs. The customer may be unsure, and the initial requirements
will often be fuzzy or incomplete. I should emphasize again and again that writing the
requirements and deriving the specification is not a strictly sequential process. Rather, we must
explore the requirements and system specification iteratively, until a satisfactory solution is
found. Even then, we may need to revisit and reexamine both if questions arise during the design
and implementation.
Although the system requirements are ultimately decided by the customer, the developer needs to
know how to ask the right questions and how to systemize the information gathered from the
customer. But, what questions to ask? A useful approach would be to be start with a catalogue of
simple representative problems that tend to occur in every real-world problem. These elementarybuilding-block problems are called “problem frames.” Each can be described in a well-defined
format, each has a well-known solution, and each has a well-known set of associated issues. We
already made initial steps in Section 2.3.1. In Section 3.3 we will see how complex problems can
be made manageable by applying problem frames. In this way, problem frames can help us bridge
the gap between system requirements and system specification.
3.1 What is a System?
“All models are wrong, but some are useful.” —George E. P. Box
“There is no property absolutely essential to one thing. The same property, which figures as the essence of
a thing on one occasion, becomes a very inessential feature upon another.” —William James
In Chapter 2 we introduced system-to-be, or more accurately the software-to-be, as the software
product that a software engineer (or a team of engineers) sets out to develop. Apart from the
system, the rest of the world (“environment”) has been of concern only as far as it interacts with
Ivan Marsic

Rutgers University
172
the system and it was abstracted as a set of actors. By describing different interaction scenarios as
a set of use cases, we were able to develop a software system in an incremental fashion.
However, there are some limitations with this approach. First, by considering only the “actors”
that the system directly interacts with, we may leave out some parts of the environment that have
no direct interactions with the software-to-be but are important to the problem and its solution.
Consider, for example, the stock market fantasy league system and the context within which it
operates (Figure 1-32). Here, the real-world stock market exchange does not interact with our
software-to-be, so it would not be considered an “actor.” Conceivably, it would not even be
mentioned in any of the use cases, because it is neither an initiating nor a participating actor! I
hope that the reader would agree that this is strange—the whole project revolves about a stock
exchange and yet the stock exchange may not appear in the system description at all.
Second, starting by focusing on interaction scenarios may not be the easiest route in describing
the problem. Use cases describe the sequence of user’s (actor) interaction with the system. I
already mentioned that use cases are procedural rather than object-oriented. The focus on
sequential procedure may not be difficult to begin with, but it requires being on a constant watch
for any branching off of the “main success scenario.” Decision making (branching points) may be
difficult to detect—it may be hard to conceive what could go wrong—particularly if not guided
by a helpful representation of the problem structure.
The best way to start conceptual modeling may be with how users and customers prefer to
conceptualize their world, because the developer needs to have a great deal of interaction with
customers at the time when the problem is being defined. This may also vary across different
application domains.
In this chapter I will present some alternative approaches to problem description (i.e.,
requirements and specification), which may be more involved but are believed to offer easier
routes to solving large-scale and complex problems.
3.1.1
World Phenomena and Their Abstractions
The key to solving a problem is in understanding the problem. Because problems are in the real
world, we need good abstractions of world phenomena. Good abstractions will help us to
represent accurately the knowledge that we gather about the world (that is, the “application
domain,” as it relates to our problem at hand). In object-oriented approach, key abstractions are
objects and messages and they served us well in Chapter 2 in understanding the problem and
deriving the solution. We are not about to abandon them now; rather, we will broaden our
horizons and perhaps take a slightly different perspective.
Usually we partition the world in different parts (or regions, or domains) and consider different
phenomena, see Figure 3-1. A phenomenon is a fact, or object, or occurrence that appears or is
perceived to exist, or to be present, or to be the case, when you observe the world or some part of
it. We can distinguish world phenomena by different criteria. Structurally, we have two broad
categories of phenomena: individuals and relations among individuals. Logically, we can
distinguish causal vs. symbolic phenomena. In terms of behavior, we can distinguish
deterministic vs. stochastic phenomena. Next I describe each kind briefly.
Chapter 3 
173
Modeling and System Specification
WORLD
Part/Domain J
Part/Domain I
Phenomena in Part i
Phenomena in Part j
Shared
phenomena
Phenomena in Part k
Part/Domain K
Figure 3-1: World partitioning into domains and their phenomena.
I should like to emphasize that this is only one possible categorization, which seems suitable for
software engineering; other categorizations are possible and have been proposed. Moreover, any
specific identification of world phenomena is evanescent and bound to become faulty over time,
regardless of the amount of effort we invest in deriving it. I already mentioned in Section 1.1.1
the effect of the second law of thermodynamics. When identifying the world phenomena, we
inevitably make approximations. Certain kinds of information are regarded as important and the
rest of the information is treated as unimportant and ignored. Due to the random fluctuations in
the nature and society, some of the phenomena that served as the basis for our separation of
important and unimportant information will become intermingled thus invalidating our original
model. Hence the ultimate limits to what our modeling efforts can achieve.
Individuals
An individual is something that can be named and reliably distinguished from other individuals.
Decisions to treat certain phenomena as individuals are not objective—they depend on the
problem at hand. It should be clear by now that the selected level of abstraction is relative to the
observer. We choose to recognize just those individuals that are useful to solving the problem and
are practically distinguishable. We will choose to distinguish three kinds of individual: events,
entities, and values.
 An event is an individual happening, occurring at a particular point in time. Each event is
indivisible and instantaneous, that is, the event itself has no internal structure and takes no time to
happen. Hence, we can talk about “before the event” and “after the event,” but not about “during
the event.” An example event is placing a trading order; another example event is executing a
stock trading transaction; yet another example is posting a stock price quotation. Further
discussion of events is in Section 3.1.3.
Ivan Marsic

Rutgers University
174
Set of all pairs of persons
Set of all 3-tuples of persons
Set of love triangles
Set of neighbors
(a)
(b)
Figure 3-2: Example relations: Neighbors(Person_i, Person_j) and InLoveTriangle(Person_i,
Person_j, Person_k).
 An entity is an individual with distinct existence, as opposed to a quality or relation. An entity
persists over time and can change its properties and states from one point in time to another.
Some entities may initiate events; some may cause spontaneous changes to their own states; some
may be passive.
Software objects and abstract concepts modeled in Chapter 2 are entities. But entities also include
real-world objects. The entities are determined by what part of the world is being modeled. A
financial-trader in our investment assistant case study (Section 1.3.2) is an entity; so is his
investment-portfolio; a listed-stock is also an entity. They belong to entity classes trader,
portfolio, and stock, respectively.
 A value is an intangible individual that exists outside time and space, and is not subject to
change. The values we are interested in are such things as numbers and characters, represented by
symbols. For example, a value could be the numerical measure of a quantity or a number
denoting amount on some conventional scale, such as 7 kilograms.
In our case study (Section 1.3.2), a particular stock price is a number of monetary units in which
a stock share is priced—and is therefore a value. Examples of value classes include integer,
character, string, and so on.
Relations
“I have an infamously low capacity for visualizing relationships, which made the study of geometry and all
subjects derived from it impossible for me.” —Sigmund Freud
We say that individuals are in relation if they share a certain characteristic. To define a relation,
we also need to specify how many individuals we consider at a time. For example, for any pair of
people, we could decide that they are neighbors if their homes are less than 100 meters apart from
each other. Given any two persons, Person_i and Person_j, if they pass this test then the relation
holds (is true); otherwise it does not hold (is false). All pairs of persons that pass the test are said
to be in the relation Neighbors(Person_i, Person_j). The pairs of persons that are neighbors form
a subset of all pairs of persons as shown in Figure 3-2(a).
Relations need not be established on pairs of individuals only. We can consider any number of
individuals and decide whether they are in a relation. The number n of considered individuals can
be any positive integer n  2 and it must be fixed for every test of the relation; we will call it an ntuple. We will write relation as RelationName(Individual1, …, Individualn). When one of the
Chapter 3 
Modeling and System Specification
175
individuals remains constant for all tests of the relation, we may include its name in the relation’s
name. For example, consider the characteristic of wearing eyeglasses. Then we can test whether a
Person_i is in relation Wearing(Person_i, Glasses), which is a subset of all persons. Because
Glasses remain constant across all tests, we can write WearingGlasses(Person_i), or simply
Bespectacled(Person_i). Consider next the so-called “love triangle” relation as an example for n =
3. Obviously, to test for this characteristic we must consider exactly three persons at a time; not
two, not four. Then the relation InLoveTriangle(Person_i, Person_j, Person_k) will form a set of
all triplets (3-tuples) of persons for whom this characteristic is true, which is a subset of all
3-tuples of persons as shown in Figure 3-2(b). A formal definition of relation will be given in
Section 3.2.1 after presenting some notation.
We will consider three kinds of relations: states, truths, and roles.
 A state is a relation among individual entities and values, which can change over time. I will
describe states in Section 3.1.2, and skip them for now.
 A truth is a fixed relation among individuals that cannot possibly change over time. Unlike
states, which change over time, truths remain constant. A bit more relaxed definition would be to
consider the relations that are invariable on the time-scale that we are interested in. Example
time-scales could be project duration or anticipated product life-cycle. When stating a truth, the
individuals are always values, and the truth expresses invariable facts, such as GreaterThan(5, 3)
or StockTickerSymbol(“Google, Inc.,” “GOOG”). It is reasonably safe to assume that company
stock symbols will not change (although mergers or acquisitions may affect this!).
 A role is a relation between an event and individual that participate in it in a particular way.
Each role expresses what you might otherwise think of as one of the “arguments” (or
“parameters”) of the event.
Causal vs. Symbolic Phenomena
 Causal phenomena are events, or roles, or states relating entities. These are causal phenomena
because they are directly produced or controlled by some entity, and because they can give rise to
other phenomena in turn.
 Symbolic phenomena are values, and truths and states relating only values. They are called
symbolic because they are used to symbolize other phenomena and relationships among them. A
symbolic state that relates values—for example, the data content of a disk record—can be
changed by external causation, but we do not think of it as causal because it can neither change
itself nor cause change elsewhere.
Deterministic vs. Stochastic Phenomena
 Deterministic phenomena are the causal phenomena for which the occurrence or nonoccurrence can be established with certainty.
 Stochastic phenomena are the causal phenomena that are governed by a random distribution of
probabilities.
Ivan Marsic

3.1.2
States and State Variables
Rutgers University
176
A state describes what is true in the world at each particular point in time. The state of an
individual represents the cumulative results of its behavior. Consider a device, such as a digital
video disc (DVD) player. How the device reacts to an input command depends not only upon that
input, but also upon the internal state that the device is currently in. So, if the “PLAY” button is
pushed on a DVD player, what happens next will depend on various things, such as whether or
not the player is turned on, contains a disc, or is already playing. These conditions represent
different states of a DVD player.
By considering such options, we may come up with a list of all states for a DVD player, like this:
State 1: NotPowered
State 2: Powered
State 3: Loaded
State 4: Playing
(the player is not powered up)
(the player is powered up)
(a disc is in the tray)
We can define state more precisely as a relation on a set of objects, which simply selects a subset
of the set. For the DVD player example, what we wish to express is “The DVD player’s power is
off.” We could write Is(DVDplayer, NotPowered) or IsNotPowered(DVDplayer). We will settle
on this format: NotPowered(DVDplayer). NotPowered(x) is a subset of DVD players x that are
not powered up. In other words, NotPowered(x) is true if x is currently off. Assuming that one
such player is the one in the living room, labeled as DVDinLivRm, then
NotPowered(DVDinLivRm) holds true if the player in the living room is not powered up.
Upon a closer examination, we may realize that the above list of states implies that a nonpowered-up player never contains a disc in the tray. If you are charged to develop software for the
DVD player, you must clarify this. Does this mean that the disc is automatically ejected when the
power-off button is pushed? If this is not the case or the issue is yet unresolved, we may want to
redesign our list of DVD player states as:
State 1: NotPoweredEmpty
State 2: NotPoweredLoaded
State 3: PoweredEmpty
State 4: PoweredLoaded
State 5: Playing
(the player is not powered up and it contains no disc)
(the player is not powered up but a disc is in the tray)
(the player is powered up but it contains no disc)
(the player is powered up and a disc is in the tray)
At this point one may realize that instead of aggregate or “global” system states it may be more
elegant to discern different parts (sub-objects) of the DVD player and, in turn, consider the state
of each part (Figure 3-3). Each part has its “local” states, as in this table
System part (Object)
State relations
Power button
{Off, On}
Disc tray
{Empty, Loaded}
Play button
{Off, On}
…
…
Note that the relation Off(b) is defined on the set of buttons. Then these relations may be true:
Off(PowerButton) and Off(PlayButton). Similar holds for On(b).
Given the states of individual parts, how can we define the state of the whole system? Obviously,
we could say that the aggregate system state is defined by the states of its parts. For example, one
Chapter 3 
177
Modeling and System Specification
DVD player
DVD player
(a)
Power
button
Disc
tray
Play
button
…
(b)
Figure 3-3: Abstractions of a DVD player at different levels of detail: (a) The player as a
single entity. (b) The player seen as composed of several entities.
state of the DVD player is { On(PowerButton), Empty(), Off(PlayButton), … }. Note that the
relation Empty() is left without an argument, because it is clear to which object it refers to. In this
case we could also write Empty without parentheses. The arrangement of the relations in this
“state tuple” is not important as long as it is clear what part each relation refers to.
The question now arises, is every combination of parts’ states allowed? Are these parts
independent of each other or there are constraints on the state of one part that are imposed by the
current states of other parts? Some states of parts of a composite domain may be mutually
exclusive. Going back to the issue posed earlier, can the disc tray be in the “loaded” state when
the power button is in the “off” state? Because these are parts of the same system, we must make
explicit any mutual dependencies of the parts’ states. We may end up with a list of valid system
state tuples that does not include all possible tuples that can be constructed.
Both representations of a system state (single aggregate state vs. tuple of parts’ states) are correct,
but their suitability depends on what kind of details you care to know about the system. In
general, considering the system as a set of parts that define state tuples presents a cleaner and
more modular approach than a single aggregate state.
In software engineering, we care about the visible aspects of the software system. In general,
visible aspects do not necessarily need to correspond to “parts” of the system. Rather, they are
any observable qualities of the system. For example, domain-model attributes identified in
Section 2.5 represent observable qualities of the system. We call each observable quality a state
variable. In our first case-study example, variables include the lock and the bulb. Another
variable is the counter of the number of attempts at opening the lock. Yet another variable is the
amount of timer that counts down the time elapsed since the lock was open, to support auto-lock
functionality. The state variables of our system can be summarized as in this table
Variable
State relations
Door lock
{Disarmed, Armed}
Bulb
{Lit, Unlit}
Counter of failed attempts {0, 1, …, maxNumOfAttempts}
Auto-lock timer
{0, 1, …, autoLockInterval}
In case of multiple locks and/or bulbs, we have a different state variable for every lock/bulb,
similar to the above example of DVD player buttons. So, the state relations for backyard and front
door locks could be defined as Disarmed(Backyard) and Disarmed(Front).
Ivan Marsic

Rutgers University
178
The situation with numeric relations is a bit trickier. We could write 2(Counter) to mean that the
counter is currently in state “2,” but this is a bit awkward. Rather, just for the sake of convenience
I will write Equals(Counter, 2) and similarly Equals(Timer, 3).
System state is defined as a tuple of state variables containing any valid combination of state
relations. State is an aggregate representation of the system characteristics that we care to know
about looking from outside of the system. For the above example, an example state tuple is:
{Disarmed(Front), Lit, Armed(Backyard), Equals(Counter, 0), Equals(Timer, 0) }.
One way to classify states is by what the object is doing in a given state:
 A state is a passive quality if the object is just waiting for an event to happen. For the DVD
player described earlier, such states are “Powered” and “Loaded.”
 A state is an active quality if the object is executing an activity. When the DVD player is in the
“Playing” state it is actively playing a disc.
A combination of these options is also possible, i.e., the object may be executing an activity and
also waiting for an event.
The movements between states are called transitions and are most often caused by events
(described in Section 3.1.3). Each state transition connects two states. Usually, not all pairs of
states are connected by transitions—only specific transitions are permissible.
Example 3.1
Identifying Stock Exchange States (First Attempt)
Consider our second case study on an investment assistant system (Section 1.3.2), and suppose that we
want to identify the states of the stock exchange. There are many things that we can say about the
exchange, such as where it is located, dimensions of the building, the date it was built, etc. But, what
properties we care to know as it relates to our problem? Here are some candidates:

What are the operating hours and is the exchange currently “open” or “closed?”

What stocks are currently listed?

For each listed stock, what are the quoted price (traded/bid/ask) and the number of offered shares?

What is the current overall trading volume?

What is the current market index or average value?
The state variables can be summarized like so:
Variable
Operating condition (or gate condition)
ith stock price
ith stock number of offered shares
Trading volume
Market index/average
State relations
{Open, Closed}
any positive real number
{0, 1, 2, 3, …}
{0, 1, 2, 3, …}
any positive real number
The asterisk in the table indicates that the prices are quoted up to a certain number of decimal places
and there is a reasonable upper bound on the prices. In other words, this is a finite set of finite values.
Obviously, this system has a great many of possible states, which is, nonetheless, finite. An improvised
graphical representation is shown in Figure 3-4. (UML standard symbols for state diagrams are
described later in Section 3.2.2.)
An example state tuple is: { Open, Equals(Volume, 783014), Equals(Average, 1582), Equals(Price_1,
74.52), Equals(Shares_1, 10721), Equals(Price_2, 105.17), Equals(Shares_2, 51482), … }. Note that
the price and number of shares must be specified for all the listed stocks.
Chapter 3 
179
Modeling and System Specification
Market
index
Market
gate
Open
Closed
1.00
Stock_1_Price
Stock_1_Shares
Stock_2_Price
Stock_2_Shares
1.00
1
1.00
1
1.01
2
1.01
2
1.02
3
1.02
3
1.01
1.02
( prices and num. of shares
for all listed stocks )
Figure 3-4: Graphical representation of states for Example 3.1. The arrows indicate the
permissible paths for transitioning between different states.
As the reader should know by now, the selection of state phenomena depends on the observer and
observer’s problem at hand. An alternative characterization of a market state is presented later in
Example 3.2.
Observables vs. Hidden Variables
States Defined from Observable Phenomena
State is an abstraction, and as such it is subjective—it depends on who is making the
abstraction. There are no “objective states”—every categorization of states is relative
to the observer. Of course, the same observer can come up with different abstractions.
The observer can also define new states based on observable phenomena; such states
are directly observed. Consider, for example, a fruit states: “green,” “semiripe,” “ripe,”
“overripe,” and “rotten.” The state of “ripeness” of a fruit is defined based on observable
parameters such as its skin color and texture, size, scent, softness on touch, etc. Similarly, a
“moving” state of an elevator is defined by observing its position over subsequent time moments
and calculating the trend.
For the auto-lock timer discussed earlier, we can define the states “CountingDown” and “Idle”
like so:
CountingDown(Timer)
Idle(Timer)
The symbol




Example 3.2


The relation Equals(Timer, ) holds true for  decreasing with time
The relation Equals(Timer, ) holds true for  remaining constant with time
means that this is a defined state.
Identifying Stock Exchange States (Second Attempt)
Ivan Marsic

Rutgers University
180
OrderPending
OrderExecuted
x2
x 1
x
x 1
x2
Figure 3-5: Graphical representation of states for Example 3.2. Microstates from Figure 3-4
representing the number of offered shares are aggregated into two macrostates.
Let us revisit Example 3.1. Upon closer examination, one may conclude that the trader may not find
very useful the variables identified therein. In Section 1.3.2, we speculated that what trader really cares
about is to know if a trading opportunity arises and, once he places a trading order, tracking the status
of the order. Let us assume that the trading decision will be made based on the trending direction of the
stock price. Also assume that, when an upward trend of Stock_i’s price triggers a decision to buy, a
market order is placed for x shares of Stock_i. To summarize, the trader wants to represent the states of
two things:

“Stock tradability” states (“buy,” “sell,” “hold”) are defined based on considering a time window
of recent prices for a given stock and interpolating a line. If the line exhibits an upward trend, the
stock state is Buy. The states Sell and Hold are decided similarly. A more financially astute trader
may use some of the technical analysis indicators (e.g., Figure 1-23), instead of the simple
regression line.

“Order status” states (“pending” vs. “executed”) are defined based on whether there are sufficient
shares offered so the buying transaction can be carried out. We have to be careful here, because a
selling transaction can be executed only if there are willing buyers. So, the buy and sell orders
have the same states, defined differently.
Then the trader could define the states of the market as follows:

Buy  The regression line of the relation Equals(Price_i(t), p), for t = tcurrent  Window, …, tcurrent  2,
tcurrent  1, tcurrent, has a positive slope

Sell  The regression line of the relation Equals(Price_i(t), p), for t = tcurrent  Window, …, tcurrent, has a
negative slope

Hold  The regression line of the relation Equals(Price_i(t), p), for t = tcurrent  Window, …, tcurrent, has
a zero slope

SellOrderPending  The relation Equals(Shares_i, y) holds true for all values of y less than x

SellOrderExecuted  The relation Equals(Shares_i, y) holds true for all values of y greater than or
equal to x
What we did here, essentially, is to group a large number of detailed states from Example 3.1 into few
aggregate states (see Figure 3-5). These grouped states help simplify the trader’s work.
It is possible to discern further nuances in each of these states. For example, two sub-states of the state
Sell could be distinguished as when the trader should sell to avert greater loss vs. when he may wish to
take profit at a market top. The most important point to keep in mind is the trader’s goals and strategies
for achieving them. This is by no means the only way the trader could view the market. A more
proficient trader may define the states in terms of long vs. short trading positions (see Section 1.3.2,
Figure 1-22). Example states could be:
GoLong – The given stock is currently suitable for taking a long position
181
Modeling and System Specification
Elevation of the ball
Chapter 3 
2 Event: Kick
3 State:
Ball flying
Event: Kick
1
State:
Ball standing
4
Event: Splash
Event: Splash
State:
5 Ball floating
Time
Figure 3-6: Events take place at transitions between the states.
GoShort – The given stock is currently suitable for taking a long position
GoNeutral – The trader should hold or avoid the given stock at this time
The states that are directly observable at a given level of detail (coarse graining) will be called
microstates. A group of microstates is called a macrostate (or superstate). The states defined in
Example 3.2 are macrostates.
Sometimes our abstraction may identify simultaneous (or concurrent) activities that object
executes in a given state. For example, when the DVD player is in the “Playing” state it may be
simultaneously playing a disc (producing video output) and updating the time-progress display.
Section 3.2.2 describes UML state machine diagrams as a standardized graphical notation for
representing states and transitions between them.
3.1.3
Events, Signals, and Messages
Event definition requires that events are indivisible—any happening (or performance, or action)
that has an internal time structure must be regarded as two or more events. The motivation for this
restriction is to avoid having intermediate states: an event represents a sharp boundary between
two different states. We also need to assume that no two events occur simultaneously. All events
happen sequentially, and between successive events there are intervals of time in which nothing
happens—that is, there are no events. Events and intervals alternate: each event ends one interval
and begins another. Consider the example in Figure 3-6. By examining the time diagram we
partition time into intervals (“states”) and identify what point (“event”) separates two intervals.
Then we name the resulting five phenomena as shown in Figure 3-6. We cannot have an
Ivan Marsic

Rutgers University
182
uninterrupted sequence of events—this would simply be a wrong model and would require
refining the time scale to identify the intervals between successive events.
The developer may need to make a choice of what to treat as a single event. Consider the homeaccess control case study (Section 1.3.1). When the tenant is punching in his identification key,
should this be treated as a single event, or should each keystroke be considered a different event?
The answer depends on whether your problem statement requires you to treat it one way or
another. Are there any exceptions that are relevant to the problem, which may arise between
different keystrokes? If so, then we need to treat each keystroke as an event.
The reader may wonder about the relationship between events and messages, or operations in
object-oriented approach. The notion of event as defined above is more general, because it is not
limited to object orientation. The notion of message implies that a signal is sent from one entity to
another. Unlike a message, an event is something that happens—it may include one or more
individuals but it is not necessarily directed from one individual to another. Events just mark
transitions between successive states. The advantage of this view is that we can avoid specifying
processing detail at an early stage of problem definition. Use case analysis (Section 2.4.3) is
different in that it requires making explicit the sequential processing procedure (“scenarios”),
which leads to system operations.
Another difference is that events always signify state change—even for situations where system
remains in the same state, there is an explicit description of an event and state change. Hence,
events depend on how the corresponding state set is already defined. On the other hand, messages
may not be related to state changes. For example, an operation that simply retrieves the value of
an object attribute (known as accessor operation) does not affect the object’s state.
Example events:
listStock – this event marks that it is first time available for trading – marks transition between
price states; marks a transition between number-of-shares-available states
splitStock – this event marks a transition between price states; marks transition between numberof-shares-available states
submitOrder – this event marks a transition between the states of a trading order; also marks a
transition between price states (the indicative price of the stock gets updated); also marks a
transition between number-of-shares-available states, in case of a sell-order
matchFound – this event marks a transition between the states of a trading order when a matching
order(s) is(are) found; also marks a transition between price states (the traded price of the stock
gets updated); also marks a transition between number-of-shares-available states
The above events can also mark change in “trading volume” and “market index/average.” The
reader may have observed that event names are formed as verb phrases. The reason for this is to
distinguish events from states. Although this is reminiscent of messages in object-oriented
approach, events do not necessarily correspond to messages, as already discussed earlier.
Example 3.3
Identifying Stock Exchange Events
Consider Example 3.2, where the states Buy, Sell, or Hold, are defined based on recent price
movements. The events that directly lead to transitioning between these states are order placements by
other traders. There may be many different orders placed until the transition happens, but we view the
Chapter 3 
183
Modeling and System Specification
submit
InPreparation
matched
Pending
archive
Executed
Archived
Figure 3-7: Graphical representation of events marking state transitions of a trading order.
transitioning as an indivisible event—the moment when the regression line slope exceeds a given
threshold value. The events can be summarized like so:
Event
Description
trade
Causes transition between stock states Buy, Sell, or Hold
submit
Causes transition between trading-order states
InPreparation  OrderPending
matched
Causes transition between trading-order states
OrderPending  OrderExecuted
…
…
…
…
The events marking a trading order transitions are shown in Figure 3-7. Other possible events include
bid and offer, which may or may not lead to transitions among the states of a trading order. We will
consider these in Section 3.2.2.
3.1.4
Context Diagrams and Domains
Now that we have defined basic phenomena, we can start the problem domain analysis by placing
the planned system in a context—the environment in which it will work. For this we use context
diagrams, which are essentially a bit more than the commonplace “block diagrams.” Context
diagrams are not part of UML; they were introduced by Michael Jackson [1995] based on the
notation dating back to structured analysis in 1970s. The context diagram represents the context
of the problem that the developer sets out to solve. The block diagrams we encountered in Figure
1-20(b) and Figure 1-32 are essentially context diagrams. Based on the partitioning in Figure 3-1,
we show different domains as rectangular boxes and connect them with lines to indicate that they
share certain phenomena. Figure 3-8 is Figure 1-20(b) redrawn as a context diagram, with some
details added. Our system-to-be, labeled “machine,” subsumes the broker’s role and the figure
also shows abstract concepts such as portfolio, trading order, and ith stock. Jackson uses the term
“machine” to avoid the ambiguities of the word “system,” some of which were discussed in
Section 2.4.2. We use all three terms, “system-to-be,” “software-to-be,” and “machine.”
A context diagram shows parts of the world (Figure 3-1) that are relevant to our problem and only
the relevant parts. Each box in a context diagram represents a different domain. A domain is a
part of the world that can be distinguished because it is conveniently considered as a whole, and
can be considered—to some extent—separately from other parts of the world. Each domain is a
different subject matter that appears in the description of the problem. A domain is described by
the phenomena that exist or occur in it. In every software development problem there are at least
two domains: the application domain (or environment, or real world—what is given) and the
Ivan Marsic

Rutgers University
Investment
portfolio
184
Trading
order
Context diagram symbols:
A box with a double stripe
is a machine domain
Machine
(SuD)
Trader
Stock
exchange
ith stock
A box with a single stripe
is a designed domain
A box with no stripe
is a given domain
Bank
Figure 3-8: Context diagram for our case study 2: investment advisory system.
machine (or system-to-be—what is to be constructed). Some of the domains in Figure 3-8
correspond to what we called “actors” in Chapter 2. However, there are other subject matters, as
well, such as “Investment portfolio.”
To simplify, we decide that all the domains in the context diagram are physical. In Figure 3-8,
while this may be clear for other domains, even “Investment portfolio” should be a physical
domain. We assume that the corresponding box stands for the physical representation of the
information about the stocks that the trader owns. In other words, this is the representation stored
in computer memory or displayed on a screen or printed on paper. The reason for emphasizing
physical domains and physical interactions is because the point of software development is to
build systems that interact with the physical world and help the user solve problems.
Domain Types
Domains can be distinguished as to whether they are given or are to be designed. A given domain
is a problem domain whose properties are given—we are not allowed to design such a domain. In
some cases the machine can influence the behavior of a given domain. For example, in Figure 3-8
executing trading orders influences the behavior of the stock exchange (given domain). A
designed domain is a problem domain for which data structures and, to some extent, its data
content need to be determined and constructed. An example is the “Investment portfolio” domain
in Figure 3-8.
Often, one kind of problem is distinguished from another by different domain types. To a large
degree these distinctions arise naturally out of the domain phenomena. But it is also useful to
make a broad classification into three main types.
 A causal domain is one whose properties include predictable causal relationships among its
causal phenomena.
A causal domain may control some or all or none of the shared phenomena at an interface with
another domain.
 A biddable domain usually consists of people. The most important characteristic of a biddable
domain is that it lacks positive predictable internal causality. That is, in most situations it is
impossible to compel a person to initiate an event: the most that can be done is to issue
instructions to be followed.
Chapter 3 
185
Modeling and System Specification
enable
disable
Play
button
notify
start
stop
activate
shut down
Disc tray
eject (?)
Power
button
activate
shut down
Display
eject
load
enable
disable
notify
Eject
button
Figure 3-9: Domains and shared phenomena in the problem of controlling a DVD player.
 A lexical domain is a physical representation of data—that is, of symbolic phenomena.
Shared Phenomena
So far we considered world phenomena as belonging to particular domains. Some phenomena are
shared. Shared phenomena, viewed from different domains, are the essence of domain interaction
and communication. You can think of the domains as seeing the same event from different points
of view.
Figure 3-9 shows
3.1.5
Systems and System Descriptions
Now that we have defined domains as distinguishable parts of the world, we can consider any
domain as a system. A system is an organized or complex whole, an assemblage of things or
parts interacting in a coordinated way. All systems are affected by events in their environment
either internal and under the organization’s control or external and not controllable by the
organization.
Behavior under Perturbations: We need to define the initial state, other equilibrium states, and
state transitions.
Most of real-world problems require a dynamical model to capture a process which changes over
time. Depending on the application, the particular choice of model may be continuous or discrete
(using differential or difference equations), deterministic or stochastic, or a hybrid. Dynamical
systems theory describes properties of solutions to models that are prevalent across the sciences.
It has been quite successful, yielding geometric descriptions of phase portraits that partition state
space into region of solution trajectories with similar asymptotic behavior, characterization of the
statistical properties of attractors, and classification of bifurcations marking qualitative changes of
dynamical behavior in generic systems depending upon parameters. [Strogatz, 1994]
S. Strogatz, Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry
and Engineering. Perseus Books Group, 1994.
Ivan Marsic

Rutgers University
186
Given an external perturbation or stimulus, the system responds by traversing a set of transient
states until it settles at an equilibrium state. An equilibrium state may involve stable oscillations,
e.g., a behavior driven by an internal clock.
In mechanics, when an external force acts on an object, we describe its behavior through a set of
mathematical equations. Here we describe it as a sequence of (discrete) action-reaction or
stimulus-response events, in plain English.
Figure 3-x shows the state transition diagram. Action “turnLightOff” is marked with question
mark because we are yet to arrive at an acceptable solution for this case. The state [disarmed,
unlit] is not shown because the lock is not supposed to stay for a long in a disarmed state—it will
be closed shortly either by the user or automatically.
3.2 Notations for System Specification
“… psychologically we must keep all the theories in our heads, and every theoretical physicist who is any
good knows six or seven different theoretical representations for exactly the same physics. He knows that
they are all equivalent, and that nobody is ever going to be able to decide which one is right at that level,
but he keeps them in his head, hoping that they will give him different ideas for guessing.”
—Richard Feynman, The Character of Physical Law
3.2.1
Basic Formalisms for Specifications
“You can only find truth with logic if you have already found truth without it.”
—Gilbert Keith Chesterton, The Man who was Orthodox
“Logic: The art of thinking and reasoning in strict accordance with the limitations and incapacities of the
human misunderstanding.” —Ambrose Bierce, The Devil’s Dictionary
This section reviews some basic discrete mathematics that often appears in specifications. First I
present a brief overview of sets notation. A set is a well-defined collection of objects that are
called members or elements. A set is completely defined by its elements. To declare that object x
is a member of a set A, write x  A. Conversely, to declare that object y is not a member of a set
A, write x  A. A set which has no members is the empty set and is denoted as { } or .
Sets A and B are equal (denoted as A = B) if they have exactly the same members. If A and B are
not equal, write A  B. A set B is a subset of a set A if all of the members of B are members of A,
and this is denoted as B  A. The set B is a proper subset of A if B is a subset of A and B  A,
which is denoted as B  A.
The union of two sets A and B is the set whose members belong to A, B or both, and is denoted as
A  B. The intersection of two sets A and B is the set whose members belong to both A and B,
and is denoted as A  B. Two sets A and B are disjoint if their intersections is the empty set: A 
B = . When B  A, the set difference A \ B is the set of members of A which are not members of
B.
Chapter 3 
Modeling and System Specification
187
The members of a set can themselves be sets. Of particular interest is the set that contains all the
subsets of a given set A, including both  and A itself. This set is called the power set of set A
and is denoted (A), or A, or 2A.
The ordered pair x, y is a pair of objects in which x is the first object and y is the second object.
Two ordered pairs x, y and a, b are equal if and only if x = a and y = b. We define Cartesian
product or cross product of two sets A and B (denoted as A  B) as the set of all ordered pairs
x,y where x  A and y  B. We can define the n-fold Cartesian product as A  A  …  A.
Recall the discussion of relations among individuals in Section 3.1.1. An n-ary relation R on A,
for n  1, is defined as a subset of the n-fold Cartesian product, R  A  A  …  A.
Boolean Logic
The rules of logic give precise meaning to statements and so they play a key role in
specifications. Of course, all of this can be expressed in a natural language (such as English) or
you can invent your own syntax for describing the system requirements and
specification. However, if these descriptions are expressed in a standard and
p
q pq
predictable manner, not only they can be easily understood, but also
T T
T
automated tools can be developed to understand such descriptions. This
T F
F
allows automatic checking of descriptions.
F T
T
F
F
T
Propositions are the basic building block of logic. A proposition is a
declarative sentence (a sentence that declares a fact) that is either true or Truth table
false, but not both. We already saw in Section 1.3 that a proposition is a for p  q.
statement of a relation among concepts, given that the truth value of the
statement is known. Examples of declarative sentence are “Dogs are mammals” and “one plus
one equals three.” The first proposition is true and the second one is false. The sentence “Write
this down” is not a proposition because it is not a declarative sentence. Also, the sentence “x is
smaller than five” is not a proposition because it is neither true nor false (depends on what x is).
The conventional letters used to denote propositions are p, q, r, s, … These are called
propositional variables or statement variables. If a proposition is true, its truth value is denoted
by T and, conversely, the truth value of a false proposition is denoted by F.
Many statements are constructed by combining one or more propositions, using logical operators,
to form compound propositions. Some of the operators of propositional logic are shown on top of
Table 3-1.
A conditional statement or, simply a conditional, is obtained by combining two propositions p
and q to a compound proposition “if p, then q.” It is also written as p  q and can be read as “p
implies q.” In the conditional statement p  q, p is called the premise (or antecedent or
hypothesis) and q is called the conclusion (or consequence). The conditional statement p  q is
false when the premise p is true and the conclusion q is false, and true otherwise. It is important
to note that conditional statements should not be interpreted in terms of cause and effect. Thus,
when we say “if p, then q,” we do not mean that the premise p causes the conclusion q, but only
that when p is true, q must be true as well1.
1
This is different from the if-then construction used in many programming languages. Most programming
languages contain statements such as if p then S, where p is a proposition and S is a program segment of
Ivan Marsic

Rutgers University
188
Table 3-1: Operators of the propositional and predicate logics.
Propositional Logic

conjunction (p and q)

implication (if p then q)

disjunction (p or q)

biconditional (p if and only if q)

negation (not p)

equivalence (p is equivalent to q)
Predicate Logic (extends propositional logic with two quantifiers)

universal quantification (for all x, P(x))

existential quantification (there exists x, P(x))
The statement p  q is a biconditional, or bi-implication, which means that p  q and q  p.
The biconditional statement p  q is true when p and q have the same truth value, and is false
otherwise.
So far we have considered propositional logic; now let us briefly introduce predicate logic. We
saw ealier that the sentence “x is smaller than 5” is not a proposition because it is neither true nor
false. This sentence has two parts: the variable x, which is the subject, and the predicate, “is
smaller than 5,” which refers to a property that the subject of the sentence can have. We can
denote this statement by P(x), where P denotes the predicate “is smaller than 5” and x is the
variable. The sentence P(x) is also said to be the value of the propositional function P at x. Once a
specific value has been assigned to the variable x, the statement P(x) becomes a proposition and
has a truth value. In our example, by setting x = 3, P(x) is true; conversely, by setting x = 7, P(x)
is false2.
There is another way of creating a proposition from a propositional function, called
quantification. Quantification expresses the extent to which a predicate is true over a range of
elements, using the words such as all, some, many, none, and few. Most common types of
quantification are universal quantification and existential quantification, shown at the bottom of
Table 3-1.
The universal quantification of P(x) is the proposition “P(x) is true for all values of x in the
domain,” denoted as x P(x). The value of x for which P(x) is false is called a counterexample of
x P(x). The existential quantification is the proposition “There exists a value of x in the domain
such that P(x) is true,” denoted as x P(x).
In constructing valid arguments, a key elementary step is replacing a statement with another
statement of the same truth value. We are particularly interested in compound propositions
formed from propositional variables using logical operators as given in Table 3-1. Two types of
compound propositions are of special interest. A compound proposition that is always true,
regardless of the truth values of its constituent propositions is called a tautology. A simple
example is p  p, which is always true because either p is true or it is false. On the other hand, a
compound proposition that is always false is called a contradiction. A simple example is p  p,
because p cannot be true and false at the same time. Obviously, the negation of a tautology is a
one or more statements to be executed. When such an if-then statement is encountered during the
execution of a program, S is executed is p is true, but S is not executed if p is false.
2
The reader might have noticed that we already encountered predicates in Section 3.1.2 where the state
relations for objects actually are predicates.
Chapter 3 
Modeling and System Specification
189
contradiction, and vice versa. Finally, compound proposition that is neither a tautology nor a
contradiction is called a contingency.
The compound propositions p and q are said to be logically equivalent, denoted as p  q, if p  q
is a tautology. In other words, p  q if p and q have the same truth values for all possible truth
values of their component variables. For example, the statements r  s and r  s are logically
equivalent, which can be shown as follows. Earlier we stated that a conditional statement is false
only when its premise is true and its conclusion is false, and true otherwise. We can write this as
r  s  (r  s)
 r  (s)
 r  s
by the first De Morgan’s law: (p  q)  p  q
[For the sake of completeness, I state here, as well, the second De Morgan’s law: (p  q) 
pq.]
Translating sentences in natural language into logical expressions is an essential part of
specifying systems. Consider, for example, the following requirements in our second case study
on financial investment assistant (Section 1.3.2).
Example 3.4
Translating Requirements into Logical Expressions
Translate the following two requirements for our second case study on personal investment assistant
(Table 2-2) into logical expressions:
REQ1. The system shall support registering new investors by providing a real-world email, which
shall be external to our website. Required information shall include a unique login ID and a
password that conforms to the guidelines, as well as investor’s first and last name and other
demographic information. Upon successful registration, the system shall set up an account with
a zero balance for the investor.
REQ2. The system shall support placing Market Orders specified by the action (buy/sell), the stock to
trade, and the number of shares. The current indicative (ask/bid) price shall be shown and
updated in real time. The system shall also allow specifying the upper/lower bounds of the
stock price beyond which the investor does not wish the transaction executed. If the action is to
buy, the system shall check that the investor has sufficient funds in his/her account. When the
market order matches the current market price, the system shall execute the transaction
instantly. It shall then issue a confirmation about the outcome of the transaction (known as
“order ticket”), which contains: the unique ticket number, investor’s name, stock symbol,
number of shares, the traded share price, the new portfolio state, and the investor’s new
account balance.
We start by listing all the declarative sentences that can be extracted from the requirements. REQ1
yields the following declarative sentences. Keep in mind that these are not necessarily propositions
because we still do not know whether they have truth value.
Label
a
b
c
d
e
f
g
h
Declarative sentence (not necessarily a proposition!)
The investor can register with the system
The email address entered by the investor exists in real world
The email address entered by the investor is external to our website
The login ID entered by the investor is unique
The password entered by the investor conforms to the guidelines
The investor enters his/her first and last name, and other demographic info
Registration is successful
Account with zero balance is set up for the investor
Ivan Marsic

Rutgers University
190
Next we need to ascertain their truth value. Recall that the specifications state what is true about the
system at the time it is delivered to the customer. The truth value of a must be established by the
developer before the system is delivered. The truth values of b, c, d, and e depends on what the
investor will enter. Hence, these are propositional functions at investor’s input. Consider the sentence
b. Assuming that email denotes the investor’s input and B denotes the predicate in b, the propositional
function is B(email). Similarly, c can be written as C(email), d as D(id), and e as E(pwd). The system
can and should evaluate these functions at runtime, during the investor registration, but the
specification refers to the system deployment time, not its runtime. I will assume that the truth of
sentence f is hard to ascertain so the system will admit any input values and consider f true.
We have the following propositions derived from REQ1:
REQ1 represented as a set of propositions
a
( email)( id)( pwd) [B(email)  C(email)  D(id)  E(pwd)  g]
f
gh
The reader should be reminded that conditional statements in logic are different from if-then
constructions in programming languages. Hence, g  h does not describe a cause-effect sequence of
instructions such as: when registration is successful, do set up a zero-balance account. Rather, this
simply states that when g is true, h must be true as well.
The system correctly implements REQ1 for an assignment of truth values that makes all four
propositions true. Note that it would be wrong to simply write (b  c  d  e)  g instead of the
second proposition above, for this does not correctly reflect the reality of user choice at entering the
input parameters.
Extracting declarative sentences from REQ2 is a bit more involved than for REQ1. The two most
complex aspects of REQ2 seem to be about ensuring the sufficiency of funds for the stock purchase
and executing the order only if the current price is within the bounds (in case the trader specified the
upper/lower bounds). Let us assume that the ticker symbol selected by the trader is denoted by SYM
and its current ask price at the exchange is IP (for indicative price). Note that unlike the email and
password in REQ1, here we can force the user to select a valid ticker symbol by displaying only
acceptable options. The number of shares (volume) for the trade specified by the investor is denoted as
VOL. In case the investor specifies the upper/lower bounds, let their values be denoted as UB and LB,
respectively. Lastly, the investor’s current account balance is denoted as BAL.
Here is a partial list of propositions needed to state these two constraints:
Label Propositions (partial list)
m
The action specified by the investor is “buy”
n
The investor specified the upper bound of the “buy” price
o
The investor specified the lower bound of the “sell” price
The above table contains propositions because their truth value can be established independent of the
user’s choice. For example, the developer should allow only two choices for trading actions, “buy” or
“sell,” so m means that the investor selected “sell.” In case the investor specifies the upper/lower
bounds, the system will execute the transaction only if [n  m  (IP  UB)]  [o  m  (LB  IP)].
To verify that the investor’s account balance is sufficient for the current trade, the system needs to
check that [n  (VOL  IP  BAL)]  [n  (VOL  UB  BAL)].
The additional declarative sentences extracted from REQ2 are:
Label Propositions (they complete the above list)
p
The investor requests to place a market order
q
The investor is shown a blank ticket where the trade can be specified (action, symbol, etc.)
r
The most recently retrieved indicative price is shown in the currently open order ticket
s
The symbol SYM specified by the investor is a valid ticker symbol
t
The current indicative price that is obtained from the exchange
Chapter 3 
u
v
w
x
Modeling and System Specification
191
The system executes the trade
The system calculates the player’s account new balance
The system issues a confirmation about the outcome of the transaction
The system archives the transaction
We have the following propositions derived from REQ2:
REQ2 represented as a set of propositions
pqr
s
y = v  {(n  o)  [(o  p  o  q)  ( IP)(LB  IP  UB)]}
z = m  {[n  (VOL  IP  BAL)]  [n  (VOL  UB  BAL)]}
yzu
uvwx
Again, all of the above propositions must evaluate to true for the system to correctly implement REQ2.
Unlike REQ1, we have managed to restrict the user choice and simplify the representation of REQ2. It
is true that by doing this we went beyond mere problem statement and imposed some choices on the
problem solution, which is generally not a good idea. But in this case I believe these are very simple
and straightforward choices. It requires the developer’s judgment and experience to decide when
simplification goes too far into restricting the solution options, but sometimes the pursuit of purity only
brings needless extra work.
System specifications should be consistent, which means that they should not contain conflicting
requirements. In other words, if the requirements are represented as a set of propositions, there
should be an assignment of truth values to the propositional variables that makes all requirements
propositions true.
Example…
In Section 3.2.3 we will see how logic plays role in the part of the UML standard called Object
Constraint Language (OCL). Another notation based on Boolean logic is TLA+, described in
Section 3.2.4.
Finite State Machines
The behavior of complex objects and systems depends not only on their immediate input, but also
on the past history of inputs. This memory property, represented as a state, allows such systems
to change their actions with time. A simple but important formal notation for describing such
systems is called finite state machines (FSMs). FSMs are used extensively in computer science
and data networking, and the UML standard extends the FSMs into UML state machine diagrams
(Section 3.2.2).
There are various ways to represent a finite state machine. One way is to make a table showing
how each input affects the state the machine is in. Here is the state table for the door lock used in
our case-study example
Present state
Ivan Marsic

Rutgers University
192
lock
unlock
Closed
unlock
Open
lock
Figure 3-10: State transition diagram for a door lock.
Armed
Disarmed
lock
Armed
Armed
Input
unlock Disarmed Disarmed
Here, the entries in the body of the table show the next state the machine enters,
depending on the present state (column) and input (row).
We can also represent our machine graphically, using a transition diagram, which is a
directed graph with labeled edges. In this diagram, each state is represented by a circle.
Arrows are labeled with the input for each transition. An example is shown in Figure
3-10. Here the states “Disarmed” and “Armed” are shown as circles, and labeled arrows
indicate the effect of each input when the machine is in each state.
A finite state machine is formally defined to consist of a finite set of states S, a finite set
of inputs I, and a transition function with S  I as its domain and S as its codomain (or
range) such that if s  S and i I, the f(s, i) is the state the machine moves to when it is
in state s and is given input i. Function f can be a partial function, meaning that it can be
undefined for some values of its domain. In certain applications, we may also specify an
initial state s0 and a set of final (or accepting) states S  S, which are the states we
would like the machine to end in. Final states are depicted in state diagrams by using
double concentric circles. An example is shown in Figure 3-11, where M =
maxNumOfAttempts is the final state: the machine will halt in this state and needs to be
restarted externally.
A string is a finite sequence of inputs. Given a string i1i2 … in and the initial state s0, the
machine successively computes s1 = f(s0, i1), then s2 = f(s1, i2), and so on, finally ending up with
state sn. For the example in Figure 3-11, the input string iiv transitions the FSM through the states
s0s1s2s0. If sn  S, i.e., it is an accepting state, then we say that the string is accepted; otherwise it
is rejected. It is easy to see that in Figure 3-11, the input string of M i’s (denoted as iM) will be
accepted. We say that this machine recognizes this string and, in this sense, it recognizes the
Start
v
i
0
v = (input-key  Valid-keys)
i = (input-key  Valid-keys)
M = maxNumOfAttempts
i
1
2
M
v
v
Figure 3-11: State transition diagram for the counter of unsuccessful lock-opening attempts.
Chapter 3 
193
Modeling and System Specification
lock
(a)
unlock / beep
Closed
unlock
Open
lock / beep
lock
(b)
unlock [key  Valid-keys] / beep
Closed
unlock
Open
lock / beep
Figure 3-12: State transition diagram from Figure 3-10, modified to include output labels
(a) and guard labels (b).
attempted intrusion.
A slightly more complex machine is an FSM that yields output when it transitions to the next
state. Suppose that, for example, the door lock in Figure 3-10 also produces an audible signal to
let the user know that it is armed or disarmed. The modified diagram is shown in Figure 3-12(a).
We use a slash symbol to separate the input and the output labels on each transition arrow. (Note
that here we choose to produce no outputs when the machine receives duplicate inputs.)
We define a finite state machine with output to consist of a finite set of states S, a finite set of
inputs I, a finite set of outputs O, along with a function f : S  I  S that assigns to each (state,
input) pair a new state and another function g : S  I  O that assigns to each (state, input) pair
an output.
We can enrich the original FSM model by adding new features. Figure 3-12(b) shows how we
can add guards to transitions. The full notation for transition descriptions is then
input[guard]/output, where each element is optional. A guard is a Boolean proposition that
permits or blocks the transition. When a guard is present, the transition takes place if the guard
evaluates to true, but the transition is blocked if the guard is false. Section 3.2.2 describes how
UML adds other features to extend the FSM model into UML state machine diagrams.
3.2.2
UML State Machine Diagrams
One of the key weaknesses of the original finite-state-machines model (described in the preceding
section) in the context of system and software specification is the lack of modularization
mechanisms. When considering the definitions of states and state variables in Section 3.1.2,
FSMs are suitable for representing individual simple states (or microstates). UML state machine
diagrams provide a standardized diagrammatic notation for state machines and also incorporate
extensions, such as macrostates and concurrent behaviors.
Ivan Marsic

Rutgers University
194
trade
IPO
planned
initial-listing
bankruptcy,
acquisition,
merger, …
Traded
Delisted
IPO = initial public offering
Figure 3-13: UML state machine diagram showing the states and transitions of a stock.
Basic Notation
In every state machine diagram, there must be exactly one default initial state, which we
designate by writing an unlabeled transition to this state from a special icon, shown as a filled
circle. An example is shown in Figure 3-13. Sometimes we also need to designate a stop state. In
most cases, a state machine associated with an object or the system as a whole never reaches a
stop state—the state machine just vanishes when the object it represents is destroyed. We
designate a stop state by drawing an unlabeled state transition from this state to a special icon,
shown as a filled circle inside a slightly larger hollow circle.3 Initial and final states are called
pseudostates.
Transitions between pairs of states are shown by directed arrows. Moving between states is
referred to as firing the transition. A state may transition to itself, and it is common to have many
different state transitions from the same state. All transitions must be unique, meaning that there
will never be any circumstances that would trigger more than one transition from the same state.
There are various ways to control the firing of a transition. A transition with no annotation is
referred to as a completion transition. This simply means that when the object completes the
execution of an activity in the source state, the transition automatically fires, and the target state is
entered.
In other cases, certain events have to occur for the transition to fire. Such events are annotated on
the transition. (Events were discussed is Section 3.1.3.) In Figure 3-13, one may argue that
bankruptcy or acquisition phenomena should be considered states rather than events, because
company stays in bankruptcy for much longer than an instant of time. The correct answer is
relative to the observer. Our trader would not care how long the company will be in bankruptcy—
the only thing that matters is that its stock is not tradable anymore starting with the moment the
bankruptcy becomes effective.
We have already seen for FSMs that a guard condition may be specified to control the transition.
These conditions act as guards so that when an event occurs, the condition will either allow the
transition (if the condition is true) or disallow the transition (if the condition is false).
State Activities: Entry, Do, and Exit Activities
I already mentioned that states can be passive or active. In particular, an activity may be specified
to be carried out at certain points in time with respect to a state:
3
The Delisted state in Figure 3-13 is the stop state with respect to the given exchange. Although investors
can no longer trade shares of the stock on that exchange, it may be traded on some other markets
Chapter 3 
195
Modeling and System Specification
view
archive
Pending
submit
InPreparation
do: check_price+supply [buy]
check_price+demand [sell]
matched
Executed
Archived
cancel,
reject
data
Entry
trade
Cancelled
Figure 3-14: Example of state activities for a trading order. Compare with Figure 3-7.
Traded
trade
IPO
planned
initiallisting
trade
trade
trade
Buy
trade
Hold
trade
composite state
nested
state
Sell
bankruptcy,
acquisition,
merger, …
Delisted
trade
trade
Figure 3-15: Example of composite and nested states for a stock. Compare with Figure 3-13.
 Perform an activity upon entry of the state
 Do an activity while in the state
 Perform an activity upon exit of the state
An example is shown in Figure 3-14.
Composite States and Nested States
UML state diagrams define superstates (or macrostates). A superstate is a complex state that is
further refined by decomposition into a finite state machine. A superstate can also be obtained by
aggregation of elementary states, as already seen in Section 3.1.2.
Suppose now that we wish to extend the diagram in Figure 3-13 to show the states Buy, Sell, and
Hold, which we defined in Example 3.2. These states are a refinement of the Traded state within
which they are nested, as shown in Figure 3-15. This nesting is depicted with a surrounding
boundary known as a region and the enclosing boundary is called a composite state. Given the
composite state Traded with its three substates, the semantics of nesting implies an exclusive OR
(XOR) relationship. If the stock is in the Traded state (the composite state), it must also be in
exactly one of the three substates: Buy, Hold, or Sell.
Nesting may be to any depth, and thus substates may be composite states to other lower-level
substates. For simplicity in drawing state transition diagrams with depth, we may zoom in or
zoom out relative to a particular state. Zooming out conceals substates, as in Figure 3-13, and
Ivan Marsic

Rutgers University
196
Figure 3-16: Example of concurrency in states.
zooming in reveals substates, as in Figure 3-15. Zoomed out representation may improve
comprehensibility of a complex state machine diagram.
Concurrency
Figure 3-16
Applications
State machine diagrams are typically used to describe the behavior of individual objects.
However, they can also be used to describe the behavior of any abstractions that the developer is
currently considering. We may also provide state machine diagrams for the entire system under
consideration. During the analysis phase of the development lifecycle (described in Section 2.5),
we are considering the event-ordered behavior of the system as a whole; hence, we may use state
machine diagrams to represent the behavior of the system. During the design phase (described in
Section 2.6), we may use state machine diagrams to capture dynamic behavior of individual
classes or of collaborations of classes.
In Section 3.3 we will use state machine diagrams to describe problem domains when trying to
understand and decompose complex problems into basic problems.
3.2.3
UML Object Constraint Language (OCL)
“I can speak French but I cannot understand it.” —Mark Twain
The UML standard defines Object Constraint Language (OCL) based on Boolean logic. Instead
of using mathematical symbols for operators (Table 3-1), OCL uses only ASCII characters which
makes it easier for typing and computer processing. It also makes OCL a bit wordy in places.
OCL is not a standalone language, but an integral part of the UML. An OCL expression needs to
be placed within the context of a UML model. In UML diagrams, OCL is primarily used to write
constraints in class diagrams and guard conditions in state and activity diagrams. OCL
expressions, known as constraints, are added to express facts about elements of UML diagrams.
Chapter 3 
197
Modeling and System Specification
Table 3-2: Basic predefined OCL types and operations on them.
Type
Boolean
Integer
Real
String
Values
true, false
1, 48, 3, 84967, …
0.5, 3.14159265, 1.e+5
'With more exploration comes more text.'
Operations
and, or, xor, not, implies, if-then-else
*, , , /, abs()
*, , , /, floor()
concat(), size(), substring()
Any implementation derived from such a design model must ensure that each of the constraints
always remains true.
We should keep in mind that for software classes there is no notion of a computation to specify in
the sense of having well-defined start and end points. A class is not a program or subroutine.
Rather, any of object’s operations can be invoked at arbitrary times with no specific order. And
the state of the object can be an important factor in its behavior, rather than just input-output
relations for the operation. Depending on its state, the object may act differently for the same
operation. To specify the effect of an operation on object’s state, we need to be able to describe
the present state of the object which resulted from any previous sequence of operations invoked
on it. Because object’s state is captured in its attributes and associations to other objects, OCL
constraints usually relate to these properties of objects.
OCL Syntax
OCL’s syntax is similar to object-oriented languages such as C++ or Java. OCL expressions
consist of model elements, constraints, and operators. Model elements include class attributes,
operations, and associations. However, unlike programming languages OCL is a pure
specification language, meaning that an OCL expression is guaranteed to be without side effects.
When an OCL expression is evaluated, it simply returns a value. The state of the system will
never change because of the evaluation of an OCL expression, even though an OCL expression
can be used to specify a state change, such as in a post-condition specification.
OCL has four built-in types: Boolean, Integer, Real, and String. Table 3-2 shows example values
and some examples of the operations on the predefined types. These predefined value types are
independent of any object model and are part of the definition of OCL.
When writing an OCL contract, the first step is to decide the context, which is the software class
for which the OCL expression is applicable. Within the given class context, the keyword self
refers to all instances of the class. Other model elements can be obtained by navigating using the
dot notation from the self object. Consider the example of the class diagram in Figure 2-35
(Section 2.6). To access the attribute numOfAttempts_ of the class Controller, we write
self.numOfAttempts_
Due to encapsulation, object attributes frequently must be accessed via accessor methods. Hence,
we may need to write self.getNumOfAttempts().
Ivan Marsic

Rutgers University
(a) Local attribute
198
(b) Directly related class
Class_A
– attribute1
– attribute2
– …
Class_A
(c) Indirectly related class
Class_A
*
assocBA
*
assocBA
*
assocAB
*
assocAB
Class_B
Class_B
*
assocCB
*
assocBC
Class_C
Figure 3-17: Three basic types of navigation in a UML class diagram. (a) Attributes of class
A accessed from an instance of class A. (b) Accessing a set of instances of class B from an
instance of class A. (c) Accessing a set of instances of class C from an instance of class A.
Starting from a given context, we can navigate associations on the class diagram to refer to other
model elements and their properties. The three basic types of navigation are illustrated in Figure
3-17. In the context of Class_A, to access its local attribute, we write self.attribute2.
Similarly, to access instances of a directly associated class we use the name of the opposite
association-end in the class diagram. So in Figure 3-17(b), in the context of Class_A, to access
the set of instances of Class_B, we write self.assocAB. Lastly in Figure 3-17(c), in the
context of Class_A, to access instances of an indirectly associated class Class_C, we write
self.assocAB.assocBC. (This approach should not come as a surprise to the reader familiar
with an object programming language, such as Java or C#.)
We already know from UML class diagrams that object associations may be individual objects
(association multiplicity equals 1) or collections (association multiplicity  1). Navigating a oneto-one association yields directly an object. Figure 2-35 shows a single LockCtrl (assuming
that a single lock device is controlled by the system). Assuming that this association is named
lockCtrl_ as in Listing 2.2, the navigation self.lockCtrl_ yields the single object
lockCtrl_ : LockCtrl. However, if the Controller were associated with multiple
locks, e.g., on front and backyard doors, then this navigation would yield a collection of two
LockCtrl objects.
OCL specifies three types of collections:

OCL sets are used to collect the results of navigating immediate associations with one-tomany multiplicity.

OCL sequences are used when navigating immediate ordered associations.

OCL bags are used to accumulate the objects when navigating indirectly related objects.
In this case, the same object can show up multiple times in the collection because it was
accessed via different navigation paths.
Note that in the example in Figure 3-17(c), the expression self.assocAB.assocBC
evaluates to the set of all instances of class Class_C objects associated with all instances of
class Class_B objects that, in turn, are associated with class Class_A objects.
Chapter 3 
Modeling and System Specification
199
Table 3-3: Summary of OCL operations for accessing collections.
OCL Notation
Meaning
EXAMPLE OPERATIONS ON ALL OCL COLLECTIONS
c->size()
Returns the number of elements in the collection c.
c->isEmpty()
Returns true if c has no elements, false otherwise.
c1->includesAll(c2)
Returns true if every element of c2 is found in c1.
c1->excludesAll(c2)
Returns true if no element of c2 is found in c1.
c->forAll(var | expr) Returns true if the Boolean expression expr true for all
elements in c. As an element is being evaluated, it is bound
to the variable var, which can be used in expr. This
implements universal quantification .
c->forAll(var1, var2
Same as above, except that expr is evaluated for every
| expr)
possible pair of elements from c, including the cases where
the pair consists of the same element.
c->exists(var | expr) Returns true if there exists at least one element in c for
which expr is true. This implements existential
quantification .
c->isUnique(var |
Returns true if expr evaluates to a different value when
expr)
applied to every element of c.
c->select(expr)
Returns a collection that contains only the elements of c for
which expr is true.
EXAMPLE OPERATIONS SPECIFIC TO OCL SETS
s1->intersection(s2)
Returns the set of the elements found in s1 and also in s2.
s1->union(s2)
Returns the set of the elements found either s1 or s2.
s->excluding(x)
Returns the set s without object x.
EXAMPLE OPERATION SPECIFIC TO OCL SEQUENCES
seq->first()
Returns the object that is the first element in the sequence
seq.
To distinguish between attributes in classes from collections, OCL uses the dot notation for
accessing attributes and the arrow operator -> for accessing collections. To access a property of a
collection, we write the collection’s name, followed by an arrow ->, and followed by the name of
the property. OCL provides many predefined operations for accessing collections, some of which
are shown in Table 3-3.
Constants are unchanging (non-mutable) values of one of the predefined OCL types (Table 3-2).
Operators combine model elements and constants to form an expression.
OCL Constraints and Contracts
Contracts are constraints on a class that enable the users of the class, implementers, and
extenders to share the same assumptions about the class. A contract specifies constraints on the
class state that must be valid always or at certain times, such as before or after an operation is
invoked. The contract is between the class implementer about the promises of what can be
Ivan Marsic

Rutgers University
200
expected and the class user about the obligations that must be met before the class is used. There
are three types of constraints in OCL: invariants, preconditions, and postconditions.
One important characterization of object states is describing what remains invariant throughout
the object’s lifetime. This can be described using an invariant predicate. An invariant must
always evaluate to true for all instance objects of a class, regardless of what operation is invoked
and in what order. An invariant applies to a class attribute.
In addition, each operation can be specified by stating a precondition and a postcondition. A
precondition is a predicate that is checked before an operation is executed. A precondition
applies to a specific operation. Preconditions are frequently used to validate input parameters to
an operation.
A postcondition is a predicate that must be true after an operation is executed. A postcondition
also applies to a specific operation. Postconditions are frequently used to describe how the
object’s state was changed by an operation.
We already encountered some preconditions and postconditions in the context of domain models
(Section 2.5.4). Subsequently, in Figure 2-35 we assigned the domain attributes to specific
classes. Therein, we used an informal, ad-hoc notation. OCL provides a formal notation for
expressing constraints. For example, one of the constraints for our case study system is that the
maximum allowed number of failed attempts at disarming the lock is a positive integer. This
constraint must be always true, so we state it as an invariant:
context Controller inv:
self.getMaxNumOfAttempts() > 0
Here, the first line specifies the context, i.e., the model element to which the constraint applies, as
well as the type of the constraint. In this case the inv keyword indicates the invariant constraint
type. In most cases, the keyword self can be omitted because the context is clear.
Other possible types of constraint are precondition (indicated by the pre keyword) and
postcondition (indicated by the post keyword). A precondition for executing the operation
enterKey() is that the number of failed attempts is less than the maximum allowed number:
context Controller::enterKey(k : Key) : boolean pre:
self.getNumOfAttempts()  self.getMaxNumOfAttempts()
The postconditions for enterKey()are that (Poc1) a failed attempt is recorded, and (Poc2) if
the number of failed attempts reached the maximum allowed number, the system becomes
blocked and the alarm bell is sounded. The first postcondition (Poc1) can be restated as:
(Poc1) If the provided key is not element of the set of valid keys, then the counter of failed
attempts after exiting from enterKey() must be by one greater than its value before
entering enterKey().
The above two postconditions (Poc1) and (Poc2) can be expressed in OCL as:
context Controller::enterKey(k : Key) : Boolean
-- postcondition (Poc1):
post: let allValidKeys : Set = self.checker.validKeys()
if allValidKeys.exists(vk | k = vk) then
Chapter 3 
Modeling and System Specification
201
getNumOfAttempts() = getNumOfAttempts()@pre
else
getNumOfAttempts() = getNumOfAttempts()@pre + 1
-- postcondition (Poc2):
post: getNumOfAttempts() >= getMaxNumOfAttempts() implies
self.isBlocked() and self.alarmCtrl.isOn()
There are three features of OCL used in stating the first postcondition above that the reader
should note. First, the let expression allows one to define a variable (in this case
allValidKeys of the OCL collection type Set) which can be used in the constraint.
Second, the @pre directive indicates the value of an object as it existed prior to the operation.
Hence,
getNumOfAttempts()@pre
denotes
the
value
returned
by
getNumOfAttempts()before
invoking
enterKey(),
and
getNumOfAttempts()denotes the value returned by the same operation after invoking
enterKey().
Third, the expressions about getNumOfAttempts() in the if-then-else operation are
not assignments. Recall that OCL is not a programming language and evaluation of an OCL
expression will never change the state of the system. Rather, this just evaluates the equality of the
two sides of the expression. The result is a Boolean value true or false.
SIDEBAR 3.1: The Dependent Delegate Dilemma
 The class invariant is a key concept of object-oriented programming, essential for reasoning
about classes and their instances. Unfortunately, the class invariant is, for all but non-trivial
examples, not always satisfied. During the execution of the method that client object called on
the server object (“dependent delegate”), the invariant may be temporarily violated. This is
considered acceptable because in such an intermediate state the server object is not directly
usable by the rest of the world—it is busy executing the method that client called—so it does
not matter that its state might be inconsistent. What counts is that the invariant will hold before
and after the execution of method calls.
However, if during the executing of the server’s method the server calls back a method on the
client, then the server may catch the client object in an inconsistent state. This is known as the
dependent delegate dilemma and is difficult to handle. The interested reader should check
[Meyer, 2005] for more details.
The OCL standard specifies only contracts. Although not part of the OCL standard, nothing
prevents
us
from
specifying
program
behavior
using
Boolean
logic.
[ give example ]
3.2.4
TLA+ Notation
This section presents TLA+ system specification language, defined by Leslie Lamport. The book
describing TLA+ can be downloaded from http://lamport.org/. There are many other specification
languages, and TLA+ reminds in many ways of Z (pronounced Zed, not Zee) specification
Ivan Marsic
1
2
3
4
5
6
7
8
9
10
11

Rutgers University
202
AccessController
The set of valid keys.
A ValidateKey(k) step checks if k is a valid key.
MODULE
validKeys,
ValidateKey( _ )
ASSUME validKeys  STRING
ASSUME  key  STRING : ValidateKey(key)  BOOLEAN
VARIABLE status
TypeInvariant ̂ status  [lock : {“disarmed”, “armed”}, bulb : {“lit”, “unlit”}]
CONSTANTS
Init ̂  TypeInvariant
 status.lock = “armed”
 status.bulb = “unlit”
The initial predicate.
12
13
14
Unlock(key) ̂  ValidateKey(key)
 status.lock = “disarmed”
 status.bulb = “lit”
Only if the user enters a valid key, then
unlock the lock and
turn on the light (if not already lit).
15
16
Lock ̂  status.lock = “armed”
 UNCHANGED status.bulb
Anybody can lock the doors
but not to play with the lights.
17
18
19
20
21
22
Next ̂ Unlock(key)  Lock
The next-state action.
Spec ̂ Init  [Next]status
The specification.
THEOREM
Spec 
TypeInvariant
Type correctness of the specification.
Figure 3-18: TLA+ specification of the cases study system.
language. My reason for choosing TLA+ is that it uses the language of mathematics, specifically
the language of Boolean algebra, rather than inventing another formalism.
A TLA+ specification is organized in a module, as in the following example, Figure 3-18, which
specifies our home access case study system (Section 1.3.1). Observe that TLA+ language
reserved words are shown in SMALL CAPS and comments are shown in a highlighted text. A
module comprises several sections

Declaration of variables, which are primarily the manifestations of the system visible to
an outside observer

Definition of the behavior: the initial state and all the subsequent (next) states, which
combined make the specification

The theorems about the specification
The variables could include internal, invisible aspects of the system, but they primarily address
the external system’s manifestations. In our case-study of the home access controller, the
variables of interest describe the state of the lock and the bulb. They are aggregated in a single
status record, lines 6 and 7.
The separator lines 8 and 20 are a pure decoration and can be omitted. Unlike these, the module
start and termination lines, lines 1 and 22, respectively, have semantic meaning and must appear.
Chapter 3 
Modeling and System Specification
203
Lines 2 and 3 declare the constants of the module and lines 4 and 5 list our assumptions about
these constants. For example, we assume that the set of valid passwords is a subset of all
character strings, symbolized with STRING. Line 5 essentially says that we expect that for any key
k, ValidateKey(k) yields a BOOLEAN value.
TypeInvariant in line 7 specifies all the possible values that the system variable(s) can assume in
a behavior that satisfies the specification. This is a property of a specification, not an assumption.
That is why it is stated as a theorem at the end of the specification, line 21.
The definition of the initial system state appears in lines 9 and 10.
Before defining the next state in line 17, we need to define the functions that could be requested
of the system. In this case we focus only on the key functions of disarming and arming the lock,
Disarm and Arm, respectively, and ignore the rest (see all the use cases in Section 2.2). Defining
these functions is probably the most important part of a specification.
The variable status with an apostrophe symbol represents the state variable in the next step, after
an operation takes place.
3.3 Problem Frames
“Computers are useless. They can only give you answers.” —Pablo Picasso
“Solving a problem simply means representing it so as to make the solution transparent.”
—Herbert Simon, The Sciences of the Artificial
Problem frames were proposed by Michael Jackson [1995; 2001] as a way for understanding and
systematic describing the problem as a first step towards the solution. Problem frames decompose
the original complex problem into simple, known subproblems. Each frame captures a problem
class stylized enough to be solved by a standard method and simple enough to present clearly
separated concerns.
We have an intuitive feeling that a problem of data acquisition and display is different from a
problem of text editing, which in turn is different from writing a compiler that translates source
code to machine code. Some problems combine many of these simpler problems. The key idea of
problem frames is to identify the categories of simple problems, and to devise a methodology for
representing complex problems in terms of simple problems.
There are several issues to be solved for successful formulation of a problem frame methodology.
First we need to identify the frame categories. One example is the information frame, which
represents the class of problems that are primarily about data acquisition and display. We need to
define the notation to be used in describing/representing the frames. Then, given a complex
problem, we need to determine how to decompose it into a set of problem frames. Each individual
frame can then be considered and solved independently of other frames. A key step in solving a
frame is to address the frame concerns, which are generic aspects of each problem type that need
to be addressed for solving a problem of a particular type.
Ivan Marsic

(a)
Rutgers University
The
Machine
204
a
Problem
Domain
b
The
Requirement
Domain properties
seen by the requirement
(b)
The
Machine
Specification
a
Problem
Domain
Domain properties
seen by the machine
b
The
Requirement
Requirement
a: specification interface phenomena
b: requirement interface phenomena
Figure 3-19: (a) The Machine and the Problem Domain. (b) Interfaces between the problem
domain, the requirements and the machine.
Finally, we need to determine how to compose the individual solutions into the overall solution
for the original problem. We need to determine how individual frames interact with each other
and we may need to resolve potential conflicts of their interaction.
3.3.1
Problem Frame Notation
We can picture the relationship between the computer system to be developed and the real world
where the problem resides as in Figure 3-19. The task of software development is to construct the
Machine by programming a general-purpose computer. The machine has an interface a consisting
of a set of phenomena—typically events and states—shared with the Problem Domain. Example
phenomena are keystrokes on a computer keyboard, characters and symbols shown on a computer
screen, signals on the lines connecting the computer to an instrument, etc.
The purpose of the machine is described by the Requirement, which specifies that the machine
must produce and maintain some relationship among the phenomena of the problem domain. For
example, to disarm the lock device when a correct code is presented, or to ensure that the figures
printed on a restaurant check correctly reflect the patron’s consumption.
Phenomena a shared by a problem domain and the machine are called specification phenomena.
Conversely, phenomena b articulate the requirements and are called the requirement phenomena.
Although a and b may be overlapping, they are generally distinct. The requirement phenomena
are the subject matter of the customer’s requirement, while the specification phenomena describe
the interface at which the machine can monitor and control the problem domain.
A problem diagram as in Figure 3-19 provides a basis for problem analysis because it shows you
what you are concerned with, and what you must describe and reason about in order to analyze
the problem completely. The key topics of your descriptions will be:
Chapter 3 
Modeling and System Specification
205
 The requirement that states what the machine must do. The requirement is what your customer
would like to achieve in the problem domain. Its description is optative (it describes the
option that the customer has chosen). Sometimes you already have an exact description of the
requirement, sometimes not. For example, requirement REQ1 given in Table 2-2 states
precisely how users are to register with our system.
 The domain properties that describe the relevant characteristics of each problem domain. These
descriptions are indicative because they describe what is true regardless of the machine’s
behavior. For example, Section 1.3.2 describes the functioning of financial markets, which
we must understand to implement a useful system that will provide investment advice.
 The machine specification. Like the requirement, this is an optative description: it describes the
machine’s desired behavior at its interfaces with the problem domain.
Obviously, the indicative domain properties play a key role: without a clear understanding of how
financial markets work we would never be able to develop a useful investment assistant system.
3.3.2
Problem Decomposition into Frames
Problem analysis relies on a strategy of problem
decomposition based on the type of problem domain and
the domain properties. The resulting subproblems are
treated independently of other subproblems, which is the
basis of effective separation of concerns. Each
subproblem has its own machine (specification), problem
domain(s), and requirement. Each subproblem is a
projection of the full problem, like color separation in
printing, where colors are separated independently and
then overlaid (superimposed) to form the full picture.
Jackson [2001] identifies five primitive problem frames,
which serve as the basic units of problem decomposition.
These are (i) required behavior, (ii) commanded
behavior, (iii) information display, (iv) simple
workpieces, and (v) transformation. They differ in their
requirements, domain characteristics, domain involvement (whether the domain is controlled,
active, inert, etc.), and the frame concern. These problem frames correspond to the problem types
identified earlier in Section 2.3.1 (see Figure 2-11).
Each frame has a particular concern, which is a set of generic issues that need to be solved when
solving the frame:
(a) Required behavior frame concern: To describe precisely (1) how the controlled domain
currently behaves; (2) the desired behavior for the domain, as stated by the requirement;
and, (3) what the machine (software-to-be) will be able to observe about the domain state,
by way of the sensors that will be used in the system-to-be.
(b) Commanded behavior frame concern: To identify (1) all the commands that will be
possible in the envisioned system-to-be; (2) the commands that will be supported or
Ivan Marsic

Rutgers University
Control
Machine
CM!C1
CD!C2
206
Controlled
Domain
C3
C
Required
Behavior
Figure 3-20: Problem frame diagram for the required behavior frame.
allowed under different scenarios; and, (3) what should happen if the user tries to execute
a command that is not supported/allowed under the current scenario.
(c) Information display frame concern: To identify (1) the information that the machine will
be able to observe from the problem domain, by way of the sensors that will be used in
the system-to-be; (2) the information that needs to be displayed, as stated by the
requirement; and, (3) the transformations needed to process the raw observed information
to obtain displayable information.
(d) Simple workpieces frame concern: To describe precisely (1) the data structures of the
workpieces; (2) all the commands that will be possible in the envisioned system-to-be;
(3) the commands that will be supported or allowed under different scenarios; and, (4)
what should happen if the user tries to execute a command that is not supported/allowed
under the current scenario.
(e) Transformation frame concern: To describe precisely (1) the data structures of the input
data and output data; (2) how each data structure will be traversed (travelled over); and,
(3) how each element of the input data structure will be transformed to obtain the
corresponding element in the output data structure.
Identification and analysis of frame flavors, reflecting a finer classification of domain properties
The frame concern is to make the requirement, specification, and domain descriptions and to fit
them into a correctness argument for the machine to be built. Frame concerns include:
initialization, overrun, reliability, identities, and completeness. The initialization concern is to
ensure that a machine is properly synchronized with the real world when it starts.
… frame variants, in which a domain is usually added to a problem frame
Basic Frame Type 1: Required Behavior
In this scenario, we need to build a machine which controls the behavior of a part of the physical
world according to given conditions.
Figure 3-20 shows the frame diagram for the required behavior frame. The control machine is the
machine (system) to be built. The controlled domain is the part of the world to be controlled. The
requirement, giving the condition to be satisfied by the behavior of the controlled domain, is
called the required behavior.
The controlled domain is a causal domain, as indicated by the C in the bottom right corner of its
box. Its interface with the machine consists of two sets of causal phenomena: C1, controlled by
the machine, and C2, controlled by the controlled domain. The machine imposes the behavior on
the controlled domain by the phenomena C1; the phenomena C2 provide feedback.
Chapter 3 
207
Modeling and System Specification
Control
Machine
Broker
Machine
Controlled
Domain
a
Stock
Exchange
Required
Behavior
b
C
a: BM! {Execute[i]} [C1]
Order
handling
rules
b: SE! {Place[i], Cancel[i],
Executed[i], Expired[i]} [C3]
SE! {PriceQuotes, Ack[i]} [C2]
Figure 3-21: Example of a Required Behavior basic frame: handling of trading orders.
ET!E1
WP!Y2
Work
pieces
X
Y3
Editing
tool
Command
effects
US!E3
E3
User
B
Figure 3-22: Problem frame diagram for the simple workpieces frame.
An example is shown in Figure 3-21 for how a stock-broker’s system handles trading orders.
Once the user places an order, the order is recorded in the broker’s machine and from now on the
machine monitors the quoted prices to decide when the conditions for executing the order are
met. When the conditions are met, e.g., the price reaches a specified value, the controlled domain
(stock exchange) is requested to execute the order. The controlled domain will execute the order
and return an acknowledgement, known as “order ticket.”
Basic Frame Type 2: Commanded Behavior
In this scenario, we need to build a machine which allows an operator to control the behavior of a
part of the physical world by issuing commands.
Basic Frame Type 3: Information Display
In this scenario, we need to build a machine which acquires information about a part of the
physical world and presents it at a given place in a given form.
Basic Frame Type 4: Simple Workpieces
In this scenario, we need to build a machine which allows a user to create, edit, and store some
data representations, such as text or graphics. The lexical domain that will be edited may be
relatively simple to design, such as text document for taking notes. It may also be very complex,
such as creating and maintaining a “social graph” on a social networking website. A videogame is
another example of a very complex digital (lexical) domain that is edited as the users play and
issue different commands.
Figure 3-22 shows the frame diagram for the simple workpieces frame.
Ivan Marsic

Rutgers University
208
Workpieces
a
Editing tool
Trading
order
X
c
Trading
machine
Command
effects
Order
placing rules
b
d
Trader
B
User
a: TM! {Create[i]}
[E1]
b: TR! {PriceQuotes, Place[i]} [Y2]
c: TR! {Place[i], Cancel[i],
Executed[i], Expired[i]} [Y3]
Figure 3-23: Example of a Simple Workpieces basic frame: placing a trading order.
An example is shown in Figure 3-23.
Basic Frame Type 5: Transformation
In this scenario, we need to build a machine takes an input document and produces an output
document according to certain rules, where both input and output documents may be formatted
differently. For example, given the records retrieved from a relational database, the task is to
render them into an HTML document for Web browser display.
A key concern for a transformation frame problem is to define the order in which the data
structures of the input data and output data will be traversed and their elements accessed. For
example, if the input data structure is a binary tree, then it can be traversed in pre-order, in-order,
or post-order manner.
Figure 3-24 shows the key idea behind the frame decomposition. Given a problem represented as
a complex set of requirements relating to a complex application domain, our goal is to represent
the problem using a set of basic problem frames.
3.3.3
Composition of Problem Frames
Real-world problems almost always consist of combinations of simple problem frames. Problem
frames help us achieve understanding of simple subproblems and derive solutions (machines) for
these problem frames. Once the solution is reached at the local level of specialized frames, the
integration (or composition) or specialized understanding is needed to make a coherent whole.
There are some standard composite frames, consisting of compositions of two or more simple
problem frames.
Chapter 3 
209
Modeling and System Specification
Domain
1
The
Requirements
Domain
5
The
Machine
nt
rem
e
qui
ma
in
3
m
ire
qu 4
e
R
Do
ma
in
4
Ma
ch i
n
3 e
Machine
2
hine
1
Mac
t
en
Do
ain
Dom
1
Domain
2
Re
e nt
uirem
Req 1
(b)
Domain
3
Requirement
2
Domain
4
Domain
2
3
(a)
n
ai
m
Do 5
e
hin
ac 4
M
Figure 3-24: The goal of frame decomposition is to represent a complex problem (a) as a set
of basic problem frames (b).
3.3.4
Models
Software system may have world representation and this is always idealized. E.g., in our lock
system, built-in (as opposed to runtime sensed/acquired) knowledge is: IF valid key entered AND
sensing dark THEN turn the light on.
Ivan Marsic

Rutgers University
210
3.4 Specifying Goals
“Correctness is clearly the prime quality. If a system does not do what it is supposed to do, then everything
else about it matters little.” —Bertrand Meyer
The basic idea of goal-oriented requirements engineering is to start with the aggregate goal of the
whole system, and to refine it by successive steps into a goal hierarchy.
AND-OR refinements …
Problem frames can be related to goals. Goal-oriented approach distinguishes different kinds of
goal, as problem-frames approach distinguishes different problem classes. Given a problem
decomposition into basic frames, we can restate this as an AND-refinement of the goal hierarchy:
to satisfy the system requirement goal, it is necessary to satisfy each individual subgoal (of each
basic frame).
When programmed, the program “knows” its goals implicitly rather than explicitly, so it cannot
tell those to another component. This ability to tell its goals to others is important in autonomic
computing, as will be seen in Section 9.3.
State the goal as follows: given the states A=armed, B=lightOff, C=user positively identified,
D=daylight
(Goal is the equilibrium state to be reached after a perturbation.)
Initial state: AB, goal state: AB.
Possible actions: —setArmed; 1—setDisarmed; —setLit; 1—setUnlit
Preconditions for 1: C; for : D
We need to make a plan to achieve AB by applying the permitted actions.
Program goals, see also “fuzzy” goals for multi-fidelity algorithms, MFAs, [Satyanarayanan &
Narayanan, 2001]. http://www.cs.yale.edu/homes/elkin/ (Michael Elkin)
The survey “Distributed approximation,” by Michael Elkin. ACM SIGACT News, vol. 36, no. 1,
(Whole Number 134), March 2005. http://theory.lcs.mit.edu/~rajsbaum/sigactNewsDC.html
The purpose of this formal representation is not to automatically build a program; rather, it is to
be able to establish that a program meets its specification.
Chapter 3 
Modeling and System Specification
211
3.5 Summary and Bibliographical Notes
This chapter presents …
People often complain about software quality (for example Microsoft products). The issue of
software quality is complex one. Software appeal depends on what it does (functionality), how
good it is (quality), and what it costs (economy). Different people put different weights on each of
these, but in the end all three matter. Microsoft figured that the functionality they deliver is
beyond the reach of smeller software vendors who cannot produce it at a competitive price, so
they emphasized functionality. It paid off. It appears that the market has been more interested in
low-cost, feature-laden products than reliability (for the mass market kind of products). It worked
in the market, thus far, which is the ultimate test. Whether this strategy will continue to work, we
do not know. But the tradeoff between quality / functionality / economy will always be present.
Also see the virtues if the “good enough” principle extolled here:
S. Baker, “Why ‘good enough’ is good enough: Imperfect technology greases innovation—and
the whole marketplace,” Business Week, no. 4048, p. 48, September 3, 2007. Online at:
http://www.businessweek.com/magazine/content/07_36/b4048048.htm
Comment
Formal specifications have had lack of success, usually blamed on non-specialists finding such
specifications difficult to understand, see e.g., [Sommerville, 2004, p. 116; Larman, 2005, p. 65].
The usual rationale given for avoiding rigorous, mathematically driven program development is
the time-to-market argument—rigor takes too much time and that cannot be afforded in today’s
world. We are also told that such things make sense for developing safety-critical applications,
such as hospital systems, or airplane controls, but not for everyday use. Thanks to this
philosophy, we can all enjoy Internet viruses, worms, spam, spyware, and many other inventions
that are thriving on lousy programs.
The problem, software ends up being used for purposes that it was not intended for. Many of-theshelf software products end up being used in mission-critical operations, regardless of the fact
that they lack robustness necessary to support such operations.
It is worth noticing that often we don’t wear what we think is “cool”—we often wear what the
“trend setters” in our social circle, or society in general, wear [Gladwell, 2000]. But, as Petroski
[1992], echoing Raymond Loewy, observes, it has to be MAYA—most advanced, yet acceptable.
So, if hackers let the word out that some technique is cool, it shall become cool for the masses of
programmers.
Bibliographical Notes
Much of this chapter is directly inspired by the work of Michael Jackson [1995; 2001]. I have
tried to retell it in a different way and relate it to other developments in software engineering. I
Ivan Marsic

Rutgers University
212
hope I have not distorted his message in the process. In any case, the reader would do well by
consulting the original sources [Jackson, 1995; 2001].
This chapter requires some background in discrete mathematics. I tried to provide a brief
overview, but the reader may need to check a more comprehensive source. [Rosen, 2007] is an
excellent introduction to discrete mathematics and includes very nice pieces on logic and finite
state machines.
[Guttenplan, 1986]
[Woodcock & Loomes, 1988]
J. P. Bowen and M. G. Hinchey, “Ten commandments of formal methods… ten years later,”
IEEE Computer, vol. 39, no. 1, pp. 40-48, January 2006.
The original sources for problem frames are [Jackson, 1995; 2001]. The reader can also find a
great deal of useful information online at: http://www.ferg.org/pfa/ .
Problem 3.7: Elevator Control given below is based on the classical elevator problem, which first
appeared in Donald Knuth’s book, The Art of Computer Programming: Vol. 1, Fundamental
Algorithms. It is based on the single elevator in the mathematics building at the California
Institute of Technology, where Knuth was a graduate student. Knuth used the elevator problem to
illustrate co-routines in an imaginary computing machine, called MIX. A detailed discussion of
software engineering issues in elevator control is available in [Jackson, 2001].
Problems
Problem 3.1
Problem 3.2
Consider a system consisting of a button and two light bulbs, as shown in the figure. Assume that
the system starts from the initial state where both bulbs are turned off. When the button is pressed
the first time, one of the bulbs will be lit and the other remains unlit. When the button is pressed
the second time, the bulb which is currently lit will be turned off and the other bulb will be lit.
When the button is pressed the third time, both bulbs will be lit. When the button is pressed the
fourth time, both bulbs will be turned off. For the subsequent button presses, the cycle is repeated.
Chapter 3 
Modeling and System Specification
213
Name and describe all the states and events in this system. Draw the UML state diagram and be
careful to use the correct symbols.
Problem 3.3
Consider the auto-locking feature of the case study of the home access-control system. In Section
2.4 this feature is described via use cases (a timer is started when the doors are unlocked and if it
counts down to zero, the doors will be automatically locked).
Suppose now that you wish to represent the auto-locking subsystem using UML state diagrams.
The first step is to identify the states and events relevant for the auto-locking subsystem. Do the
following:
(a) Name and describe the states that adequately represent the auto-locking subsystem.
(b) Name and describe the events that cause the auto-locking subsystem to transition between
the states.
(Note: You do not need to use UML notation to draw a state diagram, just focus on identifying
the states and events.)
Problem 3.4
Suppose that in the virtual mitosis lab (described at the book website, given in Preface), you are
to develop a finite state machine to control the mechanics of the mitosis process. Write down the
state transition table and show the state diagram. See also Problem 2.21.
Problem 3.5
Consider the grocery inventory management system that uses Radio Frequency Identification
(RFID), described in Problem 2.15 (Chapter 2). Identify the two most important entities of the
software-to-be and represent their states using UML state diagrams. Do the following:
(a) List and describe the states that adequately represent the two most important entities
(b) List and describe the events that cause the entities to transition between the states
(c) Draw the UML state diagrams for both entities
Note: Consider all the requirements REQ1 – REQ7.
Problem 3.6

Ivan Marsic
Rutgers University
214
Problem 3.7: Elevator Control
Consider developing a software system to control an elevator in a building. Assume that there
will be a button at each floor to summon the elevator, and a set of buttons inside the elevator
car—one button per floor to direct the elevator to the corresponding floor. Pressing a button will
be detected as a pulse (i.e., it does not matter if the user keeps holding the button pressed). When
pressed, the button is illuminated. At each floor, there will be a floor sensor that is “on” when the
elevator car is within 10 cm of the rest position at the floor.
There will be an information panel above the elevator doors on each floor, to show waiting
people where the elevator car is at any time, so that they will know how long they can expect to
wait until it arrives.
The information panels will have two lamps representing each floor (see the figure below). A
square lamp indicates that the car is at the corresponding floor, and a round lamp indicates that
there is a request outstanding for the elevator to visit the corresponding floor. In addition, there
will be two arrow-shaped lamps to indicate the current direction of travel. For example, in the
figure below, the panel indicates that the elevator car is currently on the fifth floor, going up, and
there are outstanding requests to visit the lobby, third, fourth, and sixth floor.
After the elevator visits a requested floor, the corresponding lamp on all information panels
should be turned off. Also, the button that summoned the elevator to the floor should be turned
off.
Let us assume that the outstanding requests are served so that the elevator will first visit all the
requested floors in the direction to which it went first after the idle state. After this, it will serve
the requests in the opposite direction, if any. When the elevator has no requests, it remains at its
current floor with its doors closed.
L
2
Down
3
4
5
Requests outstanding to visit the floors
6 7
Up
L
2
Down
3
4
5
6
7
Up
Floor where the elevator currently is
Direction of travel
Button to summon the elevator to this floor
Suppose that you already have designed UML interaction and class diagrams. Your system will
execute in a single thread, and your design includes the following classes:
ElevatorMain: This class runs an infinite loop. During each iteration it checks the physical
buttons whether any has been pressed and reads the statuses of all the floor sensors. If a
button has been pressed or the elevator car arrived/departed a floor, it calls the appropriate
classes to do their work, and then starts a new iteration.
CarControl:
This
class
controls
the
movement
of
the
elevator
car.
This class has the attribute requests that lists the outstanding requests for the elevator to
visit the corresponding floors. It also has three operations:
addRequest(floorNum : int) adds a request to visit the floor floorNum;
Chapter 3 
Modeling and System Specification
215
stopAt(floorNum : int) requests the object to stop the car at the floor floorNum.
This operation calls DoorControl.operateDoors() to open the doors, let the
passengers in or out, and close the doors.
When operateDoors() returns, the CarControl object takes this as a signal that it is
safe to start moving the car from the current floor (in case there are no pending requests, the
car remains at the current floor).
InformationPanel: This class controls the display of information on the elevator
information panel. It also has the attribute requests and these operations:
arrivedAt(floorNum : int) informs the software object that the car has arrived at
the floor floorNum.
departed() which informs the object that the car has departed from the current floor.
OutsideButton: This class represents the buttons located outside the elevator on each floor
that serve to summon the elevator. The associated physical button should be illuminated
when pressed and turned off after the car visits the floor.
This class has the attribute illuminated that indicates whether the button was pressed. It
also has two operations:
illuminate() requests the object to illuminate the associated physical button (because it
was pressed);
turnOff()requests the object to turn off the associated physical button (because the
elevator car has arrived at this floor).
InsideButton: This class represents the buttons located inside the elevator car that serve to
direct the elevator to the corresponding floor. The associated physical button should be
illuminated when pressed and turned off after the car visits the floor. It has the same attributes
and operations as the class OutsideButton.
DoorControl: This class controls opening and closing of the elevator doors on each floor.
This class has the Boolean attribute doorsOpen that is set true when the associated doors
are open and false otherwise. It also has the operation:
operateDoors() : void tells the software object when to open the doors. This
operation sets a timer for a given amount of time to let the passengers in or out; after the
timer expires, the operation closes the doors automatically and returns.
Note that some classes may have multiple instances (software objects), because there are multiple
corresponding physical objects. For example, there is an information panel, outside button, and
doors at each floor. In addition, we do not have a special class to represent a floor sensor that
senses when the elevator car is in or near the rest position at the floor. The reason for this choice
is that this system is single-threaded and the ElevatorMain object will notify the interested
objects about the floor sensor status, so there is no reason to keep this information in a class
dedicated solely for this purpose.
Draw the interaction and class diagrams corresponding to the design described above.
Problem 3.8
Consider the class diagram for an online auction website given in Figure 2-48, and the system as
described in Problem 2.31 for which the solution is given on the back of this text. Suppose that
you want to specify a contract for the operation closeAuction(itemName : String) of
Ivan Marsic

Rutgers University
216
the class Controller. To close auction on an item means that no new bids will be accepted;
the item is offered to the current highest bidder. If this bidder fails to pay within the specified
time interval, the auction may be reopened.
You want to specify the preconditions that the auction for the item itemName is currently open
and the item is not reserved. The postconditions should state that the auction is closed, and the
item is reserved to the name of the highest bidder, given that there was at least one bidder. Write
this contract as statements in OCL.
You may add more classes, attributes, or operations, if you feel that this is necessary to solve the
problem, provided that you justify your modification.
Problem 3.9
Problem 3.10
Problem 3.11
Consider the automatic patient monitoring system described in Problem 2.3. Solve the
following:
(a) Identify the problem frames and describe the frame concerns for each frame.
(b) Draw the state diagram for different subsystems (problem frames). Define each
state and event in the diagrams.
(c) Explain if the system needs to behave differently when it reports abnormal vital
signs or device failures. If yes, incorporate this behavior into your state diagrams.
Problem 3.12
Derive the domain model for the patient monitoring system from Problem 3.11.
(a) Write a definition for each concept in your domain model.
(b) Write a definition for each attribute and association in your domain model.
(c) Draw the domain model.
(d) Indicate the types of concepts, such as «boundary», «control», or «entity».
Note that you are not asked to derive the use cases for this system (see Problem 2.14). The
description of the system behavior that you will generate in the solution of Problem 3.11 should
suffice for deriving its domain model.
Problem 3.13
Chapter 4
Software Measurement and Estimation
Contents
“What you measure improves.”
—Donald Rumsfeld, Known and Unknown: A Memoir
Measurement is a process by which numbers or symbols are
assigned to properties of objects. To have meaningful
assignment of numbers, it must be governed by rules or theory
(or, model). There are many properties of software that can be
measured. Similarities can be drawn with physical objects: we
can measure height, width, weight, chemical composition,
etc., properties of physical objects. The numbers obtained
through such measurement have little value by themselves—
their key value is relative to something we want to do with
those objects. For example, we may want to know the weight
so we can decide what it takes to lift an object. Or, knowing
physical dimensions helps us decide whether the object will fit
into a certain space. Similarly, software measurement is
usually done with purpose. A common purpose is for
management decision making. For example, the project
manager would like to be able to estimate the development
cost or the time it will take to develop and deliver a software
product. Similar to how knowing the object weight helps us to
decide what it takes to lift it, the hope is that by measuring
certain software properties we will be able to estimate the
necessary development effort.
4.1 Fundamentals of Measurement Theory
4.1.1 Measurement Theory
4.1.2
4.1.3
4.2 What to Measure?
4.2.1 Cyclomatic Complexity
4.2.2 Use Case Points
4.2.3
4.2.4
4.3 Measuring Module Cohesion
4.3.1 Internal Cohesion or Syntactic Cohesion
4.3.2 Semantic Cohesion
4.3.3
4.3.4
4.2.3
4.4 Psychological Complexity
4.4.1 Algorithmic Information Content
4.4.2
4.4.3
4.4.4
4.5
4.5.1
4.5.2
4.5.3
4.5.4
4.6 Summary and Bibliographical Notes
Problems
Uses of software measurements:

Estimation of cost and effort (preferably early in the lifecycle)

Feedback to improve the quality of design and implementation
Obviously, once a software product is already completed, we know how much effort it took to
complete it. The invested effort is directly known, without the need for inferring it indirectly via
some other properties of the software. However, that is too late for management decisions.
Management decisions require knowing (or estimating) effort before we start with the
development, or at least early enough in the process, so we can meaningfully negotiate the budget
and delivery terms with the customer.
217
Ivan Marsic

Rutgers University
218
Therefore, it is important to understand correctly what measurement is about:
Measured property
 [ model for estimation ] 
(e.g., number of functional features)
Estimated property
(e.g., development effort required)
Notice also that we are trying to infer properties of one entity from properties of another entity:
the entity the properties of which are measured is software (design documents or code) and the
entity the properties of which are estimated is development process (people’s effort). The
“estimation model” is usually based on empirical evidence; that is, it is derived based on
observations of past projects. For past projects, both software and process characteristics are
known. From this, we can try to calculate the correlation of, say, the number of functional
features to, say, the development effort required. If correlation is high across a range of values,
we can infer that the number of functional features is a good predictor of the development effort
required. Unfortunately, we know that correlation does not equal causation. A causal model,
which not only establishes a relationship, but also explains why, would be better, if possible to
have.
Feedback to the developer is based on the knowledge of “good” ranges for software modules and
systems: if the measured attributes are outside of “good” ranges, the module needs to be
redesigned. It has been reported based on many observations that maintenance costs run to about
70 % of all lifetime costs of software products. Hence, good design can not only speed up the
initial development, but can significantly affect the maintenance costs.
Most commonly measured characteristics of software modules and systems are related to its size
and complexity. Several software characteristics were mentioned in Section 2.5, such as coupling
and cohesion, and it was argued that “good designs” are characterized by “low coupling” and
“high cohesion.” In this chapter I will present some techniques for measuring coupling and
cohesion and quantifying the quality of software design and implementation. A ubiquitous size
measure is the number of lines of code (LOC). Complexity is readily observed as an important
characteristic of software products, but it is difficult to operationalize complexity so that it can be
measured.
taking a well-reasoned, thoughtful approach that goes beyond the simplest correlative
relationships between the most superficial details of a problem.
Although it is easy to agree that more complex software is more difficult to develop and maintain,
it is difficult to operationalize complexity so that it can be measured. The reader may already be
familiar with computational complexity measure big O (or big Oh), O(n). O(n) measures software
complexity from the machine’s viewpoint in terms of how the size of the input data affects an
algorithm’s usage of computational resources (usually running time or memory). However, the
kind of complexity measure that we need in software engineering should measure complexity
form the viewpoint of human developers.
4.1 Fundamentals of Measurement Theory
Chapter 4 
219
Software Measurement and Estimation
“It is better to be roughly right than precisely wrong.” —John Maynard Keynes
The Hawthorne effect - an increase in worker productivity produced by the psychological
stimulus of being singled out and made to feel important. The Hawthorne effect describes a
temporary change to behavior or performance in response to a change in the environmental
conditions. This change is typically an improvement. Others have broadened this definition to
mean that people’s behavior and performance change following any new or increased attention.
Individual behaviors may be altered because they know they are being studied was demonstrated
in a research project (1927–1932) of the Hawthorne Works plant of the Western Electric
Company in Cicero, Illinois.
Initial improvement in a process of production caused by the obtrusive observation of that
process. The effect was first noticed in the Hawthorne plant of Western Electric. Production
increased not as a consequence of actual changes in working conditions introduced by the plant's
management but because management demonstrated interest in such improvements (related: selffulfilling hypothesis).
Measurement Theory
7 8
9 0 1
2 3
4
5
Measurement theory is a branch of applied mathematics.
The specific theory we use is called the representational
theory of measurement. It formalizes our intuitions about
the way the world actually works.
2 3
4 5
6
4.1.1
Measurement theory allows us to use statistics and probability to understand quantitatively the
possible variances, ranges, and types of errors in the data.
Measurement Scale
In measurement theory, we have five types of scales: nominal, ordinal, interval, ratio, and
absolute.
In nominal scale we can group subjects into different categories. For example, we designate the
weather condition as “sunny,” “cloudy,” “rainy,” or “snowy.” The two key requirements for the
categories are: jointly exhaustive and mutually exclusive. Mutually exclusive means a measured
attribute can be classified into one and only one category. Jointly exhaustive means that all
categories together should cover all possible values of the attribute. If the measured attribute has
more categories than we are interested in, an “other” category can be introduced to make the
categories jointly exhaustive. Provided that categories are jointly exhaustive and mutually
exclusive, we have the minimal conditions necessary for the application of statistical analysis. For
example, we may want to compare the values of software attributes such as defect rate, cycle
time, and requirements defects across the different categories of software products.
Ordinal scale refers to the measurement operations through which the subjects can be compared
in order. An example ordinal scale is: “bad,” “good,” and “excellent,” or “star” ratings used for
products or services on the Web. An ordinal scale is asymmetric in the sense that if A > B is true
then B > A is false. It has the transitivity property in that if A > B and B > C, then A > C. Although
ordinal scale orders subjects by the magnitude of the measured property, it offers no information
Ivan Marsic

Rutgers University
220
about the relative magnitude of the difference between subjects. For example, we only know that
“excellent” is better than “good,” and “good” is better than “bad.” However, we cannot compare
that the relative differences between the excellent-good and good-bad pairs. A commonly used
ordinal scale is an n-point Likert scale, such as the Likert five-point, seven-point, or ten-point
scales. For example, a five-point Likert scale for rating books or movies may assign the following
values: 1 = “Hated It,” 2 = “Didn’t Like It,” 3 = “Neutral,” 4 = “Liked It,” and 5 = “Loved It.”
We know only that 5 > 4, 4 > 3, 5 > 2, etc., but we cannot say how much greater is 5 than 4. Nor
can we say that the difference between categories 5 and 4 is equal to that between 3 and 2. This
implies that we cannot use arithmetic operations such as addition, subtraction, multiplication and
division. Nonetheless, the assumption of equal distance is often made and the average rating
reported (e.g., product rating at Amazon.com uses fractional values, such as 3.7stars).
Interval scale indicates the exact differences between measurement points. An interval scale
requires a well-defined, fixed unit of measurement that can be agreed on as a common standard
and that is repeatable. A good example is a traditional temperature scale (centigrade or Fahrenheit
scales). Although the zero point is defined in both scales, it is arbitrary and has no meaning. Thus
we can say that the difference between the average temperature in Florida, say 80F, and the
average temperature in Alaska, say 20F, is 60F, but we do not say that 80F is four times as hot
as 20F. The arithmetic operations of addition and subtraction can be applied to interval scale
data.
Ratio scale is an interval scale for which an absolute or nonarbitrary zero point can be located.
Absolute or true zero means that the zero point represents the absence of the property being
measured (e.g., no money, no behavior, none correct). Examples are mass, temperature in degrees
Kelvin, length, and time interval. Ratio scale is the highest level of measurement and all
arithmetic operations can be applied to it, including division and multiplication.
For interval and ratio scales, the measurement can be expressed in both integer and noninteger
data. Integer data are usually given in terms of frequency counts (e.g., the number of defects that
could be encountered during the testing phase).
Absolute scale is used when there is only one way to measure a property. It is independent of the
physical properties of any specific substance. In practice, values on an absolute scale are usually
(if not always) obtained by counting. An example is counting entities, such as chairs in a room.
Some Basic Measures
Ratio
Proportion
Percentage
Rate
Six Sigma
Chapter 4 
221
Software Measurement and Estimation








Figure 4-1: Issues with subjective size measures (compare to Figure 1-10). Left side of the
hedge as seen by a pessimist; right side seen by an optimist.
4.2 What to Measure?
Given a software artifact (design document or source code), generally we can measure
1. Attributes of any representation or description of a problem or a solution. Two main
categories of representations are structure vs. behavior.
2. Attributes of the development process or methodology.
Measured aspect:

quantity (size)

complexity
If the purpose of software measurement is estimation of cost and effort, we would like to measure
at an early stage in the software life-cycle. Typically a budget allocation is set at an early phase of
a procurement process and a decision on contract price made on these budget constraints and
Ivan Marsic

Rutgers University
222
suppliers’ tender responses. Consequently, the functional decomposition of the planned system
needs to be at a high level, but must be of sufficient detail to flush out as many of the implied
requirements and hidden complexities as possible, and as early as possible. In the ideal world this
would be a full and detailed decomposition of the use cases, but this is impractical during the
estimation process, because estimates need to be produced within tight time frames.
Figure 4-1
4.2.1
Use Case Points
Intuitively, projects with many complicated requirements take more effort to design and
implement than projects with few simple requirements. In addition, the effort depends not only on
inherent difficulty or complexity of the problem, but also on what tools the developers employ
and how skilled the developers are. The factors that determine the time to complete a project
include:

Functional requirements: These are often represented with use cases (Section 2.3). The
complexity of use cases, in turn, depends on the number and complexity of the actors and
the number of steps (transactions) to execute each use case.

Nonfunctional requirements: These describe the system’s nonfunctional properties,
known as FURPS+ (see Section 2.2.1), such as security, usability, and performance.
These are also known as the “technical complexity factors.”

Environmental factors: Various factors such as the experience and knowledge of the
development team, and how sophisticated tools they will be using for the development.
An estimation method that took into account the above factors early in a project’s life cycle, and
produced a reasonable accurate estimate, say within 20% of the actual completion time, would be
very helpful for project scheduling, cost, and resource allocation.
Because use cases are developed at the earliest or notional stages of system design, they afford
opportunities to understand the scope of a project early in the software life-cycle. The Use Case
Points (UCP) method provides the ability to estimate the person-hours a software project requires
based on its use cases. The UCP method analyzes the use case actors, scenarios, nonfunctional
requirements, and environmental factors and abstracts them into an equation. Detailed use case
descriptions (Section 2.3.3) must be derived before the UCP method can be applied. The UCP
method cannot be applied to sketchy use cases. As discussed in Section 2.3.1, we can apply user
story points (described in Section 2.2.3) for project effort estimation at this very early stage.
The formula for calculating UCP is composed of three variables:
1. Unadjusted Use Case Points (UUCP), which measures the complexity of the functional
requirements
2. The Technical Complexity Factor (TCF), which measures the complexity of the
nonfunctional requirements
3. The Environment Complexity Factor (ECF), which assesses the development team’s
experience and their development environment
Chapter 4 
223
Software Measurement and Estimation
Table 4-1: Actor classification and associated weights.
Actor type
Description of how to recognize the actor type
Weight
Simple
The actor is another system which interacts with our system through
a defined application programming interface (API).
1
Average
The actor is a person interacting through a text-based user interface,
or another system interacting through a protocol, such as a network
communication protocol.
2
Complex
The actor is a person interacting via a graphical user interface.
3
Each variable is defined and computed separately using weighted values, subjective values, and
constraining constants. The subjective values are determined by the development team based on
their perception of the project’s technical complexity and the team’s efficiency. Here is the
equation:
UCP = UUCP  TCF  ECF
(4.1)
Unadjusted Use Case Points (UUCPs) are computed as a sum of these two components:
1. The Unadjusted Actor Weight (UAW), based on the combined complexity of all the actors
in all the use cases.
2. The Unadjusted Use Case Weight (UUCW), based on the total number of activities (or
steps) contained in all the use case scenarios.
The computation of these components is described next.
Unadjusted Actor Weight (UAW)
An actor in a use case might be a person, another program, a piece of hardware, etc. The weight
for each actor depends on how sophisticated is the interface between the actor and the system.
Some actors, such as a user working with a text-based command-line interface, have very simple
needs and increase the complexity of a use case only slightly. Other actors, such as a user
working with a highly interactive graphical user interface, have a much more significant impact
on the effort to develop a use case. To capture these differences, each actor in the system is
classified as simple, average, or complex, and is assigned a weight as shown in Table 4-1. This
scale for rating actor complexity was devised by expert developers based on their experience.
Notice that this is an ordinal scale (Section4.1.1). You can think of this as a scale for “star
rating,” similar to “star ratings” of books (Amazon.com), films (IMDb.com), or restaurants
(yelp.com). Your task is, using this scale, to assign “star ratings” to all actors in your system. In
our case, we can assign one, two, or three “stars” to actors, corresponding to “Simple,”
“Average,” or “Complex” actors, respectively. Table 4-2 shows my ratings for the actors in the
case study of home access control, for which the actors are described in Section 2.3.1. The UAW
is calculated by totaling the number of actors in each category, multiplying each total by its
specified weighting factor, and then adding the products we obtain:
Ivan Marsic

Rutgers University
224
Table 4-2: Actor classification for the case study of home access control (see Section 2.3).
Actor name Description of relevant characteristics
Complexity Weight
Landlord is interacting with the system via a graphical user
Landlord
Complex
3
interface (when managing users on the central computer).
Tenant is interacting through a text-based user interface
(assuming that identification is through a keypad; for
Average
2
Tenant
biometrics based identification methods Tenant would be a
complex actor).
LockDevice is another system which interacts with our
LockDevice
Simple
1
system through a defined API.
LightSwitch Same as LockDevice.
Simple
1
AlarmBell
Same as LockDevice.
Simple
1
Database
Database is another system interacting through a protocol.
Average
2
Timer
Same as LockDevice.
Simple
1
Police
Our system just sends a text notification to Police.
Simple
1
UAW(home access) = 5  Simple  2  Average  1  Complex = 51  22  13 = 12
Unadjusted Use Case Weight (UUCW)
The UUCW is derived from the number of use cases in three categories: simple, average,
and complex (see Table 4-3). Each use case is categorized based on the number of steps
(or, transactions) within its event flow, including both the main success scenario and
alternative scenarios (extensions).
The number of steps in a scenario affects the estimate. A large number of steps in a use case
scenario will bias the UUCW toward complexity and increase the UCPs. A small number of steps
will bias the UUCW toward simplicity and decrease the UCPs. Sometimes, a large number of
steps can be reduced without affecting the business process.
The UUCW is calculated by tallying the number of use cases in each category, multiplying each
total by its specified weighting factor, and then adding the products. For example, Table 4-4
computes the UUCW for the sample case study.
There is a controversy on how to count alternate scenarios (extensions). Initially, it was suggested
to ignore all scenarios except the main success scenario. The current view is that extensions
represent a significant amount of work and need to be included in effort estimation. However, it is
not agreed upon how to do the inclusion. The problem is that you cannot simply count the number
of lines in an extension scenario and add those to the lines in the main success scenario.
Chapter 4 
225
Software Measurement and Estimation
Table 4-3: Use case weights based on the number of transactions.
Use case category Description of how to recognize the use-case category
Weight
Simple
Simple user interface. Up to one participating actor (plus
initiating actor). Number of steps for the success scenario:  3. If 5
presently available, its domain model includes  3 concepts.
Average
Moderate interface design. Two or more participating actors.
Number of steps for the success scenario: 4 to 7. If presently
available, its domain model includes between 5 and 10 concepts.
10
Complex
Complex user interface or processing. Three or more
participating actors. Number of steps for the success scenario:
7. If available, its domain model includes 10 concepts.
15
As seen in UC-7: AuthenticateUser (Section 2.3), each extension starts with a result of a
transaction, rather than a new transaction itself. For example, extension 2a (“Tenant/Landlord
enters an invalid identification key”) is the result of the transaction described by step 2 of the
main success scenario (“Tenant/Landlord supplies an identification key”). So, item 2a in the
extensions section of UC-7: AuthenticateUser is not counted. The same, of course, is true for 2b,
2c, and 3a. The transaction count for the use case in UC-7: AuthenticateUser is then ten. You may
want to count 2b1 and 2b2 only once but that is more effort than is worthwhile, and they may be
separate transactions sharing common text in the use case.
Another mechanism for measuring use case complexity is counting the concepts obtained by
domain modeling (Section 2.4). Of course, this assumes that the domain model is already derived
at the time the estimate is being made. The concepts can be used in place of transactions once it
has been determined which concepts model a specific use case. As indicated in Table 4-3, a
simple use case is implemented by 5 or fewer concepts, an average use case by 5 to 10 concepts,
and a complex use case by more than 10 concepts. The weights are as before. Each type of use
case is then multiplied by the weighting factor, and the products are added up to get the UUCW.
The UUCW is calculated by tallying the use cases in each category, multiplying each count by its
specified weighting factor (Table 4-3), and then adding the products:
UUCW(home access) = 1  Simple  5  Average  2  Complex = 15  510  215 = 85
The UUCP is computed by adding the UAW and the UUCW. Based on the scores in Table 4-2
and Table 4-4, the UUCP for our case study project is UUCP = UAW + UUCW = 12 + 85 = 97.
The UUCP gives the unadjusted size of the overall system, unadjusted because it does not
account for the nonfunctional requirements (TCFs) and the environmental factors (ECFs).
Ivan Marsic

Rutgers University
226
Table 4-4: Use case classification for the case study of home access control (see Section 2.3).
Use case
Description
Category Weight
Simple user interface. 5 steps for the main success
Unlock (UC-1)
scenario. 3 participating actors (LockDevice,
Average 10
LightSwitch, and Timer).
Simple user interface. 2+3=5 steps for the all
Lock (UC-2)
scenarios. 3 participating actors (LockDevice,
Average 10
LightSwitch, and Timer).
Complex user interface. More than 7 steps for the
ManageUsers
main success scenario (when counting UC-6 or
Complex 15
(UC-3)
UC-7). Two participating actors (Tenant, Database).
ViewAccessHistory Complex user interface. 8 steps for the main success
Complex 15
(UC-4)
scenario. 2 participating actors (Database, Landlord).
AuthenticateUser
Simple user interface. 3+1=4 steps for all scenarios.
Average 10
(UC-5)
2 participating actors (AlarmBell, Police).
Complex user interface. 6 steps for the main success
AddUser (UC-6)
scenario (not counting UC-3). Two participating
Average 10
actors (Tenant, Database).
Complex user interface. 4 steps for the main success
RemoveUser
scenario (not counting UC-3). One participating actor Average 10
(UC-7)
(Database).
Simple user interface. 2 steps for the main success
Login (UC-8)
Simple
5
scenario. No participating actors.
Technical Complexity Factor (TCF)—Nonfunctional
Requirements
Thirteen standard technical factors were identified (by expert developers) to estimate the impact
on productivity of the nonfunctional requirements for the project (see Table 4-5). Each factor is
weighted according to its relative impact.
The development team should assess the perceived complexity of each technical factor from
Table 4-5 in the context of their project. Based on their assessment, they assign another “star
rating,” a perceived complexity value between zero and five. The perceived complexity value
reflects the team’s subjective perception of how much effort will be needed to satisfy a given
nonfunctional requirement. For example, if they are developing a distributed system (factor T1 in
Table 4-5), it will require more skill and time than if developing a system that will run on a single
computer. A perceived complexity value of 0 means that a technical factor is irrelevant for this
project, 3 corresponds to average effort, and 5 corresponds to major effort. When in doubt, use 3.
Each factor’s weight (Table 4-5) is multiplied by its perceived complexity factor to produce the
calculated factor. The calculated factors are summed to produce the Technical Total Factor. Table
4-6 calculates the technical complexity for the case study.
Two constants are used with the Technical Total Factor to produce the TCF. The constants limit
the impact the TCF has on the UCP equation (4.1) from a range of 0.6 (when perceived
complexities are all zero) to a maximum of 1.3 (when perceived complexities are all five), see
Figure 4-2(a).
Chapter 4 
227
Software Measurement and Estimation
Table 4-5: Technical complexity factors and their weights.
Technical factor
Description
T1
Distributed system (running on multiple machines)
Performance objectives (are response time and throughput
T2
performance critical?)
T3
End-user efficiency
T4
Complex internal processing
T5
Reusable design or code
Easy to install (are automated conversion and installation
T6
included in the system?)
Easy to use (including operations such as backup, startup, and
T7
recovery)
T8
Portable
T9
Easy to change (to add new features or modify existing ones)
T10
Concurrent use (by multiple users)
T11
Special security features
Provides direct access for third parties (the system will be used
T12
from multiple sites in different organizations)
T13
Special user training facilities are required
()
(70, 1.3)
1.2
1
1
0.8
0.8
ECF
TCF
1
1
1
0.5
0.5
2
1
1
1
1
1
(0, 1.4)
1.4
1.2
(0, 0.6)
0.6
0.4
0.4
0.2
0.2
0
0
1()
Some sources assign 2 as the weight for the performance objectives factor (T2).
1.4
0.6
Weight
2
(32.5, 0.425)
0
10
20
30
40
50
60
70
80
0
10
20
30
Technical Factor Total
Environmental Factor Total
(a)
(b)
40
Figure 4-2: Scaling constants for technical and environmental factors.
TCF values less than one reduce the UCP because any positive value multiplied by a positive
fraction decreases in magnitude: 100  0.6 = 60 (a reduction of 40%). TCF values greater than
one increase the UCP because any positive value multiplied by a positive mixed number increases
in magnitude: 100  1.3 = 130 (an increase of 30%). The constants were determined by
interviews with experienced developers, based on their subjective estimates.
Because the constants limit the TCF from a range of 0.6 to 1.3, the TCF can impact the UCP
equation from 40% (0.6) to a maximum of 30% (1.3). The formula to compute the TCF is:
13
TCF = Constant-1  Constant-2  Technical Factor Total = C1  C 2   Wi  Fi
i 1
where,
(4.2)
Ivan Marsic

Rutgers University
228
Table 4-6: Technical complexity factors for the case study of home access (see Section 2.3).
Calculated Factor
Technical
Perceived
Description
Weight
(WeightPerceived
factor
Complexity
Complexity)
Distributed, Web-based system, because
T1
2
3
23 = 6
of ViewAccessHistory (UC-4)
Users expect good performance but
T2
1
3
13 = 3
nothing exceptional
End-user expects efficiency but there are
T3
1
3
13 = 3
no exceptional demands
T4
Internal processing is relatively simple
1
1
11 = 1
T5
No requirement for reusability
1
0
10 = 0
Ease of install is moderately important
T6
0.5
3
0.53 = 1.5
(will probably be installed by technician)
T7
Ease of use is very important
0.5
5
0.55 = 2.5
No portability concerns beyond a desire
2
2
T8
22 = 4
to keep database vendor options open
T9
Easy to change minimally required
1
1
11 = 1
T10
Concurrent use is required (Section 5.3)
1
4
14 = 4
T11
Security is a significant concern
1
5
15 = 5
T12
No direct access for third parties
1
0
10 = 0
T13
No unique training needs
1
0
10 = 0
Technical Factor Total: 31
Constant-1 (C1) = 0.6
Constant-2 (C2) = 0.01
Wi = weight of ith technical factor (Table 4-5)
Fi = perceived complexity of ith technical factor (Table 4-6)
Formula (4.2) is illustrated in Figure 4-2(a). Given the data in Table 4-6, the TCF = 0.6 + (0.01 
31) = 0.91. According to equation (4.1), this results in a reduction of the UCP by 9%.
Environment Complexity Factor (ECF)
The environmental factors (Table 4-7) measure the experience level of the people on the project
and the stability of the project. Greater experience will in effect reduce the UCP count, while
lower experience will in effect increase the UCP count. One might wish to consider other external
factors, such as the available budget, company’s market position, the state of the economy, etc.
The development team determines each factor’s perceived impact based on their perception the
factor has on the project’s success. A value of 1 means the factor has a strong, negative impact for
the project; 3 is average; and 5 means it has a strong, positive impact. A value of zero has no
impact on the project’s success. For factors E1-E4, 0 means no experience in the subject, 3 means
average, and 5 means expert. For E5, 0 means no motivation for the project, 3 means average, and
5 means high motivation. For E6, 0 means unchanging requirements, 3 means average amount of
change expected, and 5 means extremely unstable requirements. For E7, 0 means no part-time
technical staff, 3 means on average half of the team is part-time, and 5 means all of the team is
Chapter 4 
229
Software Measurement and Estimation
Table 4-7: Environmental complexity factors and their weights.
Environmental factor Description
E1
Familiar with the development process (e.g., UML-based)
E2
Application problem experience
E3
Paradigm experience (e.g., object-oriented approach)
E4
Lead analyst capability
E5
Motivation
E6
Stable requirements
E7
Part-time staff
E8
Difficult programming language
Weight
1.5
0.5
1
0.5
1
2
1
1
part-time. For E8, 0 means an easy-to-use programming language will be used, 3 means the
language is of average difficulty, and 5 means a very difficult language is planned for the project.
Each factor’s weight is multiplied by its perceived impact to produce its calculated factor. The
calculated factors are summed to produce the Environmental Factor Total. Larger values for the
Environment Factor Total will have a greater impact on the UCP equation. Table 4-8 calculates
the environmental factors for the case study project (home access control), assuming that the
project will be developed by a team of upper-division undergraduate students.
To produce the final ECF, two constants are computed with the Environmental Factor Total.
Similar to the TCF constants above, these constants were determined based on interviews with
expert developers. The constants constrain the impact the ECF has on the UCP equation from
0.425 (part-time workers and difficult programming language = 0, all other values = 5) to 1.4
(perceived impact is all 0). Therefore, the ECF can reduce the UCP by 57.5% and increase the
UCP by 40%, see Figure 4-2(b). The ECF has a greater potential impact on the UCP count than
the TCF. The formula is:
8
ECF = Constant-1  Constant-2  Environmental Factor Total = C1  C 2   Wi  Fi
(4.3)
i 1
where,
Constant-1 (C1) = 1.4
Constant-2 (C2) = 0.03
Wi = weight of ith environmental factor (Table 4-7)
Fi = perceived impact of ith environmental factor (Table 4-8)
Formula (4.3) is illustrated in Figure 4-2(b). Given the data in Table 4-8, the ECF = 1.4 +
(0.0311) = 1.07. For the sample case study, the team’s modest software development
experience resulted in an average EFT. All four factors E1-E4 scored relatively low. According to
equation (4.1), this results in an increase of the UCP by 7%.
Ivan Marsic

Rutgers University
230
Table 4-8: Environmental complexity factors for the case study of home access (Section 2.3).
Calculated Factor
Environmental
Perceived
Description
Weight
(Weight
factor
Impact
Perceived Impact)
Beginner familiarity with the UMLE1
1.5
1
1.51 = 1.5
based development
Some familiarity with application
E2
0.5
2
0.52 = 1
problem
Some knowledge of object-oriented
E3
1
2
12 = 2
approach
E4
Beginner lead analyst
0.5
1
0.51 = 0.5
Highly motivated, but some team
E5
1
4
14 = 4
members occasionally slacking
E6
Stable requirements expected
2
5
25 = 5
E7
No part-time staff will be involved
0
1
10 = 0
Programming language of average
3
E8
1
13 = 3
difficulty will be used
Environmental Factor Total: 11
Calculating the Use Case Points (UCP)
As a reminder, the UCP equation (4.1) is copied here:
UCP = UUCP  TCF  ECF
From the above calculations, the UCP variables have the following values:
UUCP = 97
TCF = 0.91
ECF = 1.07
For the sample case study, the final UCP is the following:
UCP = 97  0.91  1.07 = 94.45 or 94 use case points.
Note for the sample case study, the combined effect of TCF and ECF was to increase the UUCP
by approximately 3 percent (94/97100  100 = 3%). This is a minor adjustment and can be
ignored given that many other inputs into the calculation are subjective estimates.
Discussion of the UCP Metric
Notice that the UCP equation (4.1) is not consistent with measurement theory, because the counts
are on a ratio scale and the scores for the adjustment factors are on an ordinal scale (see Section
4.1.1). However, such formulas are often used in practice.
It is worth noticing that UUCW (Unadjusted Use Case Weight) is calculated simply by adding up
the perceived weights of individual use cases (Table 4-3). This assumes that all use cases are
completely independent, which usually is not the case. The merit of linear summation of size
measures was already discussed in Sections 1.2.5 and 2.2.3.
Chapter 4 
Software Measurement and Estimation
231
UCP appears to be based on a great deal of subjective and seemingly arbitrary parameters,
particularly the weighting coefficients. For all its imperfections, UCP has become widely adopted
because it provides valuable estimate early on in the project, when many critical decisions need to
be made. See the bibliographical notes (Section 4.7) for literature on empirical evidence about the
accuracy of UCP-based estimation.
UCP measures how “big” the software system will be in terms of functionality. The software size
is the same regardless of who is building the system or the conditions under which the system is
being built. For example, a project with a UCP of 100 may take longer than one with a UCP of
90, but we do not know by how much. From the discussion in Section 1.2.5, we know that to
calculate the time to complete the project using equation (1.2) we need to know the team’s
velocity. How to factor in the team velocity (productivity) and compute the estimated number of
hours will be described later in Section 4.6.
4.2.2
Cyclomatic Complexity
One of the most common areas of complexity in a program lies in complex conditional logic (or,
control flow). Thomas McCabe [1974] devised a measure of cyclomatic complexity, intended to
capture the complexity of a program’s conditional logic. A program with no branches is the least
complex; a program with a loop is more complex; and a program with two crossed loops is more
complex still. Cyclomatic complexity corresponds roughly to an intuitive idea of the number of
different paths through the program—the greater the number of different paths through a
program, the higher the complexity.
McCabe’s metric is based on graph theory, in which you calculate the cyclomatic number of a
graph G, denoted by V(G), by counting the number of linearly independent paths within a
program. Cyclomatic complexity is
V(G) = e  n  2
(4.4)
where e is the number of edges, n is the number of nodes.
Converting a program into a graph is illustrated in Figure 4-3. It follows that cyclomatic
complexity is also equal to the number of binary decisions in a program plus 1. If all decisions are
not binary, a three-way decision is counted as two binary decisions and an n-way case (select or
switch) statement is counted as n  1 binary decisions. The iteration test in a looping statement is
counted as one binary decision.
Ivan Marsic

Rutgers University
232
CODE
(a)
FLOWCHART
if expression1 then
statement2
else
statement3
end if
statement4
T
expr1
?
statm2
GRAPH
F
n1
statm3
n2
statm4
(b)
switch expr1
case 1:
statement2
case 2:
statm3
case 3:
statm4
end switch
statm5
1
expr1
?
2
statm3
statm2
n4
3
n1
statm4
statm5
statm1
do
(c)
statement1
while expr2
end do
statement3
T
n3
expr2
?
F
statm3
n2
n3
n4
n5
n1
n2
n3
Figure 4-3: Converting software code into an abstract graph.
The cyclomatic complexity is additive. The complexity of several graphs considered as a group is
equal to the sum of individual graphs’ complexities.
There are two slightly different formulas for calculating cyclomatic complexity V(G) of a graph
G. The original formula by McCabe [1974] is
V(G) = e  n  2p
(4.5)
where e is the number of edges, n is the number of nodes, and p is the number of connected
components of the graph. Alternatively, [Henderson-Sellers & Tegarden, 1994] propose a
linearly-independent cyclomatic complexity as
VLI(G) = e  n  p  1
(4.6)
Because cyclomatic complexity metric is based on decisions and branches, which is consistent
with the logic pattern of design and programming, it appeals to software professionals. But it is
not without its drawbacks. Cyclomatic complexity ignores the complexity of sequential
statements. In other words, any program with no conditional branching has zero cyclomatic
complexity! Also, it does not distinguish different kinds of control flow complexity, such as loops
vs. IF-THEN-ELSE statements or selection statements vs. nested IF-THEN-ELSE statements.
Cyclomatic complexity metric was originally designed to indicate a program’s testability and
understandability. It allows you to also determine the minimum number of unique tests that must
be run to execute every executable statement in the program. One can expect programs with
higher cyclomatic complexity to be more difficult to test and maintain, due to their higher
complexity, and vice versa. To have good testability and maintainability, McCabe recommends
that no program module should exceed a cyclomatic complexity of 10. Many software
Chapter 4 
Software Measurement and Estimation
233
refactorings are aimed at reducing the complexity of a program’s conditional logic [Fowler, 2000;
Kerievsky, 2005].
4.3 Measuring Module Cohesion
Cohesion is defined as a measure of relatedness or consistency in the functionality of a software
unit. It is an attribute that identifies to which degree the parts within that unit belong together or
are related to each other. In an object-oriented paradigm, a class can be a unit, the data can be
attributes, and the methods can be parts. Modules with high cohesion are usually robust, reliable,
reusable, and easy to understand while modules with low cohesion are associated with
undesirable traits such as being difficult to understand, test, maintain, and reuse. Cohesion is an
ordinal type of measurement and is usually expressed as “high cohesion” or “low cohesion.”
We already encountered the term cohesion in Chapter 2, where it was argued that each unit of
design, whether it is at the modular level or class level, should be focused on a single purpose.
This means that it should have very few responsibilities that are logically related. Terms such as
“intramodular functional relatedness” or “modular strength” have been used to address the notion
of design cohesion.
4.3.1
Internal Cohesion or Syntactic Cohesion
Internal cohesion can best be understood as syntactic cohesion evaluated by examining the code
of each individual module. It is thus closely related to the way in which large programs are
modularized. Modularization can be accomplished for a variety of reasons and in a variety of
ways.
A very crude modularization is to require that each module should not exceed certain size, e.g.,
50 lines of code. This would arbitrarily quantize the program into blocks of about 50 lines each.
Alternatively, we may require that each unit of design has certain prescribed size. For example, a
package is required to have certain number of classes, or each class a certain number of attributes
and operations. We may well end up with the unit of design or code which is performing
unrelated tasks. Any cohesion here would be accidental or coincidental cohesion.
Coincidental cohesion does not usually occur in an initial design. However, as the design goes
through multiple changes and modifications, e.g., due to requirements changes or bug fixes, and
is under schedule pressures, the original design may evolve into a coincidental one. The original
design may be patched to meet new requirements, or a related design may be adopted and
modified instead of a fresh start. This will easily result in multiple unrelated elements in a design
unit.
Ivan Marsic

Rutgers University
234
An Ordinal Scale for Cohesion Measurement
More reasonable design would have the contents of a module bear some relationship to each
other. Different relationships can be created for the contents of each module. By identifying
different types of module cohesion we can create a nominal scale for cohesion measurement. A
stronger scale is an ordinal scale, which can be created by asking an expert to assess subjectively
the quality of different types of module cohesion and create a rank-ordering. Here is an example
ordinal scale for cohesion measurement:
Rank
Cohesion type
Quality
6
Functional cohesion
Good
5
Sequential cohesion

4
Communication cohesion

3
Procedural cohesion

2
Temporal cohesion

1
Logical cohesion

0
Coincidental cohesion
Bad
Functional cohesion is judged to provide a tightest relationship because the design unit (module)
performs a single well-defined function or achieves a single goal.
Sequential cohesion is judged as somewhat weaker, because the design unit performs more than
one function, but these functions occur in an order prescribed by the specification, i.e., they are
strongly related.
Communication cohesion is present when a design unit performs multiple functions, but all are
targeted on the same data or the same sets of data. The data, however, is not organized in an
object-oriented manner as a single type or structure.
Procedural cohesion is present when a design unit performs multiple functions that are
procedurally related. The code in each module represents a single piece of functionality defining
a control sequence of activities.
Temporal cohesion is present when a design unit performs more than one function, and they are
related only by the fact that they must occur within the same time span. An example would be a
design that combines all data initialization into one unit and performs all initialization at the same
time even though it may be defined and utilized in other design units.
Logical cohesion is characteristic of a design unit that performs a series of similar functions. At
first glance, logical cohesion seems to make sense in that the elements are related. However, the
relationship is really quite weak. An example is the Java class java.lang.Math, which
contains methods for performing basic numeric operations such as the elementary exponential,
logarithm, square root, and trigonometric functions. Although all methods in this class are
logically related in that they perform mathematical operations, they are entirely independent of
each other.
Ideally, object-oriented design units (classes) should exhibit the top two types of cohesion
(functional or sequential), where operations work on the attributes that are common for the class.
A serious issue with this cohesion measure is that the success of any module in attaining highlevel cohesion relies purely on human assessment.
Chapter 4 
Software Measurement and Estimation
235
Interval Scales for Cohesion Measurement
We are mainly interested in the cohesion of object-oriented units of software, such as classes.
Class cohesion captures relatedness between various members of a class: attributes and
operations (or, methods). Class cohesion metrics can be broadly classified into two groups:
1. Interface-based metrics compute class cohesion from information in method signatures
2. Code-based metrics compute class cohesion in terms of attribute accesses by methods
We can further classify code-based cohesion metrics into four sub-types based on the methods of
quantification of cohesion:
2.a) Disjoint component-based metrics count the number of disjoint sets of methods or
attributes in a given class.
2.b) Pairwise connection-based metrics compute cohesion as a function of number of
connected and disjoint method pairs.
2.c) Connection magnitude-based metrics count the accessing methods per attribute and
indirectly find an attribute-sharing index in terms of the count (instead of computing
direct attribute-sharing between methods).
2.d) Decomposition-based metrics compute cohesion in terms of recursive decompositions of
a given class. The decompositions are generated by removal of pivotal elements that keep
the class connected.
These metrics compute class cohesion using manipulations of class elements. The key elements
of a class C are its a attributes A1, …, Aa, m methods M1, …, Mm, and the list of p parameter (or,
argument) types of the methods P1, …, Pm. The following sections describe various approaches to
computing class cohesion.
Many existing metrics qualify the class as either “cohesive” or “not cohesive,” and do not capture
varying strengths of cohesion. However, this approach makes it hard to compare two cohesive or
two non cohesive classes, or to know whether a code modification increased or reduced the
degree of cohesiveness. If one wishes to compare the cohesion of two different versions of
software, it is necessary to use a metric that can calculate not just whether a module is cohesive or
not cohesive but also the degree of its cohesiveness. Assuming that both the versions of our
software are cohesive, this would enable us to judge which version is better designed and more
cohesive.
4.3.2
Interface-based Cohesion Metrics
Interface-based cohesion metrics are design metrics that help evaluate cohesion among methods
of a class early in the analysis and the design phase. These metrics evaluate the consistency of
methods in a class’s interface using the lists of parameters of the methods. They can be applied on
class declarations that only contain method prototypes (method types and parameter types) and do
not require the class implementation code. One such metric is Cohesion Among Methods of
Classes (CAMC). The CAMC metric is based on the assumption that the parameters of a method
reasonably define the types of interaction that method may implement.
Figure 4-4
Ivan Marsic

Rutgers University
236
parameter types
SerialPort
String
DeviceCtrl
1
0
activate
0
1
deactivate
0
1
getStatus
0
1
O
DeviceCtrl
# devStatuses_ : Vector
+ activate(dev : String) : boolean
+ deactivate(dev :String) : boolean
+ getStatus(dev : String) : Object
methods
(a)
(b)
Figure 4-4: Class (a) and its parameter occurrence matrix (b).
To compute the CAMC metric value, we determine a union of all parameters of all the methods
of a class. A set Mi of parameter object types for each method is also determined. An intersection
(set Pi) of Mi with the union set T is computed for all methods in the class. A ratio of the size of
the intersection (Pi) set to the size of the union set (T) is computed for all methods. The
summation of all intersection sets Pi is divided by product of the number of methods and the size
of the union set T, to give a value for the CAMC metric. Formally, the metric is
CAMC (C ) 
4.3.3
1 k l

oij 

kl i 1 j 1
kl
(4.7)
Cohesion Metrics using Disjoint Sets of Elements
An early metric of this type is the Lack of Cohesion of Methods (LCOM1). This metric counts the
number of pairs of methods that do not share their class attributes. It considers how many disjoint
sets are formed by the intersection of the sets of the class attributes used by each method. Under
LCOM1, the perfect cohesion achieved when all methods access all attributes. Because of perfect
cohesion, we expect the lack-of-cohesion value to be 0. At the opposite end of the spectrum, each
method accesses only a single attribute (assuming that m = a). In this case, we expect LCOM = 1,
which indicates extreme lack of cohesion.
A formal definition of LCOM1 follows. Consider a set of methods {Mi} (i = 1, …, m} accessing a
set of attributes {Aj} (j = 1, …, a}. Let the number of attributes accessed by each method, Mi, be
denoted as (Mi) and the number of methods which access each attribute be (Aj). Then the lack
of cohesion of methods for a class Ci is given formally as
1 a

m      Aj 
 a j 1



LCOM1(Ci ) 
m 1
 
(4.8)
This version of LCOM is labeled as LCOM1 to allow for subsequent variations, LCOM2,
LCOM3, and LCOM4. Class cohesion, LCOM3, is measured as the number of connected
Chapter 4 
Software Measurement and Estimation
237
components in the graph. LCOM2 calculates the difference between the number of method pairs
that do or do not share their class attributes. LCOM2 is classified as a Pairwise Connection-Based
metric (Section ____). See the bibliographical notes (Section 4.7) for references on LCOM
metrics.
4.3.4
Semantic Cohesion
Cohesion or module “strength” refers to the notion of a module level “togetherness” viewed at the
system abstraction level. Thus, although in a sense it can be regarded as a system design concept,
we can more properly regard cohesion as a semantic concern expressed of a module evaluated
externally to the module.
Semantic cohesion is an externally discernable concept that assesses whether the abstraction
represented by the module (class in object-oriented approach) can be considered to be a “whole”
semantically. Semantic complexity metrics evaluate whether an individual class is really an
abstract data type in the sense of being complete and also coherent. That is, to be semantically
cohesive, a class should contain everything that one would expect a class with those
responsibilities to possess but no more.
It is possible to have a class with high internal, syntactic cohesion but little semantic cohesion.
Individually semantically cohesive classes may be merged to give an externally semantically
nonsensical class while retaining internal syntactic cohesion. For example, imagine a class that
includes features of both a person and the car the person owns. Let us assume that each person
can own only one car and that each car can only be owned by one person (a one-to-one
association). Then person_id  car_id, which would be equivalent to data normalization.
However, classes have not only data but operations to perform various actions. They provide
behavior patterns for (1) the person aspect and (2) the car aspect of our proposed class. Assuming
no intersecting behavior between PERSON and CAR, then what is the meaning of our class,
presumably named CAR_PERSON? Such a class could be internally highly cohesive, yet
semantically as a whole class seen from outside the notion expressed (here of the thing known as
a person-car) is nonsensical.
4.4 Coupling
Coupling metrics are a measure of how interdependent different modules are of each other. High
coupling occurs when one module modifies or relies on the internal workings of another module.
Low coupling occurs when there is no communication at all between different modules in a
program. Coupling is contrasted with cohesion. Both cohesion and coupling are ordinal
measurements and are defined as “high” or “low.” It is most desirable to achieve low coupling
and high cohesion.
Tightly coupled vs. loosely coupled
Ivan Marsic

(a)
Rutgers University
238
(b)
Figure 4-5: Random arrays illustrate information complexity vs. depth. See text for details.
4.5 Psychological Complexity
“Then he explained that what can be observed is really determined by the theory. He said, you cannot first
know what can be observed, but you must first know a theory, or produce a theory, and then you can
define what can be observed.” —Heisenberg’s recollection of his first meeting with Einstein
“I have had my results for a long time: but I do not yet know how I am to arrive at them.”
—Karl Friedrich Gauss
One frustration with software complexity measurement is that, unlike placing a physical object on
a scale and measuring its weight, we cannot put a software object on a “complexity scale” and
read out the amount. Complexity seems to be an interpreted measure, much like person’s health
condition and it has to be stated as an “average case.” Your doctor can precisely measure your
blood pressure, but a specific number does not necessarily correspond to good or bad health. The
doctor will also measure your heart rate, body temperature, and perhaps several other parameters,
before making an assessment about your health condition. Even so, the assessment will be the
best guess, merely stating that on average such and such combination of physiological
measurements corresponds to a certain health condition. Perhaps we should define software
object complexity similarly: as a statistical inference based on a set of directly measurable
variables.
4.5.1
Algorithmic Information Content
I already mentioned that our abstractions are unavoidably approximate. The term often used is
“coarse graining,” which means that we are blurring detail in the world picture and single out
only the phenomena we believe are relevant to the problem at hand. Hence, when defining
complexity it is always necessary to specify the level of detail up to which the system is
described, with finer details being ignored.
One way of defining the complexity of a program or system is by means of its description, that is,
the length of the description. I discussed above the merits of using size metrics as a complexity
Chapter 4 
Software Measurement and Estimation
239
measure. Some problems mentioned above include: size could be measured differently; it
depends on the language in which the program code (or any other accurate description of it) is
written; and, the program description can be unnecessarily stuffed to make it appear complex. A
way out is to ignore the language issue and define complexity in terms of the description length.
Suppose that two persons wish to communicate a system description at distance. Assume they are
employing language, knowledge, and understanding that both parties share (and know they share)
beforehand. The crude complexity of the system can be defined as the length of the shortest
message that one party needs to employ to describe the system, at a given level of coarse
graining, to the distant party.
A well-known such measure is called algorithmic information content, which was introduced in
1960s independently by Andrei N. Kolmogorov, Gregory Chaitin, and Ray Solomonoff. Assume
an idealized general-purpose computer with an infinite storage capacity. Consider a particular
message string, such as “aaaaabbbbbbbbbb.” We want to know: what is the shortest possible
program that will print out that string and then stop computing? Algorithmic information
content (AIC) is defined as the length of the shortest possible program that prints out a given
string. For the example string, the program may look something like: P a{5}b{10}, which
means “Print 'a' five times and 'b' ten times.”
Information Theory
Logical Depth and Crypticity
“...I think it better to write a long letter than incur loss of time...” —Cicero
“I apologize that this letter is so long. I did not have the time to make it short.” —Mark Twain
“If I had more time, I would have written a shorter letter.”
—variously attributed to Cicero, Pascal, Voltaire, Mark Twain, George Bernard Shaw, and T.S. Elliot
“The price of reliability is the pursuit of the utmost simplicity. It is a price which the very rich may find hard
to pay.” —C.A.R. Hoare
I already mentioned that algorithmic information content (AIC) does not exactly correspond to
everyday notion of complexity because under AIC random strings appear as most complex. But
there are other aspects to consider, as well. Consider the following description: “letter X in a
random array of letters L.” Then the description “letter T in a random array of letters L” should
have about the same AIC. Figure 4-5 pictures both descriptions in the manner pioneered by my
favorite teacher Bela Julesz. If you look at both arrays, I bet that you will be able to quickly
notice X in Figure 4-5(a), but you will spend quite some time scanning Figure 4-5(b) to detect the
T! There is no reason to believe that human visual system developed a special mechanism to
recognize the pattern in Figure 4-5(a), but failed to do so for the pattern in Figure 4-5(b). More
likely, the same general pattern recognition mechanism operates in both cases, but with much less
success on Figure 4-5(b). Therefore, it appears that there is something missing in the AIC notion
Ivan Marsic

Rutgers University
240
Working memory
Intelligent
system
Chunking unit
Chunking unit
Input sequence from the world (time-varying)
Figure 4-6: A model of a limited working memory.
of complexity—an apparently complex description has low AIC. The solution is to include the
computation time.
Charles Bennett defined logical depth of a description to characterize the difficulty of going from
the shortest program that can print the description to the actual description of the system.
Consider not just the shortest program to print out the string, but a set of short programs that have
the same effect. For each of these programs, determine the length of time (or number of steps)
needed to compute and print the string. Finally, average the running times so that shorter
programs are given greater weight.
4.6 Effort Estimation
“Adding manpower to a late software project makes it later.”
—Frederick P. Brooks, Jr., The Mythical Man-Month
“A carelessly planned project will take only twice as long.”
—The law of computerdom according to Golub
“The first 90 percent of the tasks takes 10 percent of the time and the last 10 percent takes the other 90
percent.” —The ninety-ninety rule of project schedules
Chapter 4 
4.6.1
241
Software Measurement and Estimation
Deriving Project Duration from Use Case Points
Use case points (UCP) are a measure of software size (Section 4.2.1). We can use equation (1.2)
given in Section 1.2.5 to derive the project duration. For this purpose we need to know the team’s
velocity, which represents the team’s rate of progress through the use cases (or, the team’s
productivity). Here is the equation that is equivalent to equation (1.2), but using a Productivity
Factor (PF):
Duration = UCP  PF
(4.9)
The Productivity Factor is the ratio of development person-hours needed per use case point.
Experience and statistical data collected from past projects provide the data to estimate the initial
PF. For instance, if a past project with a UCP of 112 took 2,550 hours to complete, divide 2,550
by 112 to obtain a PF of 23 person-hours per use case point.
If no historical data has been collected, the developer can consider one of these options:
1. Establish a baseline by computing the UCP for projects previously completed by your
team (if such are available).
2. Use a value for PF between 15 and 30 depending on the development team’s overall
experience and past accomplishments (Do they normally finish on time? Under budget?
etc.). For a team of beginners, such as undergraduate students, use the highest value (i.e.,
30) on the first project.
A different approach was proposed by Schneider and Winters [2001]. Recall that the
environmental factors (Table 4-7) measure the experience level of the people on your project and
the stability of your project. Any negatives in this area mean that you will have to spend time
training people or fixing problems due to instability (of requirements). The more negatives you
have, the more time you will spend fixing problems and training people and less time you will
have to devote to your project.
Schneider and Winters suggested counting the number of environmental factors among E1
through E6 (Table 4-8) that have the perceived impact less than 3 and those among E7 and E8
with the impact greater than 3. If the total count is 2 or less, assume 20 hours per use case point.
If the total is 3 or 4, assume 28 hours per use case. Any total greater than 4 indicates that there are
too many environmental factors stacked against the project. The project should be put on hold
until some environmental factors can be improved.
Probably the best solution for estimating the Productivity Factor is to calculate your
organization’s own historical average from past projects. This is why collecting historic data is
important for improving effort estimation on future projects. After a project completes, divide the
number of actual hours it took to complete the project by the UCP number. The result becomes
the new PF that can be used in the future projects.
When estimating the duration in calendar time, is important to avoid assuming ideal working
conditions. The estimate should account for corporate overhead—answering email, attending
meetings, and so on. Suppose our past experience suggests a PF of 23 person-hours per use case
point and our current project has 94 use case points (as determined Section 4.2.1). Equation (4.9)
Ivan Marsic

Rutgers University
242
gives the duration as 94  23 = 2162 person-hours. Obviously, this does not imply that the project
will be completed in 2162 / 24  90 days! A reasonable assumption is that each developer will
spend about 30 hours per week on project tasks and the rest of their time will be taken by
corporate overhead. With a team of four developers, this means the team will make 4  30 = 120
hours per week. Dividing 2162 person-hours by 120 hours per week we obtain a total of
approximately 18 weeks to complete this project.
4.7 Summary and Bibliographical Notes
In this chapter I described two kinds of software measurements. One kind works with scarce
artifacts that are available early on in a project, such as customer statement of requirements for
the planned system. There is a major subjective component to these measurements, and it works
mainly based on guessing and past experience with similar projects. The purpose of this kind of
measurements is to estimate the project duration and cost of the effort, so to negotiate the terms
of the contract with the customer who is sponsoring the project.
The other kind of software measurements works with actual software artifacts, such as UML
designs or source code. It aims to measure intrinsic properties of the software and avoid
developer’s subjective guesses. Because it requires that the measured artifacts already exist in a
completed or nearly completed condition, it cannot be applied early on in a project. The purpose
of this kind of measurements is to evaluate the product quality. It can serve as a test of whether
the product is ready for deployment, or to provide feedback to the development team about the
potential weaknesses that need to be addressed.
An early project effort estimate helps managers, developers, and testers plan for the resources a
project requires. The use case points (UCP) method has emerged as one such method. It is a
mixture of intrinsic software properties, measured by Unadjusted Use Case Points (UUCP) as
well as technical (TCF) and environmental factors (ECF), which depend on developer’s
subjective estimates. The UCP method quantifies these subjective factors into equation variables
that can be adjusted over time to produce more precise estimates. Industrial case studies indicate
that the UCP method can produce an early estimate within 20% of the actual effort.
Section 4.2: What to Measure?
[Henderson-Sellers, 1996] provides a condensed review of software metrics up to the publication
date, so it is somewhat outdated. It is technical and focuses on metrics of structural complexity.
Horst Zuse, History of Software Measurement, Technische Universität Berlin, Online at:
http://irb.cs.tu-berlin.de/~zuse/sme.html
[Halstead, 1977] distinguishes software science from computer science. The premise of software
science is that any programming task consists of selecting and arranging a finite number of
program “tokens,” which are basic syntactic units distinguishable by a compiler: operators and
operands. He defined several software metrics based on these tokens. However, software science
Chapter 4 
Software Measurement and Estimation
243
has been controversial since its introduction and has been criticized from many fronts. Halstead’s
work has mainly historical importance for software measurement because it was instrumental in
making metrics studies an issue among computer scientists.
Use case points (UCP) were first described by Gustav Karner [1993], but his initial work on the
subject is closely guarded by Rational Software, Inc. Hence, the primary sources describing
Karner’s work are [Schneider & Winters, 2001] and [Ribu, 2001]. UCP was inspired by Allan
Albrecht’s “Function Point Analysis” [Albrecht, 1979]. The weighted values and constraining
constants were initially based on Albrecht, but subsequently modified by people at Objective
Systems, LLC, based on their experience with Objectory—a methodology created by Ivar
Jacobson for developing object-oriented applications.
My main sources for use case points were [Schneider & Winters, 2001; Ribu, 2001; Cohn, 2005].
[Kusumoto, et al., 2004] describes the rules for a system that automatically computes the total
UCP for given use cases. I believe these rules are very useful for a beginner human when
computing UCPs for a project.
Many industrial case studies verified the estimation accuracy of the UCP method. These case
studies found that the UCP method can produce an early estimate within 20% of the actual effort,
and often closer to the actual effort than experts or other estimation methodologies. Mohagheghi
et al. [2005] described the UCP estimate of an incremental, large-scale development project that
was within 17% of the actual effort. Carroll [2005] described a case study over a period of five
years and across more than 200 projects. After applying the process across hundreds of sizable
software projects (60 person-months average), they achieved estimating accuracy of less than 9%
deviation from actual to estimated cost on 95% of the studied projects. To achieve greater
accuracy, Carroll’s estimation method includes a risk coefficient in the UCP equation.
Section 4.3: Measuring Module Cohesion
The ordinal scale for cohesion measurement with seven levels of cohesion was proposed by
Yourdon and Constantine [1979].
[Constantine et al., 1974; Eder et al., 1992; Allen & Khoshgoftaar, 1999; Henry & Gotterbarn,
1996; Mitchell & Power, 2005]
See also: http://c2.com/cgi/wiki?CouplingAndCohesion
B. Henderson-Sellers, L. L. Constantine, and I. M. Graham, “Coupling and cohesion: Towards a
valid suite of object-oriented metrics,” Object-Oriented Systems, vol. 3, no. 3, 143-158, 1996.
[Joshi & Joshi, 2010; Al Dallal, 2011] investigated the discriminative power of object-oriented
class cohesion metrics.
Section 4.4: Coupling
Ivan Marsic

Rutgers University
244
Section 4.5: Psychological Complexity
[Bennett, 1986; 1987; 1990] discusses definition of complexity for physical systems and defines
logical depth.
Section 4.6: Effort Estimation
Problems
Problem 4.1
Problem 4.2
Problem 4.3
(CYCLOMATIC/MCCABE COMPLEXITY) Consider the following quicksort sorting algorithm:
QUICKSORT(A, p, r)
1 if p < r
2
then q  PARTITION(A, p, r)
3
QUICKSORT(A, p, q  1)
4
QUICKSORT(A, q  1, r)
where the PARTITION procedure is as follows:
PARTITION(A, p, r)
1 x  A[r]
2 ip1
3 for j  p to r  1
4
do if A[j]  x
then i  i  1
5
6
exchange A[i]  A[j]
7 exchange A[i  1]  A[r]
8 return i  1
(a) Draw the flowchart of the above algorithm.
(b) Draw the corresponding graph and label the nodes as n1, n2, … and edges as e1, e2, …
(c) Calculate the cyclomatic complexity of the above algorithm.
Chapter 4 
Software Measurement and Estimation
Problem 4.4
245
Chapter 5
Design with Patterns
Contents
“It is not the strongest of the species that survive, nor the most
intelligent, but the one most responsive to change.”
—Charles Darwin
“Man has a limited biological capacity for change. When this
capacity is overwhelmed, the capacity is in future shock.”
—Alvin Toffler
Design patterns are convenient solutions for software design
problems commonly employed by expert developers. The
power of design patterns derives from reusing proven solution
“recipes” from similar problems. In other words, patterns are
codifying practice rather than prescribing practice, or, they are
capturing the existing best practices, rather than inventing
untried procedures. Patterns are used primarily to improve
existing designs or code by rearranging it according to a
“pattern.” By reusing a pattern, the developer gains
efficiency, by avoiding a lengthy process of trials and errors in
search of a solution, and predictability because this solution is
known to work for a given problem.
Design patterns can be of different complexities and for
different purposes. In terms of complexity, the design pattern
may be as simple as a naming convention for object methods
in the JavaBeans specification (see Chapter 7) or can be a
complex description of interactions between the multiple
classes, some of which will be reviewed in this chapter. In
terms of the purpose, a pattern may be intended to facilitate
component-based development and reusability, such as in the
JavaBeans specification, or its purpose may be to prescribe the
rules for responsibility assignment to the objects in a system,
as with the design principles described in Section 2.5.
5.1 Indirect Communication: PublisherSubscriber
5.1.1 Control Flow
5.1.2 Pub-Sub Pattern Initialization
5.1.3
5.1.4
5.1.5
5.2 More Patterns
5.2.1
5.2.2
5.2.3
5.2.4
Command
Decorator
State
Proxy
5.3 Concurrent Programming
5.3.1 Threads
5.3.2 Exclusive Resource Access—Exclusion
Synchronization
5.3.3 Cooperation between Threads—Condition
Synchronization
5.3.4
5.2.3
5.4 Broker and Distributed Computing
5.4.1 Broker Pattern
5.4.2 Java Remote Method Invocation (RMI)
5.4.3
5.4.4
5.5 Information Security
5.5.1 Symmetric and Public-Key Cryptosystems
5.5.2 Cryptographic Algorithms
5.5.3 Authentication
5.5.4
5.6 Summary and Bibliographical Notes
Problems
As pointed earlier, finding effective representation(s) is a recurring theme of software
engineering. By condensing many structural and behavioral aspects of the design into a few
simple concepts, patterns make it easier for team members to discuss the design. As with any
symbolic language, one of the greatest benefits of patterns is in chunking the design knowledge.
Once team members are familiar with the pattern terminology, the use of this terminology shifts
246
Chapter 5 
247
Design with Patterns
Custodian
Custodian
initializes the pattern
Instantiation
Instantiationofofthe
the
Design
DesignPattern
Pattern
Client
Client
asks for service
collection of objects
working to provide service
Figure 5-1: The players in a design pattern usage.
the focus to higher-level design concerns. No time is spent in describing the mechanics of the
object collaborations because they are condensed into a single pattern name.
This chapter reviews some of the most popular design patterns that will be particularly useful in
the rest of the text. What follows is a somewhat broad and liberal interpretation of design
patterns. The focus is rather on the techniques of solving specific problems; nonetheless, the
“patterns” described below do fit the definition of patterns as recurring solutions. These patterns
are conceptual tools that facilitate the development of flexible and adaptive applications as well
as reusable software components.
Two important observations are in order. First, finding a name that in one or few words conveys
the meaning of a design pattern is very difficult. A similar difficulty is experienced by user
interface designers when trying to find graphical icons that convey the meaning of user interface
operations. Hence, the reader may find the same or similar software construct under different
names by different authors. For example, The Publisher-Subscriber design pattern, described in
Section 5.1, is most commonly called Observer [Gamma et al., 1995], but [Larman, 2005] calls it
Publish-Subscribe. I prefer the latter because I believe that it conveys better the meaning of the
underlying software construct1. Second, there may be slight variations in what different authors
label with the same name. The difference may be due to the particular programming language
idiosyncrasies or due to evolution of the pattern over time.
Common players in a design pattern usage are shown in Figure 5-1. A Custodian object
assembles and sets up a pattern and cleans up after the pattern’s operation is completed. A client
object (can be the same software object as the custodian) needs and uses the services of the
pattern. The design patterns reviewed below generally follow this usage “pattern.”
5.1 Indirect Communication: PublisherSubscriber
1
The Publish-Subscribe moniker has a broader use than presented here and the interested reader should
consult [Eugster et al. 2003].
Ivan Marsic

Rutgers University
Subscribers
248
Publisher
Figure 5-2: The concept of indirect communication in a Publisher/Subscriber system.
“If you find a good solution and become attached to it, the solution may become your next problem.”
—Robert Anthony
“More ideas to choose from mean more complexity … and more opportunities to choose wrongly.”
—Vikram Pandit
Publisher-subscriber design pattern (see Figure 5-2) is used to implement indirect communication
between software objects. Indirect communication is usually used when an object cannot or does
not want to know the identity of the object whose method it calls. Another reason may be that it
does not want to know what the effect of the call will be. The most popular use of the pub-sub
pattern is in building reusable software components.
1) Enables building reusable components
2) Facilitates separation of the business logic (responsibilities, concerns) of objects
Centralized vs. decentralized execution/program-control method—spreads responsibilities for
better balancing. Decentralized control does not necessarily imply concurrent threads of
execution.
The problem with building reusable components can be illustrated on our case-study example. Let
us assume that we want to reuse the KeyChecker object in an extended version of our case-study
application, one that sounds alarm if someone is tampering with the lock. We need to modify the
method unlock() not only to send message to LockCtrl but also to AlarmCtrl, or to introduce a
new method. In either case, we must change the object code, meaning that the object is not
reusable as-is.
Chapter 5 
249
Design with Patterns
Publisher
Knowing Responsibilities:
• Knows event source(s)
• Knows interested obj’s (subscribers)
Doing Responsibilities:
• Registers/Unregisters subscribers
• Notifies the subscribers of events
«interface»
Publisher
+ subscribe()
+ unsubscribe()
«interface»
Subscriber
*
subscribers
+ receive()
Subscriber
Knowing Responsibilities:
• Knows event types of interest
• Knows publisher(s)
Doing Responsibilities:
• Registers/Unregisters with publishers
• Processes received event notifications
Type1Publisher
Type2Publisher
Type1Subscriber
+ subscribe()
+ unsubscribe()
+ subscribe()
+ unsubscribe()
+ receive()
(a)
(b)
Figure 5-3: Publisher/Subscriber objects employee cards (a), and the class diagram of their
collaborations (b).
Information source acquires information in some way and we assume that this information is
important for other objects to do the work they are designed for. Once the source acquires
information (becomes “information expert”), it is logical to expect it to pass this information to
others and initiate their work. However, this tacitly implies that the source object “knows” what
the doer object should do next. This knowledge is encoded in the source object as an “IF-THENELSE” rule and must be modified every time the doer code is modified (as seen earlier in Section
2.5).
Request- vs. event-based communication, Figure 5-4: In the former case, an object makes an
explicit request, whereas in the latter, the object expresses interest ahead of time and later gets
notified by the information source. In a way, the source is making a method request on the object.
Notice also that “request-based” is also synchronous type of communication, whereas event based
is asynchronous.
Another way to design the KeyChecker object is to make it become a publisher of events
as follows. We need to define two class interfaces: Publisher and Subscriber (see Figure
5-3). The first one, Publisher, allows any object to subscribe for information that it is the
source of. The second, Subscriber, has a method, here called receive(), to let the
Publisher publish the data of interest.
Listing 5-1: Publish-Subscribe class interfaces.
public interface Subscriber {
public void receive(Content content);
}
Ivan Marsic

Rutgers University
250
Request: doSomething(info)
Info
Info
Src
Src
Doer
Doer
Request: getInfo()
Info
Info
Src
Src
info
Doer
Doer
(1) Request: subscribe()
Info
Info
Src
Src
Doer
Doer
(2) event (info)
(a)
(b)
(c)
Figure 5-4: Request- vs. event-based communication among objects. (a) Direct request—
information source controls the activity of the doer. (b) Direct request—the doer controls
its own activity, information source is only for lookup, but doer must know when is the
information ready and available. (c) Indirect request—the doer controls its own activity
and does not need to worry when the information is ready and available—it gets prompted
by the information source.
import java.util.ArrayList;
public class Content {
public Publisher source_;
public ArrayList data_;
public Content(Publisher src, ArrayList dat) {
source_ = src;
data_ = (ArrayList) dat.clone(); // for write safety...
}
// ...avoid aliasing and create a new copy
}
public interface Publisher {
public subscribe(Subscriber subscriber);
public unsubscribe(Subscriber subscriber);
}
A Content object contains only data, no business logic, and is meant to transfer data from
Publisher to Subscriber. The actual classes then implement those two interfaces. In our
example, the key Checker object would then implement the Publisher, while DeviceCtrl
would implement the Subscriber.
Listing 5-2: Refactored the case-study code of using the Publisher-Subscriber design
pattern. Here, the class DeviceCtrl implements the Subscriber interface and the
class Checker implements the Publisher interface.
public class DeviceCtrl implements Subscriber {
protected LightBulb bulb_;
protected PhotoSObs sensor_;
public DeviceCtrl(Publisher keyChecker, PhotoSObs sensor, ... ) {
sensor_ = sensor;
keyChecker.subscribe(this);
...
}
Chapter 5 
Design with Patterns
251
public void receive(Content content) {
if (content.source_ instanceof Checker) {
if ( ((String)content.data_).equals("valid") ) {
// check the time of day; if daylight, do nothing
if (!sensor_.isDaylight()) bulb_.setLit(true);
}
} else (check for another source of the event ...) {
...
}
}
}
import java.util.ArrayList;
import java.util.Iterator;
public class Checker implements Publisher {
protected KeyStorage validKeys_;
protected ArrayList subscribers_ = new ArrayList();
public Checker( ... ) { }
public subscribe(Subscriber subscriber) {
subscribers_.add(subscriber); // could check whether this
}
// subscriber already subscribed
public unsubscribe(Subscriber subscriber) {
int idx = subscribers_.indexOf(subscriber);
if (idx != -1) { subscribers_.remove(idx); }
}
public void checkKey(Key user_key) {
boolean valid = false;
... // verify the user key against the "validKeys_" database
// notify the subscribers
Content cnt = new Content(this, new ArrayList());
if (valid) { // authorized user
cnt.data.add("valid");
} else {
// the lock is being tampered with
cnt.data.add("invalid");
}
cnt.data.add(key);
for (Iterator e = subscribers_.iterator(); e.hasNext(); ) {
((Subscriber) e.next()).receive(cnt);
}
}
}
Ivan Marsic

: Controller
enterKey()
Rutgers University
k : Key
252
: Checker
: KeyStorage
: LockCtrl
: PhotoSObs
: LightCtrl
: AlarmCtrl
: Logger
k := create()
checkKey(k)
loop
sk := getNext()
compare()
alt
valid == true
for all KeyIsValid subscribers
loop
keyIsValid()
keyIsValid()
keyIsValid()
dl := isDaylight()
opt
dl == false
setLit(true)
[else]
loop
prompt:
"try again"
for all KeyIsInvalid subscribers
keyIsInvalid()
numOfAttempts++
keyIsInvalid()
opt
numOfAttempts == maxNumOfAttempts
soundAlarm()
keyIsInvalid()
Figure 5-5: Sequence diagram for publish-subscribe version of the use case “Unlock.”
Compare this with Figure 2-27.
A Subscriber may be subscribed to several sources of data and each source may provide several
types of content. Thus, the Subscriber must determine the source and the content type before it
takes any action. If a Subscriber gets subscribed to many sources which publish different content,
the Subscriber code may become quite complex and difficult to manage. The Subscriber would
contain many if()or switch() statements to account for different options. A more objectoriented solution for this is to use class polymorphism—instead of having one Subscriber, we
should have several Subscribers, each specialized for a particular source. The Subscribers may
also have more than one receive() method, each specialized for a particular data type. Here is
an example. We could implement a Switch by inheriting from the generic Subscriber
interface defined above, or we can define new interfaces specialized for our problem domain.
Listing 5-3: Subscriber interfaces for “key-is-valid” and “key-is-invalid” events.
public interface KeyIsValidSubscriber {
public void keyIsValid(LockEvent event); // receive() method
}
public interface KeyIsInvalidSubscriber {
public void keyIsInvalid(LockEvent event); // receive() method
Chapter 5 
Design with Patterns
253
}
The new design for the Unlock use case is shown in Figure 5-5, and the corresponding code
might look as shown next. Notice that here the attribute numOfAttempts belongs to the
AlarmCtrl, unlike the first implementation in Listing 2-2 (Section 2.7), where it belonged to
the Controller. Notice also that the Controller is a KeyIsInvalidSubscriber so it
can prompt the user to enter a new key if the previous attempt was unsuccessful.
Listing 5-4: A variation of the Publisher-Subscriber design from Listing 5-2 using the
subscriber interfaces from Listing 5-3.
public class Checker implements LockPublisher {
protected KeyStorage validKeys_;
protected ArrayList keyValidSubscribers_ = new ArrayList();
protected ArrayList keyInvalidSubscribers_ = new ArrayList();
public Checker(KeyStorage ks) { validKeys_ = ks; }
public void subscribeKeyIsValid(KeyIsValidSubscriber sub) {
keyValidSubscribers_.add(sub);
}
public void subscribeKeyIsInvalid(KeyIsInvalidSubscriber sub) {
keyInvalidSubscribers_.add(sub);
}
public void checkKey(Key user_key) {
boolean valid = false;
... // verify the key against the database
// notify the subscribers
LockEvent evt = new LockEvent(this, new ArrayList());
evt.data.add(key);
if (valid) {
for (Iterator e = keyValidSubscribers_.iterator();
e.hasNext(); ) {
((KeyIsValidSubscriber) e.next()).keyIsValid(evt);
}
} else { // the lock is being tampered with
for (Iterator e = keyInvalidSubscribers_.iterator();
e.hasNext(); ) {
((KeyIsInvalidSubscriber) e.next()).keyIsInvalid(evt);
}
}
}
}
public class DeviceCtrl implements KeyIsValidSubscriber {
protected LightBulb bulb_;
protected PhotoSObs photoObserver_;
public DeviceCtrl(LockPublisher keyChecker, PhotoSObs sensor, .. )
{
photoObserver_ = sensor;
Ivan Marsic

Rutgers University
254
keyChecker.subscribeKeyIsValid(this);
...
}
public void keyIsValid(LockEvent event) {
if (!photoObserver_.isDaylight()) bulb_.setLit(true);
}
}
public class AlarmCtrl implements KeyIsInvalidSubscriber {
public static final long maxNumOfAttempts_ = 3;
public static final long interAttemptInterval_ =300000; //millisec
protected long numOfAttempts_ = 0;
protected long lastTimeAtempt_ = 0;
public AlarmCtrl(LockPublisher keyChecker, ...) {
keyChecker.subscribeKeyIsInvalid(this);
...
}
public void keyIsInvalid(LockEvent event) {
long currTime = System.currentTimeMillis();
if ((currTime – lastTimeAttempt_) < interAttemptInterval_) {
if (++numOfAttempts_ >= maxNumOfAttempts_) {
soundAlarm();
numOfAttempts_ = 0; // reset for the next user
}
} else { // this must be a new user's first mistake ...
numOfAttempts_ = 1;
}
lastTimeAttempt_ = currTime;
}
}
It is of note that what we just did with the original design for the Unlock use case can be
considered refactoring. In software engineering, the term refactoring is often used to describe
modifying the design and/or implementation of a software module without changing its external
behavior, and is sometimes informally referred to as “cleaning it up.” Refactoring is often
practiced as part of the software development cycle: developers alternate between adding new
tests and functionality and refactoring the code to improve its internal consistency and clarity. In
our case, the design from Figure 2-27 has been transformed to the design in Figure 5-5.
There is a tradeoff between the number of receive() methods and the switch() statements.
On one hand, having a long switch() statement complicates the Subscriber’s code and makes
it difficult to maintain and reuse. On the other hand, having too many receive() statements
results in a long class interface, difficult to read and represent graphically.
5.1.1
Applications of Publisher-Subscriber
The Publisher-Subscriber design pattern is used in the Java AWT and Swing toolkits for
notification of the GUI interface components about user generated events. (This pattern in Java is
known as Source-Listener or delegation event model, see Chapter 7.)
Chapter 5 
Design with Patterns
255
One of the main reasons for software components is easy visualization in integrated development
environments (IDEs), so the developer can visually assemble the components. The components
are represented as “integrated circuits” in analogy to hardware design, and different receive()
/ subscribe() methods represent “pins” on the circuit. If a component has too many pins, it
becomes difficult to visualize, and generates too many “wires” in the “blueprint.” The situation is
similar to determining the right number of pins on an integrated circuit. (See more about software
components in Chapter 7.)
Here I reiterate the key benefits of using the pub-sub design pattern and indirect communication
in general:

The components do not need to know each other’s identity. For example, in the sample
code given in Listing 1-1 (Section 1.4.2), LockCtrl maintains a reference to a LightCtrl
object.

The component’s business logic is contained within the component alone. In the same
example, LockCtrl explicitly invokes the LightCtrl’s method setLit(), meaning that it
minds LightCtrl’s business. In the worst case, even the checking of the time-of-day may
be delegated to LockCtrl in order to decide when to turn the light on.
Both of the above form the basis for component reusability, because making a component
independent of others makes it reusable. The pub-sub pattern is the most basic pattern for
reusable software components as will be discussed in Chapter 7.
In the “ideal” case, all objects could be made self-contained and thus reusable by applying the
pub-sub design pattern. However, there are penalties to pay. As visible from the examples above,
indirect communication requires much more code, which results in increased demand for memory
and decreased performance. Thus, if it is not likely that a component will need to be reused or if
performance is critical, direct communication should be applied and the pub-sub pattern should
be avoided.
When to apply the pub-sub pattern? The answer depends on whether you anticipate that the
component is likely to be reused in future projects. If yes, apply pub-sub. You should understand
that decoupled objects are independent, therefore reusable and easier to understand, while highly
interleaved objects provide fast inter-object communication and compact code. Decoupled objects
are better suited for global understanding, whereas interleaved objects are better suited for local
understanding. Of course, in a large system, global understanding matters more.
5.1.2
Control Flow
Figure 5-6 highlights the difference in control flow for direct and indirect communication types.
In the former case, the control is centralized and all flows emanate from the Controller. In the
latter case, the control is decentralized, and it is passed as a token around, cascading from object
to object. These diagrams also show the dynamic (behavioral) architecture of the system.
Ivan Marsic

Rutgers University
256
: Key
: KeyStorage
getNext()
: Checker
create()
checkKey()
: Logger
: Controller
logTransaction()
setOpen()
: LockCtrl
isDaylight()
soundAlarm()
setLit()
: AlarmCtrl
: PhotoSObs
create()
: Key
: Checker
getNext()
: KeyStorage
: LightCtrl
checkKey()
(a)
: Controller
keyIsValid()
keyIsInvalid()
: LockCtrl
: AlarmCtrl
: Logger
: LightCtrl
isDaylight()
(b)
: PhotoSObs
Figure 5-6: Flow control without (a) and with the Pub-Sub pattern (b). Notice that these
UML communication diagrams are redrawn from Figure 2-27 and Figure 5-5, respectively.
Although in Figure 5-6(b) it appears as if the Checker plays a central role, this is not so because it
is not “aware” of being assigned such a role, i.e., unlike the Controller from Figure 5-6(a), this
Checker does not encode the requisite knowledge to play such a role. The outgoing method calls
are shown in dashed lines to indicate that these are indirect calls, through the Subscriber interface.
Whatever the rules of behavior are stored in one Controller or distributed (cascading) around in
many objects, the output (seen from outside of the system) is the same. Organization (internal
function) matters only if it simplifies the software maintenance and upgrading.
Chapter 5 
257
Design with Patterns
: Controller
: Checker
: LockCtrl
: LightCtrl
: AlarmCtrl
: Logger
create()
subscribeKeyIsInvalid()
A method call
that passes a
reference to the
Checker
subscribeKeyIsValid()
subscribeKeyIsValid()
subscribeKeyIsInvalid()
subscribeKeyIsValid()
subscribeKeyIsInvalid()
Figure 5-7: Initialization of the pub-sub for the lock control example.
5.1.3
Pub-Sub Pattern Initialization
Note that the “setup” part of the pattern, example shown in Figure 5-7, plays a major, but often
ignored, role in the pattern. It essentially represents the master plan of solving the problem using
the publish-subscribe pattern and indirect communication.
Most programs are not equipped to split hard problems into parts and then use divide-andconquer methods. Few programs, too, represent their goals, except perhaps as comments in their
source codes. However, a class of programs, called General Problem Solver (GPS), was
developed in 1960s by Allen Newel, Herbert Simon, and collaborators, which did have explicit
goals and subgoals and solved some significant problems [Newel & Simon, 1962].
I propose that goal representation in object-oriented programs be implemented in the setup part of
the program, which then can act at any time during the execution (not only at the initialization) to
“rewire” the object relationships.
5.2 More Patterns
Publisher-Subscriber belongs to the category of behavioral design patterns. Behavioral patterns
separate the interdependent behavior of objects from the objects themselves, or stated differently,
they separate functionality from the object to which the functionality applies. This promotes
reuse, because different types of functionality can be applied to the same object, as needed. Here I
review Command as another behavioral pattern.
Another category is structural patterns. An example structural pattern reviewed later is Proxy.
Ivan Marsic

Rutgers University
258
create( params )
doAction( params )
Client
Client
AA
Server
Server
BB
doAction( params )
execute()
Client
Client
AA
Command
Command
Receiver
Receiver
BB
unexecute()
(a)
(b)
Figure 5-8: Command pattern interposes Command (and other) objects between a client
and a server object. Complex actions about rolling back and forward the execution history
are delegated to the Command, away from the client object.
A common drawback of design patterns, particularly behavioral patterns, is that we are replacing
what would be a single method call with many method calls. This results in performance
penalties, which in certain cases may not be acceptable. However, in most cases the benefits of
good design outweigh the performance drawbacks.
5.2.1
Command
Objects invoke methods on other objects as depicted in Figure 1-22, which is abstracted in Figure
5-8(a). The need for the Command pattern arises if the invoking object (client) needs to reverse
the effect of a previous method invocation. Another reason is the ability to trace the course of the
system operation. For example, we may need to keep track of financial transactions for legal or
auditing reasons. The purpose of the Command patter is to delegate the functionality associated
with rolling back the server object’s state and logging the history of the system operation away
from the client object to the Command object, see Figure 5-8(b).
Instead of directly invoking a method on the Receiver (server object), the client object appoints a
Command for this task. The Command pattern (Figure 5-9) encapsulates an action or processing
task into an object thus increasing flexibility in calling for a service. Command represents
operations as classes and is used whenever a method call alone is not sufficient. The Command
object is the central player in the Command pattern, but as with most patterns, it needs other
objects to assist with accomplishing the task. At runtime, a control is passed to the execute()
method of a non-abstract-class object derived from Command.
Figure 5-9(c) shows a sequence diagram on how to create and execute a command. In addition to
executing requests, we may need to be able to trace the course of the system operation. For
example, we may need to keep track of financial transactions for legal or auditing reasons.
CommandHistory maintains history log of Commands in linear sequence of their execution.
Chapter 5 
259
Design with Patterns
invoker
Command
Knowing Responsibilities:
• Knows receiver of action request
• Optional: May know whether action is reversible
Doing Responsibilities:
• Executes an action
• Optional: May undo an action if it is reversible
«interface»
Command
+ execute()
Receiver1
receiver
+ doAction1()
Receiver2
custodian
invoker
ActionType2Cmd
+ execute()
+ execute()
receiver
+ doAction2()
(a)
ActionType1Cmd
(b)
cmd : Command
receiver
: CommandHistory
create(receiver, args)
accept(cmd)
(c)
execute()
doAction(args)
log(cmd)
Figure 5-9: (a) Command object employee card. (b) The Command design pattern (class
diagram). The base Command class is an interface implemented by concrete commands.
(c)Interaction diagram for creating and executing a command.
It is common to use Command pattern in operating across the Internet. For example, suppose that
client code needs to make a function call on an object of a class residing on a remote server. It is
not possible for the client code to make an ordinary method call on this object because the remote
object cannot appear in the usual compile-execute process. It is also difficult to employ remote
method invocation (Section 5.4.2) here because we often cannot program the client and server at
the same time, or they may be programmed by different parties. Instead, the call is made from the
client by pointing the browser to the file containing the servlet (a server-side software
component). The servlet then calls its method service(HttpServletRequest,
HttpServletResponse). The object HttpServletRequest includes all the information
that a method invocation requires, such as the argument values, obtained from the “environment”
variables at standardized global locations. The object HttpServletResponse carries the
result of invoking service(). This technique embodies the basic idea of the Command design
pattern. (See also Listing 5-5.)
Web services allow a similar runtime function discovery and invocation, as will be seen in
Chapter 8.
Ivan Marsic

Rutgers University
«interface»
Command
260
invoker
: CommandHistory
undo()
+ execute()
+ unexecute()
+ isReversible() : boolean
ActionType1Cmd
ActionType2Cmd
+ execute()
+ unexecute()
+ isReversible()
+ execute()
+ unexecute()
+ isReversible()
(a)
opt
: Command
receiver
isReversible()
reversible == true
unexecute()
doInverseAction()
setCurrentCmdPtr()
(b)
Figure 5-10: (a) Class diagram for commands that can be undone. (b) Interaction diagram
for undoing a (reversible) command. Compare to Figure 5-9.
Undo/Redo
The Command pattern may optionally be able to support rollback of user’s actions in an elegant
fashion. Anyone who uses computers appreciates the value of being able to undo their recent
actions. Of course, this feature assumes that a command’s effect can be reversed. In this case, the
Command interface would have two more operations (Figure 5-10(a)): isReversible() to
allow the invoker to find out whether this command can be undone; and unexecute() to undo
the effects of a previous execute() operation.
Figure 5-10(b) shows a sequence diagram on how to undo/redo a command, assuming that it is
undoable. Observe also that CommandHistory should decrement its pointer of the current
command every time a command is undone and increments it every time a command is redone.
An additional requirement on CommandHistory is to manage properly the undo/redo caches. For
example, if the user backs up along the undo queue and then executes a new command, the whole
redo cache should be flushed. Similarly, upon a context switching, both undo/redo caches should
be flushed. Obviously, this does not provide for long-term archiving of the commands; if that is
required, the archive should be maintained independently of the undo/redo caches.
In physical world, actions are never reversible (because of the laws of thermodynamics). Even an
approximate reversibility may not be realistic to expect. Consider a simple light switch. One
might thing that turning the switch off is exactly opposite of turning it on. Therefore, we could
implement a request to turn the switch off as an undo operation of the command to turn the switch
on. Unfortunately, this may not be true. For example, beyond the inability to recover the energy
lost during the period that the switch was on, it may also happen that the light bulb is burnt.
Obviously, this cannot be undone (unless the system has a means of automatically replacing a
burnt light bulb with a new one ).
In digital world, if the previous state is stored or is easy to compute, then the command can be
undone. Even here we need to beware of potential error accumulation. If a number is repeatedly
divided and then multiplied by another number, rounding errors or limited number of bits for
number representation may yield a different number than the one we started with.
Chapter 5 
261
Design with Patterns
Decorator
client
Knowing Responsibilities:
• Knows next decorator or real subject
• Has same interface as real subject
Doing Responsibilities:
• Contributes a special-case processing
• Forwards the request to next object
in chain (decorator or real subject)
«interface»
Subject
next object
+ request()
Decorator
RealSubject
+ request()
(a)
client :
+ request()
ConcreteDecorator1
ConcreteDecorator2
+ request()
+ request()
: ConcreteDecorator1
request( args )
(b)
: ConcreteDecorator2
: RealSubject
addedProcessing( )
request( args )
addedProcessing( )
request(?args‡ )
(c)
 and ‡ denote
added specialcase processing
‡
result?
result
‡
result?
moreAddedProcessing( )
moreAddedProcessing( )
Figure 5-11: (a) Decorator object employee card. (b) The Decorator design pattern (class
diagram). (c)Interaction diagram for the Decorator pattern.
5.2.2
Decorator
The Decorator pattern is used to add non-essential behavior to key objects in a software design.
The embellished class (or, decoratee) is wrapped up by an arbitrary number of Decorator classes,
which provide special-case behaviors (embellishments).
Figure 5-11
Notice that the Decorator is an abstract class (the class and method names are italicized). The
reason for this choice is to collect the common things from all different decorators into a base
decorator class. In this case, the Decorator class will contain a reference to the next decorator.
The decorators are linked in a chain. The client has a reference to the start of the chain and the
chain is terminated by the real subject. Figure 5-11(c) illustrates how a request from the client
propagates forward through the chain until it reaches the real subject, and how the result
propagates back.
To decide whether you need to introduce Decorator, look for special-case behaviors
(embellishment logic) in your design.
Consider the following example, where we wish to implement the code that will allow the user to
configure the settings for controlling the household devices when the doors are unlocked or
locked. The corresponding user interface is shown in Figure 2-2 (Section 2.2). Figure 5-12 and

Ivan Marsic
Rutgers University
262
client
nextDevice
«interface»
DeviceCtrl
Subject and
Decorator interface
+ activate()
Controller
LockCtrl
+ activate()
– disarmLock()
MusicCtrl
LightCtrl
+ activate()
– turnOnMusicPlayer()
+ activate()
– turnOnLight()
AlarmCtrl
+ activate()
– ...
RealSubject
Concrete Decorators
Figure 5-12: Example Decorator class diagram, for implementing the interface in Figure 2-2.
Figure 5-13 show UML diagrams that use the Decorator design pattern in solving this problem.
Notice the slight differences in the class diagrams in Figure 5-11(b) and Figure 5-12. As already
pointed out, the actual pattern implementation will not always strictly adhere to its generic
prototype.
In this example, the decorating functionalities could be added before or after the main function,
which is to activate the lock control. For example, in Figure 5-13 the decorating operation
LightCtrl.turnOnLight() is added before LockCtrl.activate(), but
MusicCtrl.turnOnMusicPlayer() is added after it. In this case all of these operations
are commutative and can be executed in any order. This may not always be the case with the
decorating functionalities.
5.2.3
State
The State design pattern is usually used when an object’s behavior depends on its state in a
complex way. In this case, the state determines a mode of operation. Recall that the state of a
software object is represented by the current values of its attributes. The State pattern externalizes
the relevant attributes into a State object, and this State object has the responsibility of managing
the state transitions of the original object. The original object is called “Context” and its attributes
are externalized into a State object (Figure 5-14).
A familiar example of object’s state determining its mode of operation includes tools in document
editors. Desktop computers normally have only keyboard and mouse as interaction devices. To
enable different manipulations of document objects, the document needs to be put in a proper
state or mode of operation. That is why we select a proper “tool” in a toolbar before performing a
manipulation. The selected tool sets the document state. Consider an example of a graphics
editor, such as Microsoft PowerPoint. When the user clicks the mouse pointer on a graphical
object and drags the mouse, what will happen depends on the currently selected tool. The default
Chapter 5 
263
Design with Patterns
: Controller
: MusicCtrl
: LightCtrl
: LockCtrl
: PhotoObsrv
: AlarmCtrl
enterKey( )
ref
val := check the key validity
(see sequence fragment in Figure 2-20)
alt
val == true
activate()
activate()
opt
dl := isDaylight()
dl == false
turnOnLight()
activate()
disarmLock()
turnOnMusicPlayer()
[else]
alt
numOfAttempts++
numOfAttempts == maxNumOfAttempts
denyMoreAttempts()
activate()
…
AlarmCtrl
preceded by
suitable decorators
[else]
Figure 5-13: Decorator sequence diagram for the class diagram in Figure 5-12.
tool will relocate the object to a new location; the rotation tool will rotate the object for an angle
proportional to the distance the mouse is dragged over; etc. Notice that the same action (mouse
click and drag) causes different behaviors, depending on the document state (i.e., the currently
selected tool).
The State pattern is also useful when an object implements complex conditional logic for
changing its state (i.e., the values of this object’s attributes). We say that the object is
transitioning from one state (one set of attribute values) to another state (another set of attribute
values). To simplify the state transitioning, we define a State interface and different classes that
implement this interface correspond to different states of the Context object (Figure 5-14(b)).
Each concrete State class implements the behavior of the Context associated with the state
implemented by this State class. The behavior includes calculating the new state of the Context.
Because specific attribute values are encapsulated in different concrete states, the current State
class just determines the next state and returns it to the Context. Let us assume that the UML state
diagram for the Context class is represented by the example in Figure 5-14(c). As shown in
Figure 5-14(d), when the Context receives a method call request() to handle an event, it calls
the method handle() on its currentState. The current state processes the event and
Ivan Marsic

Rutgers University
264
Context
State
+ request(evt : Event)
Knowing Responsibilities:
• Knows one set of values (state) of
attributes of the Context object
Doing Responsibilities:
• Implement behavior associated
with this state of the Context
«interface»
State
currentState
+ handle()
(b)
ConcreteStateA
+ handle()
ConcreteStateB
+ handle()
(a)
: Context
request( event-1 )
currentState : ConcreteStateA
handle( event-1 )
nextState := this
result, nextState
event-1
event-2 [condition] / action-2
result
State-A
State-B
currentState := nextState
request( event-2 )
opt
condition == true
handle( event-2 )
result, nextState
(c)
(d)
perform action-2
nextState :=
ConcreteStateB
currentState := nextState
Figure 5-14: (a) State object employee card. (b) The State design pattern (class diagram).
(c) Example state diagram for the Context object. (d)Interaction diagram for the state
diagram in (c).
performs any action associated with the current state transition. Finally, it returns the next state to
the caller Context object. The Context sets this next state as the current state and the next
request will be handled by the new current state.
5.2.4
Proxy
The Proxy pattern is used to manage or control access to an object. Proxy is needed when the
logistics of accessing the subject’s services is overly complex and comparable or greater in size
than that of client’s primary responsibility. In such cases, we introduce a helper object (called
“proxy”) for management of the subject invocation. A Proxy object is a surrogate that acts as a
stand-in for the actual subject, and controls or enhances the access to it (Figure 5-15). The proxy
object forwards requests to the subject when appropriate, depending on whether the constraint of
the proxy is satisfied.
Chapter 5 
265
Design with Patterns
Proxy
client
«interface»
Subject
Knowing Responsibilities:
• Knows the real subject of requests
• Has same interface as real subject
Doing Responsibilities:
• Intercepts & preprocesses requests
• Ensures safe, efficient & correct
access to the real subject
+ request()
Proxy
realSubject
+ request()
RealSubject
+ request()
(a)
(b)
client :
: Proxy
request( args )
opt
(c)
: RealSubject
preprocessRequest( )
constraint satisfied
 denotes possibly
preprocessed
input arguments
request( args )
result
postprocessResult( )

result?
Figure 5-15: (a) Proxy object employee card. (b) The Proxy design pattern (class diagram).
(c)Interaction diagram for the Proxy pattern.
The causes of access complexity and the associated constraints include:

The subject is located in a remote address space, e.g., on a remote host, in which case the
invocation (sending messages to it) requires following complex networking protocols.
Solution: use the Remote Proxy pattern for crossing the barrier between different
memory spaces

Different access policies constrain the access to the subject. Security policies require that
access is provided only to the authorized clients, filtering out others. Safety policies may
impose an upper limit on the number of simultaneous accesses to the subject.
Solution: use the Protection Proxy pattern for additional housekeeping

Deferred instantiation of the subject, to speed up the performance (provided that its full
functionality may not be immediately necessary). For example, a graphics editor can be
started faster if the graphical elements outside the initial view are not loaded until they
are needed; only if and when the user changes the viewpoint, the missing graphics will be
loaded. Graphical proxies make this process transparent for the rest of the program.
Solution: use the Virtual Proxy pattern for optimization in object creation
In essence we could say that proxy allows client objects to cross a barrier to server objects (or,
“subjects”). The barrier may be physical (such as network between the client and server
computers) or imposed (such as security policies to prevent unauthorized access). As a result, the
client cannot or should not access the server by a simple method call as when the barrier does not
exist. The additional functionality needed to cross the barrier is extraneous to the client’s business
logic. The proxy object abstracts the details of the logistics of accessing the subject’s services
Ivan Marsic

Rutgers University
266
Obtain user role
and credentials
[ user == sys-admin ]
Grant full access
to metadata and data
[else]
[ user == landlord ]
Grant read/write access
to all data
[else]
[ user == tenant ]
[else]
Grant read-only access to
personal data and activity
data for own apartment
Deny all access
Figure 5-16: Conditional logic for controlling access to the database of the secure home
access system.
across different barriers. It does this transparently, so the client has an illusion it is directly
communicating with the subject, and does not know that there is a barrier in the middle.
Proxy offers the same interface (set of methods and their signatures) as the real subject and
ensures correct access to the real subject. For this reason, the proxy maintains a reference to the
real subject (Figure 5-15(b)). Because of the identical interface, the client does not need to change
its calling behavior and syntax from that which it would use if there were no barrier involved.
The Remote Proxy pattern will be incorporated into a more complex Broker pattern (Section 5.4).
The rest of this section provides more detail on the Protection Proxy.
Protection Proxy
The Protection Proxy pattern can be used to implement different policies to constrain the access
to the subject. For example, a security policy may require that a defined service should be seen
differently by clients with different privileges. This pattern helps us customize the access, instead
of using conditional logic to control the service access. In other words, it is applicable where a
subset of capabilities or partial capability should be made available to different actors, based on
their roles and privileges.
For example, consider our case study system for secure home access. The sequence diagram for
use case UC-5: Inspect Access History is shown in Figure 2-26. Before the Controller calls the
method accessList := retrieve(params : string) on Database Connection, the
system should check that this user is authorized to access the requested data. (This fragment is not
shown in Figure 2-26.) Figure 5-16 depicts the Boolean logic for controlling the access to the data
in the system database. One way to implement this scheme is to write one large conditional IFTHEN-ELSE statement. This approach would lead to a complex code that is difficult to
understand and extend if new policies or roles need to be considered (e.g., the Maintenance
Chapter 5 
267
Design with Patterns
«interface»
java.sql.Connection
Subject
client : Controller
+ createStatement( … ) : Statement
+ getMetaData() : DatabaseMetaData
…
RealSubject
request() methods
ConnectionImpl
…
tenant’s Proxy
admin’s Proxy
DBConTenant
dBase
+ createStatement( … ) : Statement
+ getMetaData() : DatabaseMetaData
…
DBConAdmin
# credentials_ : Object
# credentials_ : Object
+ createStatement( … ) : Statement
+ getMetaData() : DatabaseMetaData
…
– checkRequestAuthorized()
– createStatmProxy( … ) : Statement
+ createStatement( … ) : Statement
+ getMetaData() : DatabaseMetaData
…
– checkRequestAuthorized()
– createStatmProxy( … ) : Statement
dBc
dBc
factory
Factory
+ getDbaseConnection(credentials : Object) : java.sql.Connection
Factory pattern
for creating Connection
and wrapping with Proxy
Figure 5-17: Class diagram for the example proxy for enforcing authorized database access.
See the interactions in Figure 5-18. (Compare to Figure 5-15(b) for generic Proxy pattern.)
actor). In addition, it serves as a distraction from the main task of the client or server objects. This
is where protection proxy enters the picture in and takes on the authorization responsibility.
Figure 5-17 shows how Protection Proxy is implemented in the example of safe database access.
Each different proxy type specifies a set of legal messages from client to subject that are
appropriate for the current user’s access rights. If a message is not legal, the proxy will not
forward it to the real subject (the database connection object ConnectionImpl); instead, the
proxy will send an error message back to the caller (i.e., the client).
In this example, the Factory object acts as a custodian that sets up the Proxy pattern (see Figure
5-17 and Figure 5-18).
It turns out that in this example we need two types of proxies: (a) proxies that implement the
database connection interface, such as java.sql.Connection if Java is used; and (b)
proxies that implement the SQL statement interface, such as java.sql.Statement if Java is
used. The connection proxy guards access to the database metadata, while the statement proxy
guards access to the database data. The partial class diagram in Figure 5-17 shows only the
connection-proxy classes, and Figure 5-18 mentions the statement proxy only in the last method
call createStatmProxy(), by which the database proxy (DBConTenant) creates a
statement proxy and returns it.

Ivan Marsic
: Controller
Rutgers University
268
proxyAD :
DBConAdmin
factory : Factory
proxyLL :
DBConLlord
proxyTN :
DBConTenant
dBc :
ConnectionImpl
dBase := getDbaseConnection( credentials )
dBc := java.sql.DriverManager.getConnection(…)
[credentls == "admin"]
alt
proxyAD := create( dBc )
return proxyAD
[credentls == "landlord"]
(a)
proxyLL := create( dBc )
return proxyLL
[credentls == "tenant"]
proxyTN := create( dBc )
return proxyTN
[else]
return NULL
query := createStatement( … )
statm := createStatement( … )
(b)
createStatmProxy( statm, … )
return SQL Statement Proxy
Figure 5-18: Example of Protection Proxy setup (a) and use (b) that solves the accesscontrol problem from Figure 5-16. (See the corresponding class diagram in Figure 5-17.)
Figure 5-18
Listing 5-5: Implementation of the Protection Proxy that provides safe access to the
database in the secure home access system.
import
import
import
import
import
import
import
import
java.sql.Connection;
java.sql.DriverManager;
java.sql.ResultSet;
java.sql.Statement;
javax.servlet.ServletConfig;
javax.servlet.http.HttpServlet;
javax.servlet.http.HttpServletRequest;
javax.servlet.http.HttpServletResponse;
public class WebDataAccessServlet extends HttpServlet {
private String
// database access parameters
driverClassName = "com.mysql.jdbc.Driver",
dbURL = "jdbc:mysql://localhost/homeaccessrecords",
dbUserID = null,
dbPassword = null;
private Connection dBase = null;
Chapter 5 
Design with Patterns
269
public void init(ServletConfig config) throws ServletException {
super.init(config);
...
dbUserID = config.getInitParameter("userID");
dbPassword = config.getInitParameter("password");
Factory factory = new Factory(driverClassName, dbURL);
dBase = factory.getDbaseConnection(dbUserID, dbPassword);
}
public void service(
HttpServletRequest req, HttpServletResponse resp
) throws ServletException, java.io.IOException {
Statement statm = dBase.createStatement();
// process the request and prepare ...
String sql = // ... an SQL statement from the user's request
boolean ok = statm.execute(sql);
ResultSet result = statm.getResultSet();
// print the result into the response (resp argument)
}
}
public class Factory {
protected String dbURL_;
protected Connection dBc_ = null;
public Factory(String driverClassName, String dbURL) {
// load the database driver class
(the Driver class creates
Class.forName(driverClassName); //
an instance of itself)
dbURL_ = dbURL;
}
public Connection getDbaseConnection(
String dbUserID, String dbPassword
) {
dBc_ = DriverManager.getConnection(
dbURL_, dbUserID, dbPassword
);
Connection proxy = null;
int userType = getUserType(dbUserID, dbPassword);
switch (userType) {
case 1:
// dbUserID is a system administrator
proxy = new DBConAdmin(dBc_, dbUserID, dbPassword);
case 2:
// dbUserID is a landlord
proxy = new DBConLlord(dBc_, dbUserID, dbPassword);
case 3:
// dbUserID is a tenant
proxy = new DBConTenant(dBc_, dbUserID, dbPassword);
default:
// dbUserID cannot be identified
proxy = null;
}
return proxy;
}
}
Ivan Marsic

Rutgers University
270
// Protection Proxy class for the actual java.sql.Connection
public class DBConTenant implements Connection {
protected Connection dBc_ = null;
protected String
dbUserID = null,
dbPassword = null;;
public DBConTenant(
Connection dBc, String dbUserID, String dbPassword
) {
...
}
public Statement createStatement() {
statm = dBc_.createStatement();
return createStatmProxy(statm, credentials_);
}
private Statement createStatmProxy(
Statement statm, credentials_
) {
// create a proxy of java.sql.Statement that is appropriate
// for a user of the type "tenant"
}
}
One may wonder if we should similarly use Protection Proxy to control the access to the locks in
a building, so the landlord has access to all apartments and a tenant only to own apartment. When
considering the merits of this approach, the developer first needs to compare it to a
straightforward conditional statement and see which approach would create a more complex
implementation.
The above example illustrated the use of Proxy to implement security policies for authorized data
access. Another example involves safety policies to limit the number of simultaneous accesses.
For example, to avoid inconsistent reads/writes, the policy may allow at most one client at a time
to access the subject, which in effect serializes the access to the subject. This constraint is
implemented by passing a token among clients—only the client in possession of the token can
access the subject, by presenting the token when requesting access.
The Protection Proxy pattern is structurally identical to the Decorator pattern (compare Figure
5-11 and Figure 5-15). We can also create a chain of Proxies, same as with the Decorators (Figure
5-11(c)). The key difference is in the intent: Protection Proxy protects an object (e.g., from
unauthorized access) while Decorator adds special-case behavior to an object.
SIDEBAR 5.1: Structure and Intention of Design Patterns
Chapter 5 
Design with Patterns
271

The reader may have noticed that many design patterns look similar to one another. For
example, Proxy is structurally almost identical to Decorator. The difference between them is in
their intention—what they are used for. The intention of Decorator is to add functionality,
while the intention of Proxy is to subtract functionality, particularly for Protection Proxy.
5.3 Concurrent Programming
“The test of a first-rate intelligence is the ability to hold two opposed ideas in mind at the same time and
still retain the ability to function.” —F. Scott Fitzgerald
The benefits of concurrent programming include better use of multiple processors and easier
programming of reactive (event-driven) applications. In event-driven applications, such as
graphical user interfaces, the user expects a quick response from the system. If the (singleprocessor) system processes all requests sequentially, then it will respond with significant delays
and most of the requestors will be unhappy. A common technique is to employ time-sharing or
time slicing—a single processor dedicates a small amount of time for each task, so all of them
move forward collectively by taking turns on the processor. Although none of the tasks
progresses as fast as it would if it were alone, none of them has to wait as long as it could have if
the processing were performed sequentially. The task executions are really sequential but
interleaved with each other, so they virtually appear as concurrent. In the discussion below I
ignore the difference between real concurrency, when the system has multiple processors, and
virtual concurrency on a single-processor system. From the user’s viewpoint, there is no logical
or functional difference between these two options—the user would only see difference in the
length of execution time.
Computer process is, roughly speaking, a task being executed by a processor. A task is defined by
a temporally ordered sequence of instructions (program code) for the processor. In general, a
process consists of:

Memory, which contains executable program code and/or associated data

Operating system resources that are allocated to the process, such as file descriptors
(Unix terminology) or handles (Windows terminology)

Security attributes, such as the identity of process owner and the process’s set of
privileges

Processor state, such as the content of registers, physical memory addresses, etc. The
state is stored in the actual registers when the process is executing, and in memory
otherwise
Threads are similar to processes, in that both represent a single sequence of instructions executed
in parallel with other sequences, either by time slicing on a single processor or multiprocessing. A
process is an entirely independent program, carries considerable state information, and interacts
with other processes only through system-provided inter-process communication mechanisms.
Conversely, a thread directly shares the state variables with other threads that are part of the same
Ivan Marsic

Rutgers University
272
yield()
Runnable
Alive
Blocked
wait()
Waiting for
notification
start()
Ready
interrupt()
New
Waiting for
I/O or lock
join()
Waiting for
rendezvous
Target finishes
run() returns
Dead
notify() or
notifyAll()
Interrupted
sleep()
Sleeping
Time out
interrupt()
/ throws InterruptedException
Figure 5-19: State diagram representing the lifecycle of Java threads. (State diagram
notation is defined in Section 3.2.2.)
process, as well as memory and other resources. In this section I focus on threads, but many
concepts apply to processes as well.
So far, although I promoted the metaphor of an object-based program as a “bucket brigade,” the
objects carried their tasks sequentially, one after another, so in effect the whole system consists of
a single worker taking the guises one-by-one of different software objects. Threads allow us to
introduce true parallelism in the system functioning.
Subdividing a problem to smaller problems (subtasks) is a common strategy in problem solving.
It would be all well and easy if the subtasks were always disjoint, clearly partitioned and
independent of each other. However, during the execution the subtasks often operate on the same
resources or depend on results of other task(s). This is what makes concurrent programming
complex: threads (which roughly correspond to subtasks) interact with each other and must
coordinate their activities to avoid incorrect results or undesired behaviors.
5.3.1
Threads
A thread is a sequence of processor instructions, which can share a single address space with
other threads—that is, they can read and write the same program variables and data structures.
Threads are a way for a program to split itself into two or more concurrently running tasks. It is a
basic unit of program execution. A common use of threads is in reactive applications, having one
thread paying attention to the graphical user interface, while others do long calculations in the
background. As a result, the application more readily responds to user’s interaction.
Figure 5-19 summarizes different states that a thread may go through in its lifetime. The three
main states and their sub-states are:
1. New: The thread object has been created, but it has not been started yet, so it cannot run
Chapter 5 
Design with Patterns
273
2. Alive: After a thread is started, it becomes alive, at which point it can enter several different
sub-states (depending on the method called or actions of other threads within the same
process):
a. Runnable: The thread can be run when the time-slicing mechanism has CPU cycles
available for the thread. In other words, when there is nothing to prevent it from being
run if the scheduler can arrange it
b. Blocked: The thread could be run, but there is something that prevents it (e.g., another
thread is holding the resource needed for this thread to do its work). While a thread is in
the blocked state, the scheduler will simply skip over it and not give it any CPU time, so
the thread will not perform any operations. As visible from Figure 5-19, a thread can
become blocked for the following reasons:
i. Waiting for notification: Invoking the method wait() suspends the thread until the
thread gets the notify() or notifyAll() message
ii. Waiting for I/O or lock: The thread is waiting for an input or output operation to
complete, or it is trying to call a synchronized method on a shared object, and
that object’s lock is not available
iii. Waiting for rendezvous: Invoking the method join(target) suspends the thread
until the target thread returns from its run() method
iv. Sleeping: Invoking the method sleep(milliseconds) suspends the thread for
the specified time
3. Dead: This normally happens to a thread when it returns from its run() method. A dead
thread cannot be restarted, i.e., it cannot become alive again
The meaning of the states and the events or method invocations that cause state transitions will
become clearer from the example in Section 5.3.4.
A thread object may appear as any other software object, but there are important differences.
Threads are not regular objects, so we have to be careful with their interaction with other objects.
Most importantly, we cannot just call a method on a thread object, because that would execute the
given method from our current thread—neglecting the thread of the method’s object—which
could lead to conflict. To pass a message from one thread to another, we must use only the
methods shown in Figure 5-19. No other methods on thread objects should be invoked.
If two or more threads compete for the same “resource” which can be used by only one at a time,
then their access must be serialized, as depicted in Figure 5-20. One of them becomes blocked
while the other proceeds. We are all familiar with conflicts arising from people sharing resources.
For example, people living in a house/apartment share the same bathroom. Or, many people may
be sharing the same public payphone. To avoid conflicts, people follow certain protocols, and
threads do similarly.
Ivan Marsic

Rutgers University
274
Thread 1
Shared
Resource
Thread 2
RA
IL
RA
IL
RA
G
SIN
OSROA
D
CR
G
IN
RS
OS OA
D
CR
Step 1: Lock
Step 2: Use
IL
G
SIN
OSROA
D
CR
Step 3: Unlock
Figure 5-20: Illustration of exclusion synchronization. The lock simply ensures that
concurrent accesses to the shared resource are serialized.
5.3.2
Exclusive Resource Access—Exclusion
Synchronization
If several threads attempt to access and manipulate the same data concurrently a race condition or
race hazard exists, and the outcome of the execution depends on the particular order in which the
access takes place. Consider the following example of two threads simultaneously accessing the
same banking account (say, husband and wife interact with the account from different branches):
Thread 1
Thread 2
oldBalance = account.getBalance();
newBalance = oldBalance + deposit;
account.setBalance(newBalance);
...
oldBalance = account.getBalance();
newBalance =
oldBalance - withdrawal;
account.setBalance(newBalance);
...
The final account balance is incorrect and the value depends on the order of access. To avoid race
hazards, we need to control access to the common data (shared with other threads) and make the
access sequential instead of parallel.
DING
DING
DING
DING
I THINK I CAN,
I THINK I CAN,
I THINK I CAN...
A segment of code in which a thread may modify
shared data is known as a critical section or critical
region. The critical-section problem is to design a
protocol that threads can use to avoid interfering
with each other. Exclusion synchronization, or
mutual exclusion (mutex), see Figure 5-21, stops
different threads from calling methods on the same
object at the same time and thereby jeopardizing the
integrity of the shared data. If thread is executing in
its critical region then no other thread can be
Chapter 5 
275
Design with Patterns
thrd1 : Thread
region
Critical region;
the code fragment
can have only one
thread executing it
at once.
shared_obj : Object
thrd2 : Thread
acquire lock
acquire lock
Successful:
thrd1 locks itself
in & does the work
release lock
Unsuccessful:
thrd2 blocked and
waiting for the
shared object to
become vacant
transfer lock
thrd2 acquires
the lock &
does the work
Lock transfer
controlled by
operating system
and hardware
Figure 5-21: Exclusion synchronization pattern for concurrent threads.
executing in its critical region. Only one thread is allowed in a critical region at any moment.
Java provides exclusion synchronization through the keyword synchronized, which simply
labels a block of code that should be protected by locks. Instead of the programmer explicitly
acquiring or releasing the lock, synchronized signals to the compiler to do so. As illustrated
in Figure 5-22, there are two ways to use the keyword synchronized. First technique declares
class methods synchronized, Figure 5-22(a). If one thread invokes a synchronized
method on an object, that object is locked. Another thread invoking this or another
synchronized method on that same object will block until the lock is released.
Nesting method invocations are handled in the obvious way: when a synchronized method is
invoked on an object that is already locked by that same thread, the method returns, but the lock
is not released until the outermost synchronized method returns.
Second technique designates a statement or a block of code as synchronized. The
parenthesized expression must produce an object to lock—usually, an object reference. In the
simplest case, it could be this reference to the current object, like so
synchronized (this) {
/* block of code statements */
}
When the lock is obtained, statement is executed as if it were synchronized method on the
locked object. Examples of exclusion synchronization in Java are given in Section 5.3.4.
shared object
acquire
lock
release
lock
public class SharedClass {
...
public synchronized void
method1( ... ) {
...
}
}
(a)
acquire
lock
release
lock
public class AnyClass {
...
public void method2( ... ) {
...
synchronized (expression) {
statement
}
...
}
shared object
}
(b)
Figure 5-22: Exclusion synchronization in Java: (a) synchronized methods, and
(b) synchronized statements.
Ivan Marsic

5.3.3
Cooperation between Threads—Condition
Synchronization
Rutgers University
276
Exclusion synchronization ensures that threads “do not step on each other’s toes,” but other than
preventing them from colliding, their activities are completely independent. However, sometimes
one thread’s work depends on activities of another thread, so they must cooperate and coordinate.
A classic example of cooperation between threads is a Buffer object with methods put() and
get(). Producer thread calls put() and consumer thread calls get(). The producer must wait
if the buffer is full, and the consumer must wait if it is empty. In both cases, threads wait for a
condition to become fulfilled. Condition synchronization includes no assumption that the wait
will be brief; threads could wait indefinitely.
Condition synchronization (illustrated in Figure 5-23) complements exclusion synchronization. In
exclusion synchronization, a thread encountering an occupied shared resource becomes blocked
and waits until another thread is finished with using the resource. Conversely, in condition
synchronization, a thread encountering an unmet condition cannot continue holding the resource
on which condition is checked and just wait until the condition is met. If the tread did so, no other
thread would have access to the resource and the condition would never change—the resource
must be released, so another thread can affect it. The thread in question might release the resource
and periodically return to check it, but this would not be an efficient use of processor cycles.
Rather, the thread becomes blocked and does nothing while waiting until the condition changes,
at which point it must be explicitly notified of such changes.
In the buffer example, a producer thread, t, must first lock the buffer to check if it is full. If it is, t
enters the “waiting for notification” state, see Figure 5-19. But while t is waiting for the condition
to change, the buffer must remain unlocked so consumers can empty it by calling get().
Because the waiting thread is blocked and inactive, it needs to be notified when it is ready to go.
Every software object in Java has the methods wait() and notify() which makes possible
sharing and condition synchronization on every Java object, as explained next. The method
wait() is used for suspending threads that are waiting for a condition to change. When t finds
the buffer full, it calls wait(), which atomically releases the lock and suspends the thread (see
Figure 5-23). Saying that the thread suspension and lock release are atomic means that they
happen together, indivisibly from the application’s point of view. After some other thread notifies
t that the buffer may no longer be full, t regains the lock on the buffer and retests the condition.
The standard Java idiom for condition synchronization is the statement:
while (conditionIsNotMet) sharedObject.wait();
Such a wait-loop statement must be inside a synchronized method or block. Any attempt to
invoke the wait() or notify() methods from outside the synchronized code will throw
IllegalMonitorStateException. The above idiom states that the condition test should
always be in a loop—never assume that being woken up means that the condition has been met.
The wait loop blocks the calling thread, t, for as long as the condition is not met. By calling
wait(), t places itself in the shared object’s wait set and releases all its locks on that object. (It
is of note that standard Java implements an unordered “wait set” rather than an ordered “wait
queue.” Real-time specification for Java—RTSJ—corrects this somewhat.)
Chapter 5 
277
Design with Patterns
thrd1 : Thread
loop
shared_obj : Object
thrd2 : Thread
acquire lock
region
Successful:
Checks condition
alt
Condition is met
Does the work
release lock
[else]
wait()
Atomically releases the
lock and waits blocked
region
Lock transfer
controlled by
operating system
and hardware
acquire lock
Successful:
Does the work that can
affect the wait condition
notify()
transfer lock
release lock
Figure 5-23: Condition synchronization pattern for concurrent threads.
A thread that executes a synchronized method on an object, o, and changes a condition that
can affect one or more threads in o’s wait set must notify those threads. In standard Java, the call
o.notify() reactivates one arbitrarily chosen thread, t, in o’s wait set. The reactivated thread
then reevaluates the condition and either proceeds into the critical region or reenters the wait set.
The call to o.notifyAll() releases all threads in the o’s wait set. In standard Java this is the
only way to ensure that the highest priority thread is reactivated. This is inefficient, though,
because all the threads must attempt access while only one will succeed in acquiring the lock and
proceed.
The reader might have noticed resemblance between the above mechanism of wait/notify and the
publish/subscribe pattern of Section 5.1. In fact, they are equivalent conceptually, but there are
some differences due to concurrent nature of condition synchronization.
5.3.4
Concurrent Programming Example
The following example illustrates cooperation between threads.
Example 5.1
Concurrent Home Access
In our case-study, Figure 1-12 shows lock controls both on front and backyard doors. Suppose two
different tenants arrive (almost) simultaneously at the different doors and attempt the access, see
Figure 5-24. The single-threaded system designed in Section 2.7 would process them one-by-one,
which may cause the second user to wait considerable amount of time. A user unfamiliar with the
system intricacies may perceive this as a serious glitch. As shown in Figure 5-24, the processor is idle
most of the time, such as between individual keystrokes or while the user tries to recall the exact
Ivan Marsic
User 1

Rutgers University
278
Service
Arrival
Time
Arrival
Waiting
Service
User 2
Total interaction time for both users
Average interaction time
Figure 5-24: Single-threaded, sequential servicing of users in Example 4.1.
password after an unsuccessful attempt. Meanwhile, the second user is needlessly waiting. To improve
the user experience, let us design a multithreaded solution.
The solution is given next.
The first-round implementation in Section 2.7, considered the system with a single door lock. We
have not yet tackled the architectural issue of running a centralized or a distributed system. In the
former case, the main computer runs the system and at the locks we have only embedded
processors. We could add an extra serial port, daisy-chained with the other one, and the control
would remain as in Section 2.7. In the latter case of a distributed system, each lock would have a
proximal embedded computer. The embedded computers would communicate mutually or with
the main computer using a local area network. The main computer may even not be necessary,
and the embedded processors could coordinate in a “peer-to-peer” mode. Assume for now that we
implement the centralized PC solution with multiple serial ports. We also assume a single
photosensor and a single light bulb, for the sake of simplicity.
The first question to answer is, how many threads we need and which objects should be turned
into threads? Generally, it is not a good idea to add threads indiscriminately, because threads
consume system resources, such as computing cycles and memory space.
It may appear attractive to attach a thread to each object that operates physical devices, such as
LockCtrl and LightCtrl, but is this the right approach? On the other hand, there are only two users
(interacting with the two locks), so perhaps two threads would suffice? Let us roll back, see why
we consider introducing threads in the first place. The reason is to improve the user experience,
so two users at two different doors can access the home simultaneously, without waiting. Two
completely independent threads would work, which would require duplicating all the resources,
but this may be wasteful. Here is the list of resources they could share:

KeyStorage, used to lookup the valid keys

Serial port(s), to communicate with the devices

System state, such as the device status or current count of unsuccessful attempts
Sharing KeyStorage seems reasonable—here it is just looked up, not modified. The serial port can
also be shared because the communication follows a well-defined RS-232 protocol. However,
sharing the system state needs to be carefully examined. Sharing the current count of
Chapter 5 
Design with Patterns
279
unsuccessful attempts seems to make no sense—there must be two counters, each counting
accesses for its corresponding door.
There are several observations that guide our design. From the system sequence diagram of
Figure 2-15(a), we can observe that the system juggles two distinct tasks: user interaction and
internal processing which includes controlling the devices. There are two copies (for the two
doors) of each task, which should be able to run in parallel. The natural point of separation
between the two tasks is the Controller object, Figure 2-27, which is the entry point of the domain
layer of the system. The Controller is a natural candidate for a thread object, so two internal
processing tasks can run in parallel, possibly sharing some resources. The threaded controller
class, ControllerThd, is defined below. I assume that all objects operating the devices
(LockCtrl, LightCtrl, etc.) can be shared as long as the method which writes to the serial port is
synchronized. LockCtrl must also know which lock (front or backyard) it currently operates.
Listing 5-6: Concurrent version of the main class for home access control. Compare to
Listing 2-1.
import
import
import
import
javax.comm.*;
java.io.IOException;
java.io.InputStream;
java.util.TooManyListenersException;
public class HomeAccessControlSystem_2x extends Thread
implements SerialPortEventListener {
protected ControllerThd contrlFront_; // front door controller
protected ControllerThd contrlBack_;
// back door controller
protected InputStream inputStream_; // from the serial port
protected StringBuffer keyFront_ = new StringBuffer();
protected StringBuffer keyBack_ = new StringBuffer();
public static final long keyCodeLen_ = 4; // key code of 4 chars
public HomeAccessControlSystem_2x(
KeyStorage ks, SerialPort ctrlPort
) {
try {
inputStream_ = ctrlPort.getInputStream();
} catch (IOException e) { e.printStackTrace(); }
LockCtrl lkc = new LockCtrl(ctrlPort);
LightCtrl lic = new LightCtrl(ctrlPort);
PhotoObsrv sns = new PhotoObsrv(ctrlPort);
AlarmCtrl ac = new AlarmCtrl(ctrlPort);
contrlFront_ = new ControllerThd(
new KeyChecker(ks), lkc, lic, sns, ac, keyFront_
);
contrlBack_ = new ControllerThd(
new KeyChecker(ks), lkc, lic, sns, ac, keyBack_
);
try {
ctrlPort.addEventListener(this);
} catch (TooManyListenersException e) {
Ivan Marsic

Rutgers University
280
e.printStackTrace(); // limited to one listener per port
}
start(); // start the serial-port reader thread
}
/** The first argument is the handle (filename, IP address, ...)
* of the database of valid keys.
* The second arg is optional and, if present, names
* the serial port. */
public static void main(String[] args) {
...
// same as in Listing 2-1 above
}
/** Thread method; does nothing, just waits to be interrupted
* by input from the serial port. */
public void run() {
while (true) {
try { Thread.sleep(100); }
catch (InterruptedException e) { /* do nothing */ }
}
}
/** Serial port event handler.
* Assume that the characters are sent one by one, as typed in.
* Every character is preceded by a lock identifier (front/back).
*/
public void serialEvent(SerialPortEvent evt) {
if (evt.getEventType() == SerialPortEvent.DATA_AVAILABLE) {
byte[] readBuffer = new byte[5]; // just in case, 5 chars
try {
while (inputStream_.available() > 0) {
int numBytes = inputStream_.read(readBuffer);
// could chk if numBytes == 2 (char + lockId) ...
}
} catch (IOException e) { e.printStackTrace(); }
// Append the new char to a user key, and if the key
// is complete, awaken the corresponding Controller thread
if (inputStream_[0] == 'f') { // from the front door
// If this key is already full, ignore the new chars
if (keyFront_.length() < keyCodeLen_) {
synchronized (keyFront_) { // CRITICAL REGION
keyFront_.append(new String(readBuffer, 1,1));
// If the key just got completed,
// signal the condition to others
if (keyFront_.length() >= keyCodeLen_) {
// awaken the Front door Controller
keyFront_.notify(); //only 1 thrd waiting
}
} // END OF THE CRITICAL REGION
}
} else if (inputStream_[0] == 'b') { // from back door
if (keyBack_.length() < keyCodeLen_) {
synchronized (keyBack_) { // CRITICAL REGION
keyBack_.append(new String(readBuffer, 1, 1));
Chapter 5 
Design with Patterns
281
if (keyBack_.length() >= keyCodeLen_) {
// awaken the Back door Controller
keyBack_.notify();
}
} // END OF THE CRITICAL REGION
} // else, exception ?!
}
}
}
Each Controller object is a thread, and it synchronizes with HomeAccessControlSystem_2x via
the corresponding user key. In the above method serialEvent(), the port reader thread fills
in the key code until completed; thereafter, it ignores the new keys until the corresponding
ControllerThd processes the key and resets it in its run() method, shown below. The reader
should observe the reuse of a StringBuffer to repeatedly build strings, which works here, but
in a general case many subtleties of Java StringBuffers should be considered.
Listing 5-7: Concurrent version of the Controller class. Compare to Listing 2-2.
import javax.comm.*;
public class ControllerThd implements Runnable {
protected KeyChecker checker_;
protected LockCtrl lockCtrl_;
protected LightCtrl lightCtrl_;
protected PhotoObsrv sensor_;
protected AlarmCtrl alarmCtrl_;
protected StringBuffer key_;
public static final long maxNumOfAttempts_ = 3;
public static final long attemptPeriod_ = 600000; // msec [=10min]
protected long numOfAttempts_ = 0;
public ControllerThd(
KeyChecker kc, LockCtrl lkc, LightCtrl lic,
PhotoObsrv sns, AlarmCtrl ac, StringBuffer key
) {
checker_ = kc;
lockCtrl_ = lkc; alarmCtrl_ = ac;
lightCtrl_ = lic; sensor_ = sns; key_ = key;
Thread t = new Thread(this, getName());
t.start();
}
public void run() {
while(true) { // runs forever
synchronized (key_) { // CRITICAL REGION
// wait for the key to be completely typed in
while(key_.length() <
HomeAccessControlSystem_2x.keyCodeLen_) {
try {
key_.wait();
} catch(InterruptedException e) {
throw new RuntimeException(e);
Ivan Marsic

Rutgers University
282
}
}
} // END OF THE CRITICAL REGION
// Got the key, check its validity:
// First duplicate the key buffer, then release old copy
Key user_key = new Key(new String(key_));
key_.setLength(0); // delete code, so new can be entered
checker_.checkKey(user_key); // assume Publish-Subs. vers.
}
}
}
The reader should observe the thread coordination in the above code. We do not want the
Controller to grab a half-ready Key and pass it to the Checker for validation. The Controller will
do so only when notify() is invoked. Once it is done with the key, the Controller resets it to
allow the reader to fill it again.
You may wonder how is it that in ControllerThd.run() we obtain the lock and then loop
until the key is completely typed in—would this not exclude the port reader thread from the
access to the Key object, so the key would never be completed?! Recall that wait() atomically
releases the lock and suspends the ControllerThd thread, leaving Key accessible to the port
reader thread.
It is interesting to consider the last three lines of code in ControllerThd.run(). Copying a
StringBuffer to a new String is a thread-safe operation; so is setting the length of a
StringBuffer. However, although each of these methods acquires the lock, the lock is
released in between and another thread may grab the key object and do something bad to it. In our
case this will not happen, because HomeAccessControlSystem_2x.serialEvent()
checks the length of the key before modifying it, but generally, this is a concern.
Figure 5-25 summarizes the benefit achieved by a multithreaded solution. Notice that there still
may be micro periods of waiting for both users and servicing the user who arrived first may take
longer than in a single-threaded solution. However, the average service time per user is much
shorter, close to the single-user average service time.
Hazards and Performance Penalties
Ideally, we would like that the processor is never idle while there is a task waiting for execution.
As seen in Figure 5-25(b), even with threads the processor may be idle while there are users who
are waiting for service. The question of “granularity” of the shared resource. Or stated differently,
the key issue is how to minimize the length (i.e., processing time) of the critical region.
Solution: Try to narrow down the critical region by lock splitting or using finer-grain locks.
http://www.cs.panam.edu/~meng/Course/CS6334/Note/master/node49.html
http://searchstorage.techtarget.com/sDefinition/0,,sid5_gci871100,00.html
http://en.wikipedia.org/wiki/Thread_(computer_programming)
Chapter 5 
283
Design with Patterns
User 1
Service
Arrival
Time
(a)
Arrival
Service
Waiting
User 2
Total interaction time for both users
Average interaction time
User 1
Service
Arrival
Time
(b)
Waiting
Arrival
Service
User 2
Total interaction time for both users
Average interaction time
Figure 5-25: Benefit of a multithreaded solution. (a) Sequential servicing, copied from
Figure 5-24. (b) Parallel servicing by two threads.
Control of access to shared resources itself can introduce problems, e.g., it can cause deadlock.
5.4 Broker and Distributed Computing
“If computers get too powerful, we can organize them into a committee—that will do them in.”
—Bradley’s Bromide
Let us assume that in our case-study example of home access control the tenants want to remotely
download the list of recent accesses. This requires network communication. The most basic
network programming uses network sockets, which can call for considerable programming skills
(see Appendix B for a brief review). To simplify distributed computing, a set of software
techniques have been developed. These generally go under the name of network middleware.
When both client and server objects reside in the same memory space, the communication is
carried out by simple method calls on the server object (see Figure 1-22). If the objects reside in
different memory spaces or even on different machines, they need to implement the code for
interprocess communication, such as opening network connections, sending messages across the
network, and dealing with many other aspects of network communication protocols. This
significantly increases the complexity of objects. Even worse, in addition to its business logic,
objects obtain responsibility for communication logic which is extraneous to their main function.
Ivan Marsic

Rutgers University
284
Unmarshaling
Marshaling
Middleware
Network
Figure 5-26: Client object invokes a method on a server object. If the message sending
becomes too complex, introducing middleware offloads the communication intricacies off
the software objects. (Compare with Figure 1-22.)
Employing middleware helps to delegate this complexity away out of the objects, see Figure
5-26. A real world example of middleware is the post service—it deals with the intricacies of
delivering letters/packages to arbitrary distances. Another example is the use of different metric
systems, currencies, spoken languages, etc., in which case the functionality for conversion
between the systems is offloaded to middleware services. Middleware is a good software
engineering practice that should be applied any time the communication between objects becomes
complex and starts rivaling the object’s business logic in terms of the implementation code size.
Middleware is a collection of objects that offer a set of services related to object communication,
so that extraneous functionality is offloaded to the middleware. In general, middleware is
software used to make diverse applications work together smoothly. The process of deriving
middleware is illustrated in Figure 5-27. We start by introducing the proxies for both
communicating objects. (The Proxy pattern is described in Section 5.2.4.) The proxy object B of
B acts so to provide an illusion for A that it is directly communicating with B. The same is
provided to B by A. Having the proxy objects keeps simple the original objects, because the
proxies provide them with an illusion that nothing changed from the original case, where they
communicated directly with each other, as in Figure 5-27(a). In other words, proxies provide
location transparency for the objects—objects remain (almost) the same no matter whether they
Chapter 5 
285
Design with Patterns
(a)
Object
Object
AA
Object
Object
BB
Middleware
(b)
Object
Object
AA
Object
Object
B'
B'
Object
Object
A'
A'
Object
Object
BB
Figure 5-27: Middleware design. Objects A and B are the proxies of A and B, respectively.
reside in the same address space or in different address spaces / machines. Objects A' and B'
comprise the network middleware.
Because it is not likely that we will develop middleware for only two specific objects
communicating, further division of responsibilities results in the Broker pattern.
Ivan Marsic

Rutgers University
Middleware
(a)
CC
286
Broker component
S'
S'
BBclient
client
BBserver
server
Service
Service
C'
C'
SS
Service
Service
Broker
Knowing Responsibilities:
• Registry of name-to-reference mappings
Doing Responsibilities:
• Maintains the registry and provides lookup
• Instantiates proxies
• Network transport of request and result back
(b)
Client
+ callServer()
# useBrokerAPI()
Server
MiddlewareService
(c)
«server proxy»
Stub
+
+
#
#
request()
forwardResponse()
marshal()
unmarshal()
+
+
+
#
#
initialize()
mainEventLoop()
request()
registerService()
useBrokerAPI()
Broker
+
+
+
+
#
#
mainEventLoop()
registerService()
forwardRequest()
forwardResponse()
findServer()
findClient()
«client proxy»
Skeleton
+ forwardRequest()
# marshal()
# unmarshal()
Figure 5-28: (a) The Broker component of middleware represents the intersection of
common proxy functions, along with middleware services. (b) Broker’s employee card. (c)
The Broker pattern class diagram. The server proxy, called Stub, resides on the same host
as the client and the client proxy, called Skeleton, resides on the same host as the server.
5.4.1
Broker Pattern
The Broker pattern is an architectural pattern used to structure distributed software systems with
components that interact by remote method calls, see Figure 5-28. The proxies are responsible
only for the functions specific to individual objects for which they act as substitutes. In a
distributed system the functions that are common to all or most of the proxies are delegated to the
Broker component, Figure 5-28(b). Figure 5-28(c) shows the objects constituting the Broker
pattern and their relationships. The proxies act as representatives of their corresponding objects in
Chapter 5 
287
Design with Patterns
: Client
: Stub
: Broker
: Skeleton
: Server
callServer()
request()
marshal()
forwardRequest()
findServer()
forwardRequest()
unmarshal()
request()
response
marshal()
forwardResponse()
findClient()
forwardResponse()
unmarshal()
response
Figure 5-29: Sequence diagram for a client call to the server (remote method invocation).
the foreign address space and contain all the network-related code. The broker component is
responsible for coordinating communication and providing links to various middleware services.
Although Broker is shown as a single class in Figure 5-28(c), actual brokers consist of many
classes are provide many services.
To use a remote object, a client first finds the object through the Broker, which returns a proxy
object or Stub. As far as the client is concerned, the Stub appears and works like any other local
object because they both implement the same interface. But, in reality it only arranges the method
call and associated parameters into a stream of bytes using the method marshal(). Figure 5-29
shows the sequence diagram for remote method invocation (also called remote procedure call—
RPC) via a Broker. The Stub marshals the method call into a stream of bytes and invokes the
Broker, which forwards the byte stream to the client’s proxy, Skeleton, at the remote host. Upon
receiving the byte stream, the Skeleton rearranges this stream into the original method call and
associated parameters, using the method unmarshal(), and invokes this method on the server
which contains the actual implementation. Finally, the server performs the requested operation
and sends back the return value(s), if any.
The Broker pattern has been proven effective tool in distributed computing, because it leads to
greater flexibility, maintainability, and adaptability of the resulting system. By introducing new
components and delegating responsibility for communication intricacies, the system becomes
potentially distributable and scalable. Java Remote Method Invocation (RMI), which is presented
next, is an example of elaborating and implementing the Broker pattern.
Ivan Marsic

5.4.2
Java Remote Method Invocation (RMI)
Rutgers University
288
The above analysis indicates that the Broker component is common to all remotely
communicating objects, so it needs to be implemented only once. The proxies are object-specific
and need to be implemented for every new object. Fortunately, the process of implementing the
proxy objects can be made automatic. It is important to observe that the original object shares the
same interface with its Stub and Skeleton (if the object acts as a client, as well). Given the
object’s interface, there are tools to automatically generate source code for proxy objects. In the
general case, the interface is defined in an Interface Definition Language (IDL). Java RMI uses
plain Java for interface definition, as well. The basic steps for using RMI are:
1. Define the server object interface
2. Define a class that implements this interface
3. Create the server process
4. Create the client process
Going back to our case-study example of home access control, now I will show how a tenant
could remotely connect and download the list of recent accesses.
Step 1: Define the server object interface
The server object will be running on the home computer and currently offers a single method
which returns the list of home accesses for the specified time interval:
Listing 5-8: The Informant remote server object interface.
/* file Informant.java */
import java.rmi.Remote;
import java.rmi.RemoteException;
import java.util.Vector;
public interface Informant extends Remote {
public Vector getAccessHistory(long fromTime, long toTime)
throws RemoteException;
}
This interface represents a contract between the server and its clients. Our interface must extend
java.rmi.Remote which is a “tagging” interface that contains no methods. Any parameters
of the methods of our interface, such as the return type Vector in our case, must be serializable,
i.e., implement java.io.Serializable.
In the general case, an IDL compiler would generate a Stub and Skeleton files for the Informant
interface. With Java RMI, we just use the Java compiler, javac:
% javac Informant.java
Chapter 5 
289
Design with Patterns
Remote
UnicastRemoteObject
Informant
Client
+ getAccessHistory()
InformantImpl_Stub
InformantImpl
RemoteRef
+ getAccessHistory()
+ getAccessHistory()
Figure 5-30: Class diagram for the Java RMI example. See text for details.
Step 2: Define a class that implements this interface
The implementation class must extend class java.rmi.server.RemoteObject or one of
its
subclasses.
In
practice,
most
implementations
extend
the
subclass
java.rmi.server.UnicastRemoteObject, because this class supports point-to-point
communication using the TCP protocol. The class diagram for this example is shown in Figure
5-30. The implementation class must implement the interface methods and a constructor (even if
it has an empty body). I adopt the common convention of adding Impl onto the name of our
interface to form the implementation class.
Listing 5-9: The class InformantImpl implements the actual remote server object.
/* file InformantImpl.java (Informant server implementation) */
import java.rmi.RemoteException;
import java.rmi.server.UnicastRemoteObject;
import java.util.Vector;
public class InformantImpl extends UnicastRemoteObject
implements Informant {
protected Vector accessHistory_ = new Vector();
/** Constructor; currently empty. */
public InformantImpl() throws RemoteException { }
/** Records all the home access attempts.
* Called from another class, e.g., from KeyChecker, Listing 2-2
* @param access Contains the entered key, timestamp, etc.
*/
public void recordAccess(String access) {
accessHistory_.add(access);
}
/** Implements the "Informant" interface. */
public Vector getAccessHistory(long fromTime, long toTime)
throws RemoteException {
Vector result = new Vector();
// Extract from accessHistory_ accesses in the
// interval [fromTime, toTime] into "result"
Ivan Marsic

Rutgers University
290
return result;
}
}
Here, we first use the Java compiler, javac, then the RMI compiler, rmic:
% javac InformantImpl.java
% rmic –v1.2 InformantImpl
The first statement compiles the Java file and the second one generates the stub and skeleton
proxies, called InformantImpl_Stub.class and InformantImpl_Skel.class,
respectively. It is noteworthy, although perhaps it might appear strange, that the RMI compiler
operates on a .class file, rather than on a source file. In JDK version 1.2 or higher (Java 2),
only the stub proxy is used; the skeleton is incorporated into the server itself so that there are no
separate entities as skeletons. Server programs now communicate directly with the remote
reference layer. This is why the command line option -v1.2 should be employed (that is, if you
are working with JDK 1.2 or higher), so that only the stub file is generated.
As shown in Figure 5-30, the stub is associated with a RemoteRef, which is a class in the RMI
Broker that represents the handle for a remote object. The stub uses a remote reference to carry
out a remote method invocation to a remote object via the Broker. It is instructive to look inside
the InformantImpl_Stub.java, which is obtained if the RMI compiler is run with the
option -keep. Here is the stub file (the server proxy resides on client host):
Listing 5-10: Proxy classes (Stub and Skeleton) are automatically generated by the Java
rmic compiler from the Informant interface (Listing 5-8).
// Stub class generated by rmic, do not edit.
// Contents subject to change without notice.
Marshaling
1 public final class InformantImpl_Stub
2
extends java.rmi.server.RemoteStub
3
implements Informant, java.rmi.Remote
4 {
5
private static final long serialVersionUID = 2;
6
7
private static java.lang.reflect.Method
$method_getAccessHistory_0;
8
9
static {
10
try {
11
$method_getAccessHistory_0 =
Informant.class.getMethod("getAccessHistory", new java.lang.Class[]
{long.class, long.class});
12
} catch (java.lang.NoSuchMethodException e) {
13
throw new java.lang.NoSuchMethodError(
14
"stub class initialization failed");
15
}
16
}
17
18
// constructors
19
public InformantImpl_Stub(java.rmi.server.RemoteRef ref) {
20
super(ref);
21
}
22
Chapter 5 
Design with Patterns
291
23
// methods from remote interfaces
24
25
// implementation of getAccessHistory(long, long)
26
public java.util.Vector getAccessHistory(long $param_long_1,
long $param_long_2)
27
throws java.rmi.RemoteException
28
{
29
try {
30
Object $result = ref.invoke(this,
$method_getAccessHistory_0, new java.lang.Object[] {new
java.lang.Long($param_long_1), new java.lang.Long($param_long_2)}, 7208692514216622197L);
31
return ((java.util.Vector) $result);
32
} catch (java.lang.RuntimeException e) {
33
throw e;
34
} catch (java.rmi.RemoteException e) {
35
throw e;
36
} catch (java.lang.Exception e) {
37
throw new java.rmi.UnexpectedException("undeclared checked
exception", e);
38
}
39
}
40 }
The code description is as follows:
Line 2: shows that our stub extends the RemoteStub class, which is the common superclass
to client stubs and provides a wide range of remote reference semantics, similar to the broker
services in Figure 5-28(a).
Lines 7–15: perform part of the marshaling process of the getAccessHistory() method
invocation. Computational reflection is employed, which is described in Section 7.3.
Lines 19–21: pass the remote server’s reference to the RemoteStub superclass.
Line 26: starts the definition of the stub’s version of the getAccessHistory() method.
Line 30: sends the marshaled arguments to the server and makes the actual call on the remote
object. It also gets the result back.
Line 31: returns the result to the client.
The reader should be aware that, in terms of how much of the Broker component is revealed in a
stub code, this is only a tip of the iceberg. The Broker component, also known as Object Request
Broker (ORB), can provide very complex functionality and comprise many software objects.
Step 3: Create the server process
The server process instantiates object(s) of the above implementation class, which accept remote
requests. The first problem is, how does a client get handle of such an object, so to be able to
invoke a method on it? The solution is for the server to register the implementation objects with a
naming service known as registry. A naming registry is like a telephone directory. The RMI
Registry is a simple name repository which acts as a central management point for Java RMI. The
registry must be run before the server and client processes using the following command line:
Ivan Marsic

Rutgers University
292
% rmiregistry
It can run on any host, including the server’s or client’s hosts, and there can be several RMI
registries running at the same time. (Note: The RMI Registry is an RMI server itself.) For every
server object, the registry contains a mapping between the well-known object’s name and its
reference (usually a globally unique sequence of characters). The process of registration is called
binding. The client object can, thus, get handle of a server object by looking up its name in the
registry. The lookup is performed by supplying a URL with protocol rmi:
rmi://host_name:port_number/object_name
where host_name is the name or IP address of the host on which the RMI registry is running,
port_number is the port number of the RMI registry, and object_name is the name bound to the
server implementation object. If the host name is not provided, the default is assumed as
localhost. The default port number of RMI registry is 1099, although this can be changed as
desired. The server object, on the other hand, listens to a port on the server machine. This port is
usually an anonymous port that is chosen at runtime by the JVM or the underlying operating
system. Or, you can request the server to listen on a specific port on the server machine.
I will use the class HomeAccessControlSystem, defined in Listing 2-1, Section 2.7, as the
main server class. The class remains the same, except for several modifications:
Listing 5-11: Refactored HomeAccessControlSystem class (from Listing 2-1) to
instantiate remote server objects of type Informant.
1 import java.rmi.Naming;
2
3 public class HomeAccessControlSystem extends Thread
4
implements SerialPortEventListener {
5
...
6
private static final String RMI_REGISTRY_HOST = "localhost";
7
8
public static void main(String[] args) throws Exception {
9
...
10
InformantImpl temp = new InformantImpl();
11
String rmiObjectName =
12
"rmi://" + RMI_REGISTRY_HOST + "/Informant";
13
Naming.rebind(rmiObjectName, temp);
14
System.out.println("Binding complete...");
15
...
16
}
...
}
The code description is as follows:
Lines 3–5: The old HomeAccessControlSystem class as defined in Section 2.7.
Line 6: For simplicity’s sake, I use localhost as the host name, which could be omitted
because it is default.
Line 8: The main() method now throws Exception to account for possible RMI-related
exceptions thrown by Naming.rebind().
Line 10: Creates an instance object of the server implementation class.
Chapter 5 
Design with Patterns
293
Lines 11–12: Creates the string URL which includes the object’s name, to be bound with its
remote reference.
Line 13: Binds the object’s name to the remote reference at the RMI naming registry.
Line 14: Displays a message for our information, to know when the binding is completed.
The server is, after compilation, run on the computer located at home by invoking the Java
interpreter on the command line:
% java HomeAccessControlSystem
If Java 2 is used, the skeleton is not necessary; otherwise, the skeleton class,
InformantImpl_Skel.class, must be located on the server’s host because, although not
explicitly used by the server, the skeleton will be invoked by Java runtime.
Step 4: Create the client process
The client requests services from the server object. Because the client has no idea on which
machine and to which port the server is listening, it looks up the RMI naming registry. What it
gets back is a stub object that knows all these, but to the client the stub appears to behave same as
the actual server object. The client code is as follows:
Listing 5-12: Client class InformantClient invokes services on a remote Informant
object.
1 import java.rmi.Naming;
2
3 public class InformantClient {
4
private static final String RMI_REGISTRY_HOST = " ... ";
5
6
public static void main(String[] args) {
7
try {
8
Informant grass = (Informant) Naming.lookup(
9
"rmi://" + RMI_REGISTRY_HOST + "/Informant"
10
);
11
Vector accessHistory =
12
grass.getAccessHistory(fromTime, toTime);
13
14
System.out.println("The retrieved history follows:");
15
for (Iterator i = accessHistory; i.hasNext(); ) {
16
String record = (String) i.next();
17
System.out.println(record);
18
}
19
} catch (ConnectException conEx) {
20
System.err.println("Unable to connect to server!");
21
System.exit(1);
22
} catch (Exception ex) {
23
ex.printStackTrace();
24
System.exit(1);
25
}
26
...
27
}
...
}
Ivan Marsic

Rutgers University
294
The code description is as follows:
Line 4: Specifies the host name on which the RMI naming registry is running.
Line 8: Looks up the RMI naming registry to get handle of the server object. Because the
lookup returns a java.rmi.Remote reference, this reference must be typecast into an
Informant reference (not InformantImpl reference).
Lines 11–12: Invoke the service method on the stub, which in turn invokes this method on the
remote server object. The result is returned as a java.util.Vector object named
accessHistory.
Lines 14–18: Display the retrieved history list.
Lines 19–25: Handle possible RMI-related exceptions.
The client would be run from a remote machine, say from the tenant’s notebook or a PDA. The
InformantImpl_Stub.class file must be located on the client’s host because, although
not explicitly used by the client, a reference to it will be given to the client by Java runtime. A
security-conscious practice is to make the stub files accessible on a website for download. Then,
you set the java.rmi.server.codebase property equal to the website’s URL, in the
application which creates the server object (in our example above, this is
HomeAccessControlSystem ). The stubs will be downloaded over the Web, on demand.
The reader should notice that distributed object computing is relatively easy using Java RMI. The
developer is required to do just slightly more work, essentially to bind the object with the RMI
registry on the server side and to obtain a reference to it on the client side. All the complexity
associated with network programming is hidden by the Java RMI tools.
SIDEBAR 5.2: How Transparent Object Distribution?
 The reader who experiments with Java RMI, see e.g., Problem 5.22, and tries to implement
the same with plain network sockets (see Appendix B), will appreciate how easy it is to work
with distributed objects. My recollection from using some CORBA object request brokers was
that some provided even greater transparency than Java RMI. Although this certainly is an
advantage, there are perils of making object distribution so transparent that it becomes too
easy.
The problem is that people tended to forget that there is a network between distributed objects
and built applications that relied on fine-grained communication across the network. Too many
round-trip communications led to poor performance and reputation problems for CORBA. An
interesting discussion of object distribution issues is available in [Waldo et al., 1994] from the
same developers who authored Java RMI [Wollrath et al., 1996].
Chapter 5 
Design with Patterns
295
5.5 Information Security
“There is nothing special about security; it’s just part of getting the job done.” —Rob Short
Information security is a nonfunctional property of the system, it is an emergent property. Owing
to different types of information use, there are two main security disciplines. Communication
security is concerned with protecting information when it is being transported between different
systems. Computer security is related to protecting information within a single system, where it
can be stored, accessed, and processed. Although both disciplines must work in accord to
successfully protect information, information transport faces greater challenges and so
communication security has received greater attention. Accordingly, this review is mainly
concerned with communication security. Notice that both communication- and computer security
must be complemented with physical (building) security as well as personnel security. Security
should be thought of as a chain that is as strong as its weakest link.
The main objectives of information security are:

Confidentiality: ensuring that information is not disclosed or revealed to unauthorized
persons

Integrity: ensuring consistency of data, in particular, preventing unauthorized creation,
modification, or destruction of data

Availability: ensuring that legitimate users are not unduly denied access to resources,
including information resources, computing resources, and communication resources

Authorized use: ensuring that resources are not used by unauthorized persons or in
unauthorized ways
To achieve these objectives, we institute various safeguards, such as concealing (encryption)
confidential information so that its meaning is hidden from spying eyes; and key management
which involves secure distribution and handling of the “keys” used for encryption. Usually, the
complexity of one is inversely proportional to that of the other—we can afford relatively simple
encryption algorithm with a sophisticated key management method.
Ivan Marsic

Padlock
and shared
key copy
Rutgers University
296
Content
Shared
key copy
Message
Sender
Sender needs:
• receive securely a copy of the shared key
• positively identify the message receiver
Intermediary
Threats posed by intruder/adversary:
• forge the key and view the content
• damage/substitute the padlock
• damage/destroy the message
• observe characteristics of messages
(statistical and/or metric properties)
Receiver
Receiver needs:
• receive securely a shared key copy
• positively identify the message sender
• detect any tampering with messages
Figure 5-31: Communication security problem: Sender needs to transport a confidential
document to Receiver over an untrusted intermediary.
Figure 5-31 illustrates the problem of transmitting a confidential message by analogy with
transporting a physical document via untrusted carrier. The figure also lists the security needs of
the communicating parties and the potential threats posed by intruders. The sender secures the
briefcase, and then sends it on. The receiver must use a correct key in order to unlock the
briefcase and read the document. Analogously, a sending computer encrypts the original data
using an encryption algorithm to make it unintelligible to any intruder. The data in the original
form is known as plaintext or cleartext. The encrypted message is known as ciphertext. Without
a corresponding “decoder,” the transmitted information (ciphertext) would remain scrambled and
be unusable. The receiving computer must regenerate the original plaintext from the ciphertext
with the correct decryption algorithm in order to read it. This pair of data transformations,
encryption and decryption, together forms a cryptosystem.
There are two basic types of cryptosystems: (i) symmetric systems, where both parties use the
same (secret) key in encryption and decryption transformations; and, (ii) public-key systems, also
known as asymmetric systems, where the parties use two related keys, one of which is secret and
the other one can be publicly disclosed. I first review the logistics of how the two types of
cryptosystems work, while leaving the details of encryption algorithms for the next section.
Chapter 5 
297
Design with Patterns
Sender’s
padlock
Receiver’s
padlock
1. Sender secures the briefcase
with his/her padlock and sends
Sender
2. Receiver additionally secures
Receiver
the briefcase with his/her
padlock and returns
3. Sender removes his/her
padlock and sends again
4. Receiver removes his/her
padlock to access the content
Figure 5-32: Secure transmission via untrustworthy carrier. Note that both sender and
receiver keep their own keys with them all the time—the keys are never exchanged.
5.5.1
Symmetric and Public-Key Cryptosystems
In symmetric cryptosystems, both parties use the same key in encryption and decryption
transformations. The key must remain secret and this, of course, implies trust between the two
parties. This is how cryptography traditionally works and prior to the late 1970s, these were the
only algorithms available.
The system works as illustrated in Figure 5-31. In order to ensure the secrecy of the shared key,
the parties need to meet prior to communication. In this case, the fact that only the parties
involved know the secret key implicitly identifies one to the other.
Using encryption involves two basic steps: encoding a message, and decoding it again. More
formally, a code takes a message M and produces a coded form f(M). Decoding the message
requires an inverse function f 1 , such that f 1  f ( M )  = M. For most codes, anyone who
knows how to perform the first step also knows how to perform the second, and it would be
unthinkable to release to the adversary the method whereby a message can be turned into code.
Merely by “undoing” the encoding procedures, the adversary would be able to break all
subsequent coded messages.
In the 1970s Ralph Merkle, Whitfield Diffie, and Martin Hellman realized that this need not be
so. The weasel word above is “merely.” Suppose that the encoding procedure is very hard to
undo. Then it does no harm to release its details. This led them to the idea of a trapdoor function.
We call f a trapdoor function if f is easy to compute, but f 1 is very hard, indeed impossible for
practical purposes. A trapdoor function in this sense is not a very practical code, because the
legitimate user finds it just as hard to decode the message as the adversary does. The final twist is

Ivan Marsic
Rutgers University
“Public key”
298
Receiver distributes his/her padlock (unlocked)
to sender ahead of time, but keeps the key
“Private key”
Receiver
Receiver’s
padlock (unlocked)
Receiver’s
key
Sender uses the receiver’s padlock
to secure the briefcase and sends
Sender
Receiver removes his/her
padlock to access the content
Figure 5-33: Public-key cryptosystem simplifies the procedure from Figure 5-32.
to define f in such a way that a single extra piece of information makes the computation of f 1
easy. This is the only bit of information that must be kept secret.
This alternative approach is known as public-key cryptosystems. To understand how it works, it is
helpful to examine the analogy illustrated in Figure 5-32. The process has three steps. In the first
step, the sender secures the briefcase with his or her padlock and sends. Second, upon receiving
the briefcase, the receiver secures it additionally with their own padlock and returns. Notice that
the briefcase is now doubly secured. Finally, the sender removes his padlock and re-sends. Hence,
sender manages to send a confidential document to the receiver without needing the receiver’s
key or surrendering his or her own key.
There is an inefficiency of sending the briefcase back and forth, which can be avoided as
illustrated in Figure 5-33. Here we can skip steps 1 and 2 if the receiver distributed his/her
padlock (unlocked, of course!) ahead of time. When the sender needs to send a document, i.e.,
message, he/she simply uses the receiver’s padlock to secure the briefcase and sends. Notice that,
once the briefcase is secured, nobody else but receiver can open it, not even the sender. Next I
describe how these concepts can be implemented for electronic messages.
5.5.2
Cryptographic Algorithms
Encryption has three aims: keeping data confidential; authenticating who sends data; and,
ensuring data has not been tampered with. All cryptographic algorithms involve substituting one
thing for another, which means taking a piece of plaintext and then computing and substituting
the corresponding ciphertext to create the encrypted message.
Chapter 5 
Design with Patterns
299
Symmetric Cryptography
The Advanced Encryption Standard has a fixed block size of 128 bits and a key size of 128, 192,
and 256 bits.
Public-Key Cryptography
As stated above, f is a trapdoor function if f is easy to compute, but f 1 is very hard or
impossible for practical purposes. An example of such difficulty is factorizing a given number n
into prime numbers. An encryption algorithm that depends on this was invented by Ron Rivest,
Adi Shamir, and Leonard Adelman (RSA system) in 1978. Another example algorithm, designed
by Taher El Gamal in 1984, depends on the difficulty of the discrete logarithm problem.
In the RSA system, the receiver does as follows:
1. Randomly select two large prime numbers p and q, which always must be kept secret.
2. Select an integer number E, known as the public exponent, such that (p  1) and E have
no common divisors, and (q  1) and E have no common divisors.
3. Determine the product n = pq, known as public modulus.
4. Determine the private exponent, D, such that (ED  1) is exactly divisible by both (p  1)
and (q  1). In other words, given E, we choose D such that the integer remainder when
ED is divided by (p  1)(q  1) is 1.
5. Release publicly the public key, which is the pair of numbers n and E, K = (n, E). Keep
secret the private key, K = (n, D).
Because a digital message is a sequence of digits, break it into blocks which, when considered as
numbers, are each less than n. Then it is enough to encode block by block.
Encryption works so that the sender substitutes each plaintext block B by the ciphertext C = BE %
n, where % symbolizes the modulus operation. (The modulus of two integer numbers x and y,
denoted as x % y, is the integer remainder when x is divided by y.)
Then the encrypted message C (ciphertext) is transmitted. To decode, the receiver uses the
decoding key D, to compute B = CD % n, that is, to obtain the plaintext B from the ciphertext C.
Example 5.2
RSA cryptographic system
As a simple example of RSA, suppose the receiver chooses p = 5 and q = 7. Obviously, these are too
small numbers to provide any security, but they make the presentation manageable. Next, the receiver
chooses E = 5, because 5 and (5  1)(7  1) have no common factors. Also, n = pq = 35. Finally, the
1
144
receiver chooses D = 29, because ( pE1)D(q11)  529
46  24  6 , i.e., they are exactly divisible. The
receiver’s public key is K = (n, E) = (35, 5), which is made public. The private key K = (n, D) = (35,
29) is kept secret.
Now, suppose that the sender wants to send the plaintext “hello world.” The following table shows the
encoding of “hello.” I assign to letters a numeric representation, so that a=1, b=2, …, y=25, and z=26,
and I assume that block B is one letter long. In an actual implementation, letters are represented as
binary numbers, and the blocks B are not necessarily aligned with letters, so that plaintext “l” will not
always be represented as ciphertext “12.”
Ivan Marsic

Rutgers University
Plaintext letter
h
e
l
l
o
300
Plaintext numeric
representation
8
5
12
12
15
BE
Ciphertext BE % n
85 = 32768
55 = 3125
125 = 248832
248832
155 = 759375
85 % 35 = 8
55 % 35 = 10
125 % 35 = 17
17
155 % 35 = 15
The sender transmits this ciphertext to the receiver: 8, 10, 17, 17, 15. Upon the receipt, the receiver
decrypts the message as shown in the following table.
Ciphertext
CD
B = CD % n
8
10
17
17
15
829 = 154742504910672534362390528
100000000000000000000000000000
481968572106750915091411825223071697
481968572106750915091411825223071697
12783403948858939111232757568359375
829 % 35 = 8
5
12
12
15
Plaintext
letter
h
e
l
l
o
We can see that even this toy example produces some extremely large numbers.
The point is that while the adversary knows n and E, he or she does not know p and q, so they
cannot work out (p  1)(q  1) and thereby find D. The designer of the code, on the other hand,
knows p and q because those are what he started from. So does any legitimate receiver: the
designer will have told them. The security of this system depends on exactly one thing: the
notoriously difficulty of factorizing a given number n into primes. For example, given n = 267  1
it took three years working on Sundays for F. N. Cole to find by hand in 1903 p and q for n = pq
= 193707721  761838257287. It would be waste of time (and often combinatorially selfdefeating) for the program to grind through all possible options.
Encryption strength is measured by the length of its “key,” which is expressed in bits. The larger
the key, the greater the strength of the encryption. Using 112 computers, a graduate student
decrypted one 40-bit encrypted message in a little over 7 days. In contrast, data encrypted with a
128-bit key is 309,485,009,821,345,068,724,781,056 times stronger than data encrypted with a
40-bit key. RSA Laboratories recommends that the product of p and q be on the order of 1024
bits long for corporate use and 768 bits for use with less valuable information.
5.5.3
Authentication
5.5.4
Program Security
A virus is malicious code carried from one computer to another by some medium—often an
“infected” file. Any operating system that allows third-party programs to run can theoretically run
viruses. Some operating systems are more secure than others; earlier versions of Microsoft
Chapter 5 
Design with Patterns
301
Windows did not even provide something as simple as maintain memory space separation. Once
on a computer, a virus is executed when its carrier file is “opened” in some meaningful way by
software on that system. When the virus executes, it does something unwanted, such as causing
software on the host system to send more copies of infected files to other computers over the
network, infecting more files, and so on. In other words, a virus typically maximizes its
likelihood of being passed on, making itself contagious.
Viral behavior relies on security vulnerabilities that exist in software running on the host system.
For example, in the past, viruses have often exploited security vulnerabilities in Microsoft Office
macro scripting capabilities. Macro viruses are no longer among the most common virus types.
Many viruses take advantage of Trident, the rendering engine behind Internet Explorer and
Windows Explorer that is also used by almost every piece of Microsoft software. Windows
viruses often take advantage of image-rendering libraries, SQL Server’s underlying database
engine, and other components of a complete Windows operating system environment as well.
Viruses are typically addressed by antivirus software vendors. These vendors produce virus
definitions used by their antivirus software to recognize viruses on the system. Once a specific
virus is detected, the software attempts to quarantine or remove the virus—or at least inform the
user of the infection so that some action may be taken to protect the system from the virus.
This method of protection relies on knowledge of the existence of a virus, however, which means
that most of the time a virus against which you are protected has, by definition, already infected
someone else’s computer and done its damage. The question you should be asking yourself at this
point is how long it will be until you are the lucky soul who gets to be the discoverer of a new
virus by way of getting infected by it.
It’s worse than that, though. Each virus exploits a vulnerability — but they don’t all have to
exploit different vulnerabilities. In fact, it’s common for hundreds or even thousands of viruses to
be circulating “in the wild” that, between them, only exploit a handful of vulnerabilities. This is
because the vulnerabilities exist in the software and are not addressed by virus definitions
produced by antivirus software vendors.
These antivirus software vendors’ definitions match the signature of a given virus — and if
they’re really well-designed might even match similar, but slightly altered, variations on the virus
design. Sufficiently modified viruses that exploit the same vulnerability are safe from recognition
through the use of virus definitions, however. You can have a photo of a known bank robber on
the cork bulletin board at the bank so your tellers will be able to recognize him if he comes in —
but that won’t change the fact that if his modus operandi is effective, others can use the same
tactics to steal a lot of money.
By the same principle, another virus can exploit the same vulnerability without being recognized
by a virus definition, as long as the vulnerability itself isn’t addressed by the vendor of the
vulnerable software. This is a key difference between open source operating system projects and
Microsoft Windows: Microsoft leaves dealing with viruses to the antivirus software vendors, but
open source operating system projects generally fix such vulnerabilities immediately when
they’re discovered.
Thus, the main reason you don’t tend to need antivirus software on an open source system, unless
running a mail server or other software that relays potentially virus-laden files between other
systems, isn’t that nobody’s targeting your open source OS; it’s that any time someone targets it,
Ivan Marsic

Rutgers University
302
chances are good that the vulnerability the virus attempts to exploit has been closed up — even if
it’s a brand-new virus that nobody has ever seen before. Any half-baked script-kiddie has the
potential to produce a new virus that will slip past antivirus software vendor virus definitions, but
in the open source software world one tends to need to discover a whole new vulnerability to
exploit before the “good guys” discover and patch it.
Viruses need not simply be a “fact of life” for anyone using a computer. Antivirus software is
basically just a dirty hack used to fill a gap in your system’s defenses left by the negligence of
software vendors who are unwilling to invest the resources to correct certain classes of security
vulnerabilities.
The truth about viruses is simple, but it’s not pleasant. The truth is that you’re being taken to the
cleaners — and until enough software users realize this, and do something about it, the software
vendors will continue to leave you in this vulnerable state where additional money must be paid
regularly to achieve what protection you can get from a dirty hack that simply isn’t as effective as
solving the problem at the source would be.
However, we should not forget that security comes at a cost.
In theory, application programs are supposed to access hardware of the computer only through
the interfaces provided by the operating system. But many application programmers who dealt
with small computer operating systems of the 1970s and early 1980s often bypassed the OS,
particularly in dealing with the video display. Programs that directly wrote bytes into video
display memory run faster than programs that didn't. Indeed, for some applications—such as
those that needed to display graphics on the video display—the operating system was totally
inadequate.
What many programmers liked most about MS-DOS was that it “stayed out of the way” and let
programmers write programs as fast as the hardware allowed. For this reason, popular software
that ran on the IBM PC often relied upon idiosyncrasies of the IBM PC hardware.
5.6 Summary and Bibliographical Notes
Design patterns are heuristics for structuring the software modules and their interactions that are
proven in practice. They yield in design for change, so the change of the computing environment
has as minimal and as local effect on the code as possible.
Key Points:

Pattern use must be need-driven: use a pattern only when you need it to improve your
software design, not because it can be used, or you simply like hitting nails with your
new hammer.
Chapter 5 

Design with Patterns
303
Using the Broker pattern, a client object invokes methods of a remote server object,
passing arguments and receiving a return value with each call, using syntax similar to
local method calls. Each side requires a proxy that interacts with the system’s runtime.
There are many known design patterns and I have reviewed above only few of the major ones.
The text that most contributed to the popularity of patterns is [Gamma et al., 1995]. Many books
are available, perhaps the best known are [Gamma et al., 1995] and [Buschmann et al., 1996].
The reader can also find a great amount of useful information on the web. In particular, a great
deal of information is available in Hillside.net’s Patterns Library: http://hillside.net/patterns/ .
R. J. Wirfs-Brock, “Refreshing patterns,” IEEE Software, vol. 23, no. 3, pp. 45-47, May/June
2006.
Section 5.1: Indirect Communication: Publisher-Subscriber
Section 5.2: More Patterns
Section 5.3: Concurrent Programming
Concurrent systems are a large research and practice filed and here I provide only the
introductory basics. Concurrency methods are usually not covered under design patterns and it is
only for the convenience sake that here they appear in the section on software design patterns. I
avoided delving into the intricacies of Java threads—by no means is this a reference manual for
Java threads. Concurrent programming in Java is extensively covered in [Lea, 2000] and a short
review is available in [Sandén, 2004].
[Whiddett, 1987]
Pthreads tutorial: http://www.cs.nmsu.edu/~jcook/Tools/pthreads/pthreads.html
Pthreads tutorial from CS6210 (by Phillip Hutto):
http://www.cc.gatech.edu/classes/AY2000/cs6210_spring/pthreads_tutorial.htm
Section 5.4: Broker and Distributed Computing
The Broker design pattern is described in [Buschmann et al., 1996; Völter et al., 2005].
Java RMI:
Sun Developer Network (SDN) jGuru: “Remote Method Invocation (RMI),” Sun Microsystems,
Inc., Online at: http://java.sun.com/developer/onlineTraining/rmi/RMI.html
http://www.javaworld.com/javaworld/jw-04-2005/jw-0404-rmi.html
http://www.developer.com/java/ent/article.php/10933_3455311_1
Ivan Marsic

Rutgers University
304
Although Java RMI works only if both client and server processes are coded in the Java
programming language, there are other systems, such as CORBA (Common Object Request
Broker Architecture), which work with arbitrary programming languages, including Java. A
readable appraisal of the state of affairs with CORBA is available in [Henning, 2006].
Section 5.5: Information Security
In an increasingly networked world, all computer users are at risk of having their personally
identifying information and private access data intercepted. Even if information is not stolen,
computing resources may be misused for criminal activities facilitated by unauthorized access to
others’ computer systems.
Kerckhoffs’ Principle states that a cryptosystem should remain secure even if everything about it
other than the key is public knowledge. The security of a system’s design is in no way dependent
upon the secrecy of the design, in and of itself. Because system designs can be intercepted, stolen,
sold, independently derived, reverse engineered by observations of the system’s behavior, or just
leaked by incompetent custodians, the secrecy of its design can never really be assumed to be
secure itself. Hence, the “security through obscurity” security model by attempting to keep
system design secret Open source movement even advocates widespread access to the design of a
system because more people can review the system’s design and detect potential problems.
Transparency ensures that the security problems tend to arise more quickly, and to be addressed
more quickly. Although an increased likelihood of security provides no guarantees of success, it
is beneficial nonetheless.
There is an entire class of software, known as “fuzzers,” that is used to quickly detect potential
security weaknesses by feeding abusive input at a target application and observing its behavior
under that stress. These are the tools that malicious security crackers use all the time to find ways
to exploit software systems. Therefore, it is not necessary to have access to software design (or its
source code) to be able to detect its security vulnerabilities. This should not be surprising, given
that software defects are rarely found by looking at source code. (Recall the software testing
techniques from Section 2.7.) Where access to source code becomes much more important is
when trying to determine why a particular weakness exists, and how to remove it. One might
conclude, then, that the open source transparency does not contribute as much to detecting
security problems as it does to fixing them.
Cryptography [Menezes et al., 1997], which is available for download, entirely, online at
http://www.cacr.math.uwaterloo.ca/hac/.
ICS 54: History of Public-Key Cryptography:
http://www.ics.uci.edu/~ics54/doc/security/pkhistory.html
http://www.netip.com/articles/keith/diffie-helman.htm
http://www.rsasecurity.com/rsalabs/node.asp?id=2248
http://www.scramdisk.clara.net/pgpfaq.html
http://postdiluvian.org/~seven/diffie.html
http://www.sans.org/rr/whitepapers/vpns/751.php
http://www.fors.com/eoug97/papers/0356.htm
Chapter 5 
Design with Patterns
305
Class iaik.security.dh.DHKeyAgreement
http://www.cs.utexas.edu/users/chris/cs378/f98/resources/iaikdocs/iaik.security.dh.DHKeyAgree
ment.html
Bill Steele, “‘Fabric’ would tighten the weave of online security,” Cornell Chronicle (09/30/10):
Fabric’s programming language, which is based on Java, builds in security as the program is
written. Myers says most of what Fabric does is transparent to the programmer.
http://www.news.cornell.edu/stories/Sept10/Fabric.html
P. Dourish, R. E. Grinter, J. Delgado de la Flor, and M. Joseph, “Security in the wild: user
strategies for managing security as an everyday, practical problem,” Personal and Ubiquitous
Computing (ACM/Springer), vol. 8, no. 6, pp. 391-401, November 2004.
M. J. Ranum, “Security: The root of the problem,” ACM Queue (Special Issue: Surviving
Network Attacks), vol. 2, no. 4, pp. 44-49, June 2004.
H. H. Thompson and R. Ford, “Perfect storm: The insider, naivety, and hostility,” ACM Queue
(Special Issue: Surviving Network Attacks), vol. 2, no. 4, pp. 58-65, June 2004.
introducing trust and its pervasiveness in information technology
Microsoft offers integrated hardware-level security such as data execution prevention, kernel
patch protection and its free Security Essentials software:
http://www.microsoft.com/security_essentials/
Microsoft's 'PassPort' Out, Federation Services In
In 2004 Microsoft issued any official pronouncements on "TrustBridge," its collection of
federated identity-management technologies slated to go head-to-head with competing
technologies backed by the Liberty Alliance.
http://www.eweek.com/c/a/Windows/Microsofts-Passport-Out-Federated-Services-In/
Problems
Problem 5.1
Problem 5.2
Consider the online auction site described in Problem 2.31 (Chapter 2). Suppose you want to
employ the Publish-Subscribe (also known as Observer) design pattern in your design solution for
Problem 2.31. Which classes should implement the Publisher interface? Which classes should
Ivan Marsic

Rutgers University
306
implement the Subscriber interface? Explain your answer. (Note: You can introduce new classes
or additional methods on the existing classes if you feel it necessary for solution.)
Problem 5.3
In the patient-monitoring scenario of Problem 2.35 (Chapter 2), assume that multiple recipients
must be notified about the patient condition. Suppose that your software is to use the PublishSubscribe design pattern. Identify the key software objects and draw a UML interaction diagram
to represent how software objects in the system could accomplish the notification problem.
Problem 5.4
Problem 5.5
Problem 5.6: Elevator Control
L
2
Down
Consider the elevator control problem defined in Problem 3.7 (Chapter 3). Your task is
to determine whether the Publisher-Subscriber design pattern can be applied in this
design. Explain clearly your answer. If the answer is yes, identify which classes are
suitable for the publisher role and which ones are suitable for the subscriber role.
Explain your choices, list the events generated by the Publishers, and state explicitly
for each Subscriber to which events it is subscribed to.
Problem 5.7
Problem 5.8
Problem 5.9
Consider the automatic patient monitoring system described in Problem 2.35. Carefully
examine the draft UML sequence diagram in Figure 2-45. Check if the given design
already uses some patterns and explain your claim. Identify as many opportunities as you
can to improve the design by applying design patterns. Consider how an unnecessary
application of some design patterns would make this design worse. Draw UML sequence
diagrams or write pseudo-code to describe the proposed design. Always describe your motivation
for adopting or rejecting design modifications.
3
4
5
6 7
Up
Chapter 5 
Design with Patterns
307
Problem 5.10
Consider the system for inventory management grocery supermarket from Problem 2.15. Suppose
you are provided with an initial software design as follows. This design is based on a basic
version of the inventory system, but the reader should be aware of extensions that are discussed in
the solution of Problem 2.15(c) and Problem 2.16. The software consists of the following classes:
ReaderIface:
This class receives messages from RFID readers that specific tags moved in or out of
coverage.
DBaseConn:
This class provides a connection to a relational database that contains data about shelf
stock and inventory tasks. The database contains several tables, including
ProductsInfo[key
=
tagID], PendingTasks[key
=
userID],
CompletedTasks, and Statistics[key = infoType] for various information
types, such as the count of erroneous messages from RFID readers and the count of
reminders sent for individual pending tasks.
Dispatcher:
This class manages inventory tasks by opening new tasks when needed and generates
notifications to the concerned store employees.
Monitor:
This class periodically keeps track of potentially overdue tasks. It retrieves the list of
pending tasks from the database and generates reminders to the concerned store
employees.
Messenger:
This class sends email notifications to the concerned store employees. (The notifications
are generated by other classes.)
Assume that email notifications are used as a supplementary tool, but the system must keep an
internal record of sent notifications and pending tasks, so it can take appropriate actions.
Notice that the current design has a single timer for the whole system. The software designer
noticed that sending notifications for overdue tasks does not need to be exactly timed in this
system. Delays up to a certain period (e.g., hour or even day) are tolerable. Maintaining many
timers would be overkill and would significantly slow down the system. It would not be able to
do important activities, such as processing RFID events in a timely fashion. Therefore, the
software is designed so that, when a new pending task is created, there is no explicit activation of
an associated timer. Instead, the task is simply added to the list of pending tasks. The Monitor
object periodically retrieves this list and checks for overdue tasks, as seen below in the design for
the use case UC-5 SendReminder.
Another simplification is to check only for “out-of-stock” events and not for “low-stock” events.
If the customer demands that “low-stock” events be included, then the design of the software-tobe will become somewhat more complex.
Ivan Marsic

Rutgers University
308
The UML sequence diagrams for all the use cases are shown in the following figures. Notice that
use cases UC-3, UC-4, and UC-6 «include» UC-7: Login (user authentication), which is not
shown to avoid clutter.
UC1: RemoveItem
rfid :
ReaderIface
receive(event)
opt
: Messenger
: Dispatcher
: DBaseConn
prodInfo := getProductInfo( tagID )
[prodInfo == nil] recordStatistics( "error: unknown tagID" )
return
decrement( prodCount )
alt
[prodCount < 0]
recordStatistics( "error: negative prodCount" )
send( "error: negative prodCount" )
email to store manager
return
[prodCount  Threshold]
recordProductInfo( updated product count )
createTask( "out-of-stock" )
recordPendingTask( task info )
send( “alert: out-of-stock" )
email to store manager
[else]
Pending Task Info:
Task-type = "out-of-stock"
Assigned-time = current time
Assigned-to = Store Manager
recordProductInfo( updated product count )
return
In the design for UC-1, the system may check if a pending task for the given product already
exists in the database; if yes, it should not generate a new pending task for the same product.
Chapter 5 
309
Design with Patterns
UC2: AddItem
rfid :
ReaderIface
receive(event)
: DBaseConn
prodInfo := getProductInfo( tagID )
[prodInfo == nil]
alt
recordStatistics( "error: unknown tagID" )
return
[else]
increment( prodCount )
recordProductInfo( updated product count )
return
UC3: ViewPendingWork
user interface :
: DBaseConn
: Dispatcher
view pending
getPendingTasks( userID )
alt
[userID == manager]
tasksList := getPendingTasks( ALL )
[userID == associate]
tasksList := getPendingTasks( associateID )
show error
show tasks
[else]
return error
return tasksList
ref
UC4: AssignReplenishTask
Extension:
Store Manager may
optionally run UC4
to assign a pending task
Ivan Marsic

Rutgers University
310
UC4: AssignReplenishTask
user interface :
: Dispatcher
: Messenger
: DBaseConn
view pending
ref
UC3: ViewPendingWork
assign task
Store Manager runs
UC3 to view pending tasks
and selects one to assign
assignTask( taskID, associateID, … )
taskInfo := getPendingTaskInfo( taskID )
[ taskInfo != nil && taskType == "out-of-stock" ]
opt
seq
Pending Task Info:
Task-type = "replenish-shelf"
Assigned-time = current time
Assigned-to = Store Associate
recordPendingTask( task info )
removePendingTask( taskID )
send( "alert: task assigned" )
email to store associate
show result
return result
The [seq] interaction fragment specifies that the interactions contained within the fragment box
must occur exactly in the given order. The reason for using this constraint in UC-4 is that the
system may crash while the task is being converted from unassigned to pending. If
removePendingTask() were called first, and the system crashed after it but before
recordPendingTask(), then all information about this task would be lost! Depending on the
database implementation, it may be possible to perform these operations as atomic for they
update the same table in the database. To deal with crash scenarios where a task ends up in both
tables, the Monitor object in UC-5 SendReminder should be extended to perform a database
clean-up after a crash. It should remove those tasks from the PendingTasks table that are
marked both as unassigned and pending.
Chapter 5 
311
Design with Patterns
UC5: SendReminder
: Monitor
wakeup
loop
: Dispatcher
: Messenger
: DBaseConn
pendingList := getPendingTasks( )
[ for every task in pendingList ]
opt
[ (currentTime  task.assignTime)  thresholdPendingInterval ]
increment task.remindAttempts
recordPendingTask( task info )
alt
[ task.remindAttempts < maxAttempts ]
receivers := task.assignedTo
receivers := ALL employees (system-wide)
sendReminder( receivers, task )
increment task.remindAttempts
recordStatistics( task reminder alert info )
send( “alert:: " + task.ID + " overdue" )
email to receivers
set sleep period
return
Note that the Monitor discards the list of pending tasks before going to sleep, so it starts every
cycle with a fresh list of pending tasks, retrieved from the database, because our assumption is
that the database contains the most current information (possibly updated by other objects).
By examining the design for the use case UC-5 SendReminder, we see that the Monitor has to do
a lot of subtractions and comparisons every time it wakes up, but this can be done at leisure
because seconds or minutes are not critical for this activity. The computing power should be
better used for other use cases. Of course, we must ensure that the Monitor still works fast enough
not to introduce delays on the order of hours or days during each cycle!
In addition, we need to handle the case where the time values are not incrementing constantly
upwards, such as when a full 24 hours passes and a new day starts, the time resets to zero. In
Java, using java.lang.System.currentTimeMillis() returns the current time in
milliseconds as a long integer.
Ivan Marsic

Rutgers University
312
In the solution of Problem 2.16 we discussed using adaptive timeout calculation to adjust the
frequency of reminders for busy periods. Another option is to have shorter sleep periods, but
during each wakeup, process only part of the list of pending tasks, and leave the rest for
subsequent wakeups. Then cycle again from the head of the list. This way, the reminders will be
spread over time and not all reminders will be generated at once (avoid generating one “bulk”
notification each period).
UC6: ReplenishCompleted
: Dispatcher
close(taskID)
: Messenger
: DBaseConn
taskInfo := getPendingTask( taskID )
[taskInfo == nil]
alt
return error
[else] prodInfo := getProductInfo( taskInfo.getProductID( ) )
[prodCount  Threshold]
alt
return error
[else]
seq
recordCompletedTask( taskInfo )
removePendingTask( taskID )
send( "alert: task completed" )
email to store manager
return
The logic of UC-6 is that it first retrieves the task, checks if such a task exists, makes sure it is
really done, and finally marks it as completed. The [seq] interaction fragment specifies that the
interactions contained within the fragment box must occur exactly in the given order. Similar to
UC-4 AssignReplenishTask, this constraint is needed in case the system crashes while the task is
being closed. If removePendingTask() were called first, and the system crashed after it but
before recordCompletedTask(), then all information about this task would be lost! These
operations cannot be performed as atomic, because they work on different tables in the database.
To deal with crash scenarios where a task ends up in both tables, the Monitor object in UC-5
should be modified to perform a database clean-up after a crash. It should remove those tasks
from the PendingTasks table that are already in the CompletedTasks table.
Chapter 5 
Design with Patterns
313
Notice also that the Monitor runs in a separate thread, so while UC-6 is in the process of closing a
task, the Monitor may send an unnecessary reminder about this task (in UC-5).
C
arefully examine the existing design and identify as many opportunities as you can to
improve the design by applying design patterns. Note that the existing design ignores the
issue of concurrency, but we will leave the multithreading issue aside for now and focus only on
the patterns that improve the quality of software design. (The concurrency issues will be
considered later in Problem 5.20.)
(a) If you introduce a pattern, first provide arguments why the existing design may be
problematic.
(b) Provide as much details as possible about how the pattern will be implemented and how the
new design will work (draw UML sequence diagrams or write pseudo-code).
(c) Explain how the pattern improved the design (i.e., what are the expected benefits compared to
the original design).
If considering future evolution and extensions of the system when proposing a modification, then
describe explicitly what new features will likely be added and how the existing design would be
inadequate to cope with resulting changes. Then introduce a design pattern and explain how the
modified version is better.
If you believe that the existing design (or some parts of it) is sufficiently good then explain how
the application of some design patterns would make the design worse. Use concrete examples and
UML diagrams or pseudo-code to illustrate and refer to specific qualities of software design.
Problem 5.11
Problem 5.12
Problem 5.13
Problem 5.14
Problem 5.15
In Section 5.3, it was stated that the standard Java idiom for condition synchronization is the
statement:
while (condition) sharedObject.wait();
(a) Is it correct to substitute the yield() method call for wait()? Explain your answer
and discuss any issues arising from the substitution.
Ivan Marsic 
Rutgers University
314
(b) Suppose that if substitutes for while, so we have:
if (condition) sharedObject.wait()
Is this correct? Explain your answer.
Problem 5.16
Parking lot occupancy monitoring, see Figure 5-34. Consider a parking lot with the total number
of spaces equal to capacity. There is a single barrier gate with two poles, one for the entrance
and the other for the exit. A computer in the barrier gate runs a single program which controls
both poles. The program counts the current number of free spaces, denoted by occupancy, such
that
0  occupancy  capacity
When a new car enters, the occupancy is incremented by one; conversely, when a car exits, the
occupancy is decremented by one. If occupancy equals capacity, the red light should turn
on to indicate that the parking is full.
In order to be able to serve an entering and an exiting patron in parallel, you should design a
system which runs in two threads. EnterThread controls the entrance gate and ExitThread
controls the exit gate. The threads share the occupancy counter so to correctly indicate the
parking-full state. Complete the UML sequence diagram in Figure 5-35 that shows how the two
threads update the shared variable, i.e., occupancy.
FULL
Figure 5-34: Parking lot occupancy monitoring, Problem 5.16.
Chapter 5 
315
Design with Patterns
: EnterThread
: SharedState
: ExitThread
Car
Car
requestEntry()
requestExit()
Figure 5-35: UML diagram template for parking lot occupancy monitoring, Problem 5.16.
Cook 2
Pickup
counter
Supplier
Waiter
Egg tray
Cook 1
Figure 5-36: Concurrency problem in a restaurant scenario, Problem 5.17.
Hint: Your key concern is to maintain the consistent shared state (occupancy) and indicate
when the parking-full sign should be posted. Extraneous actions, such as issuing the ticket for an
entering patron and processing the payment for an exiting patron, should not be paid attention—
only make a high-level remark where appropriate.
Problem 5.17
Consider a restaurant scenario shown in Figure 5-36. You are to write a simulation in Java such
that each person runs in a different thread. Assume that each person takes different amount of
time to complete their task. The egg tray and the pickup counter have limited capacities, Neggs and
Nplates, respectively. The supplier stocks the egg tray but must wait if there are no free slots.
Likewise, the cooks must hold the prepared meal if the pickup counter is full.
Problem 5.18
A priority inversion occurs when a higher-priority thread is waiting for a lower-priority thread to
finish processing of a critical region that is shared by both. Although higher-priority threads
normally preempt lower-priority threads, this is not possible when both share the same critical
region. While the higher-priority thread is waiting, a third thread, whose priority is between the
first two, but it does not share the critical region, preempts the low-priority thread. Now the
Ivan Marsic 
Rutgers University
316
higher-priority thread is waiting for more than one lower-priority thread. Search the literature and
describe precisely a possible mechanism to avoid priority inversion.
Problem 5.19
Assume that the patient device described in Problem 2.3 (at the end of Chapter 2) runs in a multithreaded mode, where different threads acquire and process data from different sensors. (See also
Problem 2.35 and its solution on the back of this book.) What do you believe is the optimal
number of threads? When designing this system, what kind of race conditions or other
concurrency issues can you think of? Propose a specific solution for each issue that you identify
(draw UML sequence diagrams or write pseudo-code).
Problem 5.20
Consider the supermarket inventory management system from Problem 5.10. A first observation
is that the existing design ignores the issue of concurrency—there will be many users
simultaneously removing items, and/or several associates may be simultaneously restocking the
shelves. Also, it is possible that several employees may simultaneously wish to view pending
tasks, assign replenishment tasks, or report replenishment completed. Clearly, it is necessary to
introduce multithreading even if the present system will never be extended with new features.
Modify the existing design and introduce multithreading.
Problem 5.21
Problem 5.22
Use Java RMI to implement a distributed Publisher-Subscriber design pattern.
Requirements: The publisher and subscribers are to be run on different machines. The naming
server should be used for rendezvous only; after the first query to the naming server, the publisher
should cache the contact information locally.
Handle sudden (unannounced) departures of subscribers by implementing a heartbeat protocol.
Problem 5.23
Suppose that you are designing an online grocery store. The only supported payment method will
be using credit cards. The information exchanges between the parties are shown in Figure 5-37.
After making the selection of items for purchase, the customer will be prompted to enter
information about his/her credit card account. The grocery store (merchant) should obtain this
information and relay it to the bank for the transaction authorization.
In order to provide secure communication, you should design a public-key cryptosystem as
follows. All messages between the involved parties must be encrypted for confidentiality, so that
only the appropriate parties can read the messages. Even the information about the purchased
items, payment amount, and the outcome of credit-card authorization request should be kept
confidential. Only the initial catalog information is not confidential.
Chapter 5 
317
Design with Patterns
Customer
Merchant
Bank
enter selection (“items catalog“)
place order (“selected items")
enter credit card info (“payment amount“)
process payment (“card info")
approve transaction (“card info“, “payment amount")
notify outcome (“result value“)
notify outcome (“result value“)
Figure 5-37: Information exchanges between the relevant parties. The quoted variables in
the parentheses represent the parameters that are passed on when the operation is invoked.
The credit card information must be encrypted by the customer so that only the bank can read it—
the merchant should relay it without being able to view the credit card information. For the sake
of simplicity, assume that all credit cards are issued by a single bank.
The message from the bank containing binary decision (“approved” or “rejected”) will be sent to
the merchant, who will forward it securely to the customer. Both the merchant and customer
should be able to read it.
Answer the following questions about the cryptosystem that is to be developed:
(a) What is the (minimum) total number of public-private key pairs ( K i , K i ) that must be
issued? In other words, which actors need to possess a key pair, or perhaps some actors
need more than one pair?
(b) For each key pair i, specify which actor should issue this pair, to whom the public key
K i should be distributed, and at what time (prior to each shopping session or once for
multiple sessions). Provide an explanation for your answer!
(c) For each key pair i, show which actor holds the public key K i and which actor holds the
private key K i .
(d) For every message in Figure 5-37, show exactly which key K i / K i should be used in the
encryption of the message and which key should be used in its decryption.
Ivan Marsic 
Rutgers University
318
Problem 5.24
In the Model-View-Controller design pattern, discuss the merits of having Model subscribe to the
Controller using the Publish-Subscribe design pattern? Argue whether Controller should
subscribe to the View?
Chapter 6
XML and Data Representation
Contents
“Description is always a bore, both to the describer and the
describee.”
—Disraeli
XML defines a standard way to add markup to documents,
which facilitates representation, storage, and exchange of
information. A markup language is a mechanism to identify
parts of a document and to describe their logical relationships.
XML stands for “eXtensible Markup Language” (extensible
because it is not a fixed set of markup tags like HTML—
HyperText Markup Language). The language is standardized
by the World Wide Web Consortium (W3C), and all relevant
information is available at: http://www.w3.org/XML/.
Structured information contains both content (words, pictures,
etc.) and metadata describing what role that content plays (for
example, content in a section heading has a different meaning
from content in a footnote, which means something different
than content in a figure caption or content in a database table,
etc.).
XML is not a single, predefined markup language: it is a metalanguage—language for defining other languages—which lets
you design your own markup language. A predefined markup
language like HTML defines a specific vocabulary and
grammar to describe information, and the user is unable to
modify or extend either of these. Conversely, XML, being a
meta-language for markup, lets you design a new markup
language, with tags and the rules of their nesting that best suite
your problem domain.
6.1 Structure of XML Documents
6.1.1
6.1.2
6.1.3
6.1.4
Syntax
Document Type Definition (DTD)
Namespaces
XML Parsers
6.2 XML Schemas
6.2.1
6.2.2
6.2.3
6.2.4
6.2.5
XML Schema Basics
Models for Structured Content
Datatypes
Reuse
RELAX NG Schema Language
6.3 Indexing and Linking
6.3.1 XPointer and Xpath
6.3.2 XLink
6.3.3
6.2.4
6.4 Document Transformation and XSL
6.4.1
6.4.2
6.4.3
6.4.4
6.5
6.5.1
6.5.2
6.5.3
6.5.4
6.6
6.6.1
6.6.2
6.6.3
6.7 Summary and Bibliographical Notes
Problems
Why cover XML in a basic software engineering text? Because so far we dealt only with program
development, but neglected data. [Brooks’s comment about code vs. data.] If you are writing
software, you are inevitably representing structured information, whether it is a configuration file,
documentation, or program’s input/output data. You need to specify how data is represented and
exchanged, i.e., the data format. But there is more to it.
319
Ivan Marsic

Rutgers University
320
The perception and nature of the term “document” has changed over the time. In the past, a
document was a container of static information and it was often an end result of an application.
Recently, the nature of documents changed from passive to active. The documents themselves
became “live” applications. Witness the client-side event-handling scripting languages in Web
browsers, such as JavaScript. Moreover, great deal of a program’s business logic can be encoded
separately, as data, rather than hard-coded as instructions. Cf. data-driven design and reflection,
Chapter 7 below.
Structured Documents and Markup
letter’s body
recipient’s
address
sender’s
address
The word “document” refers not only to traditional documents, like a book or an article, but also
to the numerous of other “data collections.” These include vector graphics, e-commerce
transactions, mathematical equations, object meta-data, server APIs, and great many other kinds
of structured information. Generally, structured document is a document in which individual parts
are identified and they are arranged in a certain pattern. Documents having structured information
include both the content as well as what the content stands for. Consider these two documents:
Unstructured Document
Structured Document
<letter>
<sender>
Mr. Charles Morse
<name>Mr. Charles Morse</name>
13 Takeoff Lane
<address>
Talkeetna, AK 99676
<street>13 Takeoff Lane</street>
<city>Talkeetna</city> <state>AK</state>
29 February, 1997 date
<postal-code>99676</postal-code>
</address>
Mrs. Robinson
</sender>
1 Entertainment Way
<date>29 February, 1997</date>
Los Angeles, CA 91011
<recipient>
<name>Mrs. Robinson</name>
Dear Mrs. Robinson, salutation
<address>
<street>1 Entertainment Way</street>
Here’s part of an update on my first day at
<city>Los Angeles</city> <state>CA</state>
the edge. I hope to relax and sow my wild
<postal-code>91011</postal-code>
oats.
</address>
</recipient>
<salutation>Dear Mrs. Robinson,</ salutation>
<body>
Here’s part of an update …
</body>
closing
<closing>Sincerely,</closing>
<signature>Charlie</signature>
signature
</letter>
You probably guessed that the document on the left is a correspondence letter, because you are
familiar with human conventions about composing letters. (I highlighted the prominent parts of
the letter by gray boxes.) Also, postal addresses adhere to certain templates as well. Even if you
never heard of a city called Talkeetna, you can quickly recognize what part of the document
appears to represent a valid postal address. But, if you are a computer, not accustomed to human
conventions, you would not know what the text contains. On the other hand, the document on the
Chapter 6 
XML and Data Representation
321
right has clearly identified (marked) parts and their sub-parts. Parts are marked up with tags that
indicate the nature of the part they surround. In other words, tags assign meaning/semantics to
document parts. Markup is a form of metadata, that is, document’s dictionary capturing
definitions of document parts and the relationships among them.
Having documents marked up enables automatic data processing and analysis. Computer can be
applied to extract relevant and novel information and present to the user, check official forms
whether or not they are properly filled out by users, etc. XML provides a standardized platform to
create and exchange structured documents. Moreover, XML provides a platform for specifying
new tags and their arrangements, that is, new markup languages.
XML is different from HTML, although there is a superficial similarity. HTML is a concrete and
unique markup language, while XML is a meta-language—a system for creating new markup
languages. In a markup language, such as HTML, both the tag set and the tag semantics are fixed
and only those will be recognized by a Web browser. An <h1> is always a first level heading and
the tag <letter> is meaningless. The W3C, in conjunction with browser vendors and the
WWW community, is constantly working to extend the definition of HTML to allow new tags to
keep pace with changing technology and to bring variations in presentation (stylesheets) to the
Web. However, these changes are always rigidly confined by what the browser vendors have
implemented and by the fact that backward compatibility is vital. And for people who want to
disseminate information widely, features supported only by the latest release of a particular
browser are not useful.
XML specifies neither semantics nor a tag set. The tags and grammar used in the above example
are completely made up. This is the power of XML—it allows you to define the content of your
data in a variety of ways as long as you conform to the general structure that XML requires. XML
is a meta-language for describing markup languages. In other words, XML provides a facility to
define tags and the structural relationships between them. Since there is no predefined tag set,
there cannot be any preconceived semantics. All of the semantics of XML documents will be
defined by the applications that process them.
A document has both a logical and a physical structure. The logical structure allows a document
to be divided into named parts and sub-parts, called elements. The physical structure allows
components of the document, called entities, to be named and stored separately, sometimes in
other data files so that information can be reused and non-textual data (such as images) can be
included by reference. For example, each chapter in a book may be represented by an element,
containing further elements that describe each paragraph, table and image, but image data and
paragraphs that are reused (perhaps from other documents) are entities, stored in separate files.
XML Standard
XML is defined by the W3C in a number of related specifications available here:
http://www.w3.org/TR/. Some of these include:

Extensible Markup Language (XML), current version 1.1 (http://www.w3.org/XML/Core/) –
Defines the syntax of XML, i.e., the base XML specification.
Ivan Marsic

Rutgers University
322

Namespaces (http://www.w3.org/TR/xml-names11) – XML namespaces provide a simple
method for qualifying element and attribute names used in Extensible Markup Language
documents by associating them with namespaces identified by URI references.

Schema (http://www.w3.org/XML/Schema) – The schema language, which is itself
represented in XML, provides a superset of the capabilities found in XML document type
definitions (DTDs). DTDs are explained below.

XML Pointer Language (XPointer) (http://www.w3.org/TR/xptr/) and XML Linking
Language (XLink) (http://www.w3.org/TR/xlink/) – Define a standard way to represent links
between resources. In addition to simple links, like HTML’s <A> tag, XML has
mechanisms for links between multiple resources and links between read-only resources.
XPointer describes how to address a resource, XLink describes how to associate two or
more resources.

XML Path Language (XPath) (http://www.w3.org/TR/xpath20/) – Xpath is a language for
addressing parts of an XML document, designed to be used by both XSLT and XPointer.

Extensible Stylesheet Language (XSL) (http://www.w3.org/TR/xsl/) and XSL
Transformations (XSLT) (http://www.w3.org/TR/xslt/) – Define the standard stylesheet
language for XML.
Unlike programming languages, of which there are many, XML is universally accepted by all
vendors. The rest of the chapter gives a brief overview and relevant examples.
6.1 Structure of XML Documents
6.1.1
Syntax
Syntax defines how the words of a language are arranged into phrases and sentences and how
components (like prefixes and suffixes) are combined to make words. XML documents are
composed of markup and content—content (text) is hierarchically structured by markup tags.
There are six kinds of markup that can occur in an XML document: elements, entity references,
comments, processing instructions, marked sections, and document type declarations. The
following subsections introduce each of these markup concepts.
Elements
Elements indicate logical parts of a document and they are the most common form of markup. An
element is delimited by tags which are surrounded by angle brackets (“<”, “>” and “</”, “/>”).
The tags give a name to the document part they surround—the element name should be given to
convey the nature or meaning of the content. A non-empty element begins with a start-tag,
<tag>, and ends with an end-tag, </tag>. The text between the start-tag and end-tag is called
the element’s content. In the above example of a letter document, the element
Chapter 6 
323
XML and Data Representation
<salutation>Dear Mrs. Robinson,</salutation> indicates the salutation part of
the letter. Rules for forming an element name are:

Must start with a letter character

Can include all standard programming language identifier
[0-9A-Za-z] as well as underscore _, hyphen -, and colon :

Is case sensitive, so <name> and <Name> are different element names
characters,
i.e.,
Some elements may be empty, in which case they have no content. An empty element can begin
and end at the same place in which case it is denoted as <tag/>. Elements can contain subelements. The start tag of an element can have, in addition to the element name, related (attribute,
value) pairs. Elements can also have mixed content where character data can appear alongside
subelements, and character data is not confined to the deepest subelements. Here is an example:
<salutation>Dear <name>Mrs. Robinson</name>, </salutation>
Notice the text appearing between the element <salutation> and its child element <name>.
Attributes
Attributes are name-value pairs that occur inside start-tags after the element name. A start tag can
have zero or more attributes. For example,
<date format="English_US">
is an element named date with the attribute format having the value English_US, meaning
that month is shown first and named in English. Attribute names are formed using the same rules
as element names (see above). In XML, all attribute values must be quoted. Both single and
double quotes can be used, provided they are correctly matched.
Entities and Entity References
XML reserves some characters to distinguish markup from plain text (content). The left angle
bracket, <, for instance, identifies the beginning of an element’s start- or end-tag. To support the
reserved characters as part of content and avoid confusion with markup, there must be an
alternative way to represent them. In XML, entities are used to represent these reserved
characters. Entities are also used to refer to often repeated or varying text and to include the
content of external files. In this sense, entities are similar to macros.
Every entity must have a unique name. Defining your own entity names is discussed in the
section on entity declarations (Section 6.1.2 below). In order to use an entity, you simply
reference it by name. Entity references begin with the ampersand and end with a semicolon, like
this &entityname;. For example, the lt entity inserts a literal < into a document. So to
include the string <non-element> as plain text, not markup, inside an XML document all
reserved characters should be escaped, like so &lt;non-element&gt;.
A special form of entity reference, called a character reference, can be used to insert arbitrary
Unicode characters into your document. This is a mechanism for inserting characters that cannot
be typed directly on your keyboard.
Ivan Marsic

Rutgers University
324
Character references take one of two forms: decimal references, &#8478;, and hexadecimal
references, &#x211E;. Both of these refer to character number U+211E from Unicode (which is
the standard Rx prescription symbol).
Comments
A comment begins with the characters <!-- and ends with -->. A comment can span multiple
lines in the document and contain any data except the literal string “--.” You can place
comments anywhere in your document outside other markup. Here is an example:
<!-- ********************
My comment is imminent.
-->
Comments are not part of the textual content of an XML document and the parser will ignore
them. The parser is not required to pass them along to the application, although it may do so.
Processing Instructions
Processing instructions (PIs) allow documents to contain instructions for applications that will
import the document. Like comments, they are not textually part of the XML document, but this
time around the XML processor is required to pass them to an application.
Processing instructions have the form: <?name pidata?>. The name, called the PI target,
identifies the PI to the application. For example, you might have <?font start italic?>
and <?font end italic?>, which indicate the XML processor to start italicizing the text
and to end, respectively.
Applications should process only the targets they recognize and ignore all other PIs. Any data that
follows the PI target is optional; it is for the application that recognizes the target. The names
used in PIs may be declared as notations in order to formally identify them. Processing instruction
names beginning with xml are reserved for XML standardization.
CDATA Sections
In a document, a CDATA section instructs the parser to ignore the reserved markup characters. So,
instead of using entities to include reserved characters in the content as in the above example of
&lt;non-element&gt;, we can write:
<![CDATA[ <non-element> ]]>
Between the start of the section, <![CDATA[ and the end of the section, ]]>, all character data
are passed verbatim to the application, without interpretation. Elements, entity references,
comments, and processing instructions are all unrecognized and the characters that comprise them
are passed literally to the application. The only string that cannot occur in a CDATA section is
“]]>”.
Chapter 6 
XML and Data Representation
325
Document Type Declarations (DTDs)
Document type declarations (DTDs) are reviewed in Section 6.1.2 below. DTD is used mainly to
define constraints on the logical structure of documents, that is, the valid tags and their
arrangement/ordering.
This is about as much as an average user needs to know about XML. Obviously, it is simple and
concise. XML is designed to handle almost any kind of structured data—it constrains neither the
vocabulary (set of tags) nor the grammar (rules of how the tags combine) of the markup language
that the user intends to create. XML allows you to create your own tag names. Another way to
think of it is that XML only defines punctuation symbols and rules for forming “sentences” and
“paragraphs,” but it does not prescribe any vocabulary of words to be used. Inventing the
vocabulary is left to the language designer.
But for any given application, it is probably not meaningful for tags to occur in a completely
arbitrary order. From a strictly syntactic point of view, there is nothing wrong with such an XML
document. So, if the document is to have meaning, and certainly if you are writing a stylesheet or
application to process it, there must be some constraint on the sequence and nesting of tags,
stating for example, that a <chapter> that is a sub-element of a <book> tag, and not the other
way around. These constraints can be expressed using an XML schema (Section 6.2 below).
XML Document Example
The letter document shown initially in this chapter can be represented in XML as follows:
Listing 6-1: Example XML document of a correspondence letter.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
<?xml version="1.0" encoding="UTF-8"?>
<!-- Comment: A personal letter marked up in XML. -->
<letter language="en-US" template="personal">
<sender>
<name>Mr. Charles Morse</name>
<address kind="return">
<street>13 Takeoff Lane</street>
<city>Talkeetna</city><state>AK</state>
<postal-code>99676</postal-code>
</address>
</sender>
<date format="English_US">February 29, 1997</date>
<recipient>
<name>Mrs. Robinson</name>
<address kind="delivery">
<street>1 Entertainment Way</street>
<city>Los Angeles</city><state>CA</state>
<postal-code>91011</postal-code>
</address>
</recipient>
<salutation style="formal">Dear Mrs. Robinson,</ salutation>
<body>
Here's part of an update ...
</body>
<closing>Sincerely,</closing>
Ivan Marsic

Rutgers University
326
XML language for letters, variant 1
<address kind="return">
<street>13 Takeoff Lane</street>
<city>Talkeetna</city>
<state>AK</state>
<postal-code>99676</postal-code>
</address>
XML language for letters, variant 2
<address kind="return“
street="13 Takeoff Lane“
city="Talkeetna"
state="AK“
zip="99676" />
Translator
Figure 6-1: Different XML languages can be defined for the same domain and/or concepts.
In such cases, we need a “translator” to translate between those languages.
26
27
<signature>Charlie</signature>
</letter>
Line 1 begins the document with a processing instruction <?xml ... ?>. This is the XML
declaration, which, although not required, explicitly identifies the document as an XML
document and indicates the version of XML to which it was authored.
A variation on the above example is to define the components of a postal address (lines 6–9 and
14–17) as element attributes:
<address kind="return" street="13 Takeoff Lane" city="Talkeetna"
state="AK" postal-code="99676" />
Notice that this element has no content, i.e., it is an empty element. This produces a more concise
markup, particularly suitable for elements with well-defined, simple, and short content.
One quickly notices that XML encourages naming the elements so that the names describe the
nature of the named object, as opposed to describing how it should be displayed or printed. In this
way, the information is self-describing, so it can be located, extracted, and manipulated as
desired. This kind of power has previously been reserved for organized scalar information
managed by database systems.
You may have also noticed a potential hazard that comes with this freedom—since people may
define new XML languages as they please, how can we resolve ambiguities and achieve common
understanding? This is why, although the core XML is very simple, there are many XML-related
standards to handle translation and specification of data. The simplest way is to explicitly state
the vocabulary and composition rules of an XML language and enforce those across all the
involved parties. Another option, as with natural languages, is to have a translator in between, as
illustrated in Figure 6-1. The former solution employs XML Schemas (introduced in Section 6.2
below), and the latter employs transformation languages (introduced in Section 6.4 below).
Well-Formedness
A text document is an XML document if it has a proper syntax as per the XML specification.
Such document is called a well-formed document. An XML document is well-formed if it
conforms to the XML syntax rules:

Begins with the XML declaration <?xml ... ?>

Has exactly one root element, called the root or document, and no part of it can appear in
the content of any other element
Chapter 6 
327
XML and Data Representation

Contains one or more elements delimited by start-tags and end-tags (also remember that
XML tags are case sensitive)

All elements are closed, that is all start-tags must match end-tags

All elements must be properly nested within each
<outer><inner>inner content</inner></outer>

All attribute values must be within quotations

XML entities must be used for special characters. Each of the parsed entities that are
referenced directly or indirectly within the document is well-formed.
other,
such
as
Even if documents are well-formed they can still contain errors, and those errors can have serious
consequences. XML Schemas (introduced in Section 6.2 below) provide further level of error
checking. A well-formed XML document may in addition be valid if it meets constraints
specified by an associated XML Schema.
Document- vs. Data-Centric XML
Generally speaking, there are two broad application areas of XML technologies. The first relates
to document-centric applications, and the second to data-centric applications. Because XML can
be used in so many different ways, it is important to understand the difference between these two
categories. (See more at http://www.xmleverywhere.com/newsletters/20000525.htm)
Initially, XML’s main application was in semi-structured document representation, such as
technical manuals, legal documents, and product catalogs. The content of these documents is
typically meant for human consumption, although it could be processed by any number of
applications before it is presented to humans. The key element of these documents is semistructured marked-up text. A good example is the correspondence letter in Listing 6-1 above.
By contrast, data-centric XML is used to mark up highly structured information such as the
textual representation of relational data from databases, financial transaction information, and
programming language data structures. Data-centric XML is typically generated by machines and
is meant for machine consumption. It is XML’s natural ability to nest and repeat markup that
makes it the perfect choice for representing these types of data.
Key characteristics of data-centric XML:

The ratio of markup to content is high. The XML includes many different types of tags.
There is no long-running text.

The XML includes machine-generated information, such as the submission date of a
purchase order using a date-time format of year-month-day. A human authoring an XML
document is unlikely to enter a date-time value in this format.

The tags are organized in a highly structured manner. Order and positioning matter,
relative to other tags. For example, TBD

Markup is used to describe what a piece of information means rather than how it should
be presented to a human.
Ivan Marsic

Rutgers University
328
An interesting example of data-centric XML is the XML Metadata Interchange (XMI), which is
an OMG standard for exchanging metadata information via XML. The most common use of XMI
is as an interchange format for UML models, although it can also be used for serialization of
models of other languages (metamodels). XMI enables easy interchange of metadata between
UML-based modeling tools and MOF (Meta-Object Facility)-based metadata repositories in
distributed heterogeneous environments. For more information see here:
http://www.omg.org/technology/documents/formal/xmi.htm.
6.1.2
Document Type Definition (DTD)
Document Type Definition (DTD) is a schema language for XML inherited from SGML, used
initially, before XML Schema was developed. DTD is one of ways to define the structure of
XML documents, i.e., the document’s metadata.
Syntactically, a DTD is a sequence of declarations. There are four kinds of declarations in XML:
(1) element type declarations, used to define tags; (2) attribute list declarations, used to define tag
attributes; (3) entity declarations, used to define entities; and, (4) notation declarations, used to
define data type notations. Each declaration has the form of a markup representation, starting with
a keyword followed by the production rule that specifies how the content is created:
<!keyword production-rule>
where the possible keywords are: ELEMENT, ATTLIST (for attribute list), ENTITY, and
NOTATION. Next, I describe these declarations.
Element Type Declarations
Element type declarations identify the names of elements and the nature of their content, thus
putting a type constraint on the element. A typical element type declaration looks like this:
<!ELEMENT
<!ELEMENT
Declaration type
chapter
title
Element name
(title, paragraph+, figure?)>
(#PCDATA)>
Element’s content model (definition of allowed
content: list of names of child elements)
The first declaration identifies the element named chapter. Its content model follows the
element name. The content model defines what an element may contain. In this case, a chapter
must contain paragraphs and title and may contain figures. The commas between element names
indicate that they must occur in succession. The plus after paragraph indicates that it may be
repeated more than once but must occur at least once. The question mark after figure indicates
that it is optional (it may be absent). A name with no punctuation, such as title, must occur
exactly once. The following table summarizes the meaning of the symbol after an element:
Kleene symbol
Meaning
none
The element must occur exactly once
?
The element is optional (zero or one occurrence allowed)
*
The element can be skipped or included one or more times
+
The element must be included one or more times
Declarations for paragraphs, title, figures and all other elements used in any content model must
also be present for an XML processor to check the validity of a document. In addition to element
Chapter 6 
329
XML and Data Representation
names, the special symbol #PCDATA is reserved to indicate character data. The PCDATA stands
for parseable character data.
Elements that contain only other elements are said to have element content. Elements that contain
both other elements and #PCDATA are said to have mixed content. For example, the definition
for paragraphs might be
<!ELEMENT paragraph (#PCDATA | quote)*>
The vertical bar indicates an “or” relationship, the asterisk indicates that the content is optional
(may occur zero or more times); therefore, by this definition, paragraphs may contain zero or
more characters and quote tags, mixed in any order. All mixed content models must have this
form: #PCDATA must come first, all of the elements must be separated by vertical bars, and the
entire group must be optional.
Two other content models are possible: EMPTY indicates that the element has no content (and
consequently no end-tag), and ANY indicates that any content is allowed. The ANY content model
is sometimes useful during document conversion, but should be avoided at almost any cost in a
production environment because it disables all content checking in that element.
Attribute List Declarations
Elements which have one or more attributes are to be specified in the DTD using attribute list
type declarations. An example for a figure element could be like so
<!ATTLIST
Declaration type
figure
caption
scaling
Name of the
associated
element
CDATA
CDATA
Names of attributes
Data type
#REQUIRED
#FIXED "100%">
Keyword or default value
Repeat for each attribute of the element
The CDATA as before stands for character data and #REQUIRED means that the caption attribute
of figure has to be present. Other marker could be #FIXED with a value, which means this
attribute acts like a constant. Yet another marker is #IMPLIED, which indicates an optional
attribute. Some more markers are ID and enumerated data type like so
<!ATTLIST person sibling (brother | sister) #REQUIRED>
Enumerated attributes can take one of a list of values provided in the declaration.
Entity Declarations
As stated above, entities are used as substitutes for reserved characters, but also to refer to often
repeated or varying text and to include the content of external files. An entity is defined by its
name and an associated value. An internal entity is the one for which the parsed content
(replacement text) lies inside the document, like so:
<!ENTITY
Declaration type
substitute
Entity name
"This text is often repeated.">
Entity value (any literal) – single or double quotes can
be used, but must be properly matched
Once the above example entity is defined, it can be used in the XML document as
&substitute; anywhere where the full text should appear. Entities can contain markup as
Ivan Marsic

Rutgers University
330
well as plain text. For example, this declaration defines &contact; as an abbreviation for
person’s contact information that may be repeated multiple times in one or more documents:
<!ENTITY contact '<a href="mailto:[email protected]">
e-mail</a><br>
<a href="732-932-4636.tel">telephone</a>
<address>13 Takeoff Lane<br> Talkeetna, AK 99676</address>
'>
Conversely, the content of the replacement text of an external entity resides in a file separate from
the XML document. The content can be accessed using either system identifier, which is a URI
(Uniform Resource Identifier, see Appendix C) address, or a public identifier, which serves as a
basis for generating a URI address. Examples are:
<!ENTITY
<!ENTITY
contact
surrogate
Declaration type
Entity name
SYSTEM "http://any.company.com/contact.xml">
PUBLIC "-//home/mrsmith/text"
SYSTEM or PUBLIC identifier, followed by the external ID (URI or other)
Notation Declarations
Notations are used to associate actions with entities. For example, a PDF file format can be
associated with the Acrobat application program. Notations identify, by name, the format of these
actions. Notation declarations are used to provide an identifying name for the notation. They are
used in entity or attribute list declarations and in attribute specifications. This is a complex and
controversial feature of DTD and the interested reader should seek details elsewhere.
DTD in Use
A DTD can be embedded in the XML document for which it describes the syntax rules and this is
called an internal DTD. The alternative is to have the DTD stored in one or more separate files,
called external DTD. External DTDs are preferable since they can be reused in different XML
documents by different users. The reader should be by now aware of the benefits of modular
design, a key one being able to (re-)use modules that are tested and fixed by previous users.
However, this also means that if the reused DTD module is changed, all documents that use the
DTD must be tested against the new DTD and possibly modified to conform to the changed DTD.
In an XML document, external DTDs are referred to with a DOCTYPE declaration in the second
line of the XML document (after the first line: <?xml ... ?>) as seen in Listing 6-3 below.
The following fragment of DTD code defines the production rules for constructing book
documents.
Listing 6-2: Example DTD for a postal address element. File name: address.dtd
1
2
3
4
5
6
<!ELEMENT
<!ATTLIST
<!ELEMENT
<!ELEMENT
<!ELEMENT
<!ELEMENT
address (street+, city, state, postal-code)>
address kind (return | delivery) #IMPLIED>
street (#PCDATA)>
city (#PCDATA)>
state (#PCDATA)>
postal-code (#PCDATA)>
Chapter 6 
XML and Data Representation
331
Line 1 shows the element address definition, where all four sub-elements are required, and the
street sub-element can appear more than once. Line 2 says that address has an optional
attribute, kind, of the enumerated type.
We can (re-)use the postal address declaration as an external DTD, for example, in an XML
document of a correspondence letter as shown in Listing 6-3.
Listing 6-3: Example correspondence letter that uses an external DTD.
1
2
3
4
4a
5
6
7
8
9
10
11
12
13
14
<?xml version="1.0"?> <!-- Comment: Person DTD -->
<!DOCTYPE letter SYSTEM "http://any.website.net/address.dtd" [
<!ELEMENT letter (sender?, recipient+, body)>
<!ATTLIST letter language (en-US | en-UK | fr) #IMPLIED
template (personal | business) #IMPLIED>
<!ELEMENT sender (name, address)>
<!ELEMENT recipient (name, address)>
<!ELEMENT name (#PCDATA)>
<!ELEMENT body ANY>
]>
<letter language="en-US" template="personal">
<sender>
<name>Mr. Charles Morse</name>
<address kind="return">
. . .
<!-- continued as in Listing 6-1 above -->
In the above DTD document, Lines 2 – 9 define the DTD for a correspondence letter document.
The complete DTD is made up of two parts: (1) the external DTD subset, which in this case
imports a single external DTD named address.dtd in Line 2; and (2) the internal DTD subset
contained between the brackets in Lines 3 – 8. The external DTD subset will be imported at the
time the current document is parsed. The address element is used in Lines 5 and 6.
The content of the body of letter is specified using the keyword ANY (Line 8), which means that a
body element can contain any content, including mixed content, nested elements, and even other
body elements. Using ANY is appropriate initially when beginning to design the DTD and
document structure to get quickly to a working version. However, it is a very poor practice to use
ANY in finished DTD documents.
Limitations of DTDs
DTD provided the first schema for XML documents. Their limitations include:

Language inconsistency since DTD uses a non-XML syntax

Failure to support namespace integration

Lack of modular vocabulary design

Rigid content models (cannot derive new type definitions based on the old ones)

Lack of integration with data-oriented applications
Ivan Marsic

Rutgers University
332
http://any.website.net/book
http://any.website.net/person
title
title
figure
author
chapter
name
phone
address
caption
paragraph
gender
email
Figure 6-2: Example XML namespaces providing context to individual names.

Conversely, XML Schema allows much more expressive and precise specification of the
content of XML documents. This flexibility also carries the price of complexity.
W3C is making efforts to phase DTDs out. XML Schema is described in Section 6.2 below.
6.1.3
Namespaces
Inventing new languages is an arduous task, so it will be beneficial if we can reuse (parts of) an
existing XML language (defined by a schema). Also, there are many occasions when an XML
document needs to use markups defined in multiple schemas, which may have been developed
independently. As a result, it may happen that some tag names may be non-unique.
For example, the word “title” is used to signify the name of a book or work of art, a form of
nomenclature indicating a person’s status, the right to ownership of property, etc. People easily
figure out context, but computers are very poor at absorbing contextual information. To simplify
the computer’s task and give a specific meaning to what might otherwise be an ambiguous term,
we qualify the term with and additional identifier—a namespace identifier.
An XML namespace is a collection of names, used as element names or attribute names, see
examples in Figure 6-2. The C++ programming language defines namespaces and Java package
names are equivalent to namespaces. Using namespaces, you can qualify your elements as
members of a particular context, thus eliminating the ambiguity and enabling namespace-aware
applications to process your document correctly. In other words:
Qualified name (QName) = Namespace identifier + Local name
A namespace is declared as an attribute of an element. The general form is as follows:
<bk:tagName
xmlns
mandatory
:bk
prefix
= "http://any.website.net/book" />
namespace name
There are two forms of namespace declarations due to the fact that the prefix is optional. The first
form binds a prefix to a given namespace name. The prefix can be any string starting with a letter,
followed by any combination of digits, letters, and punctuation signs (except for the colon “:”
since it is used to separate the mandatory string xmlns from the prefix, which indicates that we
are referring to an XML namespace). The namespace name, which is the attribute value, must be
a valid, unique URI. However, since all that is required from the name is its uniqueness, a URL
Chapter 6 
XML and Data Representation
333
such as http://any.website.net/schema also serves the purpose. Note that this does
not have to point to anything in particular—it is merely a way to uniquely label a set of names.
The namespace is in effect within the scope of the element for which it is defined as an attribute.
This means that the namespace is effective for all the nested elements, as well. The scoping
properties of XML namespaces are analogous to variable scoping properties in programming
languages, such as C++ or Java. The prefix is used to qualify the tag names, as in the following
example:
Scope of "bk2"
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Scope of "bk"
Listing 6-4: Example of using namespaces in an XML document.
<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
<book>
<bk:cover xmlns:bk="http://any.website.net/book">
<bk:title>A Book About Namespaces</bk:title>
<bk:author>Anonymous</bk:title>
<bk:isbn number="1881378241"/>
</bk:cover>
<bk2:chapter xmlns:bk2="http://any.website.net/book"
ch_name="Introduction">
<bk2:paragraph>
In this chapter we start from the beginning.
...
</bk2:paragraph>
. . .
</bk2:chapter>
As can be seen, the namespace identifier must be declared only in the outermost element. In our
case, there are two top-level elements: <bk:cover> and <bk:chapter>, and their embedded
elements just inherit the namespace attribute(s). All the elements of the namespace are prefixed
with the appropriate prefix, in our case “bk.” The actual prefix’s name is not important, so in the
above example I define “bk” and “bk2” as prefixes for the same namespace (in different
scopes!). Notice also that an element can have an arbitrary number of namespace attributes, each
defining a different prefix and referring to a different namespace.
In the second form, the prefix is omitted, so the elements of this namespace are not qualified. The
namespace attribute is bound to the default namespace. For the above example (Listing 6-4), the
second form can be declared as:
1
2
3
4
5
6
7
8
Scope of default n.s.
Listing 6-5: Example of using a default namespace.
<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
<book>
<cover xmlns="http://any.website.net/book">
<title>A Book About Namespaces</title>
<author>Anonymous</title>
<isbn number="1881378241"/>
</cover>
. . .
Notice that there can be at most one default namespace declared within a given scope. In Listing
6-5, we can define another default namespace in the same document, but its scope must not
overlap with that of the first one.
Ivan Marsic

6.1.4
XML Parsers
Rutgers University
334
The parsers define standard APIs to access and manipulate the parsed XML data. The two most
popular parser APIs are DOM (Document Object Model) based and SAX (Simple API for XML).
See Appendix E for a brief review of DOM.
SAX and DOM offer complementary paradigms to access the data contained in XML documents.
DOM allows random access to any part of a parsed XML document. To use DOM APIs, the
parsed objects must be stored in the working memory. Conversely, SAX provides no storage and
presents the data as a linear stream. With SAX, if you want to refer back to anything seen earlier
you have to implement the underlying mechanism yourself. For example, with DOM an
application program can import an XML document, modify it in arbitrary order, and write back
any time. With SAX, you cannot perform the editing arbitrarily since there is no stored document
to edit. You would have to edit it by filtering the stream, as it flows, and write back immediately.
Event-Oriented Paradigm: SAX
SAX (Simple API for XML) is a simple, event-based API for XML parsers. The benefit of an
event-based API is that it does not require the creation and maintenance of an internal
representation of the parsed XML document. This makes possible parsing XML documents that
are much larger than the available system memory would allow, which is particularly important
for small terminals, such as PDAs and mobile phones. Because it does not require storage behind
its API, SAX is complementary to DOM.
Chapter 6 
335
XML and Data Representation
Initiating the parser
. . .
Parser parser =
ParserFactory.makeParser
("com.sun.xml.parser.Parser");
parser.setDocumentHandler(new
DocumentHandlerImpl());
parser.parse (input);
. . .
Event triggering in SAX parser:
DocumentHandler Interface
public void startDocument()throws
SAXException{}
public void endDocument()throws
SAXException{}
public void startElement(String name,
AttributeList attrs) throws
SAXException{}
public void endElement(String
name)throws SAXException{}
public void characters(char buf [],
int offset, int len)throws
SAXException{}
<?xml ...>
startDocument
startElement
<element attr1=“val1”>
characters
This is a test.
endElement
</element>
Document
Handler
<element attr1=“val2”/>
endDocument
end of the document
Figure 6-3: SAX parser Java example.
SAX provides events for the following structural information for XML documents:

The start and end of the document

Document type declaration (DTD)

The start and end of elements

Attributes of each element

Character data

Unparsed entity declarations

Notation declarations

Processing instructions
Object-Model Oriented Paradigm: DOM
DOM (Document Object Model)
Practical Issues
Additional features relevant for both event-oriented and object-model oriented parsers include:

Validation against a DTD
Ivan Marsic

Rutgers University
336

Validation against an XML Schema

Namespace awareness, i.e., the ability to determine the namespace URI of an element or
attribute
These features affect the performance and memory footprint of a parser, so some parsers do not
support all the features. You should check the documentation for the particular parser as to the list
of supported features.
6.2 XML Schemas
Although there is no universal definition of schema, generally scholars agree that schemas are
abstractions or generalizations of our perceptions of the world around us, which is molded by our
experience. Functionally, schemas are knowledge structures that serve as heuristics which help us
evaluate new information. An integral part of schema is our expectations of people, place, and
things. Schemas provide a mechanism for describing the logical structure of information, in the
sense of what elements can or should be present and how they can be arranged. Deviant news
results in violation of these expectations, resulting in schema incongruence.
In XML, schemas are used to make a class of documents adhere to a particular interface and thus
allow the XML documents to be created in a uniform way. Stated another way, schemas allow a
document to communicate meta-information to the parser about its content, or its grammar. Metainformation includes the allowed sequence and arrangement/nesting of tags, attribute values and
their types and defaults, the names of external files that may be referenced and whether or not
they contain XML, the formats of some external (non-XML) data that may be referenced, and the
entities that may be encountered. Therefore, schema defines the document production rules.
XML documents conforming to a particular schema are said to be valid documents. Notice that
having a schema associated with a given XML document is optional. If there is a schema for a
given document, it must appear before the first element in the document.
Here is a simple example to motivate the need for schemas. In Section 6.1.1 above I introduced
an XML representation of a correspondence letter and used the tags <letter>, <sender>,
<name>, <address>, <street>, <city>, etc., to mark up the elements of a letter. What if
somebody used the same vocabulary in a somewhat different manner, such as the following?
Listing 6-6: Variation on the XML example document from Listing 6-1.
1
2
3
4
5
6
7
8
9
10
11
12
<?xml version="1.0" encoding="UTF-8"?>
<letter>
<sender>Mr. Charles Morse</sender>
<street>13 Takeoff Lane</street>
<city>Talkeetna, AK 99676</city>
<date>29.02.1997</date>
<recipient>Mrs. Robinson</recipient>
<street>1 Entertainment Way</street>
<city>Los Angeles, CA 91011</city>
<body>
Dear Mrs. Robinson,
Chapter 6 
337
XML and Data Representation
http://www.w3.org/2001/XMLSchema
http://any.website.net/letter
letter
schema
string
name
sender
element
address
complexType
sequence
recipient
boolean
salutation
street
city
This is the vocabulary that
XML Schema provides to define
your new vocabulary
<?xml version="1.0" encoding="UTF-8"?>
<lt:letter xmlns:lt ="http://any.website.net/letter"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://any.website.net/letter
http://any.website.net/letter/letter.xsd"
lt:language="English_US" lt:template="personal">
<lt:sender>
...
</lt:letter>
An instance document that conforms to the “letter” schema
Figure 6-4: Using XML Schema. Step 1: use the Schema vocabulary to define a new XML
language (Listing 6-7). Step 2: use both to produce valid XML documents (Listing 6-8).
13
14
15
16
17
18
Here's part of an update ...
Sincerely,
</body>
<signature>Charlie</signature>
</letter>
We can quickly figure that this document is a letter, although it appears to follow different rules
of production than the example in Listing 6-1 above. If asked whether Listing 6-6 represents a
valid letter, you would likely respond: “It probably does.” However, to support automatic
validation of a document by a machine, we must precisely specify and enforce the rules and
constraints of composition. Machines are not good at handling ambiguity and this is what
schemas are about. The purpose of a schema in markup languages is to:

Allow machine validation of document structure

Establish a contract (how an XML document will be structured) between multiple parties
who are exchanging XML documents
There are many other schemas that are used regularly in our daily activities. Another example
schema was encountered in Section 2.3.3—the schema for representing the use cases of a
system-to-be, Figure 2-13.
6.2.1
XML Schema Basics
XML Schema provides the vocabulary to state the rules of document production. It is an XML
language for which the vocabulary is defined using itself. That is, the elements and datatypes that
Ivan Marsic

Rutgers University
338
are used to construct schemas, such as <schema>, <element>, <sequence>, <string>,
etc., come from the http://www.w3.org/2001/XMLSchema namespace, see Figure 6-4.
The XML Schema namespace is also called the “schema of schemas,” for it defines the elements
and attributes used for defining new schemas.
The first step involves defining a new language (see Figure 6-4). The following is an example
schema for correspondence letters, an example of which is given in Listing 6-1 above.
Listing 6-7: XML Schema for correspondence letters (see an instance in Listing 6-1).
1 <?xml version="1.0" encoding="UTF-8"?>
2 <xsd:schema xmlns:xsd="http://www.w3.org/2001/XMLSchema"
targetNamespace="http://any.website.net/letter"
2a
xmlns="http://any.website.net/letter"
2b
elementFormDefault="qualified">
2c
<xsd:element name="letter">
3
<xsd:complexType>
4
5
<xsd:sequence>
<xsd:element name="sender" type="personAddressType"
6
6a
minOccurs="1" maxOccurs="1"/>
<xsd:element name="date" type="xsd:date" minOccurs="0"/>
7
8
<xsd:element name="recipient" type="personAddressType"/>
<xsd:element name="salutation" type="xsd:string"/>
9
10
<xsd:element name="body" type="xsd:string"/>
11
<xsd:element name="closing" type="xsd:string"/>
12
<xsd:element name="signature" type="xsd:string"/>
13
</xsd:sequence>
<xsd:attribute name="language" type="xsd:language"/>
14
15
<xsd:attribute name="template" type="xsd:string"/>
16
</xsd:complexType>
17
</xsd:element>
<xsd:complexType name="personAddressType">
18
19
<xsd:sequence>
20
<xsd:element name="name" type="xsd:string"/>
<xsd:element ref="address"/>
21
22
</xsd:sequence>
23
</xsd:complexType>
24
<xsd:element name="address">
25
<xsd:complexType>
26
<xsd:sequence>
27
<xsd:element name="street" type="xsd:string"
27a
minOccurs="1" maxOccurs="unbounded"/>
28
<xsd:element name="city" type="xsd:string"/>
29
<xsd:element name="state" type="xsd:string"/>
30
<xsd:element name="postal-code" type="xsd:string"/>
31
</xsd:sequence>
32
</xsd:complexType>
33
</xsd:element>
34 </xsd:schema>
The graphical representation of document structure defined by this schema is shown in Figure
6-5. The explanation of the above listing is as follows:
Line 1: This indicates that XML Schemas are XML documents.
Chapter 6 
339
XML and Data Representation
anonymous type
lt:personAddressType
lt:sender
lt:name
?
lt:date
lt:sender
lt:address
lt:recipient
lt:letter
lt:salutation
anonymous type
lt:body
+
lt:street
lt:closing
lt:city
lt:signature
?
?
lt:address
lt:state
lt:language
lt:postal-code
lt:template
XML Schema symbols
<sequence>
<element> immediately within <schema>, i.e. global
<element> not immediately within <schema>, i.e. local
<choice>
<element> has sub-elements (not shown)
<all>
Kleene operators:
<element> has sub-elements (shown)
(no indicator)
Required
One and only one
?

+
!
Optional
None or one (minOccurs = 0, maxOccurs = 1)
Optional, repeatable
Required, repeatable
Unique
None, one, or more (minOccurs = 0, maxOccurs = )
One or more (minOccurs = 1, maxOccurs = )
element values must be unique
<group> of elements
<element> reference
<attribute> of an <element>
<attributeGroup>
Figure 6-5: Document structure defined by correspondence letters schema (see Listing 6-7).
NOTE: The symbolic notation is inspired by the one used in [McGovern et al., 2003].
Line 2: Declares the xsd: namespace. A common convention is to use the prefix xsd: for
elements belonging to the schema namespace. Also notice that all XML Schemas have
<schema> as the root element—the rest of the document is embedded into this element.
Line 2a: Declares the target namespace as http://any.website.net/letter—the
elements defined by this schema are to go in the target namespace.
Line 2b: The default namespace is set to http://any.website.net/letter—same
as the target namespace—so the elements of this namespace do not need the namespace
qualifier/prefix (within this schema document).
Line 2c: This directive instructs the instance documents which conform to this schema that
any elements used by the instance document which were declared in this schema must be
namespace qualified. The default value of elementFormDefault (if not specified) is
Ivan Marsic

Rutgers University
340
"unqualified". The corresponding directive about qualifying the attributes is
attributeFormDefault, which can take the same values.
Lines 3–17: Define the root element <letter> as a compound datatype
(xsd:complexType) comprising several other elements. Some of these elements, such as
<salutation> and <body>, contain simple, predefined datatype xsd:string. Others,
such as <sender> and <recipient>, contain compound type personAddressType
which is defined below in this schema document (lines 18–23). This complex type is also a
sequence, which means that all the named elements must appear in the sequence listed.
The letter element is defined as an anonymous type since it is defined directly within the
element definition, without specifying the attribute “name” of the <xsd:complexType>
start tag (line 4). This is called inlined element declaration. Conversely, the compound type
personAddressType, defined as an independent entity in line 18 is a named type, so it
can be reused by other elements (see lines 6 and 8).
Line 6a: The multiplicity attributes minOccurs and maxOccurs constrain the number of
occurrences of the element. The default value of these attributes equals to 1, so line 6a is
redundant and it is omitted for the remaining elements (but, see lines 7 and 27a). In general,
an element is required to appear in an instance document (defined below) when the value of
minOccurs is 1 or more.
Line 7: Element <date> is of the predefined type xsd:date. Notice that the value of
minOccurs is set to 0, which indicates that this element is optional.
Lines 14–15: Define two attributes of the element <letter>, that is, language and
template. The language attribute is of the built-in type xsd:language (Section 6.2.3
below).
Lines 18–23: Define our own personAddressType type as a compound type comprising
person’s name and postal address (as opposed to a business-address-type). Notice that the
postal <address> element is referred to in line 21 (attribute ref) and it is defined
elsewhere in the same document. The personAddressType type is extended as
<sender> and <recipient> in lines 6 and 8, respectively.
Lines 24–33: Define the postal <address> element, referred to in line 21. Of course, this
could have been defined directly within the personAddressType datatype, as an
anonymous sub-element, in which case it would not be reusable. (Although the element is not
reused in this schema, I anticipate that an external schema may wish to reuse it, see Section
6.2.4 below.)
Line 27a: The multiplicity attribute maxOccurs is set to “unbounded,” to indicate that the
street address is allowed to extend over several lines.
Notice that Lines 2a and 2b above accomplish two different tasks. One is to declare the
namespace URI that the letter schema will be associated with (Line 2a). The other task is to
define the prefix for the target namespace that will be used in this document (Line 2b). The reader
may wonder whether this could have been done in one line. But, in the spirit of the modularity
principle, it is always to assign different responsibilities (tasks) to different entities (in this case
different lines).
Chapter 6 
341
XML and Data Representation
The second step is to use the newly defined schema for production of valid
instance documents (see Figure 6-4). An instance document is an XML
document that conforms to a particular schema. To reference the above
schema in letter documents, we do as follows:
conforms-to
Schema
document
Instance
documents
Listing 6-8: Referencing a schema in an XML instance document (compare to Listing 6-1)
1 <?xml version="1.0" encoding="UTF-8"?>
2 <!-- Comment: A personal letter marked up in XML. -->
3 <lt:letter xmlns:lt ="http://any.website.net/letter"
3a
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
3b
xsi:schemaLocation="http://any.website.net/letter
3c
http://any.website.net/letter/letter.xsd"
3d
lt:language="en-US" lt:template="personal">
4
<lt:sender>
...
<!-- similar to Listing 6-1 -->
10
</lt:sender>
...
<!-- similar to Listing 6-1 -->
25 </lt:letter>
The above listing is said to be valid unlike Listing 6-1 for which we generally only know that it is
well-formed. The two documents (Listings 6-1 and 6-8) are the same, except for referencing the
letter schema as follows:
Step 1 (line 3): Tell a schema-aware XML processor that all of the elements used in this
instance document come from the http://any.website.net/letter namespace. All
the element and attribute names will be prefaced with the lt: prefix. (Notice that we could
also use a default namespace declaration and avoid the prefix.)
Step 2 (line 3a): Declare another namespace, the XMLSchema-instance namespace, which
contains a number of attributes (such as schemaLocation, to be used next) that are part of
a schema specification. These attributes can be applied to elements in instance documents to
provide additional information to a schema-aware XML processor. Again, a usual convention
is to use the namespace prefix xsi: for XMLSchema-instance.
Step 3 (lines 3b–3c): With the xsi:schemaLocation attribute, tell the schema-aware
XML processor to establish the binding between the current XML document and its schema.
The attribute contains a pair of values. The first value is the namespace identifier whose
schema’s location is identified by the second value. In our case the namespace identifier is
http://any.website.net/letter and the location of the schema document is
http://any.website.net/letter/letter.xsd. (In this case, it would suffice to
only have letter.xsd as the second value, since the schema document’s URL overlaps
with the namespace identifier.) Typically, the second value will be a URL, but specialized
applications can use other types of values, such as an identifier in a schema repository or a
well-known schema name. If the document used more than one namespace, the
xsi:schemaLocation attribute would contain multiple pairs of values (all within a
single pair of quotations).
Notice that the schemaLocation attribute is merely a hint. If the parser already knows about
the schema types in that namespace, or has some other means of finding them, it does not have to
go to the location you gave it.
XML Schema defines two aspects of an XML document structure:
Ivan Marsic

Rutgers University
342
1. Content model validity, which tests whether the arrangement and embedding of tags is
correct. For example, postal address tag must have nested the street, city, and postal-code
tags. A country tag is optional.
2. Datatype validity, which is the ability to test whether specific units of information are of
the correct type and fall within the specified legal values. For example, a postal code is a
five-digit number. Data types are the classes of data values, such as string, integer, or
date. Values are instances of types.
There are two types of data:
1. Simple types are elements that contain data but not attributes or sub-elements. Examples
of simple data values are integer or string, which do not have parts. New simple
types are defined by deriving them from existing simple types (built-in’s and derived).
2. Compound types are elements that allow sub-elements and/or attributes. An example is
personAddressType type defined in Listing 6-7. Complex types are defined by
listing the elements and/or attributes nested within them.
6.2.2
Models for Structured Content
As noted above, schema defines the content model of XML documents—the legal building blocks
of an XML document. A content model indicates what a particular element can contain. An
element can contain text, other elements, a mixture of text and elements, or nothing at all. Content
model defines:

elements that can appear in a document

attributes that can appear in a document

which elements are child elements

the order of child elements

the multiplicity of child elements

whether an element is empty or can include text

data types for elements and attributes

default and fixed values for elements and attributes
This section reviews the schema tools for specifying syntactic and structural constraints on
document content. The next section reviews datatypes of elements and attributes, and their value
constraints.
XML Schema Elements
XML Schema defines a vocabulary on its own, which is used to define other schemas. Here I
provide only a brief overview of XML Schema elements that commonly appear in schema
documents. The reader should look for the complete list here: http://www.w3.org/TR/2004/RECxmlschema-1-20041028/structures.html.
The <schema> element defines the root element of every XML Schema.
Chapter 6 
XML and Data Representation
Syntax of the <schema> element
<schema
id=ID ……………………………………………
attributeFormDefault=qualified | unqualified
elementFormDefault=qualified | unqualified
blockDefault=(#all |
list of (extension | restriction | substitution))
finalDefault=(#all |
list of (extension | restriction | list | union))
targetNamespace=anyURI …………………
version=token
xmlns=anyURI ………………………………
any attributes
>
((include | import | redefine | annotation),
(((simpleType | complexType | group |
attributeGroup) | element | attribute |
notation), annotation))
343
Description (attributes are optional unless stated else)
Specifies a unique ID for the element.
The form for attributes declared in the target namespace of this
schema. The value must be "qualified" or "unqualified". Default is
"unqualified". "unqualified" indicates that attributes from the target
namespace are not required to be qualified with the namespace
prefix. "qualified" indicates that attributes from the target namespace
must be qualified with the namespace prefix.
The form for elements declared in the target namespace of this
schema. The value must be "qualified" or "unqualified". Default is
"unqualified". "unqualified" indicates that elements from the target
namespace are not required to be qualified with the namespace
prefix. "qualified" indicates that elements from the target namespace
must be qualified with the namespace prefix.
A URI reference of the namespace of this schema.
Required. A URI reference that specifies one or more namespaces for
use in this schema. If no prefix is assigned, the schema components
of the namespace can be used with unqualified references.
Kleene operators ?, , and  are defined in Figure 6-5.
</schema>
The <element> element defines an element. Its parent element can be one of the following:
<schema>, <choice>, <all>, <sequence>, and <group>.
Description (all attributes are optional)
Syntax of the <element> element
<element
id=ID
name=NCName ………………………………
ref=QName ……………………………………
type=QName …………………………………
substitutionGroup=QName
default=string …………………………………
fixed=string
form=qualified|unqualified
maxOccurs=nonNegativeInteger|unbounded
Specifies a name for the element. This attribute is required if the
parent element is the schema element.
Refers to the name of another element. This attribute cannot be used
if the parent element is the schema element.
Specifies either the name of a built-in data type, or the name of a
simpleType or complexType element.
This value is automatically assigned to the element when no other
value is specified. (Can only be used if the element’s content is a
simple type or text only).
Specifies the maximum number of times this element can occur in the
parent element. The value can be any number >= 0, or if you want to
Ivan Marsic

Rutgers University
minOccurs=nonNegativeInteger ……………
nillable=true|false
abstract=true|false
block=(#all|list of (extension|restriction))
final=(#all|list of (extension|restriction))
any attributes
>
annotation?,((simpleType |
complexType)?,(unique | key | keyref)))
344
set no limit on the maximum number, use the value "unbounded".
Default value is 1.
Specifies the minimum number of times this element can occur in the
parent element. The value can be any number >= 0. Default is 1.
Kleene operators ?, , and  are defined in Figure 6-5.
</element>
The <group> element is used to define a collection of elements to be used to model compound
elements. Its parent element can be one of the following: <schema>, <choice>,
<sequence>, <complexType>, <restriction> (both <simpleContent> and
<complexContent>),
<extension>
(both
<simpleContent>
and
<complexContent>).
Description (all attributes are optional)
Syntax of the <group> element
<group
id=ID
name=NCName ………………………………
ref=QName ……………………………………
maxOccurs=nonNegativeInteger | unbounded
minOccurs=nonNegativeInteger
any attributes
>
Specifies a name for the group. This attribute is used only when the
schema element is the parent of this group element. Name and ref
attributes cannot both be present.
Refers to the name of another group. Name and ref attributes cannot
both be present.
(annotation?, (all | choice | sequence))
</group>
The <attributeGroup> element is used to group a set of attribute declarations so that they
can be incorporated as a group into complex type definitions.
Description (all attributes are optional)
Syntax of <attributeGroup>
<attributeGroup
id=ID
name=NCName ……………………………
ref=QName …………………………………
any attributes
>
(annotation?), ((attribute | attributeGroup),
anyAttribute?))
Specifies the name of the attribute group. Name and ref attributes
cannot both be present.
Refers to a named attribute group. Name and ref attributes cannot both
be present.
Chapter 6 
XML and Data Representation
345
</attributeGroup>
The <annotation> element specifies schema comments that are used to document the
schema. This element can contain two elements: the <documentation> element, meant for
human consumption, and the <appinfo> element, for machine consumption.
Simple Elements
A simple element is an XML element that can contain only text. It cannot contain any other
elements or attributes. However, the “only text” restriction is ambiguous since the text can be of
many different types. It can be one of the built-in types that are included in the XML Schema
definition, such as boolean, string, date, or it can be a custom type that you can define
yourself as will be seen Section 6.2.3 below. You can also add restrictions (facets) to a data type
in order to limit its content, and you can require the data to match a defined pattern.
Examples of simple elements are <salutation> and <body> elements in Listing 6-7 above.
Groups of Elements
XML Schema enables collections of elements to be defined and named, so that the elements can
be used to build up the content models of complex types. Un-named groups of elements can also
be defined, and along with elements in named groups, they can be constrained to appear in the
same order (sequence) as they are declared. Alternatively, they can be constrained so that only
one of the elements may appear in an instance.
A model group is a constraint in the form of a grammar fragment that applies to lists of element
information items, such as plain text or other markup elements. There are three varieties of model
group:

Sequence element <sequence> (all the named elements must appear in the order
listed);

Conjunction element <all> (all the named elements must appear, although they can
occur in any order);

Disjunction element <choice> (one, and only one, of the elements listed must appear).
6.2.3
Datatypes
In XML Schema specification, a datatype is defined by:
(a) Value space, which is a set of distinct values that a given datatype can assume. For
example, the value space for the integer type are integer numbers in the range
[4294967296, 4294967295], i.e., signed 32-bit numbers.
Ivan Marsic

Rutgers University
346
(b) Lexical space, which is a set of allowed lexical representations or literals for the datatype.
For example, a float-type number 0.00125 has alternative representation as 1.25E3.
Valid literals for the float type also include abbreviations for positive and negative
infinity (INF) and Not a Number (NaN).
(c) Facets that characterize properties of the value space, individual values, or lexical items.
For example, a datatype is said to have a “numeric” facet if its values are conceptually
quantities (in some mathematical number system). Numeric datatypes further can have a
“bounded” facet, meaning that an upper and/or lower value is specified. For example,
postal codes in the U.S. are bounded to the range [10000, 99999].
XML Schema has a set of built-in or primitive datatypes that are not defined in terms of other
datatypes. We have already seen some of these, such as xsd:string which was used in Listing
6-7. More will be exposed below. Unlike these, derived datatypes are those that are defined in
terms of other datatypes (either primitive types or derived ones).
Simple Types: <simpleType>
These types are atomic in that they can only contain character data and cannot have attributes or
element content. Both built-in simple types and their derivations can be used in all element and
attribute declarations. Simple-type definitions are used when a new data type needs to be defined,
where this new type is a modification of some other existing simpleType-type.
Table 6-1 shows a partial list of the Schema-defined types. There are over 40 built-in simple
types and the reader should consult the XML Schema specification (see
http://www.w3.org/TR/xmlschema-0/, Section 2.3) for the complete list.
Table 6-1: A partial list of primitive datatypes that are built into the XML Schema.
Name
string
byte
unsignedByte
boolean
short
int
integer
long
float
double
duration
dateTime
Examples
My favorite text example
128, 1, 0, 1, …, 127
0, …, 255
0, 1, true, false
5, 328
7, 471
2, 435
4, 123456
0, 0, INF, INF, 1E4,
1.401298464324817e45,
3.402823466385288e38,
NaN
Comments
A signed byte value
Derived from unsignedShort
May contain either true or false, 0 or 1
Signed 16-bit integer
Signed 32-bit integer
Same as int
Signed 64-bit integer
Conforming to the IEEE 754 standard for 32bit single precision floating point number.
Note the use of abbreviations for positive and
negative infinity (INF), and Not a Number
(NaN)
Conforming to the IEEE 754 standard for 640, 0, INF, INF, 1E4,
4.9e324, 1.797e308, NaN bit double precision floating point numbers
P1Y2M3DT10H30M12.3S 1 year, 2 months, 3 days, 10 hours, 30
minutes, and 12.3 seconds
1997-03-31T13:20:00.000- March 31st 1997 at 1.20pm Eastern Standard
Chapter 6 
05:00
gYear
1997-03-31
13:20:00.000,
13:20:00.000-05:00
1997
gDay
QName
language
---31
lt:sender
en-GB, en-US, fr
ID
this-element
IDREF
this-element
date
time
347
XML and Data Representation
Time which is 5 hours behind Coordinated
Universal Time
The “g” prefix signals time periods in the
Gregorian calendar.
the 31st day
XML Namespace QName (qualified name)
valid values for xml:lang as defined in
XML 1.0
An attribute that identifies the element; can be
any string that confirms to the rules for
assigning the <element> names.
IDREF attribute type; refers to an element
which has the ID attribute with the same value
A straightforward use of built-in types is the direct declaration of elements and attributes that
conform to them. For example, in Listing 6-7 above I declared the <signature> element and
template attribute of the <letter> element, both using xsd:string built-in type:
<xsd:element name="signature" type="xsd:string"/>
<xsd:attribute name="template" type="xsd:string"/>
New simple types are defined by deriving them from existing simple types (built-in’s and
derived). In particular, we can derive a new simple type by restricting an existing simple type, in
other words, the legal range of values for the new type are a subset of the existing type’s range of
values. We use the <simpleType> element to define and name the new simple type. We use
the restriction element to indicate the existing (base) type, and to identify the facets that constrain
the range of values. A complete list of facets is provided below.
Facets and Regular Expressions
We use the “facets” of datatypes to constrain the range of values.
Suppose we wish to create a new type of integer called zipCodeType whose range of values is
between 10000 and 99999 (inclusive). We base our definition on the built-in simple type
integer, whose range of values also includes integers less than 10000 and greater than 99999.
To define zipCodeType, we restrict the range of the integer base type by employing two
facets called minInclusive and maxInclusive (to be introduced below):
Listing 6-9: Example of new type definition by facets of the base type.
<xsd:simpleType name="zipCodeType">
<xsd:restriction base="xsd:integer">
<xsd:minInclusive value="10000"/>
<xsd:maxInclusive value="99999"/>
</xsd:restriction>
</xsd:simpleType>
Table 6-2 and Table 6-3 list the facets that are applicable for built-in types. The facets identify
various characteristics of the types, such as:
Ivan Marsic

Rutgers University
348

length, minLength, maxLength—the exact, minimum and maximum character
length of the value

pattern—a regular expression pattern for the value (see more below)

enumeration—a list of all possible values (an example given in Listing 6-10 below)

whiteSpace—the rules for handling white-space in the value

minExclusive, minInclusive, maxExclusive, maxInclusive—the range
of numeric values that are allowed (see example in Listing 6-9 above)

totalDigits—the maximum allowed number of decimal digits in numeric values

fractionDigits—the number of decimal digits after the decimal point
As indicated in the tables, not all facets apply to all types.
Table 6-2: XML Schema facets for built-in simple types. Indicated are the facets that apply
to the particular type.
Simple Types
string
base64Binary
hexBinary
integer
positiveInteger
negativeInteger
nonNegativeInteger
nonPositiveInteger
decimal
boolean
time
dateTime
duration
date
Name
QName
anyURI
ID
IDREF
length
♦
♦
♦
minLength
♦
♦
♦
maxLength
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
Facets
pattern
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
enumeration
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
whitespace
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
Table 6-3: XML Schema facets for built-in ordered simple types.
Facets
maxInclusive maxExclusive minInclusive minExclusive
♦
♦
♦
♦
integer
♦
♦
♦
♦
positiveInteger
♦
♦
♦
♦
negativeInteger
♦
♦
♦
♦
nonNegativeInteger
♦
♦
♦
♦
nonPositiveInteger
♦
♦
♦
♦
decimal
Simple Types
totalDigits
♦
♦
♦
♦
♦
♦
fractionDigits
♦
♦
♦
♦
♦
♦
Chapter 6 
349
XML and Data Representation
time
dateTime
duration
date
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
♦
The pattern facet shown in Table 6-2 is particularly interesting since it allows specifying a
variety of constraints using regular expressions. The following example (Listing 6-10) shows how
to define the datatype for representing IP addresses. This datatype has four quads, each restricted
to have a value between zero and 255, i.e., [0-255].[0-255].[0-255].[0-255]
Listing 6-10: Example of IP address type definition via the pattern facet.
<xsd:simpleType name="IPaddress">
<xsd:restriction base="xsd:string">
<xsd:pattern
value="(([1-9]?[0-9]|1[0-9][0-9]|2[0-4][0-9]|25[0-5])\.){3}
([1-9]?[0-9]|1[0-9][0-9]|2[0-4][0-9]|25[0-5])"/>
</xsd:restriction>
</xsd:simpleType>
Note that in the value attribute above, the regular expression has been split over three lines.
This is for readability purposes only; in practice the regular expression would all be on one line.
Selected regular expressions with examples are given in Table 6-4.
Table 6-4: Examples of regular expressions.
Regular Expression
Section \d
Example
Section 3
Regular Expression
Chapter\s\d
Chapter&#x020;\d Chapter 7
a*b
b, ab, aab, aaab,
...
[xyz]b
xb, yb, zb
(hi){2} there
(hi\s){2} there
a?b
b, ab
(a|b)+x
a+b
[a-c]x
ab, aab, aaab, ... a{1,3}x
ax, bx, cx
a{2,}x
[-ac]x
-x, ax, cx
\w\s\w
[ac-]x
ax, cx, -x
[a-zA-Z-[Ok]]*
[^0-9]x
any non-digit char followed
by x
\.
\Dx
any non-digit char followed
by x
\n
.abc
Example
Chapter followed by a
blank followed by a digit
hihi there
hi hi there
any (one) char followed by
abc
ax, bx, aax,
bbx, abx,
bax,...
ax, aax, aaax
aax, aaax,
aaaax, ...
word character
(alphanumeric plus dash)
followed by a space
followed by a word
character
A string comprised of any
lower and upper case
letters, except "O" and "k"
The period "." (Without the
backward slash the period
means "any character")
linefeed
Ivan Marsic

Rutgers University
350
Compound Types: <complexType>
Compound or complex types can have any kind of combination of element content, character
data, and attributes. The element requires an attribute called name, which is used to refer to the
<complexType> definition. The element then contains the list of sub-elements. You may have
noticed that in the example schema (Listing 6-7), some attributes of the elements from Listing 6-1
were omitted for simplicity sake. For example, <salutation> could have a style attribute,
with the value space defined as {"informal", "formal", "business", "other"}. To
accommodate this, <salutation> should be defined as a complex type, as follows:
Listing 6-11: Upgraded XML Schema for the <salutation> element. This code
replaces line 9 Listing 6-7. The rest remains the same.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
<xsd:element name="salutation">
<xsd:complexType>
<xsd:simpleContent>
<xsd:extension base="xsd:string">
<xsd:attribute name="style" use="required">
<xsd:simpleType>
<xsd:restriction base="xsd:string">
<xsd:enumeration value="informal"/>
<xsd:enumeration value="formal"/>
<xsd:enumeration value="business"/>
<xsd:enumeration value="other"/>
</xsd:restriction>
</xsd:simpleType>
</xsd:attribute>
</xsd:extension>
</xsd:simpleContent>
</xsd:complexType>
</xsd:element>
The explanation of the above listing is as follows:
Line 2: Uses the <complexType> element to start the definition of a new (anonymous)
type.
Line 3: Uses a <simpleContent> element to indicate that the content model of the new
type contains only character data and no elements.
Lines 4–5: Derive the new type by extending the simple xsd:string type. The extension
consists of adding a style attribute using attribute declaration.
Line 6: The attribute style is a simpleType derived from xsd:string by restriction.
Lines 7–12: The attribute value space is specified using the enumeration facet. The attribute
value must be one of the listed salutation styles. Note that the enumeration values specified
for a particular type must be unique.
The content of a <complexType> is defined as follows (see also Figure 6-6):
1. Optional <annotation> (schema comments, which serve as inline documentation)
2. This must be accompanied by one of the following:
Chapter 6 
351
XML and Data Representation
type name can be given by attribute name
?
xsd:notation
xsd:simpleContent
xsd:complexType
xsd:group
xsd:complexContent
xsd:sequence
?
xsd:choice
xsd:all
xsd:attribute

xsd:attributeGroup
?
xsd:anyAttribute
Figure 6-6: Structure of the <complexType> schema element. Symbols follow the notation
introduced in Figure 6-5.
a. <simpleContent> (which is analogous to the <simpleType> element—used
to modify some other “simple” data type, restricting or extending it in some
particular way—see example in Listing 6-11 above)
b. <complexContent> (which is analogous to the <complexType> element—
used to create a compound element)
c. In sequence, the following:
i.
Zero or one from the following grouping terms:
1. <group> — Commonly used to declare a collection of elements that are
referenced from more than one place within the same schema of by other
schemas (hence, this is a global declaration). The personAddressType
type in Listing 6-7 could have been done this way
2. <sequence> — All the named elements must appear in the order listed
3. <choice> — One, and only one, of the elements listed must appear
4. <all> — All the named elements must appear, but order is not important
ii.
Followed by any number of either
1. <attribute>
2. <attributeGroup>
iii.
Then zero or one <anyAttribute> (enables attributes from a given
namespace to appear in the element)
Ivan Marsic

Rutgers University
352
In the example, Listing 6-7, we used both inlined element declaration with anonymous type as
well as named type, which was then used to declare an element. An element declaration can have
a type attribute, or a complexType child element, but it cannot have both a type attribute
and a complexType child element. The following table shows the two alternatives:
Element A references the complexType
foo:
Element A has the complexType definition
inlined in the element declaration:
<xsd:element name="A" type="foo"/>
<xsd:complexType name="foo">
<xsd:sequence>
<xsd:element name="B" .../>
<xsd:element name="C" .../>
</xsd:sequence>
</xsd:complexType>
<xsd:element name="A">
<xsd:complexType>
<xsd:sequence>
<xsd:element name="B" .../>
<xsd:element name="C" .../>
</xsd:sequence>
</xsd:complexType>
</xsd:element>
6.2.4
Reuse
We can roughly split up reuse mechanisms into two kinds: basic and advanced. The basic reuse
mechanisms address the problem of modifying the existing assets to serve the needs that are
perhaps different from what they were originally designed for. The basic reuse mechanisms in
XML Schema are:

Element references

Content model groups

Attribute groups

Schema includes

Schema imports
6.2.5
RELAX NG Schema Language
What we reviewed above is the World Wide Web Consortium’s standard XML Schema. There
are several other alternative schema languages proposed for XML. One of them is RELAX NG.
Its home page (http://www.relaxng.org/) states that RELAX NG is a schema language for XML. The
claims are that RELAX NG:
* is simple
* is easy to learn
* has both an XML syntax and a compact non-XML syntax
* does not change the information set of an XML document
Chapter 6 
XML and Data Representation
353
* supports XML namespaces
* treats attributes uniformly with elements so far as possible
* has unrestricted support for unordered content
* has unrestricted support for mixed content
* has a solid theoretical basis
* can partner with a separate datatyping language (such W3C XML Schema Datatypes)
You could write your schema in RELAX NG and use Trang (Multi-format schema converter
based on RELAX NG) to convert it to XML Schema. See online at:
http://www.thaiopensource.com/relaxng/trang.html
6.3 Indexing and Linking
Links in HTML documents are tagged with <A HREF="http://...">, where the value of
the HREF attribute refers to a target document.
6.3.1
XPointer and Xpath
XML Pointer Language (XPointer) is based on the XML Path Language (XPath), which supports
addressing into the internal structures of XML documents. XPointer allows references to
elements, attributes, character strings, and other parts of XML documents. XPointer referencing
works regardless of whether the referenced objects bear an explicit ID attribute (an attribute
named id, such as id="section4"). It allows for traversals of a document tree and choice of
its internal parts based on various properties, such as element types, attribute values, character
content, and relative position.
XPointers operate on the tree defined by the elements and other markup constructs of an XML
document. An XPointer consists of a series of location terms, each of which specifies a location,
usually relative to the location specified by the prior location term. Here are some examples of
location paths:

child::para
selects the para element children of the context node

child::*
selects all element children of the context node

child::text() selects all text node children of the context node

child::node() selects all the children of the context node, whatever their node type

attribute::name selects the name attribute of the context node

attribute::*
selects all the attributes of the context node
Ivan Marsic

Rutgers University
354

para
matches any para element

*
matches any element

chapter|appendix

olist/item

appendix//para matches any para element with an appendix ancestor element

/
matches the root node

text()
matches any text node

items/item[position()>1] matches any item element that has a items parent
and that is not the first item child of its parent

item[position() mod 2 = 1]
odd-numbered item child of its parent

@class

div[@class="appendix"]//p
matches any p element with a div ancestor
element that has a class attribute with value appendix
matches any chapter element and any appendix element
matches any item element with an olist parent
would be true for any item element that is an
matches any class attribute (not any element that has a class attribute)
The following example is a combination of a URL and an XPointer and refers to the seventh child
of the fourth section under the root element:
http://www.foo.com/bar.html#root().child(4,SECTION).child(7)
6.3.2
XLink
A link is an explicit relationship between two or more data objects or parts of data objects. A
linking element is used to assert link existence and describe link characteristics.
XML Linking Language (XLink) allows elements to be inserted into XML documents in order to
create and describe links between resources. In HTML, a link is unidirectional from one resource
to another and has no special meaning, except it brings up the referred document when clicked in
a browser. XLink uses XML syntax to create structures that can describe the simple
unidirectional hyperlinks of today’s HTML as well as more sophisticated multidirectional and
typed links. With XLink, a document author can do the following, among others:

Associate semantics to a link by giving a “role” to the link.

Define a link that connects more than two resources.

Define a bidirectional link.
A link is an explicit relationship between two or more data objects or portions of data objects. A
linking element is used to assert link existence and describe link characteristics. Linking elements
are recognized based on the use of a designated attribute named xml:link. Possible values are
“simple” and “extended” (as well as “locator”, “group”, and “document”, which
identify other related types of elements). An element that includes such an attribute should be
treated as a linking element of the indicated type. The following is an example similar to the
HTML A link:
Chapter 6 
355
XML and Data Representation
<A xml:link="simple" href="http://www.w3.org/XML/XLink/0.9">
The XLink<A>
An example of an extended link is:
<xlink:extended xmlns:xlink="http://www.w3.org/XML/XLink/0.9"
role="resources"
title="Web Resources"
showdefault="replace"
actuatedefault="user">
<xlink:locator href="http://www.xml.com"
role="resource"
title="XML.com"/>
<xlink:locator href="http://www.mcp.com"
role="resource"
title="Macmillan"/>
<xlink:locator href="http://www.netscape.com"
role="resource"
title="Netscape Communications"/>
<xlink:locator href="http://www.abcnews.com"
role="resource"
title="ABC News"/>
Link Behavior
XLink provides behavior policies that allow link authors to signal certain intentions as to the
timing and effects of link traversal. These include:

Show: The show attribute is used to express a policy as to the context in which a resource
that is traversed to should be displayed or processed. It may take one of three values:
embed, replace, new.

Actuate: The actuate attribute is used to express a policy as to when traversal of a link
should occur. It may take one of two values: auto, user.

Behavior: The behavior attribute is used to provide detailed behavioral instructions.
6.4 Document Transformation and XSL
“If at first you don't succeed, transform your data.”
—The law of computability applied to social sciences
As explained above, XML is not a fixed tag set (like HTML) so the tags do not carry a fixed,
application-specific meaning. A generic XML processor has no idea what is “meant” by the
XML. Because of this, a number of other standards to process the XML files are developed.
Extensible Stylesheet Language (XSL) is one such standard. XML markup usually does not
include formatting information. The information in an XML document may not be in the form in
Ivan Marsic

Rutgers University
356
which it is desired to be presented. There must be something in addition to the XML document
that provides information on how to present or otherwise process the XML. XSL transforms and
translates XML data from one XML format into another. It is designed to help browsers and other
applications display XML. Stated simply, a style sheet contains instructions that tell a processor
(such as a Web browser, print composition engine, or document reader) how to translate the
logical structure of a source document into a presentational structure.
The XML/XSL relationship is reminiscent of the Model-View-Controller design pattern [Gamma
et al., 1995], which separates the core data from the way it gets visualized. Likewise, XSL
enables us to separate the view from the actual data represented in XML. This has following
advantages:
Reuse of data: When the data is separate you do not need to transform the actual data to represent
it in some other form. We can just use a different view of the data.
Multiple output formats: When view is separate from the data we can have multiple output
formats for the same data e.g. the same XML file can be viewed using XSL as VRML, HTML,
XML (of some other form)
Reader’s preferences: The view of the same XML file can be customized with the preferences of
the user.
Standardized styles: Within one application domain there can be certain standard styles which are
common throughout the developer community.
Freedom from content authors: A person not so good at presentation can just write the data and
have a good presenter to decide on how to present the data.
Different ways of displaying an XML files are shown in Figure 6-7.
XSL can act as a translator, because XSL can translate XML documents that comply with two
different XML schemas. XSL is an unfortunate name, since you may think it deals only with
stylesheets. That is not true, it is much more general and as I said, XSL can translate any XML
document to any other XML document.
Chapter 6 
357
XML and Data Representation
XSL
General form of a template rule:
<xsl:template match="pattern">
... action ...
</xsl:template>
XML
Transformation Engine
(XSL Processor)
HTML /
text /
XML
Figure 6-7. How XSL transformation works.
XSL Example
The following example shows an original XML document transformed to an HTML document.
Listing 6-12: Example XSL document.
Original XML source:
1
2
<?xml version='1.0'?>
<para>This is a <emphasis>test</emphasis>.</para>
XSL stylesheet:
1
2
3
4
5
6
7
8
9
10
11
12
13
<?xml version='1.0'?>
<xsl:stylesheet
xmlns:xsl="http://www.w3.org/1999/XSL/Format" version="1.0">
<xsl:template match="para">
<p><xsl:apply-templates/></p>
</xsl:template>
<xsl:template match="emphasis">
<i><xsl:apply-templates/></i>
</xsl:template>
</xsl:stylesheet>
Resultant HTML source:
1
2
<?xml version="1.0" encoding="utf-8"?>
<p>This is a <i>test</i>.</p>
XGMML (eXtensible Graph Markup and Modeling Language) 1.0 Draft
http://www.cs.rpi.edu/~puninj/XGMML/draft-xgmml.html
Ivan Marsic

Rutgers University
358
"Mark Pilgrim returns with his latest Dive into XML column, "XML on the Web Has Failed,"
claiming that XML on the Web has failed miserably, utterly, and completely. Is Mark right or
wrong? You be the judge."
http://www.xml.com/pub/a/2004/07/21/dive.html
6.5 Summary and Bibliographical Notes
As a historical footnote, XML is derived from SGML (Standard Generalized Markup Language),
which is a federal (FIPS 152) and international (ISO 8879) standard for identifying the structure
and content of documents.
I have no intention of providing a complete coverage of XML since that would require more than
a single book and would get us lost in the mind numbing number of details. My main focus is on
the basic concepts and providing enough details to support meaningful discussion. I do not expect
that anybody would use this text as an XML reference. The reader interested in further details
should consult the following and other sources.
XML is defined by the W3C in a number of related specifications available here:
http://www.w3.org/TR/. A great source of information on XML is http://www.xml.com/.
The standard information about HTTP is available here: http://www.w3.org/Protocols/
HTML standard information is available here: http://www.w3.org/MarkUp/
XML Tutorial online at: http://www.w3schools.com/xml/default.asp
Reference [Lee & Chu, 2000] reviews several alternative XML schema languages.
A book by Eric van der Vlist, RELAX NG, O’Reilly & Associates, is available online at:
http://books.xmlschemata.org/relaxng/page1.html .
Chapter 6 
XML and Data Representation
359
Problems
Problem 6.1
Problem 6.2
Write the XML Schema that defines the production rules for the instance document shown in
Listing 6-13 below. The parameters are specified as follows.
Possible values for the attribute student status are “full time” and “part time” and it is required
that this attribute appears.
The student identification number must be exactly 9 digits long and its 4th and 5th digits must
always be a zero ( ַ◌ ַ◌ ַ◌ 00 ַ◌ ַ◌ ַ◌ ַ◌ ). (According to the US Social Security Administration, a
number with a zero in the 4th and 5th digits will never be assigned as a person’s SSN. Hence, you
can easily distinguish the difference between the student id and the SSN by scanning the 4th and
5th digits.)
The school number must be a two-digit number (including numbers with the first digit equal to
zero).
The graduation class field should allow only Gregorian calendar years.
The curriculum number must be a three-digit number between 100 and 999.
The student grade field is optional, but when present it can contain only one of the following
values: “A,” “B+,” “B,” “C+,” “C,” “D,” and “F.”
All elements are required, unless stated otherwise. As for the non-specified parameters, make
your own (reasonable) assumptions. Write down any assumptions you make.
Listing 6-13: Instance XML document containing a class roster.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
<?xml version="1.0" encoding="UTF-8"?>
<!-- ***** associate here your schema with this document ***** -->
<class-roster>
<class-name> Introduction to Software Engineering </class-name>
<index> 61202 </index>
<semester> Spring 2006 </semester>
<enrollment> 58 </enrollment>
<student status="full-time">
<student-id> 201000324 </student-id>
<name>
<first-name> Jane </first-name>
<last-name> Doe </last-name>
</name>
<school-number> 14 </school-number>
<graduation-class> 2006 </graduation-class>
<curriculum> 332 </curriculum>
<grade> A </grade>
Ivan Marsic
18
19
20
21
22

Rutgers University
</student>
<student status="part-time">
...
</student>
</class-roster>
Problem 6.3
360
Chapter 7
Software Components
Contents
“The Organism Principle: When a system evolves to become
more complex, this always involves a compromise: if its parts
become too separate, then the system’s abilities will be
limited—but if there are too many interconnections, then each
change in one part will disrupt many others.”
—Marvin Minsky, The Emotion Machine
Software engineers have always envied hardware engineers
for their successful standardization of hardware design and the
power of integration and reusability achieved by the
abstraction of VLSI chips. There have been many attempts in
the past to bring such standardization to software design.
Recently, these seem to be achieving some success, most
probably because the level of knowledge in software
engineering has achieved the needed threshold.
There is currently a strong software industry trend towards
standardizing software development through software
components. Components are reusable pieces of software.
Components can be GUI widgets, non-visual functions and
services, applets or large-scale applications. Each component
can be built by different developers at separate times.
Components enable rapid development using third party
software: independent components are used without
modifications as building blocks to form composite
applications. Components can be composed into:

Composite components

Applets (small client-side applications)

Applications

Servlets (small server-side applications)
7.1 Components, Ports, and Events
7.1.1
7.1.2
7.1.3
7.1.4
7.1.5
7.1.6
7.2 JavaBeans: Interaction with Components
7.2.1 Property Access
7.2.2 Event Firing
7.2.3 Custom Methods
7.2.4
7.3 Computational Reflection
7.3.1 Run-Time Type Identification
7.3.2 Reification
7.3.3 Automatic Component Binding
7.3.4
7.2.3
7.4 State Persistence for Transport
7.4.1
7.4.2
7.4.3
7.4.4
7.5 A Component Framework
7.5.1 Port Interconnections
7.5.2
7.5.3
7.5.4
7.6
7.6.1
7.6.2
7.6.3
7.7 Summary and Bibliographical Notes
Problems
Although a single class may not be a useful unit of reuse, a component that packages a number of
services can be. Components enable medium-grained reuse.
The composition can be done in visual development tools, since the components expose their
features to a builder tool. A builder tool that lets you:
361
Ivan Marsic

Rutgers University
362

Graphically assemble the components into more complex components or applications

Edit the properties of components

Specify how information in a component is to be propagated to other components
The component developer has to follow specific naming conventions (design pattern), which help
standardize the interaction with the component. In this way the builder tool can automatically
detect the component’s inputs and outputs and present them visually. If we visualize a bean as an
integrated circuit or a chip, the interaction methods can be visualized as the pins on the chip.
Two major component architectures are JavaBeans from Sun [Sun Microsystems, JavBeans] and
ActiveX controls from Microsoft [Denning, 1997]. Here I first review the JavaBeans component
model, which is a part of the Java Development Kit version 1.1 or higher. JavaBeans component
model is a specification standard, not an implementation in a programming language. There is not
a class or interface in Java called Bean. Basically, any class can be a bean—the bean developer
just has to follow a set of design patterns, which are essentially guidelines for naming the
methods to interact with the bean.
Software components, as any other software creation, comprise state and behavior.
Two key issues in component development are

How to interact with a component

How to transfer a component’s state from one machine to another
Programming business logic of reusable components is the same as with any other software
objects and thus it is not of concern in a component standard.
7.1 Components, Ports, and Events
“Before software can be reusable it first has to be usable.”
—Ralph Johnson
The hardware-software component analogy is illustrated in Figure 7-1. Component communicates
with the rest of the world only via its ports using events. This simplification and uniformity of the
component model is promoted as the main reason for introducing components as opposed to
software objects. Objects succeeded in encapsulation of state and behavior (see Section 1.4), but
have not had much success on the issue of reuse. It is claimed that the main reason for this is that
object interconnections are often concealed and difficult to identify. We can easily determine the
“entry points” of objects, i.e., the points through which other objects invoke the given object,
which are its methods for well-designed objects. However, it is difficult to pinpoint the “exit
points” of objects—the points through which the object invokes the other objects—without
carefully examining the source code. Consider the following example (in Java):
class MyClass {
...
Chapter 7 
363
Software Components
AMPLIFIER NO. 2
VCC+
Component
14
13
12
11
10
9
8
1
2
3
4
5
6
7
Port
NONINV
OUT- INV
PUT INPUT INPUT
8
7
6
5
1
2
3
4
OUT- INV NON- VCC–
PUT INPUT INV
INPUT
AMPLIFIER NO. 1
(a)
(b)
(c)
Figure 7-1. Hardware analogy for software component abstraction. (a) Software component
corresponds to a hardware chip. (b) A component has attached ports (pins), each with a
distinctive label. (c) Events or “signals” arriving at the ports are processed within the
component and the results may be output on different ports.
public void doSomething(AnotherClass obj1) {
...
obj1.getObj2().callMethod(args);
...
}
}
Here, the method getObj2() returns an object of a different class, which we would not know
that is being involved without careful code examination. Hence the importance of enforcing
uniform style for getting into and out of components, i.e., via their ports.
7.2 JavaBeans: Interaction with Components
The simplest thing to do with a software component is to retrieve its state or alter it by explicitly
setting the new state or invoking a behavior of the component.
Reusable software components are usually shipped around in a compiled code, rather than in
source code. Given a compiled component (bytecode or binary code), the goal is to uncover its
public methods in order to be able to interact with it. The process of discovering the features of a
class is known as introspection. The bean developer can help the introspection process in two
ways:

Implicitly, by adhering to certain conventions (design patterns) in writing the code for a
Java bean

Explicitly, by specifying explicit additional information about the bean
Ivan Marsic

Rutgers University
364
The second way should be used in case bean contains interaction methods that do not follow the
design patterns.
Related to introspection is reflection, the support Java provides for examining type information at
run time. Given an object or class name, you can use the class Class to discover:

The fields of the class

Constructors

Methods

Other properties (class name, isPrimitive, etc.)
The reader may wonder why anyone would want to write a program that does this; why not look
up the needed information when the program is written? Why wait until run time? The answer is
that this capability allows the other applications to discover the way to interact with a bean that
was developed by a third party. Reflection is used to gain information about components, but the
component developer is allowed to specify more information to help with characterizing the
component.
7.2.1
Property Access
Properties define the bean’s appearance and behavior characteristics. For properties that need to
be exposed, the developer must provide:

Setter method void set<PropertyName>(Type newvalue)
only property, e.g., password, or

Getter method Type get<PropertyName>()

Both setter and getter
// write-
// read-only property, or
// read-write property.
In addition to naming the accessor methods according to these conventions, the developer may
also provide property editors for certain properties. For example, to a property may determine the
bean’s background color. The user may enter the value for this property as a hexadecimal number
“1b8fc0,” but it is difficult or impossible for the user to visualize how this color15 looks.
Instead, the developer may supply a graphical color chooser for the property editor, which is
much more convenient for the user.
7.2.2
Event Firing
The delegation-based event model was introduced with the JavaBeans framework [SunJavaBeans]. In this model, there is no central dispatcher of events; every component that
generates events dispatches its own events as they happen. The model is a derivative of the
Publisher-Subscriber pattern. Events are identified by their class instead of their ID and are either
propagated or delegated from an “event source” to an “event listener.”
15
In case you are looking at a black-and-white print, the background color around the text is magenta.
Chapter 7 
365
Software Components
According to the delegation model, whenever an event for which an object declared itself as a
source gets generated, the event is multicast to all the registered event listeners. The source object
delegates of “fires” the events to the set of listeners by invoking a method on the listeners and
passing the corresponding event object. Only objects interested in a particular event need to deal
with the event and no super-event handler is required. This is also a better way of passing events
to distributed objects.
EventSource — Any object can declare itself as a source of certain types of events. An event
source has to either follow standard beans design patterns when giving names to the methods or
use the BeanInfo class to declare itself as a source of certain events. When the source wishes to
delegate a specific event type, it must first define a set of methods that enable listener(s) to
register with the source. These methods take the form of set<EventType>Listener for
unicast and/or add<EventType>Listener for multicast delegation. [The source must also
provide the methods for the listeners de-register.]
EventListener — An object can register itself as a listener of a specific type of events
originating from an event source. A listener object should implement the
<EventType>Listener interface for that event type, which inherits from the generic
java.util.EventListener interface. The “Listener” interface is typically defined only by
few methods, which makes it easy to implement the interface.
7.2.3
Custom Methods
In addition to the information a builder tool can discover from the bean’s class definition through
reflection, the bean developer can provide it with explicit additional information. A bean can be
customized for a specific application either programmatically, through Java code, or visually,
through GUI interfaces hosted by application builder tools. In the former case, the developer
specifies additional information by providing a BeanInfo object. In the latter case, the
developer can provide customized dialog boxes and editing tools with sophisticated controls.
These customization tools are called property editors and customizers, and they are packaged as
part of the bean, by providing the PropertyEditor and Customizer classes.
The BeanInfo class should be named as <BeanName>BeanInfo, for example, for the Foo
bean the BeanInfo class should be named FooBeanInfo.
class FooBeanInfo extends SimpleBeanInfo {
:
:
with methods:
getBeanDescriptor()
getIcon()
getMethodDescriptors()
getPropertyDescriptors()
//
//
//
//
has
for
for
can
class and customizer
displaying the bean in the palette
providing more information than
be gained through reflection alone
A property descriptor can provide a PropertyEditor, in case the developer does not want to use the
standard property editor for that property type (or there is not one available).
In addition, the developer can provide a Customizer in the bean descriptor. Customizers are used
to customize the entire bean, not just a property and they are not limited to customizing
Ivan Marsic

Rutgers University
366
Metadata
Introspection
Reification
Data
Figure 7-2: Computational reflection consists of two phases: (i) an introspection phase,
where data is analyzed to produce suitable metadata, and (ii) a reification phase, where
changes in the metadata alter the original behavior of the data it represents.
properties. There is no “design pattern” for Customizers. You must use the BeanInfo to use a
customizer and you need to use the BeanDescriptor constructor to specify the customizer
class. More information about bean customization is available in [Johnson, 1997].
7.3 Computational Reflection
Computational reflection is a technique that allows a system to maintain information about itself
(meta-information) and use this information to change its behavior (adapt). As shown in Figure
7-2, computational reflection refers to the capability to introspect and represent meta-level
information about data or programs, to analyze and potentially modify the meta-level
representation of the data, and finally to reify such changes in the metadata so that the original
data or programs behave differently. It should be noted that the notion of data is universal in that
it includes data structures in a program used in the source code.
This is achieved by processing in two well-defined levels: functional level (also known as base
level or application level) and management (or meta) level. Aspects of the base level are
represented as objects in the meta-level, in a process known as reification (see below). Meta-level
architectures are discussed in Section 2.2 (??) and reflective languages in Section 2.3. Finally,
Section 2.4 shows the use of computational reflection in the structuring and implementation of
system-oriented mechanisms.
http://cliki.tunes.org/Reflection
Chapter 7 
7.3.1
Software Components
367
Run-Time Type Identification
If two processes communicate externally to send and receive data, what happens when the data
being sent is not just a primitive or an object whose type is known by the receiving process? In
other words, what happens if we receive an object but do not know anything about it—what
instance variables and methods it has, etc. Another way to pose the question: What can we find
out about the type of an object at run-time?
A simple way to solve this problem is to check for all possible objects using instanceof, the
operator that lets you test at run-time, whether or not an object is of a given type. A more
advanced way is supported by the java.lang.reflect package, which lets you find out
almost anything you want to know about an object’s class at run-time.
An important class for reflection is the class Class, which at first may sound confusing. Each
instance of the class Class encapsulates the information about a particular class or interface.
There is one such object for each class or interface loaded into the JVM.
There are two ways to get an instance of class Class from within a running program:
1. Ask for it by name using the static method forName():
Class fooClass = Class.forName("Foo");
This method will return the Class object that describes the class Foo
2. Ask an instance of any Object for its class:
Foo f = new Foo();
Class fooClass = f.getClass();
As a side note, this construct is legal:
Class classClass = Class.forName("Class");
It returns back the instance of Class that describes the class Class.
Once you have a Class object, you can call methods on it to find out information about the
class. None of the methods are mutators, so you cannot change the class at run-time. However,
you can use it to create new instance of a class, and to call methods on any instance. Some of the
methods available in class Class are:
Constructor getConstructor(Class[] paramTypes);
Constructor[] getConstructors();
Field getField(String name);
Field[] getFields();
Method getMethod(String name, Class[] paramTypes);
Method[] getMethods();
boolean isInstance(Object obj);
boolean isPrimitive();
String getName();
Ivan Marsic

Rutgers University
caller
368
kernel
op : Operation
: MetaObject
obj : Object
res : Result
method(arg)
create(obj, method, arg)
op
handle(op)
method(arg)
result
create(op, result)
res
handle(res)
result
Figure 7-3. Reifying an operation. See text for explanation.
String toString();
The return types Constructor, Field, and Method are defined in the package
java.lang.reflect.
7.3.2
Reification
For the meta-level to be able to reflect on several objects, especially if they are instances of
different classes, it must be given information regarding the internal structure of objects. This
meta-level object must be able to find out what are the methods implemented by an object, as
well as the fields (attributes) defined by this object. Such base-level representation, that is
available for the meta-level, is called structural meta-information. The representation, in form of
objects, of abstract language concepts, such as classes and methods, is called reification.
Base-level behavior, however, cannot be completely modeled by reifying only structural aspects
of objects. Interactions between objects must also be materialized as objects, so that meta-level
objects can inspect and possibly alter them. This is achieved by intercepting base-level operations
such as method invocations, field value inquiries or assignments, creating operation objects that
represent them, and transferring control to the meta level, as shown in Figure 7-3. In addition to
receiving reified base-level operations from the reflective kernel, meta-level objects should also
be able to create operation objects, and this should be reflected in the base level as the execution
of such operations.
A reflective kernel is responsible for implementing an interception mechanism. The method
invocation is reified as an operation object and passed for the callee’s meta-object to reflect upon
(handle). Eventually, the meta-object requests the kernel to deliver the operation to the callee’s
replication object, by returning control (as in the diagram) or performing some special meta-level
action. Finally, the result of the operation is reified and presented to the meta-object.
Chapter 7 
Software Components
369
Show example of reflection using DISCIPLE Commands in Manifold
Reflection enables dynamic (run-time) evolution of programming systems, transforming
programs statically (at compile-time) to add and manage such features as concurrency,
distribution, persistence, or object systems, or allowing expert systems to reason about and adapt
their own behavior. In a reflective application, the base level implements the main functionality
of an application, while the meta level is usually reserved for the implementation of management
requirements, such as persistence [28,34], location transparency [26], replication [8,18], fault
tolerance [1,2,9,10] and atomic actions [35,37]. Reflection has also been shown to be useful in the
development of distributed systems [6,20,36,40,41] and for simplifying library protocols [38].
http://www.dcc.unicamp.br/~oliva/guarana/docs/design-html/node6.html#transparency
A recent small project in Squeak by Henrik Gedenryd to develop a "Universal Composition"
system for programs. It essentially involves a graph of meta-objects describing sourcecomposition operations which can be eagerly or lazily (statically or dynamically) driven, resulting
in partial evaluation or forms of dynamic composition such as Aspect-Oriented Programming
(AOP).
7.3.3
Automatic Component Binding
Components can be seen as pieces of a jigsaw puzzle, like molecular binding—certain molecule
can bind only a molecule of a corresponding type.
7.4 State Persistence for Transport
A class is defined by the Java bytecode and this is all that is necessary to create a fresh new
instance of the class. As the new instance (object) interacts with their objects, its state changes.
The variables assume certain values. If we want a new instance resume from this state rather than
from the fresh (initial) state, we need a mechanism to extract the object state. The mechanism is
known as object serialization or in the CORBA jargon it is called object externalization [OMGCORBA-Services].
Object serialization process transforms the object graph structure into a linear sequence of bytes,
essentially a byte array. The array can be stored on a physical support or sent over the network.
The object that can be serialized should implement the interface java.io.Serializable.
The object essentially implements the methods writeObject() and readObject(). These
methods define how to convert the component attributes, which represented by programmerdefined data types, to a one-dimensional bytestream.
When restoring the object, we need to have its class (bytecode) because class definition is not
stored in the serialized state. If the receiving process does not know the object's class, it will
Ivan Marsic

Rutgers University
370
throw the java.io.SerializationError exception. This may be a problem if we are
sending the object to a server which is running all the time and cannot load new classes, so its
class loader cannot know about the newly defined class. The solution is to use the method:
byte[] getBytes();
which is available on every java.lang.Object object, i.e., on every Java object. The method
returns a byte array—a primitive data type, which can be used by the server to reconstruct the
object there.
JAR file contains:

Manifest file

Classes (next to the class name, there is a Boolean variable “bean” which can be true or
false

Bean customizers
Beans provide for a form of state migration since the bean state can be “canned” (serialized) and
restored on a remote machine (unlike an applet which always starts in an initial state after landing
on a remote machine). However, this is not object migration since the execution state (program
counters for all threads) would need to be suspended and resumed. Persistency is more meant to
preserve the state that resulted from external entities interacting with the bean, rather than the
state of the execution, which would be possible to resume on another machine.
7.5 A Component Framework
“All parts should go together without forcing. You must remember that the parts you are reassembling
were disassembled by you. Therefore, if you can’t get them together again, there must be a reason. By all
means, do not use a hammer.”
—IBM maintenance manual, 1925
Here I present a component framework that I designed, which is inspired by several component
frameworks existing in research laboratories. Compared to JavaBeans, this framework is more
radical in enforcing the component encapsulation and uniformity of communication with it.
7.5.1
Port Interconnections
Options for component wiring are illustrated in Figure 7-4. The ports of different components can
be connected directly, one-on-one. In the simplest case, output port of a component directly
connects to an input port of another component. Another useful analogy is wire, which is a
broadcast medium to which multiple ports can be connected. Similar effect could be achieved by
the Publisher-Subscriber pattern (see Section 4.1 above), but the Wire abstraction appears to be
more elegant in the context of components and ports.
In the spirit of object-oriented encapsulation, components as defined here share nothing—they
have no shared state variables. The communication is entirely via messages. Of course,
Chapter 7 
A
371
Software Components
B
D
E
I
G
H
F
C
Wire
(a)
Composite
component
(b)
Prefix
ports
(c)
Figure 7-4. Options for wiring the ports. (a) Component ports can be directly “soldered”
one-on-one. (b) The abstraction of wire provides a broadcast medium where the event from
any output connected to the wire appears on all input ports connected to the same wire. (c)
Prefix ports are used in wiring composite components.
communication via ports is limited to the inter-component communication. Components contain
one or more objects, which mutually communicate via regular method calls. Similarly,
components make calls to the runtime environment and vice versa, via method calls. The only
requirement that is enforced is that components communicate with each other via ports only.
Components can be composed into more complex, composite components, as in Figure 7-1(c),
where each operational amplifier within the chip could be a separate “component.” Composing
components is illustrated in Figure 7-4(c), where component H contains component I. Notice that
one of the output ports of component G connects to an input port of component H which is
directly connected to an input port of component I. Regular input port cannot be connected to
another input port. Similar is true for output ports. To support forming chains of ports of the same
type, we introduce a prefix port sub-type, shown in Figure 7-4(c).
Typical event communication and processing in a single-threaded system is illustrated in the
sequence diagram in Figure 7-5. In this example, component A receives Event 1 on the input port
a1, processes it and generates event 2 on the output port a2. This event is received and processed
by component B. The dashed lines at the bottom of the figure indicate how the thread returns after
this sequential processing is completed.
All components and ports are named and addressable using a Unix-type path. For example, full
path format for a port on a nested component is as:
container_component_name/inner_component_name@port_name
Component names are separated by forward slashes (/) and the port name is preceded by “at” sign
(@). Thus, the components and their ports form a tree structure.
Design Issues
It was debated whether to strictly enforce the Port model for communicating with Components.
Currently, actions that require computation go via Ports. Conversely, access to state variables
(component properties) is via setProperty()/getProperty() methods. So, if component
has a handle/reference to another component (which normally should not happen!), it can invoke
these methods. Of course, the state access could also go over the Ports, but convenience was
Ivan Marsic

Rutgers University
a1 : PortIn
receive(Event1)
372
A : Component
a2 : PortOut
b1 : PortIn
B : Component
process(Event1)
send(Event2)
receive(Event2)
process(Event2)
Figure 7-5. Typical event communication and processing in a single-threaded system.
preferred over compliance. Further deliberation may be warranted to evaluate the merits and
hazards of the current solution.
Another issue is, what if, in Figure 7-5, component A needs some return value from component
B? This could probably be achieved by connecting B’s output port to A’s another input port, but is
this the most elegant/efficient solution?
7.5.2
Levels of Abstraction
A key question with components is the right “size”, the level of abstraction. Low level of
specialization provides high generality and reusability, but also low functionality thus resulting in
productivity gain.
Cf. Lego blocks: Although it is possible to build anything using the simplest rectangular blocks,
Lego nevertheless provides specialized pre-made pieces for certain purposes. The point is not
whether anything can be built with universal components; the point is whether it is cost effective
to do so.
Recall Example 3.1 (??) about the simplified domain model of the virtual biology lab (described
at the book website, given in Preface). We compared the solution presented in Problem 2.12 that
uses abstract objects modeled after the physical objects against a simplified solution that uses
only abstract geometric figures (lines, circles, polygons, etc.). True, the simplified model is not
adequate because classes Cell or Nucleus have relatively strong semantic meaning, whereas class
Circle can stand for both of these and many other things. However, one may wonder whether we
need to represent every detail in the problem domain by a different class. Consider human bodies
composed of cells—there are only 255 or so specialized sorts of cell [Alberts, et al., 1989]. For
comparison, Java libraries have thousands different classes. Of course, each cell has many
complex elements. One may wonder whether it is possible to have a similar, biologically inspired,
hierarchical abstraction and composition of functions. UML packages contain classes, but they
Chapter 7 
Software Components
373
are not functional abstraction. Work on software architectures is in early stages and may
eventually offer the path for component abstraction.
7.6 Summary and Bibliographical Notes
Cite a book on Java Beans.
Computational reflection was introduced in [Smith, 1982]. A review is available in [Maes, 1987].
The component design presented in Section 7.5 is derived from the current literature, mostly the
references [Allan et al., 2002; Bramley et al., 2000; Chen & Szymanski, 2001; Hou et al., 2005;
Szymanski & Chen, 2002; Tyan et al., 2005]. Additional information is available from the
Common Component Architecture (CCA) Forum at http://www.cca-forum.org/.
Problems
Chapter 8
Web Services
Contents
A Web service is a software function or resource offered as a
service; remotely accessed over the “web”; and, has
programmatic access using standard protocols. A Web service
is an interface that describes a collection of operations that are
network accessible through standardized XML messaging.
Web services fulfill a specific task or a set of tasks. A Web
service is described using a standard, formal XML notion,
called its service description that provides all the details
necessary to interact with the service including message
formats, transport protocols and location.
8.1 Service Oriented Architecture
While there are several definitions available it can be broadly
agreed that a Web service is a platform and implementation
independent software component that can be,
8.3 WSDL for Web Service Description

Described using a service description language

Published to a registry of services

Discovered through a standard mechanism

Invoked through a declared API, usually over a
network

Composed with other services
8.1.1
8.1.2
8.1.3
8.1.4
8.1.5
8.1.6
8.1.7
8.2 SOAP Communication Protocol
8.2.1
8.2.2
8.2.3
8.2.4
The SOAP Message Format
The SOAP Section 5 Encoding
SOAP Communication Styles
Binding SOAP to a Transport Protocol
8.3.1 The WSDL 2.0 Building Blocks
8.3.2 Defining a Web Service’s Abstract
Interface
8.3.3 Binding a Web Service Implementation
8.3.4 Using WSDL to Generate SOAP Binding
8.3.5 Non-functional Descriptions and Beyond
WSDL
8.4 UDDI for Service Discovery and Integration
8.4.1
8.4.2
8.4.3
8.4.4
8.5 Developing Web Services with Axis
Web services are characterized as loose coupling of
applications—clients and service providers need not be known
a priori. The underlying mechanism that enables this is:
publish-find-bind, or sometimes called find-bind-execute. The
application can be developed without having to code or
compile in what services you need. Similarly, when service
provider deploys a service, it does not need to know its clients.
In summary, (1) Service publishes its description; (2) Client
finds service based on description; and, (3) Client binds itself
to the service.
374
8.5.1 Server-side Development with Axis
8.5.2 Client-side Development with Axis
8.5.3
8.5.4
8.6 OMG Reusable Asset Specification
8.6.1
8.6.2
8.6.3
8.7 Summary and Bibliographical Notes
Problems
Chapter 8 
Web Services
375
The only common thing across Web services is the data format (ontology). There is no API’s
involved, no remote service invocation. Each “method” is a different service; invocation is
governed by the “service contract.” A web site (portal) provides a collection of Web services.
Tom Gruber, What is an Ontology? Online at: http://www-ksl.stanford.edu/kst/what-is-anontology.html
A closer look at Microsoft's new Web services platform, "Indigo," from the same source.
http://www.eweek.com/article2/0,1759,1763162,00.asp
Web services essentially mean using SOAP Remote Procedure Call. Web services function in the
information “pull” mode. The reason for this is firewalls, which allow only pulling of the
information with Web services. Although HTTP 1.1 allows keeping state on the server, it is still a
“pull” mode of communication.
The “pull” mode is not appropriate for distributing real-time information; for this we need the
“push” mode. If we want to use Web services in the “push” mode, we have two options:
1. Use tricks
2. Web services pushing unsolicited notification
In the first case, we have an independent broker to which the clients that are to receive
notifications should connect. The broker is in the trusted part of the network and it helps to push
information to the client of the Web service.
In the second case, the Web service standard needs to be modified to allow pushing information
from the server. An important issue that needs to be considered is whether this capability can lead
to denial-of-service attacks.
Peer-to-peer communication is also an issue because of firewalls.
So, What the Heck Are Web Services?
http://www.businessweek.com/technology/content/feb2005/tc2005028_8000_tc203.htm
A "Milestone" for Web Services
http://www.businessweek.com/technology/content/feb2005/tc2005028_4104_tc203.htm
GENERAL:
http://www.businessweek.com/technology/index.html
Web Services Leaders Submit Key Messaging Spec: A group of leading Web services
proponents, including Microsoft and Sun Microsystems, on Tuesday announced the joint
submission of a key Web services specification to the World Wide Web Consortium (W3C).
http://www.eweek.com/article2/0,1759,1634158,00.asp
Read how this could signal a turning point in the battle over Web services specifications
intellectual property rights.
Ivan Marsic

Rutgers University
Pu
1.
376
er
ist
g
e
/R
sh
i
l
b
2.
Fin
d/ S
ea
rc
Discovery Agency /
Registry
h
Service
Description
3. Bind/Use
Service Provider
Service Customer
Figure 8-1: Service Oriented Architecture—components and relationships.
8.1 Service Oriented Architecture
Service Oriented Architecture (SOA) is an architectural style whose goal is to achieve loose
coupling among interacting software agents. Typically, service oriented architecture contains
three components (see Figure 8-1): a service provider, a service customer, and a service registry.
A service is a unit of work done by a service provider to achieve desired end results for a service
customer. Both provider and customer are roles played by software agents on behalf of their
users. SOA is a simple but powerful concept which can be applied to describe a wide variety of
Web service implementations.
A service provider creates a service description, publishes that service description to one or more
service registries and receives Web service invocation requests from one or more service
consumers. It is important to note that the service provider publishes the description of how the
service behaves and not the service code. This service description informs the service customer
everything it needs to know in order to properly understand and invoke the Web service. More
information on service description is available in Section 8.3 below. A service customer finds
service descriptions published to one or more service registries and use service descriptions to
bind or to invoke Web services hosted by service providers.
Chapter 8 
377
Web Services
data tracking
and machine learning
: Backend
discovery agency
: UDDI, WSIL
Delphi method facilitator
: Service Requestor
: SOAP Runtime
stock forecast expert
: Service Provider
http
: Transport
http
: Transport
: SOAP Runtime
forecaster description
: WSDL
create
obtain
publish
initialize
find
get
recommendation
invoke
do
forecast
Figure 8-2: Web services dynamic interactions.
8.2 SOAP Communication Protocol
SOAP (Simple Object Access Protocol or Service Oriented Architecture Protocol) is the
communication protocol for Web services. It is intended for exchanging structured information
(based on XML) and is relatively simple (lightweight). Most commonly it runs over HTTP
(Appendix C), but it can run over a variety of underlying protocols. It has been designed to be
independent of any particular programming model and other implementation-specific semantics.
A key advantage of SOAP is that, because it is XML based, it is programming-language,
platform, and hardware independent.
SOAP, as any other communication protocol, governs how communication happens and how data
is represented on the wire. The SOAP framework is an implementation of the Broker design
pattern (Section 5.4) and there are many similarities between SOAP and Java RMI (or CORBA).
This section describes SOAP version 1.2, which is the current SOAP specification. The older
SOAP version 1.1 is somewhat different.
SOAP defines the following pieces of information, which we will look at in turn:

The way the XML message is structured

The conventions representing a remote procedure call in that XML message
Ivan Marsic

Rutgers University
SOAP envelope
SOAP header
header block
378
<Envelope>
<Header>
header blocks
</Header>
<Body>
SOAP body
body block
body blocks
</Body>
attachment blocks
attachment block
Identifies message as
a SOAP message
(required)
</Envelope>
Processing instructions
Context information
(optional)
Actual message content
(required)
Arbitrary content
(optional)
Figure 8-3: Schematic representation of a SOAP message. Highlighted are the required
elements.

A binding to HTTP, to ensure that the XML message is transported correctly

The conventions for conveying an error in message processing back to the sender
8.2.1
The SOAP Message Format
A unit of communication in SOAP is a message. A SOAP message is an ordinary XML document
containing the following elements (Figure 8-3):

A required Envelope element that identifies the XML document as a SOAP message

An optional Header element that contains the message header information; can include
any number of header blocks (simply referred to as headers); used to pass additional
processing or control information (e.g., authentication, information related to transaction
control, quality of service, and service billing and accounting-related data)

A required Body element that contains the remote method call or response information;
all immediate children of the Body element are body blocks (typically referred to simply
as bodies)

An optional Fault element that provides information about errors that occurred while
processing the message
SOAP messages are encoded using XML and must not contain DTD references or XML
processing instructions. Figure 8-4 illustrates the detailed schema for SOAP messages using the
notation introduced in Figure 6-5. If a header is present in the message, it must be the first
immediate child of the Envelope element. The Body element either directly follows the
Header element or must be the first immediate child of the Envelope element if no header is
present.
Because the root element Envelope is uniquely identified by its namespace, it allows
processing tools to immediately determine whether a given XML document is a SOAP message.
The main information the sender wants to transmit to the receiver should be in the body of the
message. Any additional information needed for intermediate processing or added-value services
(e.g., authentication, security, transaction control, or tracing and auditing) goes into the header.
This is the common approach for communication protocols. The header contains information that
Chapter 8 
379
Web Services
Envelope
soap-env:Envelope
Header
?
soap-env:Header
?
soap-env:Header
?
soap-env:Body
?
##other:any
##other:anyAttribute
##other:any
Body
soap-env:mustUnderstand
soap-env:role
?
soap-env:Body
Global attributes:
?
Fault
soap-env:Fault
##other:anyAttribute
Code
soap-env:encodingStyle
soap-env:relay
##other:any
Reason
?
Node
?
Role
?
Detail
Figure 8-4: The XML schema for SOAP messages. The graphical notation is given in Figure
6-5. (SOAP version 1.2 schema definition available at: http://www.w3.org/2003/05/soap-envelope).
can be used by intermediate nodes along the SOAP message path. The payload or body is the
actual message being conveyed. This is the reason why the header is optional.
Each of the SOAP elements Envelope, Header, or Body can include arbitrary number of
<any> elements. Recall that the <any> element enables us to extend the XML document with
elements not specified by the schema. Its namespace is indicated as ##other, which implies
elements from any namespace that is not the namespace of the parent element, that is, soapenv.
An example SOAP message containing a SOAP header block and a SOAP body is given as:
Listing 8-1: Example of a SOAP message.
1 <soap-env:Envelope
2
xmlns:soap-env="http://www.w3.org/2003/05/soap-envelope"
3
xmlns:xsd="http://www.w3.org/2001/XMLSchema"
4
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance">
5
<soap-env:Header>
6
<ac:alertcontrol xmlns:ac="http://example.org/alertcontrol"
7
soap-env:mustUnderstand="1">
8
<ac:priority>high</ac:priority>
9
<ac:expires>2006-22-00T14:00:00-05:00</ac:expires>
10
</ac:alertcontrol>
11
</soap-env:Header>
12
<soap-env:Body>
13
<a:notify xmlns:a="http://example.org/alert">
14
<a:note xsi:type="xsd:string">
Ivan Marsic

Rutgers University
380
15
Reminder: meeting today at 11AM in Rm.601
16
</a:note>
17
</a:notify>
18
</soap-env:Body>
19 </soap-env:Envelope>
The above SOAP message is a request for alert to a Web service. The request contains a text note
(in the Body) and is marked (in the Header) to indicate that the message is high priority, but
will become obsolete after the given time. The details are as follows:
Lines 1–2: Prefix soap-env, declared in Line 2, identifies SOAP-defined elements, namely
Envelope, Header, and Body, as well as the attribute mustUnderstand (appears in
Line 7).
Line 3: Prefix xsd refers to XML Schema elements, in particular the built-in type string
(appears in Line 14).
Line 4: Prefix xsi refers to XML Schema instance type attribute, asserting the type of the
note as an XML Schema string (appears in Line 14).
Line 7: The mustUnderstand attribute value "1" tells the Web service provider that it
must understand the semantics of the header block and that it must process the header. The
Web service requestor demands express service delivery.
Lines 12–18: The Body element encapsulates the service method invocation information,
namely the method name notify, the method parameter note, its associated data type and
its value.
SOAP message body blocks carry the information needed for the end recipient of a message. The
recipient must understand the semantics of all body blocks and must process them all. SOAP does
not define the schema for body blocks since they are application specific. There is only one
SOAP-defined body block—the Fault element shown in Figure 8-4—which is described below.
A SOAP message can pass through multiple nodes on its path. This includes the initial SOAP
sender, zero or more SOAP intermediaries, and an ultimate SOAP receiver. SOAP intermediaries
are applications that can process parts Service Requestor
Service Provider
of a SOAP message as it travels from
Receiver
Sender
Intermediary
Intermediary
the
sender
to
the
receiver.
Intermediaries can both accept and forward (or relay, or route) SOAP messages. Three key usecases define the need for SOAP intermediaries: crossing trust domains, ensuring scalability, and
providing value-added services along the SOAP message path. Crossing trust domains is a
common issue faced when implementing security in distributed systems. Corporate firewalls and
virtual private network (VPN) gateways let some requests cross the trust domain boundary and
deny access to others.
Similarly, ensuring scalability is an important requirement in distributed systems. We rarely have
a simplistic scenario where the sender and receiver are directly connected by a dedicated link. In
reality, there will be several network nodes on the communication path that will be crossed by
many other concurrent communication flows. Due to the limited computing resources, the
performance of these nodes may not scale well with the increasing traffic load. To ensure
scalability, the intermediate nodes need to provide flexible buffering of messages and routing
Chapter 8 
Web Services
381
based not only on message parameters, such as origin, destination, and priority, but also on the
state of the network measured by parameters such as the availability and load of its nodes as well
as network traffic information.
Lastly, we need intermediaries to provide value-added services in a distributed system. Example
services include authentication and authorization, security encryption, transaction management,
message tracing and auditing, as well as billing and payment processing.
SOAP Message Global Attributes
SOAP defines three global attributes that are intended to be usable via qualified attribute names
on any complex type referencing them. The attributes are as follows (more details on each are
provided below):

The mustUnderstand attribute specifies whether it is mandatory or optional that a
message receiver understands and processes the content of a SOAP header block. The
message receiver to which this attribute refers to is named by the role attribute.

The role attribute is exclusively related to header blocks. It names the application that
should process the given header block.

The encodingStyle attribute indicates the encoding rules used to serialize parts of a
SOAP message. Although the SOAP specification allows this attribute to appear on any
element of the message (including header blocks), it mostly applies to body blocks.

The relay attribute is used to indicate whether a SOAP header block targeted at a SOAP
receiver must be relayed if not processed.
The mustUnderstand attribute can have values '1' or '0' (or, 'true' or 'false').
Value '1' indicates that the target role of this SOAP message must understand the semantics of
the header block and process it. If this attribute is missing, this is equivalent to having value '0'.
This value indicates that the target role may, but does not have to, process the header block.
The role attribute carries an URI value that names the recipient of a header block. This can be
the ultimate receiver or an intermediary node that should provide a value-added service to this
message. The SOAP specification defines three roles: none, next, and ultimateReceiver. An
attribute value of http://www.w3.org/2003/05/soap-envelope/role/next
identifies the next SOAP application on the message path as the role for the header block. A
header without a role attribute is intended for the ultimate recipient of this message.
The encodingStyle attribute declares the mapping from an application-specific data
representation to the wire format. An encoding generally defines a data type and data mapping
between two parties that have different data representation. The decoding converts the wire
representation of the data back to the application-specific data format. The translation step from
one data representation to another, and back to the original format, is called serialization and
deserialization. The terms marshalling and unmarshalling may be used as alternatives. The scope
of the encodingStyle attribute is that of its owner element and that element’s descendants,
excluding the scope of the encodingStyle attribute on a nested element. More about SOAP
encoding is given in Section 8.2.2 below.
Ivan Marsic

Rutgers University
382
The relay attribute indicates whether a header block should be relayed in the forwarded
message if the header block is targeted at a role played by the SOAP intermediary, but not
otherwise processed by the intermediary. This attribute type is Boolean and, if omitted, it is
equivalent as if included with a value of “false.”
Error Handling in SOAP: The Fault Body Block
If a network node encounters problems while processing a SOAP message, it generates a fault
message and sends it back to the message sender, i.e., in the direction opposite to the original
message flow. The fault message contains a Fault element which identifies the source and
cause of the error and allows error-diagnostic information to be exchanged between participants
in an interaction. Fault is optional and can appear at most once within the Body element. The
fault message originator can be an end host or an intermediary network node which was supposed
to relay the original message. The content of the Fault element is slightly different in these two
cases, as will be seen below.
A Fault element consists of the following nested elements (shown in Figure 8-4):

The Code element specifies the failure type. Fault codes are identified via namespacequalified names. SOAP predefines several generic fault codes and allows custom-defined
fault codes, as described below.

The Reason element carries a human-readable explanation of the message-processing
failure. It is a plain text of type string along with the attribute specifying the language the
text is written in.

The Node element names the SOAP node (end host or intermediary) on the SOAP message
path that caused the fault to happen. This node is the originator of the fault message.

The Role element identifies the role the originating node was operating in at the point the
fault occurred. Similar to the role attribute (described above), but instead of identifying the
role of the recipient of a header block, it gives the role of the fault originator.

The Detail element carries application-specific error information related to the Body
element and its sub-elements.
As mentioned, SOAP predefines several generic fault codes. They must be namespace qualified
and appear in a Code element. These are:
Fault Code
Explanation
VersionMismatch
The SOAP node received a message whose version is not supported,
which is determined by the Envelope namespace. For example, the
node supports SOAP version 1.2, but the namespace qualification of
the SOAP message Envelope element is not identical to
http://www.w3.org/2003/05/soap-envelope .
DataEncodingUnknown A SOAP node to which a SOAP header block or SOAP body child
element information item was targeted was targeted does not support
the data encoding indicated by the encodingStyle attribute.
MustUnderstand
A SOAP node to which a header block was targeted could not
Chapter 8 
383
Web Services
process the header block, and the block
mustUnderstand attribute value "true".
contained
a
Sender
A SOAP message was not appropriately formed or did not contain all
required information. For example, the message could lack the proper
authentication or payment information. Resending this identical
message will again cause a failure.
Receiver
A SOAP message could not be processed due to reasons not related
to the message format or content. For example, processing could
include communicating with an upstream SOAP node, which did not
respond. Resending this identical message might succeed at some
later point in time.
SOAP allows custom extensions of fault codes through dot separators so that the right side of a
dot separator refines the more general information given on the left side. For example, the Code
element conveying a sender authentication error would contain Sender.Authentication.
SOAP does not require any further structure within the content placed in header or body blocks.
Nonetheless, there are two aspects that influence how the header and body of a SOAP message
are constructed: communication style and encoding rules. These are described next.
8.2.2
The SOAP Section 5 Encoding Rules
Encoding rules define how a particular entity or data structure is represented in XML. Connecting
different applications typically introduces the problem of interoperability: the data representation
of one application is different from that of the other application. The reader may recall the
example in Figure 6-1 that shows two different ways of representing a postal address. The
applications may even be written in different programming languages. In order for the client and
server to interoperate, it is essential that they agree on how the contents of a SOAP message are
encoded. SOAP 1.2 defines a particular form of encoding called SOAP encoding.2 This defines
how data structures in the application’s local memory, including basic types such as integers and
strings as well as complex types such as arrays and structures, can be serialized into XML. The
serialized representation allows transfer of data represented in application-specific data types
from one application to another.
The encoding rules employed in a particular SOAP message are specified by the
encodingStyle attribute, as discussed above. There is no notion of default encoding in a
SOAP message. Encoding style must be explicitly specified if the receiving application is
expected to validate the message.
SOAP does not enforce any special form of coding—other encodings may be used as well. In
other words, applications are free to ignore SOAP encoding and choose a different one instead.
For instance, two applications can simply agree upon an XML Schema representation of a data
structure as the serialization format for that data structure. This is commonly referred to as literal
encoding (see also Section 8.2.3 below).
2
There is no “official” name for SOAP encoding, but it is often referred to as SOAP Section 5 encoding,
because the rules are presented in Section 5 of the SOAP specification.
Ivan Marsic

Rutgers University
384
A typical programming language data model consists of simple types and compound types.
Compound types are based on simple types or other compound types. Dealing with simple data
types would be easy, since all these types have direct representation in XML Schema (some are
shown in Table 6-1 above). However, the story with complex types, such as arrays and arbitrary
software objects, is more complicated. XML Schema defines complex types, but these are very
general, and some degree of specialization, e.g., for arrays, could make job easier for the Web
services developer.
SOAP does not define an independent data type system. Rather, it relies on the XML Schema
type system. It adopts all XML Schema built-in primitive types and adds few extensions for
compound types. The SOAP version 1.2 types extending the XML Schema types are defined in a
separate namespace, namely http://www.w3.org/2003/05/soap-encoding.
XML elements representing encoded values may hold the XML Schema type attribute for
asserting the type of a value. For example, the sender of the SOAP encoded message in Listing
8-1 above in Line 14 explicitly asserts the type of the note element content to be a string:
14
<a:note xsi:type="xsd:string">
15
Reminder: meeting at 11AM in Rm.601
16
</a:note>
The XML elements representing encoded values may also be untyped, i.e., not contain the type
attribute:
<a:note> Reminder: meeting at 11AM in Rm.601 </a:note>
In this case, a receiver deserializing a value must be able to determine its type just by means of
the element name <a:note>. If a sender and a receiver share the same data model, and both
know that a note labeled value in an application-specific data graph is a string type, they are
able to map the note element content to the appropriate data type without explicitly asserting the
type through the XML Schema type attribute. However, in this case we cannot rely on XML
Schema to explicitly validate the content of the message.
SOAP Compound Data Types
SOAP Section 5 encoding assumes that any application-specific data is represented as a directed
graph. Consider the class diagram for an online auction site shown in Figure 2-44, Problem 2.31
in Chapter 2, which is simplified here in Figure 8-5. Suppose that the sender sends an instance of
ItemInfo called item to the receiver in a SOAP message.
A compound data type can be either a struct or an array. A struct is an element that contains
different child elements. The SellerInfo in Figure 8-5 is an example of a struct. An array is a
compound type that contains elements of the same name. The BidList in Figure 8-5 is an example
of an array since it contains a group of individual Bid entries.
When serialized to XML, the object graph of item in Figure 8-5 will be represented as follows:
Listing 8-2: Example of SOAP encoding for the object graph in Figure 8-5.
<soap-env:Envelope
xmlns:soap-env="http://www.w3.org/2003/05/soap-envelope"
xmlns:xsd="http://www.w3.org/2001/XMLSchema"
Chapter 8 
385
Web Services
item
ItemInfo
name : String
startPrice : float
reserved : boolean
bids
BidsList
1
BuyerInfo
seller
name : String
address : String
entry
1
*
SellerInfo
name : String
address : String
Bid
amount : float
bidder
1
Receiver
(Web service)
Sender
item
SOAP message
Figure 8-5: Example class diagram, extracted from Figure 2-44 (Chapter 2).
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xmlns:soap-enc="http://www.w3.org/2003/05/soap-encoding"
xmlns:ns0="http://www.auctions.org/ns"
soap-env:encodingStyle="http://www.w3.org/2003/05/soap-encoding">
<soap-env:Body>
<ns0:item>
<ns0:name xsi:type="xsd:string"> old watch </ns0:name>
<ns0:startPrice xsi:type="xsd:float"> 34.99 </ns0:startPrice>
<ns0:reserved xsi:type="xsd:boolean"> false </ns0:reserved>
<ns0:seller>
<ns0:name xsi:type="xsd:string"> John Doe </ns0:name>
<ns0:address xsi:type="xsd:string">
13 Takeoff Lane, Talkeetna, AK 99676
</ns0:address>
<ns0:bids xsi:type="soap-enc:array" soap-enc:arraySize="*">
<ns0:entry>
<ns0:amount xsi:type="xsd:float"> 35.01 </ns0:amount>
<ns0:bidder> . . . </ns0:bidder>
</ns0:entry>
<ns0:entry>
<ns0:amount xsi:type="xsd:float"> 34.50 </ns0:amount>
<ns0:bidder> . . . </ns0:bidder>
</ns0:entry>
. . .
</ns0:bids>
</ns0:seller>
</ns0:item>
</soap-env:Body>
</soap-env:Envelope>
Ivan Marsic

Rutgers University
386
<Envelope>
<Header>
header blocks
</Header>
<Body>
SOAP Message
arbitrary XML document
(same for Request or Response)
</Body>
</Envelope>
Figure 8-6: Structure of a document-style SOAP message.
Array attributes. Needed to give the type and dimensions of an array’s contents, and the offset for
partially-transmitted arrays. Used as the type of the arraySize attribute. Restricts asterisk ( * )
to first list item only. Instances must contain at least an asterisk ( * ) or a nonNegativeInteger.
May contain other nonNegativeIntegers as subsequent list items. Valid instances include: *, 1, *
2, 2 2, * 2 0.
8.2.3
SOAP Communication Styles
Generally, SOAP applications can communicate in two styles: document style and RPC style
(Remote Procedure Call style). In document-style communication, the two applications agree
upon the structure of documents exchanged between them. SOAP messages are used to transport
these documents from one application to the other. The structure of both request and response
messages is the same, as illustrated in Figure 8-6, and there are absolutely no restrictions as to the
information that can be stored in their bodies. In short, any XML document can be included in the
SOAP message. The document style is often referred to also as message-oriented style.
In RPC-style communication, one SOAP message encapsulates the request while another
message encapsulates the response, just as in document-style communication. However, the
difference is in the way these messages are constructed. As shown in Figure 8-7, the body of the
request message contains the actual operation call. This includes the name of the operation being
invoked and its input parameters. Thus, the two communicating applications have to agree upon
the RPC operation signature as opposed to the document structure (in the case of document-style
communication). The task of translating the operation signature in SOAP is typically hidden by
the SOAP middleware.
Selecting the communication style is independent from selecting whether or not the message
should be encoded (Section 8.2.2 above). The term literal is commonly used to refer to nonencoded messages. Therefore, four different combinations are possible:

document/literal: A document-style message which is not encoded.

document/encoded: A document-style message which is encoded.

rpc/literal: An RPC-style message which is not encoded.

rpc/encoded: An RPC-style message which is encoded.
Chapter 8 
387
Web Services
Remote Procedure Call
(Request message)
<Envelope>
<Header>
header blocks
</Header>
<Body>
<operationNameReturn>
<return>
return value
<outputParameter_1>
value 1
<outputParameter_2>
value 2
<outputParameter_n>
</operationNameReturn>
</Body>
</Envelope>
value n
<Envelope>
<Header>
header blocks
</Header>
<Body>
<operationName>
<inputParameter_1>
<inputParameter_2>
<inputParameter_n>
</operationName>
</Body>
</Envelope>
</return>
</outputParameter_1>
</outputParameter_2>
value 1
value 2
</inputParameter_1>
</inputParameter_2>
value n
</inputParameter_n>
Remote Procedure Call Response
(Response message)
</outputParameter_n>
Figure 8-7: Structure of a SOAP RPC-style request and its associated response message.
The document/encoded combination is rarely encountered in practice, but the other three are
commonly in use. Document-style messages are particularly useful to support cases in which
RPCs result in interfaces that are too fine grained and, therefore, brittle.
The RPC-style SOAP Communication
In the language of the SOAP encoding, the actual RPC invocation is modeled as a struct type
(Section 8.2.2 above). The name of the struct (that is, the name of the first element inside the
SOAP body) is identical to the name of the method/operation. Every in or in-out parameter of the
RPC is modeled as an accessor with a name identical to the name of the RPC parameter and the
type identical to the type of the RPC parameter mapped to XML according to the rules of the
active encoding style. The accessors appear in the same order as do the parameters in the
operation signature.
All parameters are passed by value. SOAP has no notion of passing values by reference, which is
unlike most of the programming languages. For Web services, the notion of in-out and out
parameters does not involve passing objects by reference and letting the target application modify
their state. Instead, copies of the data are exchanged. It is the up to the service client code to
create the perception that the actual state of the object that has been passed in to the client method
has been modified.
Listing 8-3: An example of a SOAP 1.2 RPC-style request/response via HTTP:
<?xml version="1.0"?>
Ivan Marsic

Rutgers University
<description name="StockQuote"
targetNamespace="http://example.com/stockquote.wsdl"
xmlns:tns="http://example.com/stockquote.wsdl"
xmlns:xsd1="http://example.com/stockquote.xsd"
xmlns:soap="http://www.w3.org/2003/05/soap-envelope"
xmlns="http://www.w3.org/ns/wsdl">
<types>
<schema targetNamespace="http://example.com/stockquote.xsd"
xmlns="http://www.w3.org/2001/XMLSchema">
<element name="TradePriceRequest">
<complexType>
<all>
<element name="tickerSymbol" type="string"/>
</all>
</complexType>
</element>
<element name="TradePrice">
<complexType>
<all>
<element name="price" type="float"/>
</all>
</complexType>
</element>
</schema>
</types>
<message name="GetLastTradePriceInput">
<part name="body" element="xsd1:TradePriceRequest"/>
</message>
<message name="GetLastTradePriceOutput">
<part name="body" element="xsd1:TradePrice"/>
</message>
<portType name="StockQuotePortType">
<operation name="GetLastTradePrice">
<input message="tns:GetLastTradePriceInput"/>
<output message="tns:GetLastTradePriceOutput"/>
</operation>
</portType>
<binding name="StockQuoteSoapBinding"
type="tns:StockQuotePortType">
<soap:binding style="document"
transport="http://schemas.xmlsoap.org/soap/http"/>
<operation name="GetLastTradePrice">
<soap:operation
soapAction="http://example.com/GetLastTradePrice"/>
<input>
<soap:body use="literal"/>
</input>
<output>
<soap:body use="literal"/>
</output>
</operation>
</binding>
388
Chapter 8 
389
Web Services
HTTP Message
SOAP Message
SOAP body
Physical (Communication Protocol) Message
POST / alertcontrol HTTP 1.1
Content-Type: text/xml; charset="utf-8"
Content-Length: 581
Out-of-message context (target URI)
Host: www.example.org
SOAPAction: notify
......
Connection: Keep-Alive
Out-of-message context (SOAPAction)
Logical SOAP Message
<soap-env:Envelope
xmlns:soap-env="http://www.w3.org/2003/05/soap-envelope"
xmlns:xsd="http://www.w3.org/2001/XMLSchema"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance">
<soap-env:Header>
......
</soap-env:Header>
In-message context (header blocks)
......
SOAP Body
<soap-env:Body>
<a:notify xmlns:a="http://example.org/alert">
<a:note xsi:type="xsd:string">
Reminder: meeting today at 11AM in Rm.601
</a:note>
</a:notify>
</soap-env:Body>
Figure 8-8: SOAP message binding to the HTTP protocol for the SOAP message example
from Listing 8-1 above.
<service name="StockQuoteService">
<documentation>My first service</documentation>
<port name="StockQuotePort" binding="tns:StockQuoteBinding">
<soap:address location="http://example.com/stockquote"/>
</port>
</service>
<description>
8.2.4
Binding SOAP to a Transport Protocol
The key issue in deciding how to bind SOAP to a particular transport protocol is about
determining how the requirements for a Web service (communication style, such as RPC vs.
document, synchronous vs. asynchronous, etc.) map to the capabilities of the transport protocol.
In particular, we need to determine how much of the overall contextual information needed to
successfully execute the Web service needs to go in the SOAP message versus in the message of
the transport protocol.
Figure 8-8 illustrates this issue on an HTTP example. In HTTP, context information is passed via
the target URI and the SOAPAction header. In the case of SOAP, context information is passed
in the SOAP header.
As a communication protocol, SOAP is stateless and one-way. Although it is possible to
implement statefull SOAP interactions so that the Web service maintains a session, this is not the
most common scenario.
Ivan Marsic

Rutgers University
390
When using HTTP for SOAP messages, the developer must decide which HTTP method
(Appendix C) is appropriate to use in HTTP request messages that are exchanged between the
service consumer and the Web service. Usually the choice is between GET and POST. In the
context of the Web as a whole (not specific to Web services), the W3C Technical Architecture
Group (TAG) has addressed the question of when it is appropriate to use GET, versus when to use
POST, in [Jacobs, 2004]. Their finding is that GET is more appropriate for safe operations such as
simple queries. POST is appropriate for other types of applications where a user request has the
potential to change the state of the resource (or of related resources). Figure 8-8 shows the HTTP
request using the POST method, which is most often used in the context of Web services.
I. Jacobs (Editor), “URIs, addressability, and the use of HTTP GET and POST,” World Wide
Web Consortium, 21 March 2004. Available at: http://www.w3.org/2001/tag/doc/whenToUseGet
8.3 WSDL for Web Service Description
A Web service publishes the description of the service not the actual service code. The service
customer uses the aspects of the service description to look or find a Web service. The service
customer uses this description since it can exactly detail what is required by the client to bind to
the Web service. Service description can be partitioned into:

Parts used to describe individual Web services.

Parts used to describe relationships between sets of Web services.
Our main focus will be on the first group, describing individual Web services. The second group
which describes relationships between sets of Web services will be briefly reviewed in Section
8.3.5 below.
The most important language for describing individual Web services currently is the Web
Services Definition Language (WSDL). WSDL has a dual purpose of specifying: (a) the Web
service interface (operation names and their signatures, used in the service invocation), and (b)
the Web service implementation (network location and access mechanism for a particular instance
of the Web service). The interface specifies what goes in and what comes out, regardless of where
the service is located or what are its performance characteristics. This is why the Web service
interface is usually referred to as the abstract part of the service specification. Conversely, the
implementation specifies the service’s network location and its non-functional characteristics,
such as performance, reliability, and security. This is why the Web service implementation is
usually referred to as the concrete part of the service specification. This section describes how
WSDL is used for describing the service interface and implementation.
WSDL grammar (schema) is defined using XML Schema. The WSDL service description is an
XML document conformant to the WSDL schema definition. WSDL provides the raw technical
description of the service’s interface including what a service does, how it is accessed and where
Chapter 8 
391
Web Services
WSDL description
Service interface definition
types
Abstract part:
interface
operation1
operation2
What types of messages
(names + data types) are
communicated with the service?
How are the methods
invoked on the service?
Service implementation definition
binding
Concrete part:
operation1
operation2
service
How will the service be used on
the network for a protocol?
SOAP-specific details are here.
Where is the service located in
the network? – endpoint host(s)
Figure 8-9: The WSDL 2.0 building blocks.
a service is located (Figure 8-9). Since the Web service can be located anywhere in the network,
the minimum information needed to invoke a service method is:
1. What is the service interface, that is, what are its methods (operations), method
signatures, and return values?
2. Where in the network (host address and port number) is the service located?
3. What communication protocol does the service understand?
8.3.1
The WSDL 2.0 Building Blocks
As seen Figure 8-9, WSDL 2.0 enables the developer to separate the description of a Web
service’s abstract functionality from the concrete details of how and where that functionality is
offered. This separation facilitates different levels of reusability and distribution of work in the
lifecycle of a Web service and the WSDL document that describes it.
Different implementations of the same Web service can be made accessible using different
communication protocols. (Recall also that SOAP supports binding to different transport
protocols, Section 8.2.4 above.)
The description of the endpoint’s functional capabilities is the abstract interface specification
represented in WSDL by the interface element. An abstract interface can support any number
of operations. An operation is defined by the set of messages that define its interface pattern.
Recall that invoking an object method involves a request message passing a set of parameters (or
arguments) as well as receiving a response message that carries the result returned by the method.
(The reader should recall the discussion of the RPC-style SOAP communication in Section 8.2.3
above.) Also, some of the method parameters may be used to pass back the output results; these
are known as in-out or out parameters. Since the operation is invoked over the network, we must
Ivan Marsic

Rutgers University
392
wsdl:DescriptionType
wsdl:import
wsdl:include

wsdl:types
wsdl:description
!
wsdl:interface
!
wsdl:binding
!
wsdl:service
targetNamespace
Figure 8-10: The XML schema for WSDL 2.0. Continued in Figure 8-12 and Figure 8-14.
(WSDL version 2.0 schema definition available at: http://www.w3.org/ns/wsdl).
specify how the forward message carries the input parameters, as well as how the feedback
message carries the result and the output parameters, or error message in case of failure.
For the abstract concepts of messages and operations, concrete counterparts are specified in the
binding element. A binding mechanism represented in WSDL by a binding element is used to
map the abstract definition of the Web service to a specific implementation using a particular
messaging protocol, data encoding model and underlying communication protocol. When the
binding is combined with an address where the implementation can b