State of the Art Analysis

State of the Art Analysis
Developing Data-Intensive Cloud
Applications with Iterative Quality
Enhancements
State of the Art Analysis
Deliverable 1.1
Deliverable 1.1. State of the art analysis
Deliverable:
Title:
Editor(s):
Contributor(s):
Reviewers:
Type (R/P/DEC):
Version:
Date:
Status:
Dissemination level:
Download page:
Copyright:
D1.1
State of the Art Analysis
Giuliano Casale (IMP), Tatiana Ustinova (IMP)
Danilo Ardagna (PMI), Matej Artac (XLAB), Simona Bernardi (ZAR),
Marcello Bersani (PMI), Giuliano Casale (IMP), Pablo Domínguez-Mayo
(ZAR), Ioan Dragan (IEAT), Mădălina Eraşcu (IEAT), Marc Gil (PRO),
Gabriel Iuhasz (IEAT), Pooyan Jamshidi (IMP), Cristophe Joubert (PRO),
José Merseguer (ZAR), Daniel Pop (IEAT), Alberto Romeu (PRO), Matteo
Rossi (PMI), Tatiana Ustinova (IMP), Darren Whigham (FLEXI).
Youssef Ridene (NETF), Ilias Spais (ATC)
Report
1.02
05-December-2015
Final version
Public
http://www.dice-h2020.eu/deliverables/
Copyright © 2015, DICE consortium – All rights reserved
DICE partners
ATC:
FLEXI:
IEAT:
IMP:
NETF:
PMI:
PRO:
XLAB:
ZAR:
Athens Technology Centre
Flexiant Limited
Institutul E Austria Timisoara
Imperial College of Science, Technology & Medicine
Netfective Technology SA
Politecnico di Milano
Prodevelop SL
XLAB razvoj programske opreme in svetovanje d.o.o.
Universidad De Zaragoza
The DICE project (February 2015-January 2018) has received funding from the European
Union’s Horizon 2020 research and innovation programme under grant agreement No. 644869
Copyright © 2015, DICE consortium – All rights reserved
2
Deliverable 1.1. State of the art analysis
Executive summary
The aim of this deliverable is to review the state-of-the-art in techniques used in the developing of Big Data
applications and related technology offerings that are available on the market. After positioning DICE in
the scope of DevOps and Big Data, the deliverable provides background on Big Data and related software
engineering trends, such as the emergence of the Lamba architecture style.
We then overviewthe state of the art on designing functional and non-functional properties in enterprise
software systems, highlighting gaps towards achieving these goals for data-intensive applications. In
particular, we survey model-driven engineering (MDE) methods, which are the most popular to combine
software design with quality analysis techniques based on formal models for performance, reliability and
verification. We extensively discuss existing UML profiles relevant to software quality assessment and
highlight their gaps in relation to modelling data intensive applications. Editors and modelling tools that
can process such UML profiles are also surveyed and compared.
We then overview the problem of deploying, monitoring and testing an enterprise cloud application, and
review existing technologies and open source tools in this area. It is found that some areas, such as nonfunctional testing, are fairly under-developed in Big Data and thus offer an opportunity for innovation.
In the last part of the deliverable we summarize some relevant Big Data technologies (e.g.,
Hadoop/MapReduce, Spark, Storm, etc.) and related commercial and open source implementations. For
each technology, we highlight quality metrics that may be considered by the DICE monitoring, prediction
and analysis tools.
Copyright © 2015, DICE consortium – All rights reserved
3
Deliverable 1.1. State of the art analysis
Table of contents
EXECUTIVE SUMMARY ........................................................................................................................................... 3
TABLE OF CONTENTS .............................................................................................................................................. 4
LIST OF FIGURES ....................................................................................................................................................... 9
LIST OF TABLES ......................................................................................................................................................... 9
INTRODUCTION ....................................................................................................................................................... 11
A.
DATA-INTENSIVE APPLICATIONS ............................................................................................................. 12
A.1. Definitions and terminology. .........................................................................................................12
A.2. Importance of Big Data for the Business .......................................................................................12
A.3. Roles in Big Data Systems .............................................................................................................13
A.4. Technology Overview ....................................................................................................................14
A.4.1
Distributed computations .......................................................................................................14
A.4.2
Distributed Storage and NoSQL solutions .............................................................................14
A.4.3
Streams and Big Data in motion ............................................................................................15
A.5. Architectural Styles for DIAs ........................................................................................................15
A.5.1
Lambda Architecture..............................................................................................................15
A.5.1.1.
Coordination Layer ........................................................................................................16
A.5.1.2.
Batch Layer ....................................................................................................................17
A.5.1.3.
Speed Layer ...................................................................................................................17
A.5.1.4.
Orchestration ..................................................................................................................17
A.5.1.5.
Serving Layer .................................................................................................................17
A.5.1.6.
Academic and industrial positioning ..............................................................................17
A.5.1.7.
Oryx: an example Lambda Architecture ........................................................................18
A.5.2
Other Architectures ................................................................................................................19
B.
DEVOPS .............................................................................................................................................................. 21
B.1. Introduction. ...................................................................................................................................21
B.2. Short introduction to DevOps ........................................................................................................22
B.3. Functional modelling .....................................................................................................................24
B.3.1
Overview ................................................................................................................................24
B.3.2
MDE and domain-specific model-driven approaches ............................................................25
B.3.3
UML, MARTE and DAM ......................................................................................................27
B.3.3.1.
Domain-Specific Modelling with UML .........................................................................28
B.3.3.2.
Modelling with MARTE ................................................................................................28
B.3.3.2.1. Specification of NFP ..................................................................................................29
B.3.3.2.2. MARTE::GQAM framework .....................................................................................29
B.3.3.3.
Modelling with DAM.....................................................................................................30
B.3.3.3.1. DAM Library ..............................................................................................................30
B.3.3.3.2. Basic Dependability Types .........................................................................................30
B.3.3.3.3. Complex Dependability Types ...................................................................................31
B.3.3.3.4. DAM UML extensions ...............................................................................................31
B.3.3.4.
Access Control Modelling with UML............................................................................31
B.4. Modelling tools ..............................................................................................................................32
B.4.1
Analysis..................................................................................................................................33
B.4.2
MARTE profile feature (import *.XMI) ................................................................................34
B.4.3
UML diagrams supported ......................................................................................................35
Copyright © 2015, DICE consortium – All rights reserved
4
Deliverable 1.1. State of the art analysis
B.4.4
Summary ................................................................................................................................35
B.5. Tools for model to model transformations .....................................................................................36
B.5.1
Palladio Component Model ...................................................................................................37
B.5.2
VIATRA2 ..............................................................................................................................37
B.5.3
UML transformation tool .......................................................................................................38
B.5.4
Other tools ..............................................................................................................................39
B.5.4.1.
CARiSMA......................................................................................................................39
B.5.4.2.
UPUPA (fUML and Profiles for Performance Analysis)...............................................39
B.5.4.3.
QVT and Related Technologies .....................................................................................39
B.5.5
Summary ................................................................................................................................40
B.6. Deployment modelling with TOSCA ............................................................................................40
B.7. Deployment tools ...........................................................................................................................41
B.7.1
Overview ................................................................................................................................41
B.7.2
Configuration management ....................................................................................................41
B.7.3
Orchestration ..........................................................................................................................42
B.7.3.1.
Ubuntu Juju ....................................................................................................................42
B.7.3.2.
Cloudify .........................................................................................................................42
B.7.3.3.
Alien4Cloud ...................................................................................................................43
B.7.3.4.
Apache Brooklyn ...........................................................................................................43
B.7.3.5.
Flexiant Cloud Orchestrator ...........................................................................................43
B.7.3.6.
Rundeck .........................................................................................................................43
B.7.3.7.
CAMF ............................................................................................................................43
B.7.3.8.
CELAR ..........................................................................................................................44
B.7.3.9.
Open-TOSCA ................................................................................................................44
B.7.3.10. Tools analysis.................................................................................................................44
B.7.4
Virtualisation and containers .................................................................................................45
B.7.5
Summary ................................................................................................................................46
B.8. Continuous integration tools ..........................................................................................................46
B.8.1
TeamCity................................................................................................................................47
B.8.2
Hudson / Jenkins ....................................................................................................................47
B.8.3
Atlassian Bamboo ..................................................................................................................47
B.8.4
Go ...........................................................................................................................................47
B.8.5
Strider CD ..............................................................................................................................48
B.8.6
BuildBot .................................................................................................................................48
B.8.7
CircleCI ..................................................................................................................................48
B.8.8
Summary ................................................................................................................................49
B.9. Versioning of software engineering artefacts. ...............................................................................49
B.9.1
Classical versioning tools.......................................................................................................49
B.9.1.1.
CVS ................................................................................................................................49
B.9.1.2.
Subversion......................................................................................................................49
B.9.1.3.
Git ..................................................................................................................................50
B.9.1.4.
Mercurial ........................................................................................................................50
B.9.2
DICE needs with respect to versioning ..................................................................................50
B.10.
Discussion ..................................................................................................................................50
C.
QUALITY ASSURANCE ................................................................................................................................... 52
C.1.
C.2.
C.3.
Non-Functional Properties in DevOps ...........................................................................................52
Performance Metrics ......................................................................................................................53
Reliability Metrics .........................................................................................................................53
Copyright © 2015, DICE consortium – All rights reserved
5
Deliverable 1.1. State of the art analysis
C.4. Safety Properties ............................................................................................................................54
C.5. Quantitative analysis for assessment of performance, reliability and safety .................................56
C.5.1
Combinatorial techniques ......................................................................................................56
C.5.2
State-based techniques ...........................................................................................................57
C.5.2.1.
Stochastic Petri Nets ......................................................................................................57
C.5.2.2.
Queueing networks ........................................................................................................57
C.5.3
Monte Carlo simulation techniques .......................................................................................58
C.6. Performance and reliability prediction tools ..................................................................................59
C.6.1
Stochastic Petri Net tools .......................................................................................................59
C.6.2
Queueing Network tools ........................................................................................................60
C.7. Formal analysis of safety and privacy properties ...........................................................................61
C.8. Tools for Formal Verification ........................................................................................................63
C.9. Software Anti-patterns ...................................................................................................................65
C.10.
Optimising Deployment Plans ...................................................................................................66
C.11.
Testing and Monitoring of Non-Functional Properties ..............................................................67
C.11.1 Testing methods .....................................................................................................................67
C.11.2 Testing tools ...........................................................................................................................69
C.11.2.1. Grinder ...........................................................................................................................69
C.11.2.2. Apache JMeter ...............................................................................................................69
C.11.2.3. Selenium ........................................................................................................................70
C.11.2.4. MODAClouds MDload ..................................................................................................70
C.11.2.5. Chaos Monkey ...............................................................................................................70
C.11.3 Monitoring tools.....................................................................................................................71
C.11.3.1. Hadoop toolkit ...............................................................................................................71
C.11.3.2. SequenceIQ ....................................................................................................................71
C.11.3.3. Hadoop Vaidya ..............................................................................................................72
C.11.3.4. Ganglia ...........................................................................................................................72
C.11.3.5. Apache Ambari ..............................................................................................................72
C.11.3.6. Apache Chukwa .............................................................................................................72
C.11.3.7. Datastax-OpsCentre for Apache Cassandra ...................................................................73
C.11.3.8. MongoDB (MMS)..........................................................................................................73
C.11.3.9. Server Density ................................................................................................................73
C.11.3.10. Manage Engine ..............................................................................................................73
C.12.
Monitoring Feedback Analysis ..................................................................................................74
C.12.1 Tools for Detecting Anomalies with Machine Learning ........................................................74
C.12.2 Distributed Machine Learning Platforms ...............................................................................75
C.12.3 Model Parameter Estimation ..................................................................................................78
C.12.4 Trace Checking ......................................................................................................................79
D.
DATA-INTENSIVE TECHNOLOGIES ........................................................................................................... 80
D.1. Overview ........................................................................................................................................80
D.2. Hadoop and Spark ..........................................................................................................................80
D.2.1
Overview ................................................................................................................................80
D.2.2
Apache Hadoop ......................................................................................................................81
D.2.2.1.
Overview ........................................................................................................................81
D.2.2.2.
Hadoop public cloud offerings .......................................................................................82
D.2.2.3.
Hadoop Quality metrics .................................................................................................83
D.2.2.4.
Hadoop meta- and QoS models .....................................................................................83
D.2.3
Apache Spark .........................................................................................................................84
Copyright © 2015, DICE consortium – All rights reserved
6
Deliverable 1.1. State of the art analysis
D.2.3.1.
Overview ........................................................................................................................84
D.2.3.2.
Spark public cloud offerings ..........................................................................................84
D.2.3.3.
Spark Quality metrics and models .................................................................................84
D.3. Streaming .......................................................................................................................................85
D.3.1
Introduction ............................................................................................................................85
D.3.2
Stream processing architecture ..............................................................................................85
D.3.3
Public cloud offerings ............................................................................................................86
D.3.4
Open source solutions. ...........................................................................................................86
D.3.4.1.
Apache Storm.................................................................................................................86
D.3.4.2.
Apache Spark Streaming ................................................................................................87
D.3.4.3.
Comparison between Storm and Spark ..........................................................................88
D.3.4.4.
Apache Samza ................................................................................................................88
D.3.4.5.
Other solutions ...............................................................................................................89
D.3.5
Message queues .....................................................................................................................90
D.3.6
General characteristics ...........................................................................................................90
D.3.6.1.
Key monitoring metrics .................................................................................................90
D.3.6.2.
Main quality assurance challenges .................................................................................90
D.3.6.3.
Reliability .......................................................................................................................90
D.3.6.4.
Scalability ......................................................................................................................91
D.3.6.5.
Efficiency .......................................................................................................................91
D.3.6.6.
Privacy ...........................................................................................................................91
D.4. NoSQL ...........................................................................................................................................91
D.4.1
High level architecture of a NoSQL database ........................................................................92
D.4.2
List of public cloud offerings available for this technology ..................................................92
D.4.3
Open source solutions ............................................................................................................95
D.4.4
Quality Assurance ..................................................................................................................95
D.4.4.1.
Key monitoring metrics .................................................................................................95
D.4.4.2.
Main quality assurance challenges. ................................................................................96
D.4.4.3.
Reliability .......................................................................................................................96
D.4.4.4.
Efficiency .......................................................................................................................96
D.4.4.5.
Safety .............................................................................................................................96
D.4.4.6.
Privacy ...........................................................................................................................96
D.4.5
Models....................................................................................................................................97
D.4.5.1.
QoS prediction models. ..................................................................................................97
D.5. Software-Defined Networking .......................................................................................................97
D.5.1
Architecture............................................................................................................................97
D.5.2
Quality Assurance in Software Defined Networking .............................................................98
D.5.3
Current Issues........................................................................................................................99
D.5.4
Controllers architecture and low levels APIs .........................................................................99
D.5.5
Languages and network policies ............................................................................................99
D.5.6
SDNs in DICE......................................................................................................................100
D.6. Cloud-based blob storage .............................................................................................................100
D.6.1
Ceph .....................................................................................................................................100
D.6.1.1.
Typical architecture of CEPH ......................................................................................100
D.6.1.2.
Cloud offerings using CEPH........................................................................................101
D.6.1.3.
Open Source solutions .................................................................................................101
D.6.1.4.
Quality assurance .........................................................................................................101
D.6.1.4.1. Key monitoring metrics ............................................................................................101
Copyright © 2015, DICE consortium – All rights reserved
7
Deliverable 1.1. State of the art analysis
D.6.1.4.2. Quality assurance challenges for CEPH literature. ..................................................102
D.6.1.4.3. Reliability and high-availability support for CEPH .................................................102
D.6.1.4.4. Scalability and performance support for CEPH .......................................................102
D.6.1.4.5. Privacy and data protection with CEPH ...................................................................102
D.6.2
Amazon Simple Service Storage (Amazon S3) ...................................................................102
D.6.2.1.
Overview ......................................................................................................................102
D.6.2.2.
Public cloud offerings of S3.........................................................................................103
D.6.2.3.
Open source solutions with S3 .....................................................................................103
D.6.2.4.
Quality assurance. ........................................................................................................103
D.6.2.4.1. Key S3 monitoring metrics .......................................................................................103
D.6.2.4.2. Main quality assurance challenges. ..........................................................................104
D.6.2.4.3. Configuration options for reliability and high-availability with S3 .........................104
D.6.2.4.4. Scalability and performance for S3 ..........................................................................104
D.6.2.4.5. Privacy and data protection within S3 ......................................................................104
D.7. In-Memory Analytics ...................................................................................................................104
D.7.1
Introduction ..........................................................................................................................104
D.7.2
Diagram showing a typical architecture of this technology. ................................................104
D.7.3
Public cloud offerings available for this technology. ...........................................................104
D.7.4
Open source solutions for adoption of this technology in private clouds. ...........................105
D.7.5
Quality assurance .................................................................................................................105
D.7.5.1.
Key monitoring metrics. ..............................................................................................105
D.7.5.2.
Main quality assurance challenges. ..............................................................................105
D.7.5.3.
Reliability .....................................................................................................................105
D.7.5.4.
Efficiency .....................................................................................................................105
D.7.5.5.
Privacy .........................................................................................................................105
D.7.6
Models..................................................................................................................................105
D.7.6.1.
Meta-models. ...............................................................................................................105
D.7.6.2.
Quality of Service prediction models. ..........................................................................106
CONCLUSION .......................................................................................................................................................... 107
REFERENCES........................................................................................................................................................... 108
Copyright © 2015, DICE consortium – All rights reserved
8
Deliverable 1.1. State of the art analysis
List of figures
Figure 1. Lambda Architecture. .....................................................................................................................16
Figure 2. Architecture of Oryx 1 [32]. ...........................................................................................................18
Figure 3. Architecture of Oryx 2 [31]. ...........................................................................................................19
Figure 4. Kappa Architecture. ........................................................................................................................20
Figure 5. Liquid Architecture [38]. ................................................................................................................20
Figure 6. DevOps key values [42]. ................................................................................................................22
Figure 7. An Overview of the Activities and Tools behind DevOps. ............................................................23
Figure 8. Typical Organisational Structures behind DevOps, an example [42]. ...........................................23
Figure 9. Sketch of UML Profile definition. ..................................................................................................28
Figure 10. DAM profile overview. ................................................................................................................30
Figure 11. DAM types. ..................................................................................................................................31
Figure 12. Main TOSCA concepts and their relations [164]. ........................................................................40
Figure 13. DICE high level vision [237]........................................................................................................51
Figure 14. SequenceIQ high-level architecture [363]. ...................................................................................71
Figure 15. Architecture of a typical streaming processing solution [500]. ....................................................85
Figure 16. Typical Storm architecture [508]. .................................................................................................87
Figure 17. Typical Spark architecture [508]. .................................................................................................87
Figure 18. Typical Samza architecture [508]. ................................................................................................89
Figure 19. Architecture of a distributed NoSQL. This Figure refers to the HBASE architecture. ................92
Figure 20. Ceph integration example [589]. ................................................................................................101
List of tables
Table 1: Examples of Big Data impact across various sectors. .....................................................................13
Table 2: Machine learning algorithms used in Oryx 1 and Oryx 2. ...............................................................19
Table 3: UML CASE Tools summary. ..........................................................................................................34
Table 4: CASE tools. .....................................................................................................................................34
Table 5: UML 2.0 Diagrams supported by modelling tools. ........................................................................35
Table 6: Summary of the evaluation framework for model-to-model transformation tools [153]. ...............36
Table 7: Evaluation framework from Table 6 applied to Palladio Component Model. .................................37
Table 8: Evaluation framework from Table 6 applied to VIATRA2 framework. .........................................38
Table 9: Evaluation framework from Table 6 applied to UML transformation tool. ....................................38
Table 10: Comparative summary of deployment orchestration tools. ...........................................................44
Table 11: Comparative summary of continuous integration tools. ................................................................48
Table 12: Comparison of SPN tools. .............................................................................................................59
Table 13: Comparison of Queueing Network tools. ......................................................................................60
Table 14: Formal verification tools. ...............................................................................................................64
Table 15: State of the art in the area of software performance anti-patterns detection..................................66
Table 16: Comparative summary of testing tools. .........................................................................................70
Table 17: Comparative summary of monitoring tools. ..................................................................................73
Table 18: Machine Learning tools. ................................................................................................................75
Table 19: Distributed ML frameworks. .........................................................................................................77
Table 20: State of the art in the area of feedback analysis tools. ...................................................................78
Table 21: Public and commercial cloud offerings for stream processing. .....................................................86
Table 22: Storm vs Spark comparison. ..........................................................................................................88
Copyright © 2015, DICE consortium – All rights reserved
9
Deliverable 1.1. State of the art analysis
Table 23: Comparison of streaming processing frameworks [508], [518]-[520]. .........................................90
Table 24: Public cloud offerings for NoSQL (Database as a Service). .........................................................93
Table 25: Open source solutions for NoSQL. ................................................................................................95
Table 26: Open source SDN solutions supporting OpenFlow. ......................................................................98
Table 27: Dreamhost pricing storage [591]. ................................................................................................101
Table 28: AWS Storage Pricing US Standard [597]. ...................................................................................103
Table 29: Data Transfer Pricing US Standard [597]. ...................................................................................103
Copyright © 2015, DICE consortium – All rights reserved
10
Deliverable 1.1. State of the art analysis
Introduction
Big Data has recently emerged as a major trend in the ICT industry. However, the heterogeneity of the
technologies and software development method in use is still a challenge for researchers and practitioners.
The DICE project aims at developing a novel UML profile and tools that will help software designers
reasoning about reliability, safety and efficiency of Big Data applications. Since the DICE methodology
will cover quality assessment, architecture enhancement, continuous testing and agile delivery, there is a
large span of technologies and tools that may act as baselines to this research effort. The goal of this
deliverable is therefore to support this investigation with an analysis on the scientific and technical state of
the art in the area of software engineering for Big Data. It is important to note that the scope of the DICE
project embraces the phases of initial design, development and testing, but not operation of production
systems, therefore this survey does not cover service management. That is, the aim of the project is to put
the developer in conditions to deliver an application to market in a short period of time, focusing on
development and pre-production testing.
The document is organised in four chapters which cover the following topics:




Chapter A provides basic concepts, introducing definitions and key architectural styles that are
central to Big Data applications.
Chapter B describes the state of the art in functional development of Big Data applications,
covering in particular aspects related to Model-Driven Engineering (MDE) and to the DevOps
paradigm, which aims at applying agile concepts across the whole service delivery chain.
Chapter C continues the survey by reviewing current quality engineering methods used in modeldriven engineering and DevOps. In particular, we describe baselines such as UML MARTE and
UML DAM that are central to the DICE UML profile.
While the previous are oriented at reviewing research efforts and relevant standards, Chapter D
covers technological aspects of Big Data, surveying relevant technologies in real-time and batch
data processing.
Out of the present investigation, our main achievement is the identification of the most relevant baselines to
the DICE work and of open problems in the area touched by the project. The reader is invited to consult
deliverable D1.2 – ‘Requirement specification’ for a detailed requirement analysis that used the output of
this survey to define the DICE goals.
Copyright © 2015, DICE consortium – All rights reserved
11
Deliverable 1.1. State of the art analysis
A.
Data-Intensive Applications
A.1.
Definitions and terminology.
DICE focuses on design, analysis and testing of Data-Intensive Applications (DIAs). Our notion of DIA is
any application that is able to handle Big Data, performing computations at a speed that is in line with the
volume, variety and velocity of such data. The National Institute of Standards and Technology (NIST) [1]
defines Big Data as follows:
‘Big Data consists of extensive datasets primarily in the characteristics of volume, variety, velocity,
and/or variability that require a scalable architecture for efficient storage, manipulation, and
analysis’
This definition concerns, on the one hand, the characteristic to be owned by datasets, which are:
1. Volume, meaning that Big Data datasets are larger than those commonly handled by conventional
databases.
2. Variety, data can possibly come from different sources and in diverse forms.
3. Velocity, referring to the fact that huge amounts of data have to be analysed in a short time.
4. Variability, that is changes in other characteristics, as meaning, veracity, etc.
On the other hand, the reported definition contains also a reference to the qualities of the architectures
needed to cope with Big Data. Such architecture must be carefully designed to be scalable and efficient,
essentially referring to a high degree of parallelism both in terms of storage and processing.
Historically, ever-growing datasets were analysed by increasingly powerful monolithic systems (vertical
scaling). This strategy recently clashed against physical limitations and required to look for a different
approach. Industry and academia started relying on horizontal scaling techniques for parallel distribution
of both storage (Petabytes and even Exabytes) and processing capacity (hundreds or thousands of nodes).
The necessity of new models and tools suitable to seamlessly provide this parallelisation brought to the
emergence of sophisticated, reliable and efficient frameworks. Such frameworks, which include, for
example, Hadoop, Spark and Storm (see Chapter D), are now commonly referred to as Big Data platforms.
The goal of DICE is to realise a model-driven approach that can support the development of applications
harnessing the capabilities of these platforms.
A.2.
Importance of Big Data for the Business
While business considerations are somewhat orthogonal to the investigation in this deliverable, we include
a section to contextualise the importance of DIAs in today’s European economy. The European
commission has stated on several occasions that ‘Big Data is the new oil’. Indeed, data are not only part of
almost all economic and social activities of our society, but have managed to be viewed as an essential
resource for all sectors, organisations, countries and regions. But why this is a reality? It is expected that by
2020 there will be more than 16 zettabytes of useful data (16 Trillion GB) [3]. Data are not only part of
almost all economic and social activities of our society like velocity, variety and socioeconomic value,
flags a paradigm shift towards a data-driven socioeconomic mode which suggests a growth of 236% per
year from 2013 to 2020. Thus, data blast is indeed a reality that Europe must both face and endeavour in an
organised, forceful, user-centric and goal-oriented approach. It is obvious that data exploitation can be the
leading spear of innovation, drive cutting-edge technologies, increase competitiveness and create social
and financial impact.
Copyright © 2015, DICE consortium – All rights reserved
12
Deliverable 1.1. State of the art analysis
Table 1 (source: Big Data Value Association [4]) provides some examples of how Big Data will impact
different sectors:
Table 1: Examples of Big Data impact across various sectors.
Sectors/Domains
Public
administration
Big Data Value
Source
EUR 150 billion to EUR 300 billion in new value (Considering EU
OECD [5], 2013
23 larger governments)
Healthcare & Social EUR 90 billion considering only the reduction of national healthcare McKinsey
Global
Care
expenditure in the EU
Institute [6], 2011
Utilities
Reduce CO2 emissions by more than 2 gigatonnes, equivalent to OECD [5], 2013
EUR 79 billion (Global figure)
Transport and
logistics
USD 500 billion in value worldwide in the form of time and fuel OECD [7], 2013
savings, or 380 megatonnes of CO2 emissions saved
Retail & Trade
60% potential increase in retailers’ operating margins possible with McKinsey
Global
Big Data
Institute [6], 2011
Geospatial
USD 800 billion in revenue to service providers and value to McKinsey
Global
consumer and business end users
Institute [6], 2011
Applications &
Services
USD 51 billion worldwide directly associated to Big Data market Various [8]
(Services and applications)
Based on the previous examples, data penetration in several sectors of our society is indeed a reality.
However, can this influence the business section? Is there a market that shows how important Big Data are
for businesses? According to several studies (e.g. International Data Corporation (IDC) [9]) the Big Data
market is growing six times faster than the overall ICT market. In addition to, the compound annual growth
rate (CAGR) of the Big Data market over the period 2013 – 2017 will be around 27%, reaching an overall
total of $50 billion. Furthermore, as identified by demosEUROPA [10], ‘Overall, by 2020, big & open data
can improve the European GDP by 1.9%, an equivalent of one full year of economic growth in the EU’. At
the same time another study concludes that the increased adoption of Big Data will have positive impact on
employment, and is expected to result in 3.75 million jobs in the EU by 2017 [11]. Thus, Big Data market
has the power to influence the business section. But in which way? Is this influence positive?
According to McKinsey institute, ‘The effective use of Big Data offers the benefits to transform
economies, and delivering a new wave of productive growth’. IDC confirms that Big Data adoption in
Europe is accelerating [12] and subsequently large companies and SMEs has recognised Big Data market
value proposition to revise existing socio-economic and business models according to its offerings. Even
though (according to IDC [13]) 30% of Western European companies will adopt Big Data by the end of
2015, there is a percentage of 70% of business actors that are eager to obtain new tools and assets to propel
them into the data-driven economy.
A.3.
Roles in Big Data Systems
In [2], NIST identifies five main roles in Big Data systems:

System Orchestrator: provides the definition of the system requirements to be fulfilled; examples
of requirements are data, architecture, resources, and business requirements. The system
Copyright © 2015, DICE consortium – All rights reserved
13
Deliverable 1.1. State of the art analysis




A.4.
orchestrator defines monitoring or auditing activities as well to verify the fulfilment of the
requirements.
Data Provider: this role makes available sources of data and information to be fed into the Big
Data system. Data provider’s activities range from Data Collection (e.g. from sensors) and
Persistence (according to several hosting models, Cloud inclusive) to Access Rights Management
and Data Availability Publication (via dictionaries and catalogues).
Big Data Application Provider: realises the data life-cycle performing all the necessary operations
and transformations on the input data to fulfil the requirements defined by the System Orchestrator.
Security and Privacy are primary concerns for the Big Data Application Provider as well.
Big Data Framework Provider: provides resources and services to the Big Data Application
Provider for the creation of Data-Intensive applications. In particular, this role has to supply the
infrastructure, the data management layer and the processing framework.
Data Consumer: this role describes the final user of a Big Data system. They receive and benefit
from the data manipulated by the application. Their activities concern research, retrieval and
download of the data as well as local analysis, visualisation and reporting.
Technology Overview
Two important technological ingredients of a DIA which make them different from canonical enterprise
applications are their emphasis on data storage and data computation mechanisms. In the following
subsections we provide a very short overview of the storage and computation mechanisms available for
DIAs, which will also be described in more details in the next chapters.
A.4.1
Distributed computations
In general, the expression ‘distributed computing’ refers to a software system where various components,
located on different machines, collaborate to achieve a common goal. Over the years many frameworks for
distributed computing have been presented; however, they had the disadvantage of being too focused on
distributed execution and little on parallel data management capabilities. With the advent of Big Data, it
was necessary to develop a new paradigm able to interact effectively with the distributed and variable data.
MapReduce is the first and most famous expression of this paradigm. Its open source implementation,
Hadoop [14], includes functions for reliability, security, workflow integration and governance.
Nonetheless, other solutions have been developed aiming at overcoming the original MapReduce
shortcomings. These solutions seek to respond to different needs and often they have to coexist in the same
environment. For this reason, resource management systems have been recently developed for such
solutions. These systems are designed to handle the available computing resources assigning them
according to specific policies. In this way, multi-tenant, multi processing models (other than MapReduce)
and high availability environments can be created. The most known resource management systems are
YARN (see Section D.2) and Mesos [15].
A.4.2
Distributed Storage and NoSQL solutions
In the classical database terminology, the basic unit of information is called Data Record and corresponds
to a single observation of a particular event associated with the system to be analysed. Data records in
classical database theory represent a piece of structured information and can be thought of as a row of a
relational database table. In the Big Data context information are not necessarily structured. They can also
be unstructured (images, videos) or semi-structured (text, documents). Thus, a different data model for
storing such data became necessary. Non-relational databases, also known as NoSQL (Not only SQL, see
Section D.4) have been developed for addressing such new needs offering, at the same time, a large degree
of scalability in terms of horizontal as well as vertical scaling.
Copyright © 2015, DICE consortium – All rights reserved
14
Deliverable 1.1. State of the art analysis
The use of NoSQL solutions in the context of Big Data is preponderant and, in many respects, it made Big
Data possible; nonetheless, it is necessary to warn the reader from falling into the easy association ‘NoSQL
= Big Data’ because recent years have witnessed the appearance of highly scalable storage systems based
on the relational model.
Moreover, due to the size of datasets and the need for parallelism in the computation, high performing file
systems such as Hadoop Distributed File System (HDFS, see section D.2) have been developed. The main
characteristics of these file systems are to be 1) distributed, i.e. files and dataset are seamlessly distributed
over several nodes. In this way different nodes can access and analyse data at the same time on the same
dataset; 2) multi-structured, meaning that a variety of data types are supported. Usually, the information is
stored at block level; 3) replicated, files and blocks are replicated several times for reliability/recovery and
to enhance data locality.
A.4.3
Streams and Big Data in motion
In the previous paragraphs, we tried to outline Big Data problems and approaches leaving aside the much
of the discussion on the latency in data management. We introduced distributed storage solutions and the
reader might presume that the only way to deal with Big Data is to collect them into huge datasets (data at
rest) and analyse them ‘a posteriori’, periodically or as a response to an event. This is not always the case,
though. Data can enter the system from different sources and examined in (almost) real-time (data in
motion). In this scenario, the data flows are referred to as data streams and they are mainly characterised
by velocity and variability. Many solutions for data streaming were devised used long before the concept of
Big Data came up. Nevertheless, the parallelisation, resource management and reliability supplied by
modern Big Data framework imposed a reconsideration and redefinition of many basic elements in data
streaming. As evidence of this, in the last couple of years we have witnessed the successful application of
concepts typically used for data at rest (Maps, Reduces, direct acyclic graphs (DAGs), Batches) to data in
motion scenarios. Spark is an example of framework aiming at unifying both scenarios under the same
programming paradigm (see Chapter D).
A.5.
Architectural Styles for DIAs
The particular focus of DIA on Big Data makes them fairly different from traditional enterprise application.
Therefore, the research and technical community have investigated in recent years novel architectural styles
to support DIAs. In particular, the data velocity of Big Data applications is application-specific and may
range from a few to millions of data items per second. Different Big Data applications enforce different
quality of service constraints regarding response time. For example, a reactive use case with high-volume
data streams may require an answer in a real-time (milliseconds) fashion.
A.5.1
Lambda Architecture
The Lambda Architecture introduced by Marz [16] is an advanced architectural style to overcome the
challenges in general for Big Data applications and more specifically on real-time stream processing. The
architectural style decomposes the problem into three key layers: (i) the batch layer focuses on fault
tolerance and optimises for precise results (ii) the speed layer is optimised for real-time response-times and
only consider the most recent data and (iii) the serving layer provides low latency views to the results of
the batch layer. Note that some other layers will be introduced later for coordination and orchestration
purposes.
Lambda Architecture enables real-time responses to query over Petabytes of data. Such a query on
traditional architectures imposes unreasonably high latency. The Lambda Architecture divides this problem
into three layers. The batch layer pre-computes the query function based on the full data set and updates the
serving layer. This operation involves high latency and by the time the view of the pre-computed query is
Copyright © 2015, DICE consortium – All rights reserved
15
Deliverable 1.1. State of the art analysis
finished it is already outdated. The speed layer only operates on the most recent data in order to provide
views for the missing time span of the batch layer. The principal structure of an application based on
Lambda Architecture is shown in Figure 1.
Figure 1. Lambda Architecture.
In more details, the Lambda Architecture [17] may be divided into five components: (i) orchestration (ii)
coordination (iii) batch layer (iv) speed layer and (v) serving layer. Figure 1 illustrates the components of
the Lambda Architecture and the data flow through the system. The orchestration component manages
resource allocation, provides scheduling functionality to the other key layers. The coordination component
facilitates the entry point for data and offers high-level services to synchronise processes within different
components. The batch layer facilitates a micro-batch processing system with high reliability and fault
tolerance. The speed layer embeds a real-time stream processing system with reasonable response times.
The service layer collects results and performance data from both the batch and the speed layer in order to
measure the key performance indicators of this architecture.
The batch layer maintains the master data set that includes all data to pre-compute the necessary results.
The data storage system has to provide the following requirements: (i) efficient writes for new data (ii)
scalability to cope with the increasing need to store more data and (iii) support for parallel processing and
the ability to partition the data. Since the data set is continuously growing the latency to pre-compute the
batch views on the whole data set becomes increasingly expensive and the time span to catch up with the
speed layer will increase. The batch layer, therefore, processes new incoming data. The serving layer
provides fast access to the pre-computed results of the batch layer. These results are outdated because of
the high latency of the batch layer. Therefore the write speed to the serving layer is less important than the
read performance. The serving layer has to provide low latency random reads in order to answer queries
efficiently. The speed layer is responsible for providing results to the most recent data and has to fulfil
certain latency constraints based on the application. This implies that it is not technically possible to
compute results based on the full master data set. Instead incremental computation is applied to the most
recent data in combination with persistent state. The speed layer is more complicated than the batch layer,
but any error is eventually compensated by the batch layer.
A.5.1.1. Coordination Layer
The coordination layer is the entry point of data into the system. A typical scenario involves reading from
an input source and delivering the messages to either the batch or speed layer. Potential design patterns for
realising coordination layer:


Persistent Messaging
In-Memory Messaging
Copyright © 2015, DICE consortium – All rights reserved
16
Deliverable 1.1. State of the art analysis
A.5.1.2. Batch Layer
The objective of the batch layer is to compute precise results that eventually replace the possible inaccurate
results of the speed layer. In case of node failures the processing may be delayed in order to guarantee this
promise. There are two possible strategies to implement the batch layer: (i) a re-computation algorithm that
periodically computes the results over the full master set or (ii) an incremental algorithm that processes
new data when it is introduced into the system. Potential design patterns for realising batch layer:



Micro-batch processing
Replay Mechanism
Precise recovery
A.5.1.3. Speed Layer
The main objective of the speed layer is to deliver results in real-time and therefore optimise the latency
and throughput to process incoming data. In comparison to the batch layer, the speed layer does not
guarantee the same high availability and fault tolerance as the batch layer. In case of node or processing
failures the speed layer drops messages to rapidly process the new incoming data instead of replaying all
messages to the last processed state. This design leads to inaccurate results in case of node failure, but these
results are eventually replaced by the results of the batch layer.
A.5.1.4. Orchestration
The Lambda Architecture embeds highly distributed services on each layer with the disadvantage of
difficult synchronisation and resource management challenges. In order to mitigate these challenges an
orchestration layer is introduced. In a typical solution, this layer leverages the Apache Hadoop YARN
framework [18], built-in Storm service [19] for resource allocation and Apache ZooKeeper [20] and Kafka
[21] for coordination.
A.5.1.5. Serving Layer
The service layer provides a set of tools to support near real-time process monitoring, localised
performance measurements and result collection. This layer can access the results of the batch layer
computations with low latency. The serving layer would index views of the batch layer and provide the
necessary interfaces to access the pre-computed data with low latency queries.
A.5.1.6. Academic and industrial positioning
Fan and Bifet [22] suggest the Lambda Architecture as one solution to the future challenges of Big Data
with regard to data mining. Robak et al. [23] explored the application of Big Data and linked data concepts
in supply chain management and listed the Lambda Architecture as one possible solution to deal with high
volume of data in this field. They argue that such architecture could provide lower latency to react in
critical situations. Bär [24] presents a possible implementation of the Lambda Architecture with no
merging based on open source software components. He evaluates the capabilities of this architecture on
the SRBench Benchmark [25] and DEBS Grand Challenge 2014 [26] task with varying data frequency
rates on an unreliable infrastructure.
Google develops a new data processing service provided on top of their cloud platform called Google
Cloud Dataflow [27]. It is similar to the Lambda Architecture by allowing clients to process data with
stream and batch processing. Google framework and Spark (see Section D.2.3) rely on parallel collections
of any size that are distributed across multiple machines in order to provide scalability.
Amazon maintains a set of loosely coupled, but well integrated tools that provide the means to realise
Lambda Architecture. Kinesis [28] is a fully managed real-time stream processing service that uses a
messaging concept similar to Kafka [21]. Kinesis integrates Amazon S3 (see Section D.6.2) and Elastic
Copyright © 2015, DICE consortium – All rights reserved
17
Deliverable 1.1. State of the art analysis
MapReduce [29]. Amazon provides the necessary tools to build a batch and speed layer, but the integration
of both layers and its challenges are not offered as a holistic solution yet. Lambdoop [30] is an industry
implementation framework that facilitates application development based on Lambda Architecture.
A.5.1.7. Oryx: an example Lambda Architecture
Throughout the DICE project, we will use Oryx [31] as a reference technology to illustrate the applicability
of some of our technical results to Big Data applications.
Oryx is an open source framework for building Big Data applications, but also includes ready to be
deployed out-of-the-box example application, which can be used as it is or as a basis for developing a
custom application. The intended application of Oryx framework is predictive analytics in real time based
on the construction of models from the incoming streaming data. The models used in analysis are built
using Machine Learning (ML) algorithms. The target areas for the application - among others - are
business, health, education and weather forecasts. The design of Oryx is based on the Lambda Architecture.
The first version of Oryx (Oryx 1) [32] was released in 2013. Its architecture consisted of two layers:
Computation Layer - where models based on incoming streaming data were built and evaluated based on
the requests from a client, and Serving Layer - a medium for accepting requests from a client, transferring
them to the Computation Layer for evaluation and returning result to the client. The Computation Layer
runs both Batch process, which is an ‘offline’ process, meaning that it does not operate in real time, rather several times a day, and model update and evaluation process, which is quick (process time measured in
seconds). All parameters are controlled via configuration files. The high-level architecture of Oryx 1 is
presented in the Figure 2.
Figure 2. Architecture of Oryx 1 [32].
The high-level architecture of Oryx 2 is presented in the Figure 3.
Copyright © 2015, DICE consortium – All rights reserved
18
Deliverable 1.1. State of the art analysis
Figure 3. Architecture of Oryx 2 [31].
Both Oryx 1 and Oryx 2 employ machine learning algorithms to construct prediction models. Brief
description of these algorithms is given in the Table 2.
Table 2: Machine learning algorithms used in Oryx 1 and Oryx 2.
Collaborative
filtering
A technique used by recommender systems (engines) to suggest various items such as
movies, music, news, books, research articles, search queries and so on to the client.
Collaborative filtering implemented in Oryx uses matrix factorisation-based approach based
on a variant of Alternating Least Squares (ALS) [34].
Classification
and Regression
Classification can be used, for example, to separate e-mails into ‘spam’ and ‘non-spam’.
Regression is used to predict a specific numeric value (e.g. temperature on a given day or
salary in certain year). Oryx employs random decision forests algorithms to solve
classification and regression problems. Classification and regression belong to the supervised
learning category of ML algorithms, which means that they require some initial data sets to
be ‘trained on’.
Clustering
Clustering is similar to classification in the sense that an object is assigned to a specific
category (group, class), but clustering is an unsupervised learning method. It does not require
an initial set to create classes to which subsequent incoming data is then compared, but rather
tries to create groups (classes) from incoming data by looking for some common features in
it. Oryx implements scalable k-means++ [35] for clustering.
A.5.2
Other Architectures
Although the Lambda Architecture has gained consensus in recent years, it has some limitations. First,
maintaining code that needs to produce the same result in two complex distributed systems is expensive.
Programming in distributed frameworks like Storm [19] and Hadoop [14] is complex. One proposed
approach to fixing this is to have a language or framework that abstracts over both the real-time and batch
framework. The developer writes code using this higher level framework and then it is translated into lowlevel stream processing and/or MapReduce code. However, this entails the operational burden of running
and debugging two systems which is going to be very high.
Copyright © 2015, DICE consortium – All rights reserved
19
Deliverable 1.1. State of the art analysis
Therefore, alternative architectures have been proposed. For example, the Kappa Architecture has been
proposed by LinkedIn [36] exploiting stream processing, that is:
1. Use messaging system like Kafka [21] that will let you retain the full history of the data you want
to be able to reprocess.
2. When one wants to do the reprocessing, start a second instance of the job that starts processing
from the beginning of the data, but put this output data to a new table.
3. When the second job has been finished, switch the query to read from the new table.
4. Stop the old version of the job, and delete the old output table.
This new architectural style of the Kappa Architecture is depicted in Figure 4.
Figure 4. Kappa Architecture.
Another recent proposal is the Liquid Architecture [37], which has two layers: a messaging layer based on
Apache Kafka [21], and a processing layer based on Apache Samza. The processing layer (i) executes jobs
for different back-end systems according to a stateful stream processing model; (ii) guarantees service
levels through resource isolation; (iii) provides low latency results; and (iv) enables incremental data
processing. A messaging layer ports the processing layer. Figure 5 shows architectural style of the Liquid
Architecture.
Figure 5. Liquid Architecture [38].
Copyright © 2015, DICE consortium – All rights reserved
20
Deliverable 1.1. State of the art analysis
B.
DevOps
B.1.
Introduction.
DevOps is a hybrid software development and operation paradigm that predicates the intermixed cooperation, collaboration and communication between Dev-Teams (i.e. development teams) and Ops-Teams
(i.e. Operations Teams). The fundamental axiom in the software operational paradigm is that Dev and Ops
teams should work together since the one cannot deliver quality without efforts and insights of the other
[39]. In recent times DevOps started playing a major role in the design, development and continuous
integration of industrial-strength software [40].
In the DICE project DevOps is addressed in a novel way through a Model-Driven Engineering (MDE)
approach. In fact, we argue that exploiting models and MDE both during the design and the operation time
to reason on the applications under consideration and to evolve it to accommodate the required qualities is
a good way to accomplish the principles suggested by DevOps.
Thus, in the rest of this chapter we focus on reviewing the methods, practices and tools that can be
applicable or adaptable, in a DevOps context, to support part of the life-cycle of Data-Intensive
Applications. More specifically:








In Section B.2 we provide a short introduction to the general DevOps philosophy and practices.
In Section B.3 we provide an overview of model-driven development and of those model-driven
approaches that appear to be closer to the DICE objectives.
In Section B.4 we present the most well-known modelling tools available, both open source and
commercial.
In Section B.5 we provide an overview of model-to-model transformation approaches. These are
very important to automate the development steps in a MDE approach.
In Section B.6 we review the standard TOSCA deployment approach and highlight the potential of
this approach for DICE, and in Section B.7 we review other deployment and management tools
well known in the market.
In Section B.8 we focus on the approaches for continuous integration.
In Section B.9 we review the approaches to keep track of versions of software and models as this is
a very important issue for continuous integration and operation approaches.
Finally, in Section B.10 we summarize the main strengths and weaknesses of the reviewed
approaches and tools with respect to the DICE context.
In each section we present a short overview of the state of the art and highlight the relevance of the specific
area being surveyed with respect to the DICE objectives.
While reviewing specific tools, we provide information on the following aspects:
1. The general dimension of the tool (website, licensing, present version, communities, FP7 projects
[41] or companies that produce the tool, number of downloads, number of users etc).
2. The formalisms/standards supported by the tool (if any).
3. The features supported by the tool for a specific formalism.
4. The compatibility/integration with tools of different application domain.
Copyright © 2015, DICE consortium – All rights reserved
21
Deliverable 1.1. State of the art analysis
B.2.
Short introduction to DevOps
The DevOps philosophy revolves around four key values, explained in Figure 6. These are:




The importance of developing and testing complex software in an environment that is as close as
possible to the production one. The aim of this principle is to ensure that developers keep in mind
and experience with the operation issues like performance, availability, stability of the service
during development.
The utility of deploying software as far as possible without losing the control of the services in
operation and their qualities.
The importance of continuously checking the quality of the software product and the associated
services not only in the specific verification and validation phases but through the entire life-cycle
of the services themselves.
The need for a continuous collaboration between developers and operators with the aim of
overcoming the typical frictions between the first ones, focusing on innovating products and the
second ones focusing on guaranteeing the stability of the resulting services.
Figure 6. DevOps key values [42].
An essential part of the benefits connected to DevOps stems from the intrinsic collaboration predicated
above and represents the possibility of continuously deliver and incrementally manage operating software
using organisational and technical integration methodologies. In the scope of said goals, DevOps
practitioners developed a line of tools to increase the speed at which Dev and Ops can cooperate, collocate
their work and cohesively coordinate their efforts. Figure 7 shows a practical overview of the activities and
tools behind the DevOps philosophy. In summary, a series of typical activities entail DevOps:
1.
2.
3.
4.
5.
Industrial-Strength Version Control.
Continuous integration of several working-copies into a mainline several times per day.
Release automation, i.e., the automated (re-)deployment of improved products.
Infrastructure automation, i.e., providing automatically adjustable infrastructures.
Application management, i.e., the orchestration of services to guarantee optimal service-levels.
The above activities are usually supported by a series of tools that are proliferating as part of the DevOps
movement. Widely known examples are:
1.
2.
3.
4.
5.
Git (see Section B.9.1.3), Ant [43] or Hudson (see Section B.8.2) (Continuous Integration).
Capistrano [44] or RPM Package Manager [45] (Continuous Deployment).
Cobbler [46] or Crowbar [47] (Infrastructure automation).
Puppet [48] and Chef [49] (Configuration Management).
Nagios [50] and Sensu [51] (Run-Time Monitoring).
Copyright © 2015, DICE consortium – All rights reserved
22
Deliverable 1.1. State of the art analysis
Figure 7. An Overview of the Activities and Tools behind DevOps.
Finally, the organisational integration to be enforced within DevOps typically uses simple combined teams
(e.g., dev + ops) or follows other more strict organisational patterns typical in DevOps success stories,
many of which are currently under investigation in the Software Engineering Institute [52] to assess their
operational effectiveness. For example, Figure 8 shows an example of organisations structure typical in
organisations that embrace DevOps. More in particular, Figure 8 shows a balanced blend between Dev- and
Ops- people across a product portfolio.
Figure 8. Typical Organisational Structures behind DevOps, an example [42].
In summary, DevOps is a movement that helps to bridge the cultural gap between development and
operations [53]. Its goal is to enable each department to be aware of the perspective of the other and push
them to change the dynamics in which they interact [54].
DevOps provides patterns to foster collaboration among project stakeholders and addresses shared goals
and incentives as well as shared processes and tools [55]. Therefore, the concept of ‘sharing’ is at the very
core of DevOps: sharing ideas, goals, issues, processes and tools. In addition, DevOps incites Devs and
Ops teams to share their skills and experiences with each others, which leads to a one team approach where
individuals have at least a basic understanding of others domains [55].
DevOps tries to extend Agile practices to operations by eliminating the wall between development and
operations and to address the structural conflict between them: both teams work together to deliver
application changes to the user at a high frequency and quality.
Copyright © 2015, DICE consortium – All rights reserved
23
Deliverable 1.1. State of the art analysis
Although improving communication between developers and operations teams contributes to solve critical
issues, it is only a portion of the wider equation; integrating the right tools is important for DevOps. Major
parts of the releasing process should be automated along the delivery process in order to facilitate
collaborative change [54].
Automation has many benefits: it ensures that the software is built the same way each time and makes parts
of the process transparent for the whole team; thus software deployment to different target environments is
made in the same manner [55].
Automation includes many steps (preparing the build, checking quality of code, launching build, running
all tests, packaging, deploying and staging the artefacts) and necessitates scripts (for building, testing,
deploying, configuring application, and configuring infrastructure). In his book, Michael Hüttermann
explores concrete patterns for automatic releasing with appropriate tools [55] (Chapters 8 and 9).
By combining different approaches of DevOps and applying well-known DevOps practices, IT
performance is strongly improved, which contribute to organisational performance as described by
PuppetLabs report [56].
In the scope of DICE, DevOps methodologies and tools represent a set of reference materials as well as a
potential target for further integrated support.
On the one hand, DevOps methodologies and tools need to be taken into account as organisational and
technical concerns which are the key to enabling continuous delivery of Data-Intensive Applications
(DIAs) by design and which govern the feedback loop between runtime and design time that is intrinsic to
the DICE approach.
On the other hand, DevOps and connected tools/methods are themselves predicating on the usage of DIAs
to further the understanding of the application lifecycle for further improvement and increased
organisational/technical agility. In this regard, DICE may be a valuable tool to study where DIAs may play
a role in finding valuable business intelligence to speed up application development and deployment.
Moreover, testing within the DICE project should make massive and careful reference to DevOps methods
and tools in order to establish guidelines and test-cases according to which DIAs may be designed and
tested for ‘DevOps-Readiness’. This may include studying the best-fit organisational and socio-technical
patterns of integration between DIAs and typical DevOps toolchains.
In addition, this may also entail studying empirically in which circumstances the DICE model-driven
assumptions fail to meet DevOps expectations (e.g. continuous integration). Finally, from a methodological
point of view, DIAs’ specification and monitoring should take into account DevOps dynamics and tools to
enable DIAs monitoring in sight of their continuous improvement.
B.3.
Functional modelling
B.3.1
Overview
From a functional perspective focused around modelling, the DICE approach to define, specify and analyse
DIAs may inherit results from a number of technological baselines. In particular, we see the following
baselines as relevant to the DICE project: (1) the ‘Model-Driven Architecture’ (MDA) standard [57], i.e.,
the Object Management Group (OMG) [58] standard for Model-Driven Engineering (MDE) [59], (2) the
MODAClouds EU project [60] and related key results (e.g. MODACloudML [61]); (3) the REMICS EU
project [62] and key results therein (e.g. Model-Driven cloud-migration techniques); (4) the Artist EU
project [63] and key results therein that share a similar purpose to REMICS; (5) the JUNIPER EU project
Copyright © 2015, DICE consortium – All rights reserved
24
Deliverable 1.1. State of the art analysis
[64] and key results therein share DICE goals and aims at an even lower level of abstraction; (6) in
general, Unified Modelling Language (UML) and its profiles called MARTE (Modelling and Analysis of
Real-Time and Embedded systems) and DAM (profile for Dependability Analysis and Modelling) (see
Section B.3.3). These last ones are of particular interest to DICE as they allow users to model performance
and other quality characteristics of applications. The rest of this section elaborates said technologies in
more detail, with a hint as to their possible role in DICE.
B.3.2
MDE and domain-specific model-driven approaches
MDE techniques [59] and MDA in particular [57] define the typical abstraction layers for the purpose of
engineering software systems using a model-centric perspective. The fundamental axiom behind this
engineering paradigm is that any engineering endeavour shall be guided by at least three compounding and
interoperating perspectives, namely: (c) Computational-Independent perspective; (b) a PlatformIndependent perspective; (c) a Platform-Specific perspective. Using these three perspectives, one or more
models can be specified to properly and systematically specify a system-to-be.
1. At the Computational-independent level, business-critical details are defined as such that intended
business scenarios and systems goals may become apparent and explicit. Typically this perspective
is consistent with requirements engineering activities such as stakeholder identification and
scenario analysis.
2. At the Platform-independent level, architectural, quality and design issues are specified using one
or more Architecture Viewpoints. The specification at this level typically uses model
transformation technologies to support consistency and analysis across multiple Views and
Viewpoints.
3. At the Platform-Specific level, design decisions are realised into well-formed designs, e.g.,
reflecting appropriate selection of design patterns, usable technological platforms and middleware
(e.g. CORBA [65]).
Within DICE, MDE and MDA play a key role in providing a fundamental specification baseline. More
specifically, the DICE profile inherits the MDA separation of concerns and logical decomposition, as well
as a model-centric approach featuring multiple model transformations both for model consistency (and
eventually technological deployment) and model analysis.
MDE and MDA have been adopted and specialised for various domains in many research projects. Among
the others, MODACloudML [61] provides a Domain-Specific Modelling Language (DSML) along with a
runtime environment in order to allow, on the one hand, to model the provisioning and deployment of
multi-cloud applications, and on the other to automate the deployment and to facilitate their runtime
management (in terms of adaptation or reconfiguration actions).
In this way MODACloudsML supports DevOps in achieving better delivery life-cycle by integrating in a
single framework both development and operation activities.
MODACloudML supports both the Infrastructure as a Service (IaaS) and the Platform as a Service (PaaS)
levels, even if it mainly focuses on the former. From the modelling perspective, MODACloudML allows
the application specification at three levels of abstraction, which aim at following the general MDA
paradigm: the Cloud-enabled Computation Independent Model (CCIM) to model an application and its data
from a high-level business perspective, (ii) the Cloud-Provider Independent Model (CPIM) to characterise
cloud concerns related to the application in a cloud-agnostic way, and (iii) the Cloud-Provider Specific
Model (CPSM) to model the deployment and provisioning activities on a specific cloud.
Copyright © 2015, DICE consortium – All rights reserved
25
Deliverable 1.1. State of the art analysis
At the CCIM level, an application is described as a set of high level services following a Service Oriented
Architecture. At this level of abstraction these models involve three main concepts: a set of services, an
orchestration, and a set of usage models. At the CPIM level MODACloudML proposes a new approach to
describe the deployment, provisioning and data models of multi-cloud systems in a provider-agnostic way,
supporting both IaaS and PaaS solutions. At the CPSM level, the design alternatives and deployment
models as well as the data models are refined to include provider-specific concerns and technologies.
MODACloudML is also inspired by component-based approaches, which facilitates separation of concerns
and reusability. In this respect, deployment models can be regarded as assemblies of components exposing
ports and bindings between these ports. In addition, MODACloudML implements the type-instance pattern,
which also facilitates reusability and abstraction.
Even if it is not specifically focused on DIAs, MODACloudML can become one of the foundational
approaches behind the definition and further elaboration of the DICE profile for at least three reasons:
1. MODACloudML follows already the logical decomposition and model-driven engineering
abstraction behind typical MDE-inspired techniques, e.g., with the division and arrangement in
three tiers of modelling and analysis.
2. MODACloudML sets to describe Cloud-based applications - it is in fact a fundamental assumption
behind DIAs that their very nature resides in the cloud. Hence, MODACloudML is a reasonable
technology upon which to draw inspiration both in terms of modelling/deployment and analysis.
3. The MODAClouds project already provides infrastructure and usable technologies/tools for the
specification, analysis and deployment of MODACloudML models, e.g., allowing the DICE profile
to be quickly developed in prototypical form and further refined by means of MODAClouds-based
technology.
Other EU projects that could offer a fundamental inspiration for DICE are REMICS [62] and Artist [63].
They look at ways in which model-driven techniques can be used to: (a) accelerate the adoption of cloudbased technology, possibly migrating legacy assets; (b) use models to drive the continuous improvement of
cloud-intensive applications such as DIAs; (c) use model-centric perspectives to evaluate cloud assets and
compare multiple cloud-vendors to compute solutions best-fitting with Service-Level Agreements (SLAs)
and stakeholder/customer concerns. Quoting from REMICS website [62], the project’s purpose is to
provide a model driven methodology and tools which significantly improve the baseline ArchitectureDriven Modernisation (ADM) [66] concept. In a similar vein, DICE could work jointly with efforts
inherited from REMICS by specifying constructs, modelling and methodological notations to enable the
systematic migration of legacy Big Data analytics into well-formed and quality-aware DIAs, enabled for
continuous evolution.
Finally, the JUNIPER EU project [64] intends to construct the platform from real-time technologies, using
real-time analysis, design and development principles, so that appropriate guarantees can then be given
with respect to Big Data processing times, performance and similar quality attributes. In its current version,
JUNIPER offers a programming model and a series of APIs to speed up the development of performanceaware Big Data applications.
On one hand, the JUNIPER programming model aims to conceptualise the development process of datacentric applications in a way that is general enough to cope with every need (in scope of the JUNIPER
scenarios) but also allows common data processing patterns to be abstracted, modelled, and optimally
deployed. The programming model proposed in JUNIPER is not a replacement of any of existing parallel
processing frameworks and programming models, such as Hadoop/MapReduce [14], Message-Passing
Copyright © 2015, DICE consortium – All rights reserved
26
Deliverable 1.1. State of the art analysis
Interface etc. Rather, the programming model shares the purpose of DICE, i.e., that of integrating said
technologies to accelerate and further support their adoption during design and development.
On the other hand, as part of the JUNIPER baseline, the EU project offers Java improvements and
primitives that tackle some missing feature like flexible parallelism, locality and hardware architecture
discovery. More in particular, JUNIPER Application Programming Interfaces (APIs) characterise the host
architecture in terms of accepted patterns and assist the development of reactionary software to exploit the
hardware. When building software for Big Data systems, their unique architectures result in interesting
challenges. The JUNIPER project identifies two main levels at which issues are observed: the cluster level
and the node level. While at the cluster level, Big Data systems are deployed ‘in the cloud’, at the node
level, the Big Data problem is to be decomposed into a ‘normal data’ problem.
From the DICE perspective, besides inheriting JUNIPER concepts and definitions as outlined above, DICE
may use JUNIPER results as a baseline for at least two purposes: (a) inspire the DICE model-driven
engineering of Big Data applications following patterns and approaches previously introduced and studied
within JUNIPER; (b) test the DICE approaches and methodologies against scenarios previously envisioned
as part of the JUNIPER challenges. The innovative aspects behind DICE appear clear from at least two
standpoints: (a) Speed - a primary DICE goal is the speed-up of development & deployment times for
Data-Intensive Applications, i.e., applications using multiple Big Data programming models at once, of
which JUNIPER is only one possible alternative; (b) Quality-Awareness - another primary DICE goal is to
provide facilities to extensively model and verify the assessment of quality aspects across Big Data
applications, where JUNIPER merely provides constructs for provisioning of quality.
B.3.3
UML, MARTE and DAM
UML is a General Purpose Modelling Language (GPML). Therefore, it can be used to model a wide range
of systems. Quoting the document [67], ‘UML is a language with a very broad scope that covers a large
and diverse set of application domains. Not all of its modelling capabilities are necessarily useful in all
domains or applications’. Conversely, Domain-Specific Modelling Languages (DSML) and DomainSpecific Languages (DSL) are conceived for addressing the needs of specific application domains. A DSL
captures the semantics of the domain and offers the syntax needed for modelling the concepts in such
domain. By using a DSL the designer does not need to learn concepts completely irrelevant for the domain
they are addressing. In this regard, UML offers a solution, the so-called UML profiling mechanism [67].
Profiling opens the possibility of creating DSLs by extending or restricting UML. A Profile is then an
adaptation of UML to fit a specific domain. Examples of UML Profiles are MARTE [68] and DAM [69].
MARTE (Modelling and Analysis of Real-Time and Embedded systems) provides support for the
specification, design, quantitative evaluation, and verification & validation of software systems. DAM
(Dependability Analysis and Modelling Profile) provides support for the dependability modelling and
analysis of software systems. A UML model annotated with the MARTE and/or DAM Profile will be
called a UML-MARTE or UML-DAM model.
Summarising, some of the benefits of using the MARTE and DAM Profiles are [68]:



Providing a common way of modelling both hardware and software aspects of a system to improve
communication between developers.
Enabling interoperability between development tools used for specification, design, verification,
code generation, etc.
Fostering the construction of models that may be used to make quantitative predictions regarding
real-time and embedded features of systems taking into account both hardware and software
characteristics.
Copyright © 2015, DICE consortium – All rights reserved
27
Deliverable 1.1. State of the art analysis
However, neither MARTE nor DAM has a direct support for expressing data location, data properties such
as volume or transfer rates or operations that move data. Hence, addressing such lack is one of the
objectives of the DICE project.
B.3.3.1. Domain-Specific Modelling with UML
UML offers different diagrams for the modelling of the structural, behavioural and distribution views of a
system. For example, the object diagram and the class diagram describe the structure of a system. The state
machine diagram, interaction diagrams, activity diagram and use cases are used for the modelling of the
system dynamic and behaviour. The component and deployment diagrams describe system distribution. A
UML diagram is made of elements, for example, a class diagram is made of classes and relationships
among them, such as associations, inheritance or dependencies. The UML package diagram is useful for
organising UML diagrams and/or UML elements. The UML model of a system is made of a set of UML
diagrams. A UML model has to conform to the UML meta-model. A meta-model is a set of related metaclasses. A meta-class is the abstraction of a set of UML elements. For example, in a UML class diagram,
each association belongs to the Relationship meta-class of UML since associations share characteristics
with other relationships. This meta-model feature is an interesting characteristic of UML since the profiling
mechanism builds on it. The UML Profiles package contains mechanisms that allow meta-classes from
existing meta-models to be extended to adapt them for different purposes. This includes the ability to tailor
the UML meta-model for different platforms (such as J2EE [70] or .NET [71]) or domains (such as realtime, business process modelling or DIA). A UML Profile is made of a set of stereotypes, a set of tags and
a set of related constraints. A stereotype is just a name that will be attached to certain elements of a UML
diagram. Stereotypes have tags, we can see them as the attributes added by the stereotype. A constraint can
be attached to a stereotype definition. It is expressed in natural language or in the Object Constraint
Language (OCL) [72] and describes restrictions for the stereotype, e.g. for expressing subsets of values for
the stereotype. Figure 9 goes deeper into the Profile definition. A Profile is a specialisation of the UML
concept of Package, which means that a Profile is a set of modelling elements, in fact the stereotypes, tags
and constraints.
Figure 9. Sketch of UML Profile definition.
Therefore, stereotypes are the cornerstone concept in a Profile. At model-specification level stereotypes are
applied to concrete UML elements, which is possible since they extend concrete meta-classes of UML, as
we can see in Figure 9.
B.3.3.2. Modelling with MARTE
The MARTE profile consists of three main parts: MARTE Foundations, MARTE Design Model and
MARTE Analysis Model.
First, the MARTE Foundations define the basic behaviour concepts for the Real-Time and Embedded
Systems (RTES) domain, such as a causality model, a common framework for annotating models with
quantitative and qualitative Non-Functional Properties (NFP, say performance, reliability and safety), the
modelling of time, the modelling of resources and their allocation concerns.
Copyright © 2015, DICE consortium – All rights reserved
28
Deliverable 1.1. State of the art analysis
Second, the MARTE Design Model addresses the modelling of RTES. It introduces a general component
model suitable for RTES, high-level concepts for designing qualitative and quantitative concerns (e.g.,
concurrency and synchronisation) and a detailed resource modelling dedicated to software and hardware.
Third, the MARTE Analysis Model enables the analysis features of the system. It consists of a Generic
Quantitative Analysis and Modelling (GQAM) profile which includes common concepts to any kind of
quantitative analysis (such as workload, scenario, resource, etc.), a Schedulability Analysis and Modelling
(SAM) profile and a Performance Analysis and Modelling (PAM) profile. SAM and PAM are
specialisation of the GQAM profile. It is important to note that the DAM profile also specialises the
GQAM profile.
In addition, MARTE contains some annexes: the Value Specification Language (VSL) profile defines an
expression language for specifying the values of constraints, properties, and stereotype attributes,
particularly related to NFP; the MARTE Library defines primitive data types, a set of predefined NFP types
and units of measures.
B.3.3.2.1. Specification of NFP
The MARTE NFP profile was designed to address with the following specification needs: how to specify
NFPs, how to attach NFPs to UML model elements, how to define relationships between different NFPs
and how to express constraints on or between NFPs. The NFP profile imports the VSL profile. The VSL
profile defines a set of language constructs and a textual grammar supporting extended values, expressions
and data types used particularly for specifying NFPs. The NFP package realises the following requirements
for NFPs:




NFP Qualitative or Quantitative nature. A quantitative property may be characterised by a set of
measures expressed in terms of magnitude and unit. A qualitative property may be denoted by a
label (e.g. periodic, sporadic and bursty for an event arrival patterns) representing an abstract
characterisation with a certain meaning for designers or tools.
Qualifiers can characterise the precision, accuracy, statistical measure (e.g. mean, maximum,
minimum and variance) or source (showing how a given value was obtained, e.g. required,
measured, estimated, simulated or calculated).
Variables and Expressions, beside concrete values, raise the level of abstractions of the specified
properties, allowing to derive ones from the others.
Trade-off between usability and flexibility. Usability suggests the merit of declaring a set of
standard property types and their available operations for a certain domain, while flexibility allows
for user-defined properties.
VSL defines the following data types: Bounded Subtype, IntervalType, CollectionType, TupleType and
ChoiceType. VSL also defines four kinds of composite value specifications: collection, interval, tuple, and
choice. The value specifications defined in VSL can be attached to stereotype attributes or used in
constraints.
B.3.3.2.2. MARTE::GQAM framework
The fundamental concepts shared by different quantitative analysis domains are joined in a single package
called Generic Quantitative Analysis Model (GQAM). This package is further specialised for
schedulability (SAM), performance (PAM) and dependability (DAM) analyses. The core GQAM concepts
describe how the system uses resources over time. It contains three main categories of concepts: resources,
behaviour and workloads. In the following, we detail the resource concepts. A resource contains common
features such as scheduling discipline, multiplicity, services. The following types of resources are
important in GQAM:
Copyright © 2015, DICE consortium – All rights reserved
29
Deliverable 1.1. State of the art analysis




ExecutionHost: a processor or other computing device on which are running processes.
CommunicationsHost: a communication network or a bus connecting processing nodes and/or
devices.
SchedulableResource: a software resource managed by the operating system, like a process or
thread pool.
CommunicationChannel: logical channel that conveys messages.
In order to be executed, a SchedulableResource must be allocated to an ExecutionHost and a
CommunicationChannel to a CommunicationHost. Services are provided by resources and by subsystems.
A subsystem service associated with an interface operation provided by a component may be identified as a
RequestedService, which in turn is a subtype of Step, and may be refined by a BehaviourScenario.
B.3.3.3. Modelling with DAM
The DAM profile addresses the dependability modelling of RTES with UML. According to the Figure 10
DAM builds on MARTE and consists of a library and a set of extensions. The latter are the stereotypes of
the profile and most of them specialise the ones of MARTE. In the following we describe the main
elements of the DAM profile.
Figure 10. DAM profile overview.
B.3.3.3.1. DAM Library
The DAM library contains basic and complex dependability types. The following MARTE packages have
been used to define the DAM library:



The NFPs MARTE sub-profile for the definition of new basic dependability types.
The VSL MARTE sub-profile for the definition of complex dependability types.
The MARTE library, where the Basic NFPs types are imported in order to reuse them (both in the
definition of complex and basic dependability types).
B.3.3.3.2. Basic Dependability Types
DAM defines the following basic dependability types:


Simple enumeration types enable to characterise the threats according to the classification in [73]
and [74].
Data types are new NFP types obtained by specialising the NFP CommonType and NFP Real types
of the MARTE library. These new types inherit from their supertypes the following properties:
o expr: An expression in MARTE VSL.
o source: The origin of the specification. It can be estimated (e.g., a metric to be estimated),
required (e.g., a requirement to be satisfied), calculated or measured.
Copyright © 2015, DICE consortium – All rights reserved
30
Deliverable 1.1. State of the art analysis
o
o
statQ: The type of statistical measure (e.g., maximum, minimum, mean).
dir: The type of the quality order relation in the allowed value domain of the NFP, for
comparative analysis purposes.
B.3.3.3.3. Complex Dependability Types
Complex dependability types are MARTE tupleTypes characterised by basic NFPs, from the MARTE
library, and/or basic dependability types. They enable to characterise both from a qualitative and a
quantitative point of views, the threats (i.e., faults, errors, failures and hazards) and the mitigation solutions
(i.e., recovery and repair strategies). As for stereotypes, a complex dependability type for DAM is prefixed
by ‘Da’. Figure 11 depicts these types.
Figure 11. DAM types.
B.3.3.3.4. DAM UML extensions
The DAM extensions provide the domain expert with a set of stereotypes to be applied at model
specification level, i.e., the stereotypes necessary to represent the dependability system view in a concrete
UML model. DAM aims at providing a small yet sufficient set of stereotypes to be actually used in
practical modelling situations. The DAM stereotypes are: DaComponent, DaConnector, DaService,
DaServiceRequest,
DaStep,
DaErrorPropRelation,
DaFaultGenerator,
DaReplacementStep,
DaReallocationStep, DaActivationStep, DaAgentGroup, DaController, DaVariant, DaAdjudicator,
DaSpare and DaRedundantStructure. For a complete description of the tags for each DAM stereotyped
refer to [75].
B.3.3.4. Access Control Modelling with UML
Over the years a number of UML profiles have been defined to represent security-related properties of
systems. Since the focus of the DICE project for what concerns issues of data protection and privacy rests
on problems related to giving access to data and information only to components that have the appropriate
rights for that. In this brief section we analyse several approaches for the modelling of access rights through
Copyright © 2015, DICE consortium – All rights reserved
31
Deliverable 1.1. State of the art analysis
UML concepts. These approaches could offer the elements through which the DICE profile includes the
concepts that are suitable for the modelling of security issues in DIAs.
SecureUML [76] is a UML profile that can be combined with domain specific languages, to provide the
latter with the concepts necessary to describe the rights to perform certain actions (e.g., read/write data). In
particular, it captures the concepts underlying Role Based Access Control (RBAC) [77] through suitable
stereotypes. In particular, these stereotypes allow users to define what resources need to be protected,
which users can access them, and what actions the users can perform on the resources.
SecAM [78] is a UML profile based on MARTE and DAM. SecAM addresses access control as well as
cryptography, resilience and security mechanisms. The access control package supports the specification of
different access control policies: Mandatory Access Control (MAC), Discretionary Access Control (DAC)
and Role-Based Access Control (RBAC). SecAM addresses confidentiality, integrity and authorisation
issues. SecAM proposes a small set of stereotypes that include: a) who is accessing to the system
(SecaSubject stereotype); b) which objects should be protected for unauthorized access or modification
(SecaObject stereotype) and c) which operations a user wants to perform on the available objects
(SecaOperation).
UMLsec [79] enables to specify security relevant information during development of security-critical
systems. In particular, it considers RBAC as access control policy and provides a tool-support [80],
implemented as Eclipse plugin, for formal security verification.
B.4.
Modelling tools
This section overviews the most relevant UML Computer-Aided Software Engineering (CASE) Tools for
making a comparative analysis. The goal is to select an appropriate UML modelling tool for the DICE IDE.
The DICE vision imposes some constraints for the modelling framework we are developing. First, any tool
to be integrated in the IDE needs to be open source and free of charge. Second, the UML modelling tool
needs to import/export XML Meta-data Interchange (XMI) [81] for MARTE Profile annotated models, so
the tool needs to support UML2 and MDA. Third, the UML modelling tool should support the largest
number of UML diagrams. In the following, we list the UML tools we considered for our comparison.
More information about UML tools can be found here:


http://modeling-languages.com/uml-tools/
http://software-talk.org/blog/2014/05/comparison-of-free-uml-tools/
Enterprise Architect (EA) [82] is a commercial software, which means that it’s not suitable for DICE
IDE. However, it is an exceptional tool with a rich set of features. So we use EA as a baseline in our
analysis since it defines a comprehensive set of requirements that will be used for comparison with free
UML modelling tools.
Papyrus [83], UML2 Modeller is an open source tool. It is based on the Eclipse framework and it is
licensed under Eclipse Public License (EPL) [84]. Papyrus primary goal is to implement the complete
standard specification of UML2. Papyrus also provides an extensive support for UML profiles. It includes
hence all the facilities for defining and applying UML profiles. The MARTE profile is available in
Papyrus. An inconvenience we discovered is that the learning curve could be higher than for the other
tools. However, this tool is well-documented [83], [85]-[92].
Modelio [93] is an Open Source UML modelling tool developed by Modeliosoft [94]. It supports UML 2.0
and Business Process Model and Notation (BPMN) [95] standards. The core Modelio software was
Copyright © 2015, DICE consortium – All rights reserved
32
Deliverable 1.1. State of the art analysis
released under the GPLv3 (a viral license which is not suitable for DICE), although there is a commercial
version too. Modelio allows the import/export of UML models from/to other tools via UML XMI format.
Modelio enables the use of the MARTE profile [96].
MOdeling Software KIT (MOSKitt) [97] is a free CASE tool, built on Eclipse which is being developed
by the Valencian Regional Ministry of Infrastructure and Transport, namely through technology provider
Prodevelop, to support the gvMétrica methodology (adapting Métrica III to its specific needs). It supports
UML2, BPMN [95] standards and database diagrams. MOSKitt is released under EPL [84]. MOSKitt gives
a framework to build Model-to-Model (M2M) transformations based on Eclipse standards, as well as
import capabilities for external UML XMI [81] format models.
ArgoUML [98] is an open source and free UML modelling tool distributed under the EPL 1.0 [84].
ArgoUML is a Java-based application that is available in ten languages. ArgoUML also provides code
generation for Java [99], C++ [100], C# [101], PHP4 and PHP5 [102]. It also enables reverse engineering
from Java. External modules have been developed to complement ArgoUML in specific areas. They
provide generation of database schemas or code in other languages like Ruby [103] or Delphi [104].
However, ArgoUML offers support only for UML 1.4 diagrams, which is not enough for DICE.
StarUML [105], [106] is a UML tool licensed under a modified version of GNU General Public License
(GPL) [107] until 2014. A rewritten version (StarUML 2) was released in 2015 under a proprietary license.
StarUML 2 is compatible with UML 2.x standard and supports totally 11 kinds of UML diagrams: Class,
Object, Use Case, Component, Deployment, Composite Structure, Sequence, Communication, Statechart,
Activity and Profile Diagram. StarUML 2 stores models in a very simple JSON format.
UML Designer by Obeo [108] supports UML 2.5 models. It uses the standard UML2 meta-model
provided by the Eclipse Foundation [109]. Obeo is a free tool (Open Source with EPL license [84]) for
prototypes and starter projects but it is necessary to pay a fee for Small/Medium/Large or critical projects.
UML Designer is based on Sirius [110]. It provides an easy way to combine UML with domain specific
modelling. UML Designer includes a MARTE Designer too (Beta release) [111] which is a graphical tool
to edit and visualise MARTE models.
MagicDraw [112] is a commercial UML modelling tool with a free educational edition. It’s written in
Java, supports UML2 [109], SysML [113], BPMN [95] and UPDM [114]. It provides MDA and code
engineering mechanism (support for J2EE [70], C# [101], C++ [100], CORBA IDL programming
languages [65], .NET [71], XML Schema [115], WSDL [116]), as well as database schema modelling,
DDL generation and reverse engineering facilities. It supports large projects and is used at Netfective
Technology as the modelling component of the BluAge product [117].
IBM® Rational® Software Architect [118] is a comprehensive design, modelling and development tool
for end-to-end software delivery. It uses the Unified Modelling Language (UML) for designing enterprise
Java® applications and web services. Rational Software Architect is built on the Eclipse open source
software framework and is extensible with a variety of Eclipse plug-ins. You can also enhance functionality
for your specific requirements with separately purchased Rational extensions.
B.4.1
Analysis
Table 3 provides summary of the tools reviewed in this section.
Copyright © 2015, DICE consortium – All rights reserved
33
Deliverable 1.1. State of the art analysis
Table 3: UML CASE Tools summary.
Name
Creator
Enterprise Architect [82]
Sparx Systems [119]
Papyrus [83]
CEA [120], Atos [121]
Modelio [93]
Modeliosoft [94]
Windows (Supports
Linux & Mac
installation)
Windows, Linux
(Java)
Windows, Linux,
Mac
Conselleria de
Windows, Linux
Infraestructuras, Territorio
(Java), Mac
y Medio Ambiente [122]
Cross-platform
Tigris.org [123]
(Java)
Obeo Model Driven
Windows, Linux,
Company [124]
Macs
Windows, Linux,
No Magic, Inc. [125]
Mac
Windows, Linux,
Architect
IBM [126]
Mac
MOSKitt [97]
ArgoUML [98]
UML designer [108]
MagicDraw [112]
Rational
[118]
Open
source
Platform/OS
Software
Software
license
Programming
language used
No
Commercial
C++ [100]
Yes
EPL [84]
Java [99]
Yes
GPL and
Commercial
Java
Yes
EPL
Java
Yes
EPL
Java
Yes
EPL
Java
No
Commercial
Java
No
Commercial
Java
B.4.2
MARTE profile feature (import *.XMI)
As MARTE [68] is the most well-known UML profile for expressing quality characteristics of software
systems, we analyse the tools with respect to their capability of supporting such profile.
MARTE has been defined via a UML2 profile. Thus, UML tools able to support MARTE have to import at
least XMI with version of UML2. Currently, three open source tools are available for system modelling
using the MARTE profile: Modelio [93], Papyrus UML [83] and MARTE Designer (Obeo, but it is still in
Beta Status) [111]. ArgoUML [98] cannot import UML2, therefore it cannot support the MARTE Profile.
Table 4 provides an overview of all tools (open source and commercial) and of their level of support for
MARTE.
Table 4: CASE tools.
Name
Enterprise
Architect
[82]
UML2
Yes
Modelio
[93]
Yes
Papyrus
[83]
Yes
MOSKitt
[97]
Yes
ArgoUML
[98]
No
MDA
[57]
Yes
Yes
XMI
[81]
Yes
Languages generated
Can be
integrated
with
Forum
MARTE User
community
[68]
Manual
(30/03/2015)
Supports
MDA
templates
Yes
and Code
Generation
templates
ActionScript [127], C,
C# [101], C++ [100],
Delphi [104], Java, PHP
[102], Python [128],
Visual Basic [129],
Eclipse &
Visual Basic .NET
Visual
[130], DDL [131], EJB,
Studio [136]
XML Schema [115],
Ada [132], VHDL
[133], Verilog [134],
WSDL [116], BPEL
[135], Corba IDL [65]
Yes
Yes
Java, C++, C#, XSD,
WSDL, SQL
Eclipse,
EMF
Yes
Ada 2005, C/C++, Java
add ins
Eclipse
Yes
2/5
[141]
Post: 2966
Topics: 639
Users: 968
[147]
(Eclipse
Forum)
Eclipse
No
[142]
[148]
No
3/5
[143]
Post: 1457
Topics: 453
Users: 254
[149]
Unknown Yes
Yes
Templates
Yes
Unknown
M2M
and
Yes M2T
HTML, CSS, Java
generation
Yes
Unknown
C++, C#, Java, PHP4,
PHP5, Ruby [103]
Copyright © 2015, DICE consortium – All rights reserved
AndroMDA
[137]
5/5
[139]
Post: 108641
Topics:
28604
Users:
153438
[146]
4/5
[140]
34
Deliverable 1.1. State of the art analysis
UML
Designer
[108]
Yes
Yes
No
No
2/5
[144]
Users: 977
[150]
Eclipse
EMF,
NetBeans
[138]
Yes
3/5
[145]
Post: 9558
Topics: 3099
Users: 1295
[151]
Eclipse
No
-
Yes
Java
Eclipse
MagicDraw
[112]
Yes
Yes
Yes
Yes
Java, C++, C#, CIL,
CORBA IDL, DDL,
EJB, XML Schema,
WSDL
Rational
Software
Architect
[118]
Yes
Yes
Yes
Yes
Java
-
B.4.3
UML diagrams supported
Even though UML 2.0 has been standardised long ago, not all tools support all its diagrams. Table 5 shows
the diagrams that are supported by each tool.
Table 5: UML 2.0 Diagrams supported by modelling tools.
Structural UML diagrams
Class diagram
Component diagram
Composite structure
diagram
Deployment diagram
Object diagram
Package diagram
Profile diagram
Behavioral UML diagrams
Activity diagram
Communication diagram
Interaction overview
diagram
Sequence diagram
State diagram
Timing diagram
Use case diagram
Other diagrams
BPMN [95] diagram
SysML [113] diagram
Database diagram
UML
ArgoUML MagicDra
Designer
RSA [118]
[98]
w [112]
[108]
EA [82]
Papyrus
[83]
Modelio
[93]
MOSKitt
[97]
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
No
No
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
No
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
No
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
No
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
No
Yes
No
No
No
No
Yes
Yes
Yes
Yes
Yes
B.4.4
Summary
In summary the tools that appear to be suitable to be used in the DICE context are Modelio [93] and
Papyrus [83] since they have the appropriate licenses and support the UML Profile mechanism. However,
the Modelio plug-in for MARTE is not supported anymore and it does not work with the last version of
Modelio designer.
The last version of Papyrus (1.1) was released in June 2015 with Eclipse Mars [152]. This version brings
new features and improvements but also tackles some known issues related to performance and userfriendliness. In particular, in a well-filled containment tree, especially a developed tree, the graphical user
interface works smoothly as long as the size of the model does not becomes excessively large, i.e., more
than 2000 visible nodes.
Copyright © 2015, DICE consortium – All rights reserved
35
Deliverable 1.1. State of the art analysis
B.5.
Tools for model to model transformations
The purpose of this section is to evaluate tools that transform software models, described using non-formal
or semi-formal languages, into formal models. The goal of the target formal model is to perform analyses
of the non-functional properties of the software system, such as performance, reliability, availability, safety
or privacy. To this end, we have reused the evaluation framework proposed by [153]. This work also
reviewed some tools. Here we have added to the original evaluation some new tools that are of interest to
DICE. The evaluation framework proposed in [153] is summarized in the Table 6:
Table 6: Summary of the evaluation framework for model-to-model transformation tools [153].
1
#
Characteristic
Description/question
1
model specification
Does the tool support specification of systems as graphical models? {Yes/No}
2
graphical notation for
model transformation
Does the tool support graphical specification of transformation? {Yes/No}
3
lexical notation for model
transformation
Does the tool support lexical specification of transformation? {Yes/No}
4
model-to-model
transformation support
Does the tool support model-to-model transformation? (e.g., from one UML model to
another?) {Yes/No}
5
model-to-text
transformation support
Does the tool support model-to-text transformation, such as generation of source code?
{Yes/No}
6
support for model analysis
Is there any support for model analysis? {Yes/No}
7
support for Quality of
Service (QoS)
management
Is there any support for managing QoS during model specification and transformation?
{Yes/No}
8
meta-model-based
Is the tool based on explicit descriptions of the meta-models of source and target
model? {Yes/No}
9
MOF1 integration
Is the tool integrated with a MOF1 (or other meta-model-based repository)? {Yes/No}
10
XMI integration
Is the tool integrated with XMI1? {Yes/No} which version(s) of XMI is supported?
{list of versions}
11
based on UML
Is the tool based on UML models as source and/or target models for transformation?
{Yes/No}
12
UML specification
Does the tool provide support for UML modelling {Yes/No}
13
UML tool integration
Can the tool be integrated with existing UML tools? Either directly, as active plug-ins
in UML tools, or indirectly through model exchange via, e.g., XMI? {Yes/No}or{names
of the set of techniques}
14
iterative and incremental
transformation support
Does the tool handle reapplication of transformation after model updates? {Yes/No}
15
Bidirectional
transformation
16
traceability
Does the tool handle traceability of transformations, i.e., can it maintain traces of the
source and targets of a transformation?
{Yes/No}
17
DSM language support
Is there support for defining domain-specific modelling languages (e.g., UML
profiling) and DSM transformations? {Yes/No}
Does the tool support bidirectional transformations? {Yes/No}
MOF - Meta Object Facility. XMI - XML Metadata Interchange [81]
Once presented the evaluation framework we apply it to the tools of interest.
Copyright © 2015, DICE consortium – All rights reserved
36
Deliverable 1.1. State of the art analysis
B.5.1
Palladio Component Model
The Palladio Component Model (PCM) [154] captures the software architecture with respect to static
structure, behaviour, deployment/allocation, resource environment/execution environment, and usage
profile. In the PCM software is described in terms of components, connectors, interfaces, individual service
behaviour models (so-called Service Effect Specifications, SEFF), servers, middleware, virtual machines,
network, the allocation of components and servers, models of the user interaction with the system etc.
Overall, the PCM captures multiple views of software systems including elements which affect the extrafunctional properties (e.g. performance, reliability etc.) of software systems [154]. Table 7 presents
evaluation framework from Table 6 applied to Palladio Component Model.
Table 7: Evaluation framework from Table 6 applied to Palladio Component Model.
1
Yes, tool called ‘PCM-Bench’, which enables software developers to create instances of the PCM meta-model
2
No
3
No info given
4
Model-to-model transformation from a PCM instance to an SRE instance are performed with Java
5
Yes, model-to-text transformation based on the openArchitectureWare (oAW) framework generates code
skeletons from PCM model instances. The implementation uses either Plain Old Java Objects (POJOs) or
Enterprise Java Beans (EJBs) ready for deployment on a J2EE [70] application server
6
Model validation by checking OCL [72]constraints
7
Yes
8
Yes, the PCM is a meta-model designed to describe component-based software architectures in order to
analyse performance properties
9
Yes
10
Model instances can be serialised to XMI-files
11
Yes
12
No
13
No
14
Yes
15
No
16
No
17
No
B.5.2
VIATRA2
The main objective of the VIATRA2 [155]-[157](VIsual Automated model TRAnsformations) framework
is to provide a general-purpose support for the entire lifecycle of engineering model transformations
including the specification, design, execution, validation and maintenance of transformations within and
between various modelling languages and domains.
Table 8 presents evaluation framework from Table 6 applied to VIATRA2 model transformation tool.
Copyright © 2015, DICE consortium – All rights reserved
37
Deliverable 1.1. State of the art analysis
Table 8: Evaluation framework from Table 6 applied to VIATRA2 framework.
1
Yes. Models and meta-models are all stored uniformly in the VPM model space, which provides a very
flexible and general way for capturing languages and models on different meta-levels and from various
domains (or technological spaces).
2
No
3
Yes, VTCL transformation language [155]
4
Yes, Intra model transformations and Inter model transformation, both of these transformation categories are
supported by the transformation language of the VIATRA2 framework.
5
Yes, VIATRA2 supports mode-to-code generation in different ways
6
[155]
7
Yes
8
Yes, standard metamodelling paradigms are integrated into VIATRA2 by import plugins.
9
Yes
10
Yes
11
Yes
12
Yes
13
Yes
14
Yes
15
No
16
Yes
17
Yes
B.5.3
UML transformation tool
UML Transformation tool (UMT) [158] is a tool to support model transformation and code generation
based on UML models in the form of XMI [81]. This is a generic tool, so it does not provide models in a
concrete formalism, but the environment for obtaining them. Table 9 presents evaluation framework from
Table 6 applied to UML transformation tool.
Table 9: Evaluation framework from Table 6 applied to UML transformation tool.
1
No. There is no support for specifying models in UMT. It relies entirely on imported models from UML
tools
2
No. There is no graphical notation for model transformation
3
Yes. UMT uses XSLT and Java as transformation languages, with possibility of extending to support
other languages
4
No
5
Yes. Model-to-text transformation is the main functional domain for UMT
6
No. There is no support for model analysis, except for very simple support for checking of a model’s
conformance to simple profiles
7
No. There is no support for management of QoS
Copyright © 2015, DICE consortium – All rights reserved
38
Deliverable 1.1. State of the art analysis
8
No. UMT only targets the UML meta-model and is not flexible with respect to changing this
9
No. There is no integration with MOF
10
Yes. UMT imports UML/XMI files from different UML tools
11
Yes
12
No. There is no support for specifying UML models. UMT relies wholly on model input from external
UML tools
13
No. There is no direct UML tool integration. Integration is indirect through XMI
14
There is lightweight support for regenerating code without overwriting previously generated and modified
code
15
No. There is no direct support for bidirectional transformation. However, there is some support for reverse
engineering of code to XMI models
16
No
17
The tool does not provide support for defining DSM languages. It provides support for transformations of
DSM languages. E.g., transforming one DSM-based model to another DSM-based mode
B.5.4
Other tools
The following tools have poor information and we could not fill the evaluation framework. So we provide a
brief description of them.
B.5.4.1. CARiSMA
The CARiSMA [80] core is independent from any particular modelling language. It is just based on the
Eclipse Modelling Framework (EMF). CARiSMA enables compliance analyses, risk analyses, and security
analyses of software models. A flexible architecture makes CARiSMA extensible for new languages and
allows users to implement their own compliance, risk, or security checks.
B.5.4.2. UPUPA (fUML and Profiles for Performance Analysis)
Upupa [159] considers non-functional properties of a software system early in the development process.
UPUPA develops a model-based analysis framework based on the Foundational Subset for
Executable UML Models fUML [160] for enabling the implementation of model-based analysis tools. This
framework enables to carry out model-based analysis of non-functional properties of a software system
based on runtime information in the form of traces obtained by executing UML models using the fUML
virtual machine. Therefore, the framework integrates UML profile applications with execution traces to
enable the consideration of additional information captured in profile applications in the model-based
analysis as required for instance in performance analysis.
B.5.4.3. QVT and Related Technologies
QVT stands for Query-View-Transformations [161], a generic, all-purpose model transformation and
manipulation language standard that goes along with Model-Driven Architecture (MDA) standard [57] to
support its purposes and intent. QVT is an OMG [58] standard and, in essence, it defines a standard way in
which model transformation shall take place, using standard elements and operational transformation
behaviour. QVT defines standard views in which model information can be presented and manipulated, e.g.
in order to migrate a Platform-Independent Specification into a Platform-Specific one. The most widely
known technology related to QVT resides within the eclipse foundation model-manipulation environment,
i.e. AMMA [162]. AMMA features Atlas Transformation Language (ATL) [163], which is a QVT-like
model transformation language, with its own abstract syntax and environment. The transformation is itself
Copyright © 2015, DICE consortium – All rights reserved
39
Deliverable 1.1. State of the art analysis
a model conforming to a specific meta-model. This, for example, permits the creation of higher order
transformations, i.e., transformations that produce ATL transformations.
B.5.5
Summary
DICE will need to implement Model-to-Model (M2M) transformations extensively. First at UML level,
from the DPIM (DICE Platform Independent Model) to the DTSM (DICE Technology Specific Model),
and from the latter to the DDSM (DICE Deployment Specific Model). Second, on the DDSM an M2M
transformation will be useful for yielding the TOSCA deployment (see Section B.6). Third, each UML
diagram in the DIA design needs to be translated into the corresponding target formal model, which
comprises a combination of formalisms and UML diagrams at different abstraction levels (DPIM, DTSM
and DDSM). The third kind of M2M transformation needs to take into account the DICE stereotypes and
tags for parameterising the formal models and for extracting the quality requirements expressed as SLAs.
However, from the analysis of tools we carried out in this section we found the following conclusions.
First, there is no tool that can be reused for M2M transformations in the context of DICE. Second, the
framework in [153] should guide the development of the M2M DICE transformation tool, which means to
try to develop a tool able to answer ‘Yes’ to as many as possible of the framework questions.
B.6.
Deployment modelling with TOSCA
TOSCA stands for ‘‘Topology and Orchestration Specification for Cloud Applications’’ [164]. In the hands
of TOSCA lies the state of the art for deployment solutions that are both technology independent and multicompliant. This intrinsic characteristic stems from the joined interplay within which TOSCA was originally
specified, i.e., the OASIS standardisation effort. Within the OASIS TOSCA Technical Committee (TC) big
industrial players (e.g., IBM, Huawei, Ericsson) defined the essential elements for the purpose of providing
easily deployable specifications for cloud applications in all its aspects, including, but not limited to,
Network Function Virtualisation, Infrastructure Monitoring and similar. Essentially, quoting from the
TOSCA specification 1.0 [164], ‘‘TOSCA [...] uses the concept of service templates to describe cloud
workloads as a topology template, [...]. TOSCA further provides a type system of node types to describe the
possible building blocks for constructing a service template, as well as relationship type to describe
possible kinds of relations’’. Figure 12 outlines the essential concepts within TOSCA and their respective
relation:
Figure 12. Main TOSCA concepts and their relations [164].
Currently, the working group around TOSCA is focusing on the following key activities:
Copyright © 2015, DICE consortium – All rights reserved
40
Deliverable 1.1. State of the art analysis
(a) Interoperation and Industrial Adoption - this activity is pursued by working on two fronts: (1) from a
more ‘‘soft’’ perspective, the group has enacted a fine-grained demoing strategy involving conferences and
practitioner events (e.g. Open Source CONvention OSCON 2015 [165]); (2) from a more technical side,
the group is working on providing additional constructs needed within industrial practice and ad-hoc node
specifications (e.g., ‘compute’ and ‘store’ nodes, etc.). Most of the above activities are carried out by the
TOSCA-TC SubCommittee (SC) on interoperability [166]. Currently the SC is concentrating on
conducting adoption, interoperation and validation experiments with industry partners.
(b) Network Function Virtualisation (NFV) - this activity is pursued by defining concepts and relations
connected to Software-Defined Networking and ad-hoc TOSCA constructs that may be compatible with
said technology. Also, in the scope of these activities the TOSCA-TC is working to define ad-hoc design
patterns (e.g. ad-hoc required and provided properties within topology templates) that match the reasoning
and logical assumptions behind NFV.
(c) Language Simplification - this activity is pursued by incrementally refining a simplified TOSCA reinterpretation in YAML (YAML Ain’t Mark-up Language) [167] - a human-readable mark-up language
capable of sensibly reducing the learning curve behind TOSCA. The YAML specification is currently at
revision 15 and has not reached stability yet.
(d) TOSCA Marketing and Dissemination - in the scope of dissemination TOSCA has defined a specific
sub-committee was entrusted with the investigation of containment technologies (e.g. Docker, see Section
B.7.4) and their relation to TOSCA, with particular focus on how such technology can be used to improve
TOSCA notations and their value proposition. TOSCA defines its own Cloud Service ARchive (CSAR)
containment technology in an abstract way, e.g. to encompass Docker or similar technological solutions. In
this regard, TOSCA can be said to profile current containment technology. This group is interoperating
strongly with an ad-hoc containment group taking care of the CSAR area of TOSCA.
In the scope of the definition of the DICE profile, TOSCA and TOSCA-ready specifications play a key role
in enabling designers at the DDSM level to realise seamless auto-generation of TOSCA-ready templates.
These templates can be in turn deployed on TOSCA-enabled orchestration engines (see Section B.7).
B.7.
Deployment tools
B.7.1
Overview
Setting up virtual machines and bare-metal machines, management of network connectivity, installing and
configuring software are inevitable parts of operating with computation centres and clouds. Performing
these tasks manually is an involved and error-prone activity, which also scales badly. Over the decades the
IT industry advanced far enough to enable most if not all of these activities to be performed automatically
through scripting.
Scripting has an advantage of making all processes repeatable. However, scripted solutions are only
applicable to the platforms (operating systems, kernel versions, installed libraries etc.) that they have
originally been built for. To abstract specifics of operating systems and their distributions, tools exist,
which offer the configuration described in a Domain-Specific Language (DSL). This frees the user to focus
on describing the nodes leaving the distribution-specific operations to the tool’s drivers (also called
providers in some of the tools).
B.7.2
Configuration management
The configuration management tools embody the automation, normally focusing on management of
existing nodes. They run stand-alone on the node that they need to configure or use a central service to
Copyright © 2015, DICE consortium – All rights reserved
41
Deliverable 1.1. State of the art analysis
obtain the instructions on the node configuration. The supported DSL enables expressing common
operations (e.g. install a package, create a file, run a programme) in a simple way, while at the same time it
is powerful enough to build and maintain complex applications running across one or more clusters. The
popular representatives are the following:





Puppet – configuration management solution [48]
Chef – automation tool [49]
SaltStack [168]
Ansible – automation tool for application deployment [169]
Vagrant – tool for configuring development environments [170]
These tools have reached maturity, and they also have a large community of developers providing support
and development of the base tools. The community of those supplying the recipes is even wider, thus the
public repositories contain at least one Chef recipe for each of the core DICE building blocks, which are
freely available for installing in private data centres.
B.7.3
Orchestration
The orchestration in computing [171] represents the next level of the automation. It manages and connects
services into applications and building workflows. It employs automation provided by configuration tools,
but it runs tasks in a particular order or at particular events, e.g., when a database engine starts and becomes
available, it configures all the depending services and runs them. This is important for applications, which
normally run in multiple tiers. The orchestration tools have a notion of the application’s topology. They
also play an important role when scaling parts of the application up or down, and when healing the
application after faults.
DICE deployment tools act very much like orchestration tools, receiving a TOSCA model at the input and
deploying the application based on the contents in TOSCA. Here are some of the more prominent and
popular representatives dealing with orchestration.
B.7.3.1. Ubuntu Juju
Juju is the offering for orchestration by Canonical [172].. It introduces and uses Juju charms [173] to
represent and handle individual nodes and services. It provides a rich command-line interface, mostly
mirrored by the functionality in its web interface console. The charms are structured bundles of files,
composed of YAML files [167] to describe the charms and specify actions, scripts and other files required
in the charm. The automation in the charms revolves around hooks and actions. Hooks define what happens
when a service installs, starts, stops, changes configuration or gets upgraded. Additional hooks handle
joining, changing, departing and breaking of relations with other charms. Actions further extend the
functionality of a charm’s instance.
A bundle is a YAML file describing the whole application, listing all the charms, their relationships and
configuration values.
Charms are available for the core Big Data services, which Canonical openly promotes [174]. A growing
number of other data-related charms exist for other services as well.
B.7.3.2. Cloudify
Cloudify is a cloud application orchestrator [175] supporting a wide variety of platforms and
infrastructures. It accepts blueprints formatted in a TOSCA-compliant YAML [167].
Copyright © 2015, DICE consortium – All rights reserved
42
Deliverable 1.1. State of the art analysis
Because of its reliance on Chef, Puppet and other configuration tools, the Big Data building blocks are well
supported. The author’s own words are that their original goal was ‘to make Big Data deployments a firstclass citizen within Cloudify’ [176]. Cloudify seems as a consolidated technology with strong points in
modelling and deployment of cloud specs. These might come in handy during DICE, especially in terms of
their usage as a case-study or TOSCA usage scenario.
B.7.3.3. Alien4Cloud
Application LIfecycle ENablement for Cloud [177] aims to help in designing applications and
collaboration during the design by employing the TOSCA YAML documents for describing applications.
They offer a Graphical User Interface (GUI) for managing blueprints (called topologies) and monitoring
deployment progress. One of the core functionalities behind this technology is, quoting from the home site:
‘Create or reuse portable TOSCA blueprints and components. Leverage your existing shell, Chef or Puppet
scripts’ [177]. These features suggest that Alien4Cloud may easily be integrated in the various
methodological and technological phases intended in the definition and application of the DICE profile.
The tool relies on third party tools such as Cloudify [175] for the actual deployment. The project is in the
early stages of the initial releases.
B.7.3.4. Apache Brooklyn
The Apache Brooklyn [178] is another project using blueprints for describing applications. Its aim at the
orchestration is to enable also the application’s runtime management with the ability to express SLA-like
policies and associated actions. The policies rely on the metrics from the monitoring services, which
Brooklyn supports. The actions can only be expressed in Java Virtual Machine (JVM). Currently Brooklyn
supports OASIS CAMP [179], but they also plan to support TOSCA.
B.7.3.5. Flexiant Cloud Orchestrator
Flexiant Cloud Orchestrator (FCO) [180] is a world-leading Cloud Orchestration Software solution. FCO
provides service providers the ability to design, create and manage their own virtual public, private or
hybrid cloud solutions.
With Flexiant Cloud Orchestrator, a data centre operator can manage an entire cloud solution, from
hardware, network and storage management through to metering, billing & customer/end-user self-service.
The FCO uses Chef recipes [49] for managing the configurations, and therefore supply the Big Data
building blocks.
B.7.3.6. Rundeck
Rundeck [181] is an open source software that is designed to automate operational procedures in cloud
environments. Rundeck allows tasks to be run on any number of nodes/Virtual Machines (VMs) using a
GUI or command line interface. It does this by working in tandem with solutions such as Chef [49] or
Puppet [48] and acts as a command and control portal that lets users execute commands using features like
node filtering and parallel execution.
B.7.3.7. CAMF
CAMF [182] focuses on three distinct management operations, particularly application description,
application deployment and application monitoring. To this end, it adopts the OASIS TOSCA open
specification for blueprinting and packaging Cloud Applications. Being the part of the Eclipse Software
Foundation, part of the CAMF code will be made freely available and open source under Eclipse Public
License v1.0 [84].
Copyright © 2015, DICE consortium – All rights reserved
43
Deliverable 1.1. State of the art analysis
B.7.3.8. CELAR
A relevant background is also the CELAR project [183] and the related tool-support within Eclipse, i.e., cEclipse. The CELAR project is an initiative specific for multi-cloud elasticity provisioning. In realising
said elasticity provisioning services, CELAR and connected tool-bases are working to implement and
gradually extend a deployment engine featuring specific TOSCA templates. As part of the Eclipse
ecosystem the complete source code of c-Eclipse is made available under the terms of the Eclipse Public
License v1.0 [84]. Similarly, a part of the CELAR project code responsible for automated deployment will
also be made freely available as well.
B.7.3.9. Open-TOSCA
Open-TOSCA is an open source initiative from the university of Stuttgart to develop open source TOSCA
modelling/reasoning and orchestration technologies including support for modelling via the Winery
modelling technology [184] as well as TOSCA containment modelling via an ad-hoc OpenTOSCA
Container and instantiation via the VinoThek self-service instantiation portal [185]. Because it is composed
of a set of technologies, Open-TOSCA does not have a clear and homogeneous open source licensing
model as a single product. Rather, individual licensing has to be evaluated for the single modules it is made
of.
B.7.3.10. Tools analysis
Table 10 provides comparative summary of deployment orchestration tools described in detail in the
sections B.7.3.1-B.7.3.9.
Table 10: Comparative summary of deployment orchestration tools.
Tool name
License
Ubuntu Juju AGPL
[186]
[172]
Cloudify
[175]
Input
format
Configuration
support
Native cloud
support
Command
Juju Charms [173] OpenStack,
line, YAML - hooks and actions [192], ...
[167]
run executables in
target environment
Apache
TOSCA
License
YAML
License 2.0
[187], UI
commercial
(currently
free)
Scripts, Chef [49],
Puppet [48] ,
SaltStack [168],
OpenStack API
[188], CloudStack
[189], custom
plugins
interactive
see Cloudify
Alien4Cloud Apache
License 2.0 GUI; YAML
[177]
First
release
Latest
release
AWS at least 2
years ago
May 2015
(recent)
OpenStack,
SoftLayer [193],
Apache CloudStack,
VMware vSphere
[194] and vCloudAir
[195]; plug-ins for
AWS,...
February
2012
(3 years
ago)
see Cloudify
Q1 2015
(in
progress)
‘many supported’,
leverages Apache
jclouds [196]
January
December
2013 (2 2014
years ago) (4 months
ago) - version
0.7.0
internally
(via
Cloudify
[175])
Apache
Brooklyn
[178]
Apache
CAMPChef, SaltStack,
License 2.0 compliant
scripts
YAML[167]
+ JVM plugins, RESTful
API
Copyright © 2015, DICE consortium – All rights reserved
December
2014
(4 months
ago)
44
Deliverable 1.1. State of the art analysis
FCO
Flexiant
custom
Cloud
license
Orchestrator
[180]
FCO
2007
-Hypervisors
supported
Virtuozzo
Xen 4
KVM
VMware vSphere
Hyper-V 2012 [197]
April 2015
(recent)
Chef, Puppet,
Jenkins (see
Section B.8.2)
Using configuration
frameworks any
cloud platform can
use.
2010
April 22 2015
V2.5
TOSCA
Public
XML (GUI
License 1.0 in Eclipse)
[84]
Chef
‘many supported’,
leverages
Apache
Java
Multi-Cloud
Toolkit
(jclouds)
[196]
under
developm
ent,
started
2013
not
yet
released
CELAR
[183]
Apache
TOSCA
License 1.0
Unknown [183]
CELAR Serverbased (for the
moment) but
environment is
under definition
under
developm
ent,
started
2013
not
yet
released
OpenTOSCA
[184], [185]
No license
Implementation
OpenStack, AWS
Artefacts Engine
supporting plugins
April
2015 (1
month
ago)
April 2015
(1
month
ago)
Rundeck
[181]
GUI,
Chef, FCO
SOAP/REST Blueprints [190],
API
FCO Triggers
[191]
Apache
GUI,
License 2.0 Command
Line
CAMF [182] Eclipse
TOSCA
B.7.4
Virtualisation and containers
Orchestration tools use description of services and applications to be deployed at a high level, but they need
to handle the actual deployment and configuration also at the low level. The possibilities of where and how
to instantiate the needed building blocks are only limited by the target environment's support and the
drivers included in the orchestration and configuration tools. Currently the most common means of
provisioning and deploying resources is to create and use virtual machines in an IaaS environment.
Virtualisation enablers are the hypervisors such as KVM, Xen or VMWare [197]. Newer hypervisors aim
to offer light-weight, simplified or faster execution from the traditional ones. A peer H2020 project
MIKELANGELO [198] is going to produce a faster, lightweight software stack for virtualisation under
widely adopted and supported OpenStack [188].
The increasingly popular alternative to virtualisation is using containers. The containers rely upon the Unix
container technology which has been available for a long time, but have only recently received a good
support for management and portability of the applications within containers. Docker [199] and LXC [200]
are only two of the representatives. Docker in particular offers to package a service or an application in a
container, creating a lower footprint on the overall execution environment and increasing the portability of
the application.
The concept opposite to the virtualisation and containers is using the whole bare-metal computer as a
computational unit (MaaS - Metal as a Service). This is possible through standards such as Intelligent
Platform Management Interface (IPMI) [201].
Copyright © 2015, DICE consortium – All rights reserved
45
Deliverable 1.1. State of the art analysis
In DICE we aim to support the ability to deploy the building blocks first in the more popular environment such as the
virtualisation in the IaaS. For the more advanced releases, we will also look into offering an ability to deploy parts of
the applications using Docker containers and the OSV [202] virtualisation.
B.7.5
Summary
In DICE we are aiming to support standards, and the OASIS TOSCA is an important one. We will base our
solution on one of the more stable and powerful TOSCA-compliant orchestration tool. At the time of the
analysis, Cloudify [175] is a tool which has been available for several years, while its support and
development are still strong. Its command-line and RESTful interface is perfect for our use, where the users
never need to see the orchestrator directly. Also its reliance on jclouds [196] as the abstraction of the IaaS
and the extendibility enable its potential use in a variety of cloud providers, including the Flexiant Cloud
Orchestrator [180] as the dedicated testbed in the DICE project. The choice is further enforced because
another solution - Alien4Cloud [177] - uses it as a basis. Alternatively, Apache Brooklyn [178] promises to
add support for TOSCA in the upcoming months, making it also worth considering.
For the low-level configuration management, we plan to use Chef [49]. The competition at this level is
high, but Chef stands out because the DICE developers have a higher familiarity with Chef, and there is a
potential for reusing pre-existing cookbooks and recipes.
B.8.
Continuous integration tools
The software engineering practice where the developers merge their development changes into the shared
mainline daily or even more frequently has been named Continuous Integration (CI) [203], [204]. The
practice enables both a higher rate of software releases as well as a greater confidence in the quality of the
produced code. The latter also depends on a well-built development and testing environment, and on the
developers adhering to the test-driven approaches of the development. The overall idea is that the
developers perform code and project validation with every change, working towards the code which does
what is expected from it, while at the same time it does not break any other parts of the application. By
often integrating changes introduced concurrently by different members of the team, the developers find
and resolve conflicts while they are smaller and easier to resolve. continuous integration is therefore an
important element of the Application Lifecycle Management and thus crucial to the DevOps ecosystem.
The code validation takes many forms, from preparing and running unit tests at the small scale to
integration tests to test multiple modules, services and even systems. This can easily include assessment of
the quality of each deployment, giving both a binary response (pass/fail) as well as softer ones to help drive
the development and the improvements. In DICE we have set out to offer the tools to perform such
evaluation, and while it will be possible to execute them manually, the native way of using them in DICE
will be via the continuous integration.
The tools implementing the continuous integration handle one or more jobs, each of which can be triggered
manually, periodically or by some service or another job. The jobs typically follow a pattern of obtaining
the latest code changes from a Version Control System (VCS) such as Git (see Section B.9.1.3) or
Subversion (SVN) (see Section B.9.1.2), saving them in their local space. Then they attempt to build the
project and save the output of the successful builds (binaries, executables, libraries etc.) in an artefact
repository. Finally, they run any verification provided by the developers with the code, indicating whether
an application functions as expected or fails at any point.
The process pipelinedescribed above issues and indication when the project baseline breaks. Certain early
criticisms of the continuous integration approach [205] pointed out that this indication only happens after
the fact, halting the development process for everyone involved. The side effect of this is that the
developers approach commits under stress and fear of breaking the build, while at the extreme case the
Copyright © 2015, DICE consortium – All rights reserved
46
Deliverable 1.1. State of the art analysis
developers start to ignore the failure notifications from the continuous integration tools. These problems
can largely be avoided by employing code review steps where the tools verify the merged commit before
the actual merge.
The continuous integration jobs can also perform fully custom actions and steps, so any project team is free
to follow their own patterns and use their own tools when building and testing their application.
Here we review the tools with active development and recent latest releases. A more comprehensive
overview is available at the ThoughtWorks website [206], but the information there is partially outdated.
B.8.1
TeamCity
TeamCity is a part of the JetBrains’ commercial offering in their Teamware suite [207]. It natively supports
Java, .NET [71], Ruby [103] and XCode [208] languages and environments. Many other languages are
supported via plug-ins.
TeamCity supports a remote run feature which performs a check of a build before it is committed into the
baseline branch. The feature works from a TeamCity plug-in in Eclipse or other IDE and does not involve
the branches in VCS.
The tool also natively supports many build agents, code coverage tools and a code change inspector for the
Java code projects. Build agents can run distributed in various locations and host environments to address
load distribution and specific requirements of multi-platform builds.
B.8.2
Hudson / Jenkins
Jenkins [209] is a highly popular open source CI tool. It is a fork of the Hudson [210] tool, and both still
exist and are actively developed, although Jenkins reportedly has a larger developer community and a
higher installation base [211]. Originally the tool had a high level of support mostly for the Java-centric
projects. However, currently a wide selection of plug-ins provide support for other types of projects as well
(e.g. Python [128], Ruby [103] etc.). The plug-ins can extend the default functionality for the majority of
aspects of the CI jobs, including the build steps and the means of sending job result notifications. It is
possible to publish the plug-ins in a Jenkins plug-in directory.
Both Hudson and Jenkins are created in Java [99]. Therefore, they can be installed in any operating system
or Linux distribution. Each installation can be a fully-featured installation or a small slave agent installation
used in a distributed build agent set-up.
B.8.3
Atlassian Bamboo
Atlassian’s Bamboo [212] is a commercial offering which works as a standalone service, hosted in the
Atlassian’s cloud or installed in the customer-controlled environment. It natively includes integration with
the JIRA issue tracking service [213], and the Stash code versioning management system [214]. Bamboo
emphasises the support for continuous delivery, providing releases in multiple environments. The
commercial nature of the project is also evident in solutions easing the continuous integration management
with grouping jobs in chained stages, simplified management of distributed version control systems’
branching, as well as general support for migrating from popular open source tools such as Jenkins [209].
Nevertheless, it offers RESTful API to support custom add-ons.
B.8.4
Go
An open source continuous integration and delivery tool from ThoughtWorks called Go [215] focuses on
handling complex pipelines, chaining jobs according to their dependencies in a directed graph with possible
Copyright © 2015, DICE consortium – All rights reserved
47
Deliverable 1.1. State of the art analysis
fan-out and fan-in. Naturally it also supports integration with other ThoughtWork projects such as the agile
project management tool Mingle [216].
B.8.5
Strider CD
The Strider CD [217] is a relatively recent solution for continuous integration and Continuous
Development. It is built on lightweight technology such as node.js [218] and promises a high level of
customisability by supporting the plug-ins. The authors focus on improved user experience and automation.
VCS branches can have different jobs attached to them.
B.8.6
BuildBot
The BuildBot [219] is a framework for continuous integration emphasising the flexibility and providing the
tools for complex projects which mix technologies and languages. The jobs are configured using Python
scripts [128]. This means they can be simple, but, if needed, provide the ability to dynamically configure
builds and jobs. In the complex software projects it is possible to use a concept of source stamping to
include dependencies from various VCS projects, and the built-in versioning system helps manage the
dependencies.
B.8.7
CircleCI
The CircleCI [220] is a hosted environment for continuous integration. It advertises the speed of the job
execution. This is achieved by the timing of the builds and their subsequent distribution into parallel builds
based on this timing. Each job gets its own Docker environment [199], so each build is clean and
independent of any previous builds. Natively it also provides support for virtual graphical frame buffer
using Xvbf [221] in Linux. Table 11 provides comparative summary of continuous integration tools
described in details in the sections B.8.1-B.8.7.
Table 11: Comparative summary of continuous integration tools.
Eclipse
plug-in
Supports
custom
plug-ins
commercial
with a limited
free license
Yes
Yes
Web GUI,
RESTful
July 2006
(9 years ago)
December 2014
(6 months ago)
Hudson [210] Eclipse EPL
license [84]
Yes
Yes
Web GUI,
RESTful
Summer 2004
(11 years ago)
January 2015
(4 months ago)
Jenkins [209] MIT License
[222]
Yes
Yes
Web GUI,
RESTful
February 2011 (4 May 2015 (recent)
years ago) forked from
Hudson
Atlassian
Bamboo
[212]
commercial
Yes
Yes
Web GUI,
RESTful
February 2007 (8 November 2011
years ago)
(6 months ago)
CircleCI
[220]
commercial,
hosted only
No
No
Web GUI,
RESTful
unknown
recent
ThougtWorks Apache
Go [215]
License 2.0
[187]
N/A
Yes
Web GUI,
RESTful
7 years ago
April 2015
(1 month ago)
Strider CD
[217]
BSD License
[223]
No
Yes
Web GUI,
RESTful
July 2013
(2 years ago)
March 2015
(3 months ago)
BuildBot
[219]
GPL 2
No
Yes
Web GUI,
RESTful
March 2006
(9 years ago)
April 2015
(2 months ago)
Tool name
TeamCity
[207]
License
Copyright © 2015, DICE consortium – All rights reserved
Job control
First release
Latest release
48
Deliverable 1.1. State of the art analysis
B.8.8
Summary
DICE DevOps support tools naturally aim to support the continuous integration and Continuous
Deployment. The tools used will represent the glue between the development and simulation work on the
one side, and automated deployment and application’s execution on the other side. The selection is wide
and strong, but DICE needs to offer an open source solution. Of the ones that comply with this requirement
and have also a strong community, active support and a high level of adoption, Jenkins [209] is certainly at
the top. It is also the solution favoured by many DICE partners for their own internal or collaborative
projects. As a good alternative, BuildBot [219] also offers a good support for highly customised solutions.
The commercial solutions will likely serve as an inspiration in terms of the features to consider and support
in the DICE continuous integration. For instance, the CircleCI’s ability to provide headless testing of
graphical interfaces (as required by the Selenium library (see Section C.11.2.3) for testing the web GUI
applications) may become useful for providing quality tests.
B.9.
Versioning of software engineering artefacts.
The purpose of versioning is to map a complex system of software components to a commonly and easily
understood name or number. This helps the users understand which functionality to expect from a certain
component’s version. Considering that few systems operate on their own, the versioning also helps in
defining which components are compatible and possible to co-operate. The actual version assignments are
ultimately the responsibility of the developers and their project leaders. No system can perform version
assignment in a fully automated way. It is also highly dependent on the purpose and type of software being
versioned [224]. However, it is a common and recommended practice to use a system which automatically
assigns revisions to each change.
DevOps methodology advocates DevOps teams to version everything in their environment: application
code, infrastructure, configuration, data, and internal system artefacts. The major aspect of the systems
providing the versioning control is that they provide history of changes in the code and the ability for the
collaborating teams to obtain a consistent view of the whole project at any time. In combination with
continuous integration it is possible to also always be able to obtain a tested and stable version from the
change history.
B.9.1
Classical versioning tools
The versioning control systems - more accurately named revision control systems or source control systems
[225] - are roughly divided into two categories: centralised systems (represented by the CVS [226] and the
Subversion [227]) and distributed systems (Git [228] and Mercurial [231]).
B.9.1.1. CVS
The Concurrent Versions System [226] is one of the earlier representatives of the centralised source control
system. It serves the basic purpose of keeping the change history, but its method of checking in single files
makes it unsuitable for modern DevOps.
B.9.1.2. Subversion
Subversion or SVN [227] is a software versioning tool from Apache distributed as free software under the
Apache License 2.0 [187]. It is mainly used for revision control of source code but also for any kind of
files, such as documentation in binary format, archives, links, etc. It is a popular system, one of the
advantages being that its use is not overly complex. Revisions compose changes of one or more files in the
project, and the Subversion assigns each revision a unique non-decreasing number. The users can also
assign tags of arbitrary unique names to particular revisions, for instance to mark stable versions.
Copyright © 2015, DICE consortium – All rights reserved
49
Deliverable 1.1. State of the art analysis
B.9.1.3. Git
Git [228] differs from SVN being a distributed control version system. It is used in very large projects
thanks to easier branch operations, a full history tree available offline, and a distributed peer-to-peer model.
Currently, it is one of the most widely used version control systems for software development.
Each Git revision receives a commit ID, an alphanumeric string the length of 40 characters which contains
a hexadecimal representation of the SHA1 [229] digest of the revision. This commit ID reliably marks the
revision and its history, which is crucial for the distributed nature of the whole system. However, the ID
cannot work well as a part of the version number like the Subversion’s revision number could. Therefore,
the developers need to keep their own custom revision numbers, e.g. by storing it in a file in the project and
having it auto-increment using a repository hook [230].
B.9.1.4. Mercurial
Mercurial [231] is also a distributed revision control tool like Git [228]. It is also used in major
developments like OpenOffice.org [232]. Its main difference - with respect to other revision control
systems - is that Mercurial is primarily implemented in Python [128], as opposed to C, and is easier to learn
than other distributed versioning system.
B.9.2
DICE needs with respect to versioning
An important issue in DICE is the ability to associate to each UML model element a unique version
number, which is consistently propagated through the DICE toolchain. For example, if one splits a
component in two, the monitoring and enhancement tools should detect that the monitoring data acquired
before this change is probably no longer representative of the application at hand.
It is generally best if any DIA component has a version number stored in some component configuration so
that it is available to all the tools used in the chain. The continuous integration tool could increment the
version number upon success, but usually a build number of a kind comes from the Git [228] or any other
version control system. But otherwise it should be up to the human users to say which version they want
the current build to have.
Since history of versions is important and results of the current version build depend on previous version
builds due to the enhancement tools, the DICE tools - with the help of the continuous integration tool could always preserve the latest results and the best results in any version. But since each result has to be
bound to a specific commit to the version control system, the best results can only serve as a kind of a
reference, while only the latest results actually matter. Existing tools to help automatically increase version
numbers depend on the principal language used in the application. For example:

Versions plug-in for Maven [233] helps manage the versions of the artefacts in a Project’s Object
Model (POM);

Python-versioneer [234] provides a number of formats and styles for generating the new version
number from any past version numbers used in the VCS;

Dist::Zilla [235] manages the version increments for the libraries written in the Perl.
B.10.
Discussion
In this chapter we gave an overview of a large set of approaches and tools that are related both to
development and operation and, in most cases, have been developed in a completely independent way by
research and practitioners.
Copyright © 2015, DICE consortium – All rights reserved
50
Deliverable 1.1. State of the art analysis
Referring to Figure 13 that shows the DICE high level vision, this state of the art analysis has been drawn
within the context of DevOps (Section B.2) that is the movement within which DICE aims at operating.
The DICE methodology (on the left hand side of the figure) will be defined keeping the DevOps principles
in mind. Moreover, it will be based on Model-Driven Engineering (Section B.3) and will take its roots from
the existing model-driven approaches focusing on cloud and DIA applications. The IDE (the box in the
Figure 13 that encloses four different tools) will be developed starting with modelling tools such as Papyrus
[83] and MOSKitt [97]. These tools allow us to let DICE users exploiting the DICE profile - which will be
built as part of the MARTE profile [68] - model DIAs and interact through proper connectors using tools
that support simulation, optimisation, creation of deployment recipes and analysis of testing results. Model
to model transformations (Section B.4) will be used to support the generation of different views on a DIA
model. Such views can either be used to support the transition from high level design of a DIA to the
selection of Data-Intensive technologies and to the deployment on the cloud, or they can be used to
transform models in a way that is suitable for specific analyses and simulation activities. TOSCA (Section
B.5) will be used as an output format representing the deployable model of a DIA. Such format will be
provided as input to the deployment and management tool that will be based on those reviewed in Section
B.6. Finally, the continuous integration tools (Section B.7) will be used as the basis to support the DICE
integration phase, keeping in mind that in our approach the emphasis is not only on integrating, building
and deploying code but also on managing and continuously evolving model. Of course, models’ and
components’ evolution results in the need for keeping track of different versions and of the relationships
between the various components versions. An analysis of the literature in this field is provided in Section
B.9 and is strictly correlated to the continuous integration approaches of Section B.8.
Figure 13. DICE high level vision [237].
Copyright © 2015, DICE consortium – All rights reserved
51
Deliverable 1.1. State of the art analysis
C.
Quality Assurance
C.1.
Non-Functional Properties in DevOps
In the previous chapter we have examined approaches to design and implementation of DIA and the
underpinning development toolchain functional properties. In this chapter we investigate non-functional
properties. In the DICE vision non-functional characteristics of the software follow the definition of the
ISO/IEC standards and may be summarised at a high-level as follows:



Reliability: The capability of a software product to maintain a specified level of performance,
including Availability and Fault tolerance.
Performance: The capability of a software product to provide appropriate performance, relative to
the amount of resources used, as described by time behaviour and resource utilisation. The terms
performance and efficiency are considered synonyms throughout.
Safety: The capability of a software product to achieve acceptable levels of risk of harm to people,
business, software, property or the environment.
The project will interface the DICE profile to existing reliability and performance tools, which will be
enhanced to be able to analyse behaviour related to data. Further, the project will also encompass
verification and trace checking capabilities to support safety analysis.
In this chapter we survey tools and methods for quality analysis. We emphasise two dimensions:


Design tools and methods. The annotation and analysis of quality requirements in UML [67].
Enhancement tools and methods. The combination of monitoring, testing and feedback analysis
techniques to improve the design and quality analysis of application prototypes.
In more details, the contents of this chapter are as follows:



First, we introduce definitions of reliability, performance, safety properties and metrics that are
relevant to the analysis. Service-Level Agreements (SLAs) are readily determined by predicating
on these quantities. We illustrate various notions of predicates in the safety properties sections,
which can be readily used also to qualify reliability and performance constraints arising in SLAs
(e.g., the 95th percentile of end-to-end response times should be smaller than a certain amount).
SLAs can be directly annotated in the UML models, thus we do not look at other forms of
specification (e.g. Web Services (WS) standards [238]).
Next, we overview UML annotations to specify reliability, performance, and safety properties.
We emphasise two existing profiles as baselines: UML MARTE [68] and UML DAM [69]. We
also overview UML extensions for access control specification and techniques to optimise
deployment plans generated from such UML specifications.
Finally, we introduce Model-Driven Testing (MDT) tools and complementary monitoring tools to
retrieve information about application behaviour during test, followed by methods to analyse the
monitoring data acquired, in particular trace checking, feedback analysis, and anomaly detection
via machine learning.
In this section, we review the definition of several metrics that will be examined as part of DICE
performance, reliability and safety predictions. Performance and reliability are the focus of simulation and
optimisation tools, whereas safety predictions will be performed by verifying properties through model
checking.
Copyright © 2015, DICE consortium – All rights reserved
52
Deliverable 1.1. State of the art analysis
C.2.
Performance Metrics
In this section, we review some basic definitions concerning performance prediction metrics. These are
standard definitions that are usually used in the context of queueing models, but are also applicable to
performance predictions obtained with other formalisms, such as Stochastic Petri Nets (see Section
C.5.2.1). We give definitions for the basic case where requests are considered of a single type (i.e. a single
class model). The generalisation for multiple types is simple and in most cases requires only adding an
index to each metric to indicate the class of requests it refers to [239].
Considering an abstract system where T is the length of time we observe a system, A is the number of
arrivals into the system, and C is the number of completions (users leaving the system) we can define the
commonly used measures [239]:


Arrival rate: A = T
Throughput: X = C/T (simply the rate of request completions)
In a system with a single resource we can measure Bk, and denote Ck the number of arrivals at the resource.
If Tk is the time the resource k was observed to be busy, we can then define two additional measures:

Utilisation, Uk = Bk/Tk (this is normally given as a percentage, e.g. 30% server utilisation)

Service time per request, Sk = Bk/Ck
From these measures we can derive the Utilisation Law as U = X*S. The above quantities can be made
specific to a given resource, for example Uk stands for the utilisation of resource k.
One of the most useful fundamental laws is Little's Law [239] which states that N, the average number of
customers in a system, is equal to the product of X, the throughput of the system, and R, the average
response time a customer stays in the system. Formally this gives: N = X*R. The formula is unchanged if
one considers a closed model, which has a fixed population N of customers. Another fundamental law of
queueing systems is the Forced Flow Law [239] which, informally, states that the throughputs in all parts
of the system must be proportional to one another. Formally the Forced Flow Law is given by: Xk = Vk*X,
where Xk is the throughput at resource k, Vk is the visit count of resource k, i.e., the mean number of times
users hit this resource. Combining both Little's Law and the Forced Flow Law allows for a wide range of
scenarios to be analysed and solved for a particular desired quantity. For example, it is possible to compute
utilisation at a server k in a distributed network directly as [239] Uk=X*Dk where Dk=Vk*Sk is called the
service demand at server k.
C.3.
Reliability Metrics
The area of reliability prediction is established and focuses on determining the value for a number of
standard metrics, which we review below.
The execution time or calendar time is appropriate to define the reliability as R(t)=Prob( >t} that is,
reliability at time t is the probability that the time to failure is greater than t or, the probability that the
system is functioning correctly during the time interval (0,t].
Considering that F(t)= 1-R(t) (i.e., unreliability) is a probability distribution function, we can calculate the
expectation of the random variable as
. This is called Mean Time to Failure
(MTTF) [240] and represents the expected time until the next failure will be observed.
Copyright © 2015, DICE consortium – All rights reserved
53
Deliverable 1.1. State of the art analysis
The failure rate (called also rate of occurrence of failures) represents the probability that a component fails
between (t,dt), assuming that it has survived until the instant t, and is defined as a function of R(t):
. The cumulative failure function denotes the average cumulative failures associated
with each point in time, E[N(t)].
Maintainability is measured by the probability that the time to repair ( ) falls into the interval (0,t] [240]
M(t) = Prob { ⪯ t }
Similarly, we can calculate the expectation of the random variable
(Mean Time To Repair), and the repair rate as
as
, that is called MTTR
.
A key reliability measure for systems that can be repaired or restored is the MTBF (Mean Time Between
Failures) [240], that is the expected time between two successive failures of a system. The system/service
reliability on-demand is the probability of success of the service when requested. When the average time
to complete a service is known, then it might be possible to convert between MTBF and reliability ondemand.
Availability is defined as the probability that the system is functioning correctly at a given instant
A(t)=Prob{state=UP,time=t}. In particular, the steady state availability can be expressed as function of
MTTF and MTTR (or MTBF):
C.4.
Safety Properties
Safety properties vary from system to system and require formalisation in terms of logical expressions that
can be verified via formal methods. The general idea of formal verification is the following: given a formal
model of the system to be analysed (let us call this model S), and given a property (a requirement), also
formalised in some suitable way (let us call this property R), we check whether R holds for S or not (in
symbols, we write this as S |= R).
There are many possible formalisms for describing S and R; S and R themselves might be described using
different formalisms [241], [242].
A (non exhaustive) list of formalisms for describing the system S are:




Finite State (Büchi) Automata
Timed Automata
Probabilistic Automata (in the broad sense of the term, including Discrete-time Markov Chains,
Continuous-Time Markov Chains, Markov Decision Processes, etc.)
Temporal logics (of many different kinds: with/without metric on time, discrete/continuous-time,
etc.).
Typically (though not always) the requirement R is described through a logic. Examples of logics that can
be used are the following:
Copyright © 2015, DICE consortium – All rights reserved
54
Deliverable 1.1. State of the art analysis





Linear Temporal Logic (LTL)
Metric Temporal Logic (MTL), and the similar TRIO
Computation Tree Logic (CTL)
Probabilistic CTL (PCTL)
Continuous Stochastic Logic (CSL)
The kinds of properties that can be expressed and verified depend on the nature of the formalisms used to
describe S and R. Some categories of properties that can be expressed are the following:

Ordering properties
For example, a safety statement may require that: ‘Every occurrence of b must be followed eventually by a’
 A possible (though not the only one) formalisation of this property in LTL is the
following:
G(b → F(a))
 Metric properties
Example: ‘Every occurrence of b must be followed eventually by a within 10 time units’
 Possible formalisation of this property in MTL:
G(b → F[0,10](a))
 Invariants/safety properties
Example: ‘the system will never enter state critical’
 Possible formalisation in LTL:
G(¬critical)
 Branching-time properties (i.e., properties of executions)
Example: ‘Any time the system is in state critical, it can eventually recover to state normal’
 Possible formalisation in CTL:
AG(critical → EF(normal))
 Probabilistic properties
Example ‘The probability of eventually reaching state critical from state normal without passing from state
warning is less than 0.1’
 Possible formalisation in PCTL:
normal → P≤0.1 (¬warning U critical)
 Real-time probabilistic properties
Example ‘The probability of eventually reaching state critical from state normal without passing from state
warning and within 5 steps is less than 0.1’
 Possible formalisation in PCTL:
normal → P≤0.1 (¬warning U≤5 critical)
Over the years, several works have identified ‘patterns’ with which requirements have been defined by
users in various domains. Examples of these works are [243] for LTL specifications, [244] for real-time
patterns, [245] for probabilistic properties, and [246] for properties of service-based applications. These
patterns might be the basis for the definition of a user-friendly interface through which DICE users can
introduce the properties to check safety without having to delve into the finer details of expressing them in
logic terms.
Copyright © 2015, DICE consortium – All rights reserved
55
Deliverable 1.1. State of the art analysis
C.5.
Quantitative analysis for assessment of performance, reliability and safety
Research in reliability and performance analysis has led to a variety of modelling techniques, each focusing
on particular levels of abstraction and/or system characteristics. Such modelling techniques are used for
predictions of reliability/performance properties, often rely on stochastic assumptions, and they help in
answering questions that are more pressing to the engineers (e.g. is the system scalable? Will the system
provide a specified response time? Is the system able to tolerate N simultaneous equipment/component
failures? When a shutdown occurs, how long does it take to recover the system? Is the system able to
provide the service during a given time period? etc.).
It is worth noting that, depending on the system characteristics and on the type of measure to be evaluated,
the analysis can be carried out under steady state or transient assumptions [247], [248]. Steady state
analysis assumes that the system has reached a stable state, so the estimated measures (e.g., mean response
time of the service, steady state availability) are not influenced by the considered system initial state. On
the contrary, transient analysis assumes that the system has not yet reached a stable state and the estimated
measures depend on the choice of the initial state. The latter type of analysis is much more challenging and
more realistic when the system may take long time to reach a stable state. It should be also carried out
when the measures of interest are functions of time, e.g., the reliability function (R=Prob{ Time To Failure
> t}) and instantaneous availability (A=Prob{‘System is up at time t’}).
Herein, we review the main modelling and analysis techniques according to the following classification:
combinatorial, state-based and Monte Carlo simulation. Combinatorial techniques do not enumerate all
possible system states to obtain a solution, they are simpler than the state-based ones. However, they do not
easily capture stochastic dependences. Hence they can be used for a quick, rough-cut analysis. Steady-state
techniques are more comprehensive than combinatorial ones, since they allow to model explicitly complex
relationships, including sequencing information. Monte Carlo simulation techniques are alternative
solutions to state-based ones when the state space of the modelled system is too large to be analysed as a
whole.
C.5.1
Combinatorial techniques
Reliability Block Diagrams (RBD) [249] and Fault Trees (FT) [250], used for reliability and availability
analysis, belong to this category.
RBD [249] is a graphical representation of the system components and connectors. An RBD consists of
blocks and lines: the blocks represent system components and the lines describe the connections between
components. If any path through the system is successful, then the system succeeds, otherwise it fails. It is
assumed that connectors have reliability equal to one (i.e. do not fail) and that both the system and its
components are in one of two states: either up or down. Hence, in an RBD, each component can be seen as
a switch that is closed when the component is up and open when the component is down. The system will
be up only when a path exists between the input and output nodes. Failure/repair times of components are
assumed independent random variables.
Fault Tree (FT) [250] is an acyclic graph with internal nodes that are logic gates (e.g. AND, OR, K-of-M),
external nodes that represent component/sub-system faults (basic events, undeveloped events) and,
possibly, transfer gates. It is a top-down technique, i.e. the FT construction begins by considering a system
failure mode - the ‘top event’ - and terminates when either the basic events which provoke such a failure
are all identified or the desired level of detail is reached. FT can be used to compute the probability of
occurrence of the top event, considering the system mission time, and may correspond to the system
Copyright © 2015, DICE consortium – All rights reserved
56
Deliverable 1.1. State of the art analysis
unreliability or unavailability computed at a given time instance1. Several variants of FT have been
proposed to enhance the modelling capabilities, such as Dynamic Fault Trees [251] that include special
purpose gates capturing sequence dependencies (functional dependency, spare and priority-AND gates).
C.5.2
State-based techniques
Markov models belong to this category: they have been extensively used for reliability, performance and
performability analysis [252], [253]. The simplest Markov model is a Markov chain, which is a Markov
process with a discrete state-space. For reliability and performance analysis Continuous-Time Markov
Chain (CTMC), i.e. a Markov chain with time taking non-negative real values, is the reference model, and
many solution algorithms exist for it [254]. One of the main drawback of CTMC numerical solution is the
so-called ‘state-explosion’ problem, that is for complex systems with large number of components, the
number of states can grow in a prohibitively manner. So, techniques aimed at reducing the size of the
model have been proposed, such as state truncation methods [255], hierarchical model solutions [256], state
lumping techniques [257] and aggregation techniques [258]. The latter do not produce exact solutions, only
approximate values. Other types of techniques use special data structures to reduce the space requirements,
such as binary and multi-valued decision diagrams [259]. Another problem that may occur when a Markov
model is characterised by transition rates with very different order of magnitude is the ‘stiffness’ (e.g.
failure rates and repair rates, failure rates and job arrival rates etc.) that may be avoided by using
aggregation techniques [258] or special solvers [260].
Markov models represent the basis of higher abstraction level modelling formalisms, like Stochastic Petri
Nets (SPN) [261] and Queueing Networks (QN) [239].
C.5.2.1. Stochastic Petri Nets
Stochastic Petri Nets (SPN) [261] are usually solved by deriving the underlying CTMC, which is
isomorphic to its reachability graph, and then by using the wide variety of aforementioned techniques. For
large complex models alternative techniques can be used to avoid the space-explosion problem, such as
discrete event simulation [262] and bounding techniques [263]. SPN have been extended to include: arcs
with multiplicity, immediate transitions, transition priorities, inhibitor arcs. The most popular class of SPN
including such extensions is Generalised Stochastic Petri Nets (GSPN) [264]. Other extensions allow more
flexible firing rules including guards and enabling functions such as Stochastic Activity Networks (SAN)
[265] or Stochastic Well-Formed Nets (SWN) [266]. On the other hand, there have been several proposals
on relaxing the ‘exponential distribution transition firing time’ assumption, making the stochastic process
associated to the SPN instead of CTMC. For example, Extended Stochastic Petri Nets (ESPN) [267] are
SPN where transitions can be characterised by general firing time distributions. Deterministic and
Stochastic Petri Nets (DSPN) [268] have immediate, exponential and deterministic transitions. Markov
Regenerative Stochastic Petri Nets (MRSPN) [269] and Concurrent Generalised Stochastic Petri Nets (CGSPN) [270] are generalisations of DSPNs. Numerical techniques based on the solution of the underlying
stochastic process have been also proposed for the aforementioned SPN variants.
C.5.2.2. Queueing networks
Some classes of Big Data technologies such as streaming systems, in-memory databases and Hadoop
clusters (see Chapter D), may be effectively modelled by means of a queueing network models. A
queueing network model is a network of resources throughout which requests are routed after spending
some time at each resource. Each queue can be used to model either a physical resource (e.g., CPU,
network bandwidth etc.) or a software server (e.g. admission control or connection pools). Nowadays,
applications are often tiered and queueing networks can easily capture the interactions between tiers.
1
Observe that for non repairable systems the two measures are equivalent.
Copyright © 2015, DICE consortium – All rights reserved
57
Deliverable 1.1. State of the art analysis
Several variants of queueing network models exist, such as product-form models [271] which can be solved
analytically, and extensions to model blocking, fork/join behaviour and priorities [252].
A popular class of queueing networks used to model complex software systems are layered queueing
networks (LQNs). An LQN adds the capability of describing request workflows and organising resource
hierarchically, in layers. LQN models of an application can be built automatically from software
architecture models expressed using formalisms such as UML or Palladio Component Models (PCM)
[154]. Several evaluation techniques exist for LQNs such as [272]-[275].
The authors in [276] provide a discussion about the advantages and disadvantages of LQNs identifying a
number of key limitations for their practical use. These include, among others, difficulties in modelling
caching, lack of methods to compute percentiles of response times, trade-off between accuracy and speed.
Evaluation techniques for LQNs that allow the computation of response time percentiles have been
presented [275]. Recently these models have been further extended to represent the operational
environment in which a system operates through an abstraction called random environment. Finally, we
mention the existence of a class of models, referred to as Queueing Petri Nets (QPNs), which add to
queueing networks the possibility to include some elements of stochastic Petri nets in the specification of
the service workflow of requests. QPNs models combine benefits of both approaches, but they require
evaluation via simulation due to the current lack of analytical results for this class of models.
C.5.3
Monte Carlo simulation techniques
A common use of Monte Carlo simulation is to express the system behaviour using simulation languages,
e.g. [277] or general-purpose programming languages. Also, simulation can be used as solution technique
for systems modelled with the high level formalisms discussed in the previous section (i.e. Stochastic Petri
nets [261], Queueing Networks [239] and their extensions).
Simulation allows to estimate performance and reliability measures by generating and analysing randomly
chosen paths through the state-space. For each measure of interest, the simulation provides the estimated
value and the confidence interval, i.e. the real (unknown) value falls into this interval with a certain
probability (i.e. confidence level). The width of the confidence interval is a measure of the accuracy of the
estimated value.
Different simulation approaches exist. When the measures of interest have a transient nature, then a
common approach is the replication where N statistically independent simulation runs are executed and the
measure is estimated considering the N independent measurements collected in the different runs. More
efficient simulation approaches are used when the aim of the analysis is the estimation of steady state
measures. Indeed, in such cases just a single simulation run can be executed where the simulation time is
‘long enough’ to bring the system into the steady state, and then use measurements taken across time
(rather than measurements taken across replications). For example, the method of batch [278] partitions the
simulation interval in successive epochs of fixed duration and for each epoch a measurement is collected.
The computational cost of the simulation depends on various factors: 1) the length of transient period, i.e.
that is the simulation time until the steady state is reached, 2) the length of the simulation time needed to
collect a ‘good sample’, and 3) the length of the epochs needed to ensure statistical independence of
successive measurements. It is worth to observe that such factors depend on the system under analysis, so
the choice of simulation input parameters is a critical point since both the statistical quality of the
simulation results and computer time that is required to simulate a desired amount of system time rely on
them. Concerning the length of the epochs, a possible rigorous solution is the regenerative points method
[279], where a regeneration point is a time instant at which the system enters a given state (e.g. in an M/G/1
queue model this is the state where there are no jobs in the queue).
Copyright © 2015, DICE consortium – All rights reserved
58
Deliverable 1.1. State of the art analysis
Moreover, statistical issues may arise in case of rare events, for example when the measure to be estimated
is a small probability, e.g. the system failure probability. So-called variance reduction techniques have been
developed to overcome them [280].
With respect to the required computer time approximate accelerated stochastic simulation approaches have
been proposed to reduce it [281].
C.6.
Performance and reliability prediction tools
In this section we revise the performance and reliability tools supporting the modelling and analysis with
the formalisms of interest for DICE, i.e. Stochastic Petri Net (SPN) tools and Queueing Network tools.
C.6.1
Stochastic Petri Net tools
There is a variety of modelling and evaluation tools for SPN, most of them can be found in [282]. The
survey in [283] provides a ranking of those tools considering a set of criteria including: multi-platform
support, open source, embedded graphical animation, structural analysis and stochastic and coloured Petri
Net support.
Table 12 compares the tools that provide support to Generalised Stochastic Petri Nets (GSPN) [264] and
their extensions, such as:

Stochastic Well Formed Nets (SWN) [266] and Coloured Generalised Stochastic Petri Nets
(CGSPN), that are a stochastic variants of Coloured Petri Net [284], and

extended Deterministic and Stochastic Petri Nets (eDSPN), that are DSPN [268] characterised by
transitions with deterministic, exponentially and generally distributed delays.
The first column of the Table 12 provides a list of the features we consider relevant for the DICE
framework. Particular attention has been paid to the quantitative analysis support.
From the quantitative point of view GreatSPN [285] and TimeNET [286] are better than PIPE [287],
considering that for reliability analysis having transient solver is a need. Observe that TimeNET supports
the simulation of rare events, which are often studied in reliability/performability models. Moreover, it
includes both sequential and distributed simulation modules. GreatSPN and TimeNET provide support to
extended classes of GSPN (transitions with general distributions). On the other hand, the advantage of
using PIPE - together with the integrated modules of PeabraiN [288] – is that it is multi-platform (Javabased). It is worth to note that PeabraiN provides two types of steady state simulator, in particular the
approximated one is used to accelerate the analysis.
Table 12: Comparison of SPN tools.
Features / Tools
GreatSPN [285]
TimeNET [286]
PIPE [287]
PeabraiN [288]
Open source (license)
proprietary, free for
academic institutions
proprietary, free for non
commercial use
GPL
GPL
PNML Format
Interchange
X
X
X
Multi-Platform
X
X
X
GSPN,
GSPN, CGSPN
Net composition
PN classes
X
X
GSPN, SWN, eDSPN
GSPN, CSPN, eDSPN
Copyright © 2015, DICE consortium – All rights reserved
59
Deliverable 1.1. State of the art analysis
CGSPN
Qualitative Analysis
X
X
X
steady state, transient
steady state, transient
steady state
Quantitative Analysis
Exact numerical
analysis
Approximated
numerical analysis
Simulation
steady state
steady state
Rare event simulation
Bounding techniques
Probabilistic
checking
model
steady state, transient
steady state
(exact,approx)
steady state
steady state
X
X
CSL-TA
C.6.2
Queueing Network tools
A similar comparison is given below in the Table 13 for queueing network modelling tools.
Table 13: Comparison of Queueing Network tools.
Features / Tools
JMT [289]
LINE [290]
LQNS [291]
GPL
BSD-3
Proprietary, Free for academic
use
XML, proprietary
same as LQNS
XML, proprietary
Any Java supported
Windows, Linux
Windows
Product-form models
X
X
X
Open models
X
Approximate
Approximate
Blocking
X
Fork-join
X
Priorities
X
Open source (license)
File Format
Multi-Platform
Random environments
X
X
X
X
X
Workflows
Partial
Partial
Comprehensive
Service phases
Partial
X
X
Transformed to single
layer
X
steady state, transient
steady state
Nested layers?
Exact numerical analysis
steady state,
transient
Approximated numerical
analysis
product-form
models
Copyright © 2015, DICE consortium – All rights reserved
60
Deliverable 1.1. State of the art analysis
Percentiles analysis
X
Simulation
X
X
X
Rare event simulation
Bounding techniques
C.7.
product-form
models
Formal analysis of safety and privacy properties
The main goal of DICE is to define a Model-Driven Engineering (MDE) [59] approach and a Quality
Assurance (QA) toolchain. One of the functionalities of the DICE IDE is to offer the ability to specify the
DIAs through UML models. From these models, the toolchain will guide the developer through the
different phases of quality analysis, formal verification being one of them.
The purpose of formal verification in DICE is to assess safety and privacy risks of Data-Intensive
Applications (DIAs). More precisely, the DIAs will be transformed into DICE models. The verification and
validation of these models will be achieved by: (1) annotating the models with safety and privacy formulas,
and (2) automatically translating the models into a suitable formal representation.
For many decades it was thought that formal verification is hopeless due to cryptic notations and nonscalable techniques, as well as hard to use because of the dedicated tools. Moreover, the existing non-trivial
case studies were not convincing enough for the software or hardware engineers.
Around 1990s perspectives for formal verification started to look more optimistic once techniques like
model checking [292] and theorem proving were adopted. Researchers and practitioners began performing
more and more industrialised case studies and thereby gaining the benefits of using formal methods.
Nowadays, the formal verification techniques and tools face another challenge, namely they must undergo
a deep technological transition to exploit the new available architectures. These new complex architectures
were imposed by the Cloud Computing (the ease of accessibility of the cloud-based computing resources)
and Big Data (large amount of data which have become easily accessible) emerging technologies.
There exist in the literature only few attempts of applying formal verification in the context of DIAs. For
example, [293] focuses on techniques for checking specifications expressed in a metric temporal logic with
aggregating modalities (using the MapReduce programming model). These specifications correspond to
quantitative metrics [294], e.g. response-time, throughput or availability. Similarly, [295] proposes an
approach for the offline monitoring of DIAs. Unlike [293], it uses metric first-order temporal logic for
specifying properties of the system actions to be checked, as well as a formal framework for slicing logs.
Their work also relies on MapReduce model. From a different perspective, [296] uses Big Data technology
(Hadoop MapReduce [14]) for parallelising and distributing model checking techniques in order to make
the verification of DIAs feasible or more efficient. None of these approaches focuses on verifying
qualitative properties, like safety and privacy of DIAs. An attempt similar to ours is presented in [297],
aiming to model the MapReduce model with Communicating Sequential Processes (CSP) formalism [298].
The authors present a very simplistic model together with an example and, as future work, it is planned to
use a dedicated model checker [299] to check safety properties of the model. No explicit safety property
was verified; however, the authors mention that the model is suitable to verify properties like ‘If MASTER2
node failed, there is no component handling this exception, the system cannot recover from the error and
the MapReduce programme cannot be finished successfully.’ Computation Tree Logic (CTL) [300] is used
2
A MASTER node performs scheduling and failure handling.
Copyright © 2015, DICE consortium – All rights reserved
61
Deliverable 1.1. State of the art analysis
in [301] as underlying formalism for representing the MapReduce model and expressing its safety
properties. The general purpose model checker Uppaal [302] is used to implement the model and verify
properties like load balancing and fault tolerance.
An architecture and technology-oriented formalisation of DIAs is presented in [303]. The authors present a
multi-formalism approach which enables the modelling of the architecture (Lambda Architecture [17]) and
application logic of DIAs as well as the environment. Although their approach uses Petri Nets for
modelling the application logic, the work is important for DICE because the authors present a two-layer
abstraction model enabling a very precise modelling of the application, but at the same time allowing to
hide the complexity when necessary. Their work could be inspirational since the DICE profile comes also
at different levels of abstraction Hence we have an example of what and how different elements, e.g.
architecture components and queries, are modelled at different layers.
From the works cited above we can learn the following: (1) there are suitable logic-based formalisms
modelling the MapReduce framework; (2) there does not exist a general level of abstraction for application
architecture, application logic and environment; (3) there is no consensus on safety properties of interest to
be verified; (4) there are no dedicated methods and tools for dealing with DIAs models, so researchers use
general-purpose techniques and tools; (5) the research in this direction is in the early stage.
Hence, the challenge of the DICE research will be to exploit and adapt existing techniques for checking
safety and privacy of DIAs. Therefore, in the remainder of this section we will focus on general-purpose
verification methods and tools that we aim to use and adapt in the framework of the DICE project.
Two of the underlying formalisms used to enable formal verification which will be exploited by us are
temporal logic and first-order logic. While temporal logic proved to be useful for formalising the
application and expressing the properties to be checked, first-order logic is the working language of
powerful state-of-the-art solvers that can be used to check the satisfiability of these properties. This is also
the strategy we aim to pursue in DICE. Hence, below we’ll present relevant research in these two directions
and how we plan to exploit them.
Using temporal logic notation (see e.g. [304]) as the language of annotating the UML models, or software,
is not appealing for any kind of actor/end user, unless they have previous experience with such notations.
To overcome this shortcoming, temporal logic description of the models should be completely hidden from
the modellers. One way to achieve this is to use a pattern-based approach to the presentation, codification
and reuse of property specifications. For finite-state verification, research has been initiated in this direction
in [305], which proposes specification abstractions having the property to be parameterisable, high-level,
and formalism-independent. It turned out that their approach was successful since it is mainly the case that
practitioners prefer guidance on how to optimise the usage of the language for solving a large class of
problems, rather than using its full expressiveness. In a similar manner, the TRIO approach [306] a firstorder linear temporal logic that supports a metric on time - allows template-based specification. TRIO
specifications of systems consist of basic predicates and arithmetic temporal terms representing elementary
phenomena of the system. The system behaviour over time is described by a set of TRIO formulae, which
state how the predicates are constrained and how they vary over time in a purely declarative fashion. This
formalism has been successfully applied to the case of embedded systems [307]. More recently,
specification patterns have been used in Service-Based Applications (SBAs), in particular industrial SBAs
in the banking domain [308]. In the framework of the DICE project we plan to use and extend the TRIO
formalism, as well as for specifying Big Data applications models.
Similar to the state of the art, in DICE we will also consider using classical model checkers like SPIN
[309], NuSMV [310] and Uppaal [302], which target continuous-time, real-time systems, and PRISM
Copyright © 2015, DICE consortium – All rights reserved
62
Deliverable 1.1. State of the art analysis
[311], which handles probabilistic models, in order to test safety properties specific for DIAs. In this
context we will also exploit the techniques introduced in [312] to include the possibility of modelling and
verifying data transmission under real-time constraints. Additionaly to this, for efficiency purposes,
Satisfiability Modulo Theories (SMT) solvers such as Z3 [313] and satisfiability solvers for temporal logics
such as Zot [314] will be used as verification engines to build formal verification techniques for UML
models with real-time constraints similar to b project [307]. However, in the DICE framework these
techniques have to be extended in order to cope with the particularities of DIAs.
For what concerns the formal verification at different levels of abstraction and granularity, the work
presented in [307], which is implemented in the CorrettoUML [315] formal verification tool for UML
models, is relevant for DICE: we plan to decouple the semantics from the predicates that represent model
elements. In this way one can change the semantics while translation from UML to the predicates remains
unaffected. Since verification at all abstraction levels will be impossible due to the complexity of the model
and the properties to be verified, we plan to adopt the so-called lightweight formal approaches [316], where
verification formalisms are applied only where necessary. Moreover, [307] introduced stereotypes to allow
actors to identify those parts of the system on which they want to focus verification activities. These
stereotypes are then taken into account during the transformation phase, which only produces formal model
for the tagged parts instead of the whole model. A lightweight formal approach combined with stereotypes
will be also used within DICE. These, together with verification at different abstraction levels, will allow us
to model the application as a set of components (e.g. nodes) performing the same kind of task (e.g. Map,
Reduce) and check the same safety property for each component type. In order to avoid this repetition, a
generalisation of the safety properties for an arbitrary number of components of the same kind is useful.
Such generalisation was applied for many case studies, for example mutual exclusion protocols [317]. We
believe that this parametricity is practically important, especially in the context of DIAs modelling.
Regarding privacy properties, research similar to DICE Model-Driven Engineering context is carried out in
[318]. The authors develop a Model-Driven Development approach for the secure medical system Selkis.
They built both the functional and security models (Platform Independent Models) of the application in
SecureUML [319], translated them into a specification in the B method [320] and then used the animation
tools of the B method in order to validate the functional model and to perform systematic tests. The
modelling of privacy properties like data confidentiality and integrity made possible the detection of
insiders’ attacks. In DICE we aim at capturing privacy properties at the UML level using SecureUML and
then formalising them in temporal logic.
We presented a short overview of existing methods and tools relevant in the context of DICE for the
verification of safety, respectively, privacy of systems. We noticed that the verification tasks are
approached from different perspectives. Due to this consideration, in DICE safety and privacy properties
will be handled by using different techniques and tools suitable to the given situation.
C.8.
Tools for Formal Verification
As was explained at the beginning of the chapter in formal verification approaches we aim to check - given
a model of the system S and a requirement R, both formally expressed - whether R holds for S or not (i.e.
whether S |= R). Depending on the formalisms used to describe S and R we can express and verify different
kinds of properties (e.g. ordering, real-time, probabilistic etc.).
Many tools have been built that implement the various formal verification approaches. In this section we
briefly list some of the most commonly used, and the candidates to be the starting point for the DICE
formal verification tools.
Copyright © 2015, DICE consortium – All rights reserved
63
Deliverable 1.1. State of the art analysis
Table 14 lists the tools of interest. For each tool we indicate its name, the formalism used to describe the
system S (the ‘System language’) and the formalism used to describe the requirement R (the ‘Property
language’). The tools’ websites can be found in the list of references.
Table 14: Formal verification tools.
Tool
SPIN
[309]
System language
Transition systems
Property Language
LTL
The input language of SPIN is called Promela, and it allows users to describe transition systems, i.e., finite
state automata. The tool allows users to check for properties such as state reachability, assertion violations,
and properties expressed in LTL; essentially, it allows for the verification of ordering properties (real-time
properties can be checked provided we assume ‘1 state transition = 1 time instant’. The verification is
based on the exhaustive exploration of the state space.
NuSMV
Transition systems
[310]
LTL/CTL
NuSMV has a custom language for describing system models to be analysed, which corresponds to
defining transition systems. It implements several decision procedures based on a symbolic representation
of the state. Depending on the decision procedure, it uses a different formalism to express the property to
be checked: CTL or LTL. As a consequence, it is capable of analysing ordering properties. An evolution
of the tool has been recently released, called nuXmv [321], which is capable of analysing infinite-state
systems and also real-time properties.
Uppaal
[302]
Timed Automata
Restricted version of CTL
Uppaal is the most popular tool for analysing real-time systems. It uses Timed Automata as input
language. To define the property to be analysed it employs a restricted variant of CTL which only allows
certain combinations of temporal operators. Essentially, it allows users to check whether certain desired
states are reachable or not. The decision procedure is based on the exhaustive exploration of the state
space of the so-called ‘region automaton’ that is derived from the timed automaton.
PRISM
[311]
Probabilistic automata
PCTL/CSL
PRISM is the reference tool for the formal verification of probabilistic systems. It allows users to use
several languages to describe the system to be analysed: Discrete-Time Markov Chains (DTMC),
Continuous-Time Markov Chains (CTMC), Markov Decision Processes (MDP), Probabilistic Timed
Automata (PTA). For systems described through DTMC, MDP and PTA, the tools allow users to
formalise the property to be checked through PCTL, whereas the properties to be analysed for CTMCs are
expressed in CSL.
Zot
[314]
Metric Temporal Logic
Metric Temporal Logic
Zot is a satisfiability checker for various kinds of temporal logics, most of which have a metric on time.
More precisely, it can decide the satisfiability of formulae of LTL, MTL/TRIO, CLTL (Constraint LTL)
and CLTLoc (Constraint LTL over clocks). To check whether property R holds for system S it reduces the
problem to one of satisfiability of a logic formula. More precisely, when both S and R are described
through temporal logic formulae, it checks whether S |= R or not by determining if S ∧ ¬R is satisfiable (S
|= R holds if S ∧ ¬R is not satisfiable). Zot employs a so-called bounded decision procedure, where the
temporal logic formula P whose satisfiability needs to be checked is translated into a formula of a
decidable logic that captures the unfolding of P over k time instants (where k is the bound). Then, the
satisfiability of the latter formula is checked through an off-the-shelf SAT of SMT (Satisfiability Modulo
Theories) solver such as, for example, Z3[322].
Copyright © 2015, DICE consortium – All rights reserved
64
Deliverable 1.1. State of the art analysis
C.9.
Software Anti-patterns
Flaws in software architecture design resulting in performance problems are known as software
Performance Anti-patterns (AP).
The first study employing automated AP detection [323] involves construction of application design model
based on monitoring data and specification of a set of rules. Each rule is a description of application
behaviour if a specific anti-pattern is present. These rules are then applied to the design model in order to
detect APs. This approach became the practice used in all subsequent work in the area. However, the work
considered only Enterprise Java Bean applications and could not be used for other technologies.
Additionally, the approach utilises run-time monitoring data, and thus is not suitable for anti-patterns
detection on the early development stages.
Cortellessa [324] in his state of the art analysis has summarised the research activities undertaken by his
research group in the area of performance anti-patterns specification, detection and solution. The paper
stressed the importance of the automation of these activities and gave a summary of model-driven
techniques employing UML or the Palladio Component Model (PCM) [154] and specifically developed
Anti-patterns Modelling Language (Aemilia). [324] also proposed a ranking system to rate anti-patterns
based on their contribution to violating requirements. This approach aims to facilitate more efficient and
automated AP solving. Finally, open issues, such as accuracy of anti-pattern specification, conflict
between anti-pattern solutions, ambiguity in formalisation, interdependencies among performance
requirements and others were outlined.
Trubiani et. al. [325] built on the work cited and discussed in [324] by adding the ranking system to the
detected anti-patterns. This system allows one to find anti-patterns causing the performance degradation
and, based on the results, run the automated search for the optimal architecture configuration that would
eliminate this ‘guilty’ anti-pattern. The authors though argue against the full automation of the anti-patterns
solution process, as human experience can be useful to cut some alternatives or give priority to other
alternatives than an optimisation algorithm would. The proposed methodology is implemented as a tool,
built on PCM and Aemilia.
Wert et. al [326], [327] have also developed an automated AP detection tool, but did not extend it to the
automated AP solution. While the tool in [325] requires the prior knowledge of the application software
architecture and PCM is used to simulate it in order to obtain performance analysis results, Wert et. al.’s
tool acquires performance analysis metrics through specifically designed set of experiments. Design of
experiments technique used in this approach is based on the assumptions about application usage profiles.
Trubiani et. al [328] investigated potential benefits of synergy between two types of software performance
analysis: Bottleneck Analysis (BA), aimed at identifying overloaded software components/hardware
resources, and Performance Anti-patterns (PA) in order to enhance the efficiency of identification and
solution of performance problems, and, as a result, present software developer with a broader set of
promising alternative solutions. For the validation of this approach authors designed a case study, where
they first executed two techniques separately, and then in sequence: PA after BA and BA after PA. The
sequence ‘bottleneck analysis after performance anti-patterns’ yielded the highest improvement on the
parameter of interest – response time of a certain operation (reduction by 90.8%). This approach also
allows for the choice of anti-patterns to be resolved or discarded, but instead of ranking system introduced
in Trubiani et. al. [325], the decision is made based on the results of BA. In conclusion, the authors
discussed limitations of BA and PA if executed in isolation, trade-offs between algorithm complexity and
its effectiveness and benefits of BA and PA synergy. However, there is still a limitation to the proposed
Copyright © 2015, DICE consortium – All rights reserved
65
Deliverable 1.1. State of the art analysis
approach as its execution time depends on a number of application-specific parameters, e.g. number of
software components/hardware resources, which may cause scaling problems if implemented in large-scale
applications. Table 15 provides a summary of the methods discussed above. Automated AP detection and
automated AP solution are chosen as dimensions for comparison to indicate if manual labour is required in
the method application. Dependence on the specific technology/platform or reliance on the specificallydeveloped modelling language might become a challenge in the uptake of a method. Finally,
implementation of a formal model in the method may be an advantage, as they are well understood and
widely used. However, this may play a negative role as well, as formal models require large parameter
space which ‘explodes’ with increase of an application’s scale.
Table 15: State of the art in the area of software performance anti-patterns detection.
Automated AP
detection
Automated AP
solution
Platform/technology
independent
Utilises
formal
model
Specific modelling
language
Parsons et. al
[323]
Y
N
N
Y
N
Wert et. al.
[326], [327]
Y
N
Y
N
N
Trubiani et. al.
[325]
Y
Y
Y
Y
Y (Aemilia)
Trubiani et. al.
2 [328]
Y
Y
Y
Y
Y
C.10.
Optimising Deployment Plans
The DICE project aims at providing an optimisation tool whose main goal is to identify a minimum cost
deployment for DIAs providing performance (e.g. job execution time is within a deadline) and reliability
(e.g. probability of no failure before the end of task is at least 99%) guarantees in a (public or private)
Infrastructe as a Service (IaaS) system. Once a model-to-model transformation is created between UML
models and the class of reliability and performance models discussed earlier, one can think of a search
procedure that alters the definition of the resources needed by the application up to finding an assignment
that satisfies the SLAs. This implies that each candidate deployment needs to be re-assessed using quality
analysis tools to determine the fitness of that assignment and possibly suggest the optimal search directions
for a new guess. The final output of this activity is the generation of a cost-optimal deployment plan from
the UML model. Evidently, the development of efficient search methods is of paramount importance in this
activity, since model-to-model transformations and solutions of formal models are both computationally
expensive. Thus, we overview advanced methods that have been proposed to efficiently search the space of
deployment allocations.
Identifying a Big Data cluster of minimum cost to operate a DIA under job completion time constraints is
far from trivial. As a matter of fact, given the variety of resource types, their performance and pricing
models available in the online catalogue of any of the Cloud providers in the market, the number of
possible Cloud configurations grows combinatorially. This makes the exploration of the solution space a
difficult and time consuming task even for determining a feasible configuration of small size systems.
What is more, since the number of VMs to be allocated to a single job depends on their type, even in the
simplistic scenario where only one type of VMs is selected the resulting problem belongs to the class of bilevel optimisation problems (i.e. a one problem is nested within another) that has been demonstrated to be
NP-Hard also for the simpler linear-linear case [329]. More details on bi-level optimisation problems can
Copyright © 2015, DICE consortium – All rights reserved
66
Deliverable 1.1. State of the art analysis
be found in [330]. Finally, the problem could result in being highly nonlinear in the constraints set. In
particular, given a certain deadline for an application, a number of VMs and the associated type, the
process of checking whether the application meets the requirements on the completion time (i.e. whether it
terminates before the deadline or not) can be nonlinear in the size of the problem. The described problem,
being intractable with exact algorithms even for very small cases, calls for a meta-heuristic solution
approach [331].
One meta-heuristic of interest to DICE is local search. The rationale behind a basic local search is the idea
to reach a good solution through a sequence of moves, each one leading from one solution to another. A
move can be described informally as an atomic operation that changes the structure of a solution it is
applied upon, generating a new (hopefully better) one. Indeed, not every move implies the generation of a
better solution. A local-search-based algorithm starts from an initial solution and drives the optimisation
process by selecting the moves to perform according to a certain strategy, which might involve moves that
can lead towards better or worse result. Local search techniques within the larger class of meta-heuristics
have proven to be highly adaptable and well suited to deal with optimisation problems where the evaluation
of the objective function is a process that may require a long time. There are several reasons for this. One is
that they are generally more reactive than other techniques. Furthermore, they tends to be more suitable to
incorporate and exploit specific knowledge of the problem, derived from performance models, to drive the
search toward more promising zones of the solution space.
C.11.
Testing and Monitoring of Non-Functional Properties
C.11.1 Testing methods
In this section, we focus on state-of-the-art research in quality testing through model-driven approaches.
Since the model-driven aspect is central to DICE, we do not review here standard testing approaches that
are agnostic of the application model-driven architecture. Furthermore, we only consider testing methods
that primarily focus on quality aspects (e.g. efficiency), ignoring functional testing, which is a mature area
[332]. Model-driven functional testing methods are easy to integrate with UML 2.0, for example via the
UML Testing Profile (UTP) distributed by OMG [333].
A systematic review performed in 2007 [334] reported that among 200 surveyed works in Model-Driven
Testing, just 10 papers covered non-functional aspects, in particular performance testing and related
dimensions (i.e. stability, resource utilisation and scalability).
Model-driven performance testing aims at using model-based reasoning to: 1) automatically generate a
performance test suite, possibly using model-to-model transformations; 2) automate the execution of the
test activities such as:




Load tests: where the test workloads are similar to the ones faced in production. Therefore the aim
is to expose the expected behaviour of the application under such workloads in order to verify
compliance with service-level objectives.
Stress tests: where the goal is to expose hardware and software bottlenecks by generating a
workload that leads to resource exhaustion. Stress testing can expose worst-case scenarios, poor
system configuration or verify the ability of self-adapting system to react to extreme conditions.
Stability tests: where the application stability is assessed by long-running tests.
Spike tests: where the system response is evaluated under unit workload spikes.
Copyright © 2015, DICE consortium – All rights reserved
67
Deliverable 1.1. State of the art analysis
In the context of web-based enterprise systems, Model-Based Testing typically relies on the following
classes of models:


User behaviour models, commonly specified via formal models such as probabilistic automata or
Markov chains, which can describe the navigation behaviour of users on a set of web pages. It is
easy to define algorithms that use these models to randomly generate user sessions for testing
purposes. Often the challenge in specifying these models is in the characterisation of the
preconditions that should hold in order to generate semantically-valid user sessions [335]. UML
sequence diagrams or PCM models [154] can be automatically translated into user behaviour
models, as done for example for PCM by the MDLoad tool developed in the MODAClouds project
[336]. Furthermore, methods to automatically extract user behaviour models have also been
recently proposed [337].
Traffic models: formal models are also commonly exploited to specify the frequency at which
requests should be sent to the application. For example, Markov-modulated models [338], [339]
have been used in recent years to automatically inject workload burstiness in a test suite. In the
literature, traffic models have not been normally related to UML or PCM modelling artefacts and
they are more often provided as input parameters to the workload generator.
Other classes of models for stress testing have also been considered in the literature, such as network
usage models that can be extracted from UML diagrams to support the generation of workloads that are
likely to expose faults related to network traffic [340].
One question left open by the methods discussed above concerns their ability to cope with the generation of
large volumes of data or high-velocity streams. We have surveyed the relevant literature. For NoSQL
datastores (see Section D.4), the main benchmark is the Yahoo! YCSB benchmark [341] which allows the
generation of user-specific datasets based on the number of records/keys with a specific distribution.
Clients can be configured to access the datastore in a random fashion or skewed with distributions that
mimic internet applications/services and with different ratios of reads/writes. Traditional databases have a
number of benchmarks mostly based on the TPC [342] suite of benchmarks, including the new TPCx-HS
Big Data benchmark specification. The size of the data is defined within the specifications and is usually
scaled from the minimum specification, e.g. in TPC-C depends on the number of warehouses. Data
generation variability is controlled by defined functions designed to be semi-random. However, it has been
shown that the data tends towards uniformity [343]. Data access is semi-random based on the same
functions for data generation, but the access footprint tends towards randomness. It is important to note that
the TPC benchmarks are specifications and their implementations are left to third parties.
Concerning stream-based systems, the unique feature of load/stress testing for stream-based application is
that the parameters that we can change to put load on the system consist of two orthogonal dimensions: (1)
data stream to the system, (2) query to retrieve information from the system. For example, for the former
we can change:


Event injection rate. Represents the number of events injected per second into the system. For
example: {1K, 5K, 10K, 50K, 100K}
Number of users. Represent the concurrent number of users submitting queries (e.g., /recommend
and /similarity in Oryx [31]) to the system.
And regarding the second dimension, we can change:

Query inter-arrival time. Represent the inter-arrival time between the query submission by a user.
Copyright © 2015, DICE consortium – All rights reserved
68
Deliverable 1.1. State of the art analysis

Query parameters. Represent the type of parameter that is passed to the query. For example in
Oryx we can pass the userID and depending on the type of user the cost of running the query could
be different.
In the research community for comparing approaches benchmarks like LRB benchmark [344] are used.
However, to the best of our knowledge there is no available tool that generates loads automatically based
on a user-defined usage pattern (like model-driven approaches for web-based applications, e.g.
MODACLouds MDLoad testing tool [336]) or probabilistic characteristics. The Starfish system [345] is the
most relevant approach. It represents a cost-based optimiser for MapReduce programmes which focuses on
the optimisation of configuration parameters for execution of these programmes on the Hadoop platform. It
relies on the profiler component that collects statistical information from executing the programmes and a
what-if engine for fine-grained cost estimation processing.
Message-Oriented Middleware (MOM) is the previous generation of technology for modern event-driven
applications like stock trading, event-based supply chain management, air traffic control and online
auctions. However, with the popularity of stream-based applications, the popularity of MOMs diminished
and they are now mostly used only in web-based non-Big Data systems. There exist a number of
proprietary and open source benchmarks for messaging platforms — for example, the SonicMQ Test
Harness [346], IBM’s Performance Harness for Java Message Service [347], Apache’s ActiveMQ JMeter
Performance Test [348] and JBoss’ Messaging Performance Framework [349].
Reliability testing tools are used to exercise an application and/or platform so that failures can be
discovered within the design and functionality. To do this, these tools are used to generate load and/or
faults within the application or platform. Test results can be used to identify bottlenecks and faults within
the structure of the deployment. Traditionally these tests are run before live deployment where faults found
will cause no issues to the production system. However, with the increase in usage of the Cloud, reliability
tests are now being run on live production as fast redeployment and inexpensive elasticity have made this
possible.
C.11.2
Testing tools
C.11.2.1. Grinder
Grinder [350] is a framework for distributed testing of Java applications. In addition to testing HTTP ports,
it can test web services (SOAP/REST), and services exposed via protocols such as JMS, JDBC, SMTP,
FTP and RMI. The tool executes test scripts written in Java, Jython [351] or Clojure [352]. Tests can be
dynamic, in the sense that the script can decide the next action to perform based on the outcome of the
previous one. The tool is mature, having been first released in 2003. It is quite popular with weekly
download count of around 400 downloads. A limitation of Grinder is the lack of an organised community
of volunteers to support this tool.
Current Version: 3
License type: Custom, BSD-style license.
C.11.2.2. Apache JMeter
JMeter [353] is a distributed load testing tool that offers similar features to Grinder in terms of load
injection capabilities. If combined with Markov4JMeter [354], the tool can also feature probabilistic
behaviour. Interestingly, the tool also features native support for MongoDB (a NoSQL database, see
Section D.4). The tool excels in extensibility, which is achieved through plug-ins. A limitation of JMeter is
that the web navigation may not be representative on websites where Javascript/AJAX play a major role for
performance, since JMeter does not feature the same range of Javascript capabilities as a complex browser.
JMeter has support of the large Apache Foundation community.
Current Version: 2.13
License type: Apache Public License 2.0 [187]
Copyright © 2015, DICE consortium – All rights reserved
69
Deliverable 1.1. State of the art analysis
C.11.2.3. Selenium
Selenium [355] is a popular library that automates the control of a web browser, such as Internet Explorer
or Firefox. It allows to execute tests using a native Java API. Furthermore, test scripts can be recorded by
adding Selenium plug-in to the browser and manually executing the operations. An extension, formerly
called SeleniumGrid and now integrated in the stable branch of the main Selenium tool, allows to perform
distributed load testing.
Current Version: 2.13
License type: Apache Public License 2.0 [187]
C.11.2.4. MODAClouds MDload
The MDload tool [356] is a Java stress testing application built upon the Selenium library. The
distinguished feature of the tool is the generation of requests driven by an engine capable of a layered
model describing user behaviour and resource topology, which can be used online to decide the next
request to issue to the system. MDload offers the capability to conduct a multi-threaded test using local
Mozilla Firefox browser. The tool acts as a closed-loop workload generator, where clients need to complete
requests before issuing new requests. Requests can also be injected via bursts, for increased realism. A
limitation of MDload is the lack of integration with the distributed stress testing capabilities of Selenium.
Current Version: 1.0
License type: BSD-3 [223]
C.11.2.5. Chaos Monkey
Chaos Monkey [357] is an open source service that runs on Amazon Web Services [192]. This tool is used
to create failures within Auto Scaling Groups [358] and to help identify failures within cloud applications.
The design is flexible and can be expanded to work with additional cloud providers and provide additional
functionality. The service has a fully configurable schedule for when actions will be taken and by default
runs on non-holiday weekdays between 9am and 3pm. A similar tool has been developed within .NET [71]
for Microsoft Azure [359] called WazMonkey [360] which offers similar functionality as Chaos Monkey
such as rebooting or reimage role instances within a given Azure deployment at random.
Current Version: 2.4.0
License type: Apache Public License 2.0 [187]
Table 16 provides comparative summary of testing tools described in detail in the sections C.11.2.1C.11.2.5.
Table 16: Comparative summary of testing tools.
Type
Distributed
Testing
User Behaviour
Model
Traffic Model
Grinder [350]
Performance
Y
Y
Y
JMeter [353]
Performance
Y
Y*
Y
Selenium [355]
Performance
Y
N
N
MDload [356]
Performance
N
Y
Y
ChaosMonkey
[357]
Reliability
Y
N
N
*: via the Markov4JMeter plug-in.
Copyright © 2015, DICE consortium – All rights reserved
70
Deliverable 1.1. State of the art analysis
C.11.3 Monitoring tools
During recent years a large number of companies/organisations tried to move part or even entire
infrastructure to the cloud. This trend can be observed due to the new easy ways of delivering applications,
process large workloads and even empower the end-user to work with some model of the cloud.
Although the trend is so that lots of organisations tend to move their applications to the cloud, there are still
problems that have to be addressed in order to ease usage of cloud technologies. One of the problems that
have to be addressed is related to the way one can monitor the cloud in order to provide the user with
aggregated information about cloud behaviour in different situations. The major problem that arises in the
context of monitoring is which metrics have to be monitored. Besides this problem one has to figure out
what sort of monitoring solution fits the cloud model deployed.
Below we present a short overview of different monitoring solutions that can be applied in different cloud
deployments. We will start by presenting solutions that are based on Hadoop toolkit [361], then we’ll
provide a short overview of technologies used for monitoring NoSQL databases (see Section D.4 for the
technology description and analysis) and some other solutions used in different applications.
C.11.3.1. Hadoop toolkit
Hadoop Performance Monitoring User Interface (UI) [361] provides a Hadoop built-in solution for quickly
finding performance bottlenecks. It also can visually represent the configuration parameters which might be
tuned for better performance. The monitoring toolkit provides a lightweight monitoring UI for the Hadoop
servers. Although it is a lightweight solution, it allows its users to find out where and why bottlenecks
occur in different instances of Hadoop. The toolkit is already present in any Hadoop distribution, hence in
order to use it one only has to start the UI and do data queries.
License type: Apache Public License 2.0 [187]
C.11.3.2. SequenceIQ
SequenceIQ [362] provides a solution for monitoring Hadoop clusters. The architecture proposed in [363]
and used in order to do monitoring is based on Elasticsearch [364], Kibana [365] and Logstash [366].
The general picture of their proposed architecture as presented in [363], is shown in the Figure 14 below:
Figure 14. SequenceIQ high-level architecture [363].
The architecture proposed by [363] has the main objective of obtaining a clear separation between
monitoring tools and some existing Hadoop deployment. For achieving this they use three Docker [199]
containers.
Copyright © 2015, DICE consortium – All rights reserved
71
Deliverable 1.1. State of the art analysis
In a nutshell, the monitoring solution consists of client and server containers. The server container takes
care of the actual monitoring tools. In this particular deployment it contains Kibana for visualisation and
Elasticsearch for consolidation of the monitoring metrics. Through the capabilities of Elasticsearch one can
horizontally scale and cluster multiple monitoring components. The client container contains the actual
deployment of the tools that have to be monitored. In this particular instance it contains Logstash, Hadoop
[14] and the Collectd [367] modules. The Logstash connects to Elasticsearch cluster as client and stores the
processed and transformed metrics data there.
License type: Apache Public License 2.0 [187]
C.11.3.3. Hadoop Vaidya
Hadoop Vaidya [368] (Vaidya in Sanskrit language means ‘one who knows’, or ‘a physician’) is a rulebased performance diagnostic tool for MapReduce jobs. It performs a post execution analysis of
map/reduce job by parsing and collecting execution statistics through job history and job configuration
files. It runs a set of predefined tests/rules against job execution statistics to diagnose various performance
problems. Each test rule detects a specific performance problem with the MapReduce job and provides a
targeted advice to the user. This tool generates an XML report based on the evaluation results of individual
test rules.
License type: Apache Public License 2.0 [187]
C.11.3.4. Ganglia
Ganglia [369] is a scalable distributed monitoring system for high-performance computing systems such as
clusters and grids. It is based on a hierarchical design targeted at federations of clusters. It leverages widely
used technologies such as XML for data representation, XDR for compact, portable data transport and
RRDtool [370] for data storage and visualisation. It uses carefully engineered data structures and
algorithms to achieve very low per-node overheads and high concurrency. The implementation is robust,
has been ported to an extensive set of operating systems and processor architectures, and is currently in use
on thousands of clusters around the world. It has been used to link clusters across university campuses and
around the world and can scale to handle clusters with 2000 nodes.
License type: BSD-License [223]
C.11.3.5. Apache Ambari
The Apache Ambari [371] project is aimed at making Hadoop management simpler by developing software
for provisioning, managing, and monitoring Apache Hadoop clusters. Ambari provides an intuitive, easyto-use Hadoop management web UI backed by its RESTful APIs.
License type: Apache Public License 2.0 [187]
C.11.3.6. Apache Chukwa
Chukwa [372] is an open source data collection system for monitoring large distributed systems. Chukwa is
built on top of the Hadoop Distributed File System (HDFS) [33] and MapReduce framework and inherits
Hadoop’s scalability and robustness. Chukwa also includes a flexible and powerful toolkit for displaying,
monitoring and analysing results to make the best use of the collected data.
License type: Apache Public License 2.0 [187]
Copyright © 2015, DICE consortium – All rights reserved
72
Deliverable 1.1. State of the art analysis
C.11.3.7. Datastax-OpsCentre for Apache Cassandra
Datastax [373] provides a solution called OpsCentre [374] which can be integrated in order to monitor
Cassandra [375] installation. Using OpsCentre one can monitor different parameters of the Cassandra
instance and also different parameters provided by the actual machines on which it runs. Also, they expose
an interactive web UI that can allow administrators to add or remove nodes from the deployment. One
interesting feature provided by the OpsCentre is the automatic load balancing. A developer API is also
provided in order to integrate with other services.
License: Free for Datastax-OpsCentre (not bundled)
C.11.3.8. MongoDB (MMS)
One way of monitoring and keeping under control a MongoDB deployment is using MMS [376]. MMS is a
cloud service developed by the MongoDB team. It provides an integrated and easy way to provision,
monitor, backup and scale MongoDB on the infrastructure of choice. It also includes out of the box support
for Amazon Web Services (AWS) [192].
License: Commercial.
C.11.3.9. Server Density
The tool provides an interesting way of monitoring different systems. Besides the monitoring itself, it
provides a way to clearly visualise the data. One of the main advantages of Server Density [377] is its
versatility. Its implementation ranges from the monitoring of the basic user profile to the monitoring of the
high-throughput (over 30TB/month) processing of time-series data.
License: Commercial.
C.11.3.10.
Manage Engine
Applications Manager enables the users to monitor the performance of Cassandra [375] and also perform
administration tasks of all the nodes in a cluster in a centralised manner. One can collect different statistical
data from the JVMs that run in a cluster. Besides Cassandra the tools offered by Manage Engine [378]
cover a wide variety of other applications that can be monitored. When it comes to NoSQL databasess it
can also monitor MongoDB instances and others.
License: Commercial.
Table 17 provides comparative summary of monitoring tools described in detail in the sections C.11.3.1C.11.3.10.
Table 17: Comparative summary of monitoring tools.
Tool
Hadoop-toolkit
[361]
License
Apache [187]
SequenceIQ [362]
Multiple
Hadoop Vaidya
[368]
Apache
Ganglia [369]
BSD [223]
Monitored technology
Hadoop [14]
Tools used
-
Hadoop
Docker [199], Collectd [367], Logstash
[366], elasticsearch and Kigtbana [365]
Hadoop
-
Clusters
-
Copyright © 2015, DICE consortium – All rights reserved
73
Deliverable 1.1. State of the art analysis
Ambari [371]
Apache
Hadoop
-
Chukwa [372]
Apache
Hadoop
-
MongoDB
-
MMS (MongoDB)
[376]
DatastaxOpsCentre [374]
First 8 servers
free
Free*
Server Density
[377]
Pay
Manage Engine
[378]
Pay
C.12.
Cassandra [375]
MongoDB, others
Cassandra,
others
-
MongoDB, -
Monitoring Feedback Analysis
C.12.1 Tools for Detecting Anomalies with Machine Learning
The goal of this section is to review Machine Learning (ML) tools that may be applied to perform anomaly
detection on the monitoring data acquired during testing. One of the most well known tools for machine
learning is Weka [379]. It is an extensive collection of ML algorithms written in Java covering
classification, regression, clustering, outlier detection etc. It also features a package system that allows its
functionality to be extended with both official and non-official packages. Weka is not designed out of the
box to run on any distributed systems such as Hadoop [14] but there is a set of wrappers [380] available that
support MapReduce operations on top of Hadoop. Another interesting package built on top of Weka is
called MOA (Massive Online Analysis) [381]. It is the most popular open source framework for data
stream mining.
Scikit-learn [382] is a general ML framework developed in Python [128] with some code written in
Cython [383] for increased performance. It is built on top of existing math and science Python packages:
NumPy [384], SciPy [385] and matplotlib [386]. It has a wide variety of machine learning algorithms
similar in scope to that of Weka. It is arguably the most mature Python-based framework. It has been used
in a wide range of scientific and commercial applications.
KNIME (Konstanz Information Miner) [387] is an open source data analytics, reporting and integration
platform. It is written in Java and based on the Eclipse project. It has two important features. First, KNIME
integrates various other open source projects ML libraries and data manipulation utilities (such as Weka,
LIBSVM [388] and R [389]). The second feature is that its user interface is based on a visual paradigm
(workflows), not API.
ELKI [390] is another specialised ML framework. It aims to provide a suitable environment for the
development and evaluation of advanced data mining algorithms and their interactions with database index
structures. One of its most interesting features is the extensive collection of outlier detection algorithms.
The R language [389] is one of the most used environments among ML practitioners as the main goal of
this system is to provide an environment for statistical computing and graphics. R is easily extensible by
using packages. These packages can be written in Java [99], C/C++ [100] and Fortran [391]. The CRAN
package page [392] contains a wide variety of specialised ML algorithms. Several packages were
developed in order to enable Big Data processing within R, most of them being available under CRAN
packages page.
Copyright © 2015, DICE consortium – All rights reserved
74
Deliverable 1.1. State of the art analysis
Mathematical libraries such as Mathematica [393] and Matlab [394] have both an extensive ML libraries
and are used both in academia and commercial products. They even possess special plug-ins/libraries that
enable the implementation of MapReduce. For Mathematica it is HadoopLink [395] while for Matlab this is
Distributed MapReduce [396].
There are tools that are designed for specific problems. One such tool is Encog [397] which focuses on
neural networks. It features a wide array of neural network types and training algorithms. Although the
most widely used version of this framework is written in Java, there are .NET [71] and C++ [100] variants
as well. Another important characteristic of Encog is that it is one of the fastest ML tools for training neural
networks [398].
In the last 10 years a significant trend in ML is the use of General-Purpose Computing on Graphics
Processing Units (GPGPU) [399], which can in some instances outperform high-end CPUs (and bring
costs of equipment required for this training from $1M down to $20,000) [400], [401]. A significant
portion of these tools is geared towards deep learning scenarios. Major deep learning software frameworks
have incorporated GPU acceleration, including Caffe [402], Torch7 [403], Theano [404], CUDAConvnet2 [405], DeepLearning4J [406], cuDNN [407] and NVIDIA’s DIGITS [408].
Most of the tools listed above are not Big Data ready in the sense that they are not designed to handle large
datasets. Also, most frameworks are not horizontally scalable. In the following section we will detail
machine learning tools which are in fact Big Data ready and almost all of them are deployed on a
distributed environment making them horizontally scalable. Table 18 provides comparative summary of
ML anomaly detection tools described in detail above.
Table 18: Machine Learning tools.
Name
License
User Community
Language
Weka [379]
GPL 3
Large
Java [99]
BSD [223]
Large
Python [128]
GPL 3
Large
Java
AGPL 3 [186]
Moderate
Java
GPL 3
Large
C, Fortran [391]
Encog [397]
Apache 2.0 [187]
Low
Java, .NET [71]
Caffe [402]
BSD 2.0
Moderate
C/C++ [100]
Apache 2.0
Low
Python
Scikit-learn [382]
KNIME [387]
ELKI [390]
R [389]
Theano [404]
C.12.2 Distributed Machine Learning Platforms
H2O [409] and MLLib [410] are two of the most extensively developed projects of fully distributed and
scalable Machine Learning (ML) tools. Both feature distributed in-memory computations and are certified
both for Spark (MLLib being the part of Apache Spark [411]) and Hadoop [14] distributions. This inmemory capability means that in some instances this framework can outperform strictly Hadoop based
frameworks [412]. One important distinction is that each H2O node (which is a single JVM process) runs
as a mapper in Hadoop. There are no combiners or reducers. Also, H2O has more built-in analytical
features and a more mature REST-based interface for R [389], Python [128] and JavaScript [99] than
MLLib.
Copyright © 2015, DICE consortium – All rights reserved
75
Deliverable 1.1. State of the art analysis
Mahout [413] is an ML framework built on top of Hadoop and features an extensive collection of ML
algorithms. The latest version (0.9) supports a variety of different execution engines such as Hadoop [14],
Spark [411], H20 [409] and Fink [414]. It is widely regarded that Mahout is well-suited for high-latency
analytics and unsupervised learning tasks.
Jubatus [415] is a distributed computing framework specifically designed for online ML of Big Data. A
loose model-sharing architecture allows it to efficiently train and share ML models by defining three
fundamental operations: Update, Mix and Analyse [416]. It supports the scaling of ML processing of up to
100,000 or more data point per second. All data is processed in memory. The main difference between
Jubatus and Hadoop/Mahout is that the former can handle real-time deep analytics tasks while the latter is
not designed for these tasks.
Another distributed stream processing framework which focuses on real-time online machine learning is
Trident-ML [417] built on top of Apache Storm [19]. It processes batches of tuples in a distributed way
which means that it can scale horizontally. However, Storm doesn’t allow state updates to append
simultaneously which hinders distributed model learning. Learning step can’t be distributed and no
distributed learning algorithms are currently implemented for Trident-ML.
The Apache Oryx 2 [31] framework is a realisation of the Lambda Architecture [17] built on top of Spark
[411] and Kafka [21]. It is a specialised framework that provides real-time large scale machine learning. It
consists of three tiers: lambda, ML and application. The Lambda tier is further split up into batch, speed
and serving tiers respectively. Currently it has only 3 end-to-end implementations that implement the batch,
speed and serving layer (collaborative filtering, k-Means clustering, classification and regression based on
random forest). Although it has only these three complete implementations, its main design goal is not that
of a traditional ML libraries but more a Lambda Architecture-based platform for MLLib [410] and Mahout
[413].
Vowpal Wabbit [418] is an open source fast out-of core learning system currently sponsored by Microsoft
Research [419]. It has an efficient implementation of online ML. Learning is accomplished through
progressive validation. The so-called hash trick presented in [420] is implemented as the core data
representation which results in significant storage compression for parameter vectors. Vowpal Wabbit also
reduces regression, multi-class, multi-label, structured prediction etc. to a weighted binary classification
problem. This allows for significant computational optimisations.
One limitation of this framework is that some problems can’t be solved with the resources of a single
machine. However, a Hadoop-compatible computational model called AllReduce [421] was developed to
tackle this problem. For example, using this model, a 1000-node cluster was able to learn a terafeature
dataset in one hour [420]. Table 19 contains a list of fully distributed and scalable ML tools.
Copyright © 2015, DICE consortium – All rights reserved
76
Deliverable 1.1. State of the art analysis
Table 19: Distributed ML frameworks3.
Distributed
Environment
Licensing
Petuum
[422]
Open source
(Sailing Lab)
DL, CL, Reg, Metrics
Learning, Topic modelling
Clusters or on Amazon
EC2 [429], Google GCE
[430]
Moderate /
Large
C++ [100]
Jubatus
[415]
LGPL v2.1
CL, Reg, AD, CU, Rec,
Graph analysis
Zookeeper [20]
Moderate
C++
MLlib
[410]
(MLBase)
Apache 2.0
[187]
Reg,
CL,Rec, CU
Spark [411]
Large
Apache 2.0
Collaborative Filtering, CL,
Hadoop [14], Spark, H2O
CU, Dimensionality
Moderate
[409]
Reduction, Topic Models
Java
Oryx 2
[31]
Apache 2.0
Rec, CL, Reg, CU
Hadoop, Spark
Low
Java
TridentML [417]
Apache 2.0
CL,Reg, CU, Feature
Extraction
Storm
Low
Java
H2O [409]
Apache 2.0
DL, Reg, CL,CU, PCA
Hadoop
Moderate
Java
GraphLab
Apache 2.0
Create
[423]
CU, CL, Reg, DL, Rec
Hadoop,
Spark, MPI
High
C++
Vowpal
Wabbit
[418]
Ms-PL
CL, Reg, CU
Hadoop
Moderate
C++
Deep
Learning
4J [424]
Apache 2.0
DL
Hadoop,
Spark,
Moderate
Java,
Scala
MLBase
[425]
MIT License
[222]
CL
Julia [431]
FlinkML
[426]
Apache 2.0
CL, Reg, CU,
Rec
Flink Hadoop
Low
Scala
[433]
N/A
CU, Frequent Pattern
Mining, AD
Hadoop
None
Java
N/A
Reg, PageRank
Hadoop
None
DML
Mahout
[413]
NIMBLE
[427]
SystemML
[428]
ML Problem
Users
Community
Name
Language
Scala,
Java
Julia [432]
The tools mentioned in this section should not be considered as an exhaustive list of all the available ML
tools. Some tools which are specialised on a specific use case were not covered. One such tool is Ellogon
[434], specifically designed for natural language processing. Instead we focused on general ML
frameworks which can be adapted with relative ease to any DICE use case.
Another important consideration when choosing the ML tools is to consider the specific challenges that
arise in the application domain. For example in anomaly detection applications, one will most likely
3
CL=Classification, CU=Clustering, Reg = Regression, Rec = Recommendation, AD = Anomaly Detection, DL = Deep Learning
Copyright © 2015, DICE consortium – All rights reserved
77
Deliverable 1.1. State of the art analysis
consider both point anomalies as well as collective and context dependent anomalies. Because of this,
easily extensible libraries such as Weka [379], Encog [397] or ELKI [390] are of key interest to the project.
At this point it is not clear if a distributed ML engine is required in the DICE context but during the project
we will look into and possibly adapt distributed ML tools to fit the project's needs.
C.12.3 Model Parameter Estimation
Part of the feedback analysis concept focuses on reducing amount of manual labour for the software
developer when studying software performance analysis results and making decisions about
improvements/changes in the application code or design.
Van Hoorn et. al [435] presented Kieker – a tool that conducts performance analysis, reconstructs software
architecture model and automatically annotates it with analysis results. Additionally it presents
performance analysis results in graphical and textual form via web-based user interface. The tool can be
applied both to concurrent and distributed systems and currently implemented for a number of specific
platforms/technologies (Java [99], .NET [71] and COM). The tool does not provide automated anomaly
detection, leaving this task to the developer.
Filling-theGap (FG) tool, proposed and developed by Perez et. al [436] is similar to van Hoorn’s [435] in
that it also can be deployed both in concurrent and distributed systems, automatically updates performance
model and presents performance analysis results both in graphical and textual form. However, FG tool
employs UML diagrams and is platform and technology-independent. This is an ongoing work and
currently the tool does not support automatic anomaly detection, neither support to improvements in the
application code or design (leaving interpretation of analysis results to the developer).
Calinescu et. al [437] developed a tool for continual verification of service-based systems. The proposed
method solves a formal model (discrete time Markov chains obtained from the software model via modelto-model transformation), automatically annotates performance model with results and verifies if all model
parameters are within required values. If some thresholds are violated, algorithm chooses another service
from the pool of available services so that all performance requirements could be satisfied. In this way
developer’s involvement is minimal and is required only if algorithm cannot find any service within a pool
to satisfy performance requirements. Then the warning is issued and developer should make a decision, e.g.
change requirements or re-factor model/application code. Table 20 provides a summary of the tools
discussed above. Automated anomaly detection, automated anomaly solution, automated model update and
the option of reporting to the developer/architect are chosen as dimensions for comparison to indicate if
manual labour is required in the tool application. Dependence on the specific technology/platform might
become a challenge in the uptake of a tool. Finally, the attempt is made to assess if these tools can be used
in Big Data applications.
Table 20: State of the art in the area of feedback analysis tools.
Automated
anomaly
detection
Automated Automated
anomaly
model
solution
update
Reports
results to the
developer?
Platform/system
independent
Can be used in Big
Data applications?
Van Hoorn
et. al [435]
N
N
Y
Y
(text, visual)
N
Likely
Perez et. al
[436]
N
N
Y
Y
(text, visual)
Y
Likely
Calinescu
et. al [437]
N
N
Y
Y (warning)
N
Unclear, due to large
parameter space
Copyright © 2015, DICE consortium – All rights reserved
78
Deliverable 1.1. State of the art analysis
C.12.4 Trace Checking
Trace checking is a procedure for evaluating a formal specification against a log of recorded events
produced by a system. Traces are produced at runtime by a monitoring engine, which collects events of the
system and saves them to a storage while running. According to Bauer et al. [438], run-time verification
consists of all the verification techniques which allow for checking whether a certain execution of a system
satisfies a given property, and trace checking is a run-time verification technique as long as only one (or a
finite number of) execution of the system is analysed at a time. Therefore, runtime verification mainly
focuses on the detection of violations of properties occurring in the observed system executions.
Trace checking can be performed online or offline, depending on the moment when the trace checking
algorithm is executed and the moment when the result of the processing is available for the analysis. The
online approach includes procedures such that the trace to verify is the current execution of the system and
the analysis carries out while the system is running. The result of the verification is immediately considered
upon the procedure terminates and allows for checking the integrity of the system and the adherence to the
requirements. The offline approach prescribes that the verification is done on previously recorded traces
which correspond to past executions of the system.
Logic-based formalisms were inherited from model-checking and became very popular also in run-time
verification to specify properties. Linear Temporal Logic (LTL) [439] is employed in [440], while
extensions of LTL over continuous-time are considered in [441], which studies monitoring procedures of
LTL properties enriched with special operators dealing with time, and in [442] and [443] which devise
monitoring algorithms for Metric Temporal Logic (MTL) specifications [444].
The standard approach to deal with trace checking is to build a monitor which verifies the correctness of
the trace with respect to a certain property (e.g., ‘the average/maximum number of events per hour in the
last ten hours is less than K’). In general, the time and memory complexity of the procedure building the
monitor is small, as the monitor is only generated once for each specification to verify. On the other hand,
logs in Big Data applications are rather extensive and procedures which formally verify properties on them
can be rapidly limited by time and memory requirements.
Dealing with huge amounts of data increased the research effort towards the design of procedures for runtime verification leveraging parallel computational mechanisms. [445]shows the relationship between trace
checking and Boolean circuits for which there are efficient algorithms to evaluate the output if their graph
satisfies some specific restrictions. Although for general LTL formulae trace checking cannot be expressed
with such a class of circuits, parallelisation with respect to sub-formulae can be efficiently achieved. The
decidability of the trace checking problem over timed trace is studied in [446].
The consolidation of Big Data approaches and frameworks, started from the advent of MapReduce in
industrial and mainstream applications, have led the community to a mature stage where parallelisation is
adopted in trace checking [295] and run-time verification [447]-[449]. All adopt first-order relations over
finite domains to represent the events in the traces which, based on the terms occurring therein, can be split
into several unrelated partitions that are then verified in parallel by independent threads. These approaches
rely on splitting traces on data but not on the structure of the formula. MapReduce for trace checking is
adopted to evaluate LTL formulae over traces in [450] and in [451]. The latter employs MapReduce to
check traces with MTL specifications by parallelising the evaluation of the formulae on their syntactic
structure. The memory scalability depends on the size of the temporal intervals occurring in the formulae.
[452] addresses this limitation by exploiting a special semantics and a parametric decomposition of
formulae which allow the evaluation of the truth value of sub-formulae to consider only bounded portions
of the trace for minimising the memory usage.
Copyright © 2015, DICE consortium – All rights reserved
79
Deliverable 1.1. State of the art analysis
D.
Data-Intensive Technologies
D.1.
Overview
This section introduces the technologies relevant to the DICE framework. For each technology we
emphasise three dimensions:

General characteristics: In this subsection the innovative aspects and typical application domains
of the technology are explained. The description is accompanied by a diagramme showing a typical
architecture of this technology. Where possible, a list of public cloud offerings available for this
technology is also reported together with advertised costs and quality. As the DICE project targets
open source framework, open source solutions for adoption of this technology in private clouds are
also described. If known, pointers to existing Chef [49] recipes to automate deployment of the
solutions are provided.

Quality assurance: For quality assurance aspects a list of key monitoring metrics used to assess
quality in this technology is given. If possible, the description focuses on unique metrics specific to
this technology. The main quality assurance challenges are described for this technology, referring
to the research literature. In case there exist a native support and configuration options for
reliability and high-availability, it is also detailed, as well as native support and configuration
options for scalability and performance, for hard deadlines and real-time processing, and for
privacy and data protection.

Models (if it applies): Finally, if there exist meta-models that explicitly address this technology,
relevant citations are provided. Also, if there exist Quality of Service (QoS) prediction models that
explicitly address this technology, relevant references are provided.
The technologies and concepts covered in this chapter are listed below:






D.2.
Hadoop/MapReduce/Spark (Section D.2).
Streaming (Section D.3).
NoSQL (Section D.4).
Software Defined Networking (Section D.5).
Cloud-based Blob Storage (Ceph and Amazon S3) (Section D.6).
In-memory Databases (e.g. HANA) (Section D.7).
Hadoop and Spark
D.2.1
Overview
Below we discuss the main technological solutions adopted in the industry to develop and execute DataIntensive applications to cope with variety (unstructured data) and volume dimensions: Apache Hadoop
[14] and Spark [411]. Both frameworks can exploit a common infrastructure for data distribution (relying
on the distributed file system HDFS [33]) and for resource management (that can be handled by the
resource negotiator YARN [18]). For introduction purposes we provide an overview of HDFS and YARN
first.
The Hadoop Distributed File System (HDFS) [33] transparently distributes data across a computational
cluster. Every machine runs a Data Node agent, which is responsible for handling blocks, i.e. data chunks
belonging to files in the file system (usually of the size around 64/128 MB), whilst a central Name Node
holds metadata that map blocks to files. For fault-tolerance reasons, blocks are replicated three times by
Copyright © 2015, DICE consortium – All rights reserved
80
Deliverable 1.1. State of the art analysis
default and possibly one replica is placed on a different rack with respect to the other two. When files are
written to HDFS, the Name Node provides a list of Data Nodes that will process the corresponding blocks.
The client then sends data to the first Data Node, which further forwards replicas in a pipelined fashion to
the subsequent Data Nodes. Name Node presents a single point of failure in the HDFS. Therefore, for
reliability reasons it is common practice to run a second Name Node in passive state. It periodically saves
snapshots of file system metadata in order to provide a failover alternative in case the active Name Node
fails.
Hadoop YARN [18], acronym for Yet Another Resource Negotiator, was developed to decouple resource
management and application semantics, as well as to address scalability issues brought by the resource
management model implemented in Hadoop 1.x. Initially, Hadoop relied on a central Job Tracker to
perform tasks ranging from capacity allocation to progress monitoring, including checking possible failures
or performance degradation, and addressing them. All these duties resulted to be overwhelming for a single
node in very large clusters. The issue is now addressed in YARN by distributing responsibilities among
several entities. A first class of components - the Node Managers - handles resources locally activating
containers, an abstraction for a slice of the resources available. The central Resource Manager has the role
of allocating resources for the execution of jobs, based on a configurable scheduling policy. However, all
the duties related to application semantics, such as restarting failed tasks, tracking progress, performing
speculative execution, among the others, are delegated to Application Masters, which are applicationspecific components. When a job is submitted for execution, the Resource Manager bootstraps a container
for the corresponding Application Master. Then, the Application Master requests a number of containers to
perform the computation. According to the current state of the cluster and the scheduling policy, the
Resource Manager provides a token-authenticated lease that the Application Master will use to activate the
obtained containers at Node Managers.
After this introductory description of the common infrastructure, we present in detail the two considered
frameworks: Hadoop and Spark.
D.2.2
Apache Hadoop
D.2.2.1. Overview
Hadoop [14] project started as an open source implementation of MapReduce framework. Hadoop
framework allows for parallel and distributed computing on large scale clusters of commodity hardware,
focusing on batch processing of huge datasets. Furthermore, it guarantees beneficial properties of faulttolerance, scalability, and automatic parallelisation and distribution of computation across the cluster at the
expense of a simple and rigid programming paradigm. MapReduce jobs are composed of two main phases,
namely, Map and Reduce. The former takes as input unstructured data from HDFS [33], filtering them and
performing a preliminary elaboration according to the instructions in a user-defined function. The
intermediate results are returned as key-value pairs, grouped by key in the Shuffle phase and distributed
across the network, so as to provide each node taking part in the Reduce phase with all the values
associated with a set of keys. In the end, every Reduce node applies a second user-defined function to
complete data elaboration and outputs to HDFS.
Upon the same infrastructure, the Apache Tez [453] project allows for a less rigid programming paradigm.
Applications can be expressed as directed acyclic graphs, where vertices represent data elaboration tasks
and edges represent data dependencies among tasks. Depending on the application graph, Tez
automatically launches tasks as soon as their input data are available.
There are four main comprehensive Hadoop distributions available for adoption in private clouds:
Hortonworks Data Platform (HDP) [454], Cloudera CDH [455], MapR [456], and IBM BigInsights [457].
Copyright © 2015, DICE consortium – All rights reserved
81
Deliverable 1.1. State of the art analysis
The first is composed of only open source software and is the only distribution that runs both on Windows
and Linux. CDH offers proprietary software alongside the main open source projects to perform cluster
management tasks, whilst MapR also adds closed source features to Hadoop. IBM BigInsights, on the other
hand, integrates software from the open source Hadoop stack with enterprise-grade IBM solutions, such as
Big R [458] and Big Sheets [459]. In literature there are not any up-to-date comparative studies on these
solutions. However, they differ in terms of installed software stack: all of them include Hive [460], Pig
[461], and HBase [462], SQL-on-Hadoop [463] projects, but CDH and BigInsights do not provide Storm
(see Section D.3.4.1) and Tez, and HDP does not provide Spark [411]. Only HDP and BigInsights provide
the Ambari [371] management platform. HDP, CDH, and MapR can be deployed with outdated Chef
cookbooks available on GitHub [464]-[466].
D.2.2.2. Hadoop public cloud offerings
This section provides an overview of the offerings available from the three public cloud providers having
the largest market shares (i.e. Amazon, Microsoft and Google) with an additional overview of Flexiant path
to release Hadoop services.
Amazon provides Elastic MapReduce (EMR) [29], which automatically deploys a Hadoop cluster on EC2
[429] virtual machines. Users are billed on a pay-per-use basis, with a price per instance and hour ranging
from $0.058 to $0.756 for general purpose instances at the time of writing [467], depending on their
computational capabilities. Data can be stored on HDFS [33] hosted on ephemeral storage, at the cost of a
lower scalability. EMR distinguishes two main classes of nodes: one performing computation and data
storage, whilst the other only contributing for computational purposes. This forces the number of instances
in the former class to never decrease to avoid data loss. Further, if the total number of data nodes is
increased, HDFS is not automatically rebalanced, hence existing datasets do not benefit from the added
capacity. Alternatively, it is possible to use EMRFS [468], a custom version of HDFS hosted on top of
Amazon S3 (see Section D.6.2), working around such limitations and allowing for data persistency. With
this approach, nodes in the EMR cluster are only in charge of computation and access of S3 space, thus
allowing for dynamically scaling. Both storage solutions are billed hourly.
Microsoft offers HDInsight [469] – a Hadoop cloud service backed by Azure [359]. The pricing
mechanism is similar to Amazon EMR, with prices per instance-hour ranging from $0.08 to $1.41.
Differently from EMR, HDInsight clusters use all the attached instances as data nodes, hence they can be
easily scaled up, but can only be scaled down by restarting the currently running jobs. This poses a threat of
data loss due to the HDFS Data Nodes that are shut down in the process. Alternatively, as for the Amazon,
it is possible to use Azure Blob [470] Hadoop-compatible containers for storage, allowing for a
dynamically variable cluster size without data loss concerns.
Finally, also Google Compute Engine [430] offers Hadoop cluster, but this service is geared towards
development or test environments and is not as straightforward as the previous alternatives.
Within DICE, Flexiant goal is to build the necessary feature set to replicate the AWS EMR service in
conjunction with Flexiant Bento Boxes [471] that can be leveraged to spin up/down the Hadoop cluster. In
particular, Flexiant Cloud Orchestrator [180] will provide the following:



The means to access an object storage service to host the input files.
The means to template a Hadoop cluster with all the required nodes roles (JobTracker,
TaskTracker, NameNode, DataNode) to provide both HDFS [33] and MapReduce functionality.
This could be leveraged by enhancing the Bento Box feature, including the ability to specify the
number of workers to create from a single OS image containing the Hadoop framework
The means to submit a job to the Hadoop cluster, ideally picking it from the object storage service.
Copyright © 2015, DICE consortium – All rights reserved
82
Deliverable 1.1. State of the art analysis


The ability to tell the Hadoop cluster where to place the output files (again into the object storage
service).
The ability to destroy the cluster once the job is complete.
The object storage service will be provided externally by software like Ceph (see Section ) or Swift [472]
but Flexiant Cloud Orchestrator will retain the Control Panel functionalities that are required to configure
the service and possibly to browse its content. The OS image used by the cluster will be prepared by
Flexiant in conjunction with a Hadoop expert partner and using popular frameworks such as the one
delivered by Hortonworks [454].
D.2.2.3. Hadoop Quality metrics
The key metrics to assess the quality of MapReduce jobs are throughput and compliance with deadlines
[473], [474].
The Hadoop framework provides reliability guarantees by monitoring the execution of Map and Reduce
tasks. In case of failure of a node, all the hosted Map tasks are automatically restarted on other nodes
processing other replicas of the failed blocks, whilst Reduce tasks are transparently restarted and fetch
again the associated values across the network. In this way, the framework efficiently handles failures recomputing only the needed portions of intermediate values or final results.
D.2.2.4. Hadoop meta- and QoS models
Meta-models and high level language for specifying Data-Intensive applications have been provided in
[475], [303], which can be further translated into Stochastic Petri Nets for performance analysis. The
authors’ proposed framework supports what-if analysis for Hadoop based systems.
Several QoS models for Hadoop and HDFS can be found in literature, which can be listed in three main
categories: empirical models mapping data size into execution times, approximate formulae evaluating jobs
executions time starting from individual tasks execution time and formal models like queueing network and
colored Petri Nets.
In the first category, Herodotou [476] provides formulae that take into account every main contribution to
total execution times. Anyway, the author did not perform any validation of the proposal. Zhang et al. [477]
use regression methods to estimate job completion times based on dataset size and available capacity.
Verma et al. [478] propose the ARIA framework for MapReduce job completion time prediction, adopting
job profiles extracted from previous executions to predict performance under different resource allocations.
Zhang et al. [479] build upon these findings to derive a new model that takes into account both userdefined functions via job profiles, and generic platform performance linked to framework code.
Malekimajd et al. [480] extend the formulae proposed in [478] to the multiple users case and apply them to
optimise Hadoop cluster configuration in public clouds. Yang and Sun [481] propose a model that
separately considers job and system impacts on performance. Their results are not accurate, but under
varying configurations show a realistic trend, which might help in optimising cluster resources utilisation.
Lin et al. [482] propose a detailed model taking into account low level aspects of the framework
mechanism, but validate it only in the simplistic case of single job execution.
At last, within formal models, Lin et al. [483] propose a tandem queues network with overlapping phases to
predict MapReduce jobs performance, which correspond to the overlapping of the Map and Shuffle phases.
Tan et al. [484] consider Reduce-intensive jobs and model the Map phase with an M/G/1 queue, whilst the
Reduce phase with a multi-server queue. Bardhan and Menascé [485] adopt mean value analysis to solve a
queueing network model and predict the Map phase duration. Vianna et al. [486] develop a hierarchical
Copyright © 2015, DICE consortium – All rights reserved
83
Deliverable 1.1. State of the art analysis
model coupling a precedence tree and a queueing network and then apply approximate mean value analysis
to estimate service time.
An attempt to model HDFS with Timed Coloured Petri Nets has been carried on in [487]. Through
simulation they can also obtain an optimal number of replicas to provide a sufficient reliability of the
system in terms of read and write successful operation. Castiglione et al. [488] used a Stochastic Petri Nets
to model MapReduce applications performance and proposed a novel mean field analysis evaluation
method.
D.2.3
Apache Spark
D.2.3.1. Overview
Apache Spark [411] started as a research project at UC Berkeley to address the performance issues related
to the use of the MapReduce paradigm in applications that imply iterative algorithms or interactive
processing [489]. Spark is an in-memory framework based on the novel concept of Resilient Distributed
Datasets (RDD), i.e. read-only, partitioned collections of records. Both for consistency and performance,
RDDs can only be created through transformations, i.e. bulk operations applied to data sources or parent
RDDs. It is also possible to request actions on RDDs, such as counting elements or extracting other
aggregate values from the datasets. Among the available transformations there are Map and Reduce tasks,
but also more complex operations such as union, groupByKey, etc. Every RDD is immutable, may be
required to persist, and stores information about its lineage. RDDs are lazily created when an action is
requested on them: the framework obtains a graph from the chain of dependencies and launches
computation accordingly, possibly pipelining subsequent elaboration on RDDs with narrow dependencies.
Given its very good performance the use of Spark is advocated for batch but also for interactive job
analysis and streaming. Further, Spark supports interfaces for Python [128], R [389], and Scala [433] (R is
partially supported for streaming analysis).
D.2.3.2. Spark public cloud offerings
Contrasting to Hadoop, the main public cloud providers do not feature readily available Spark services. On
the other hand, both Amazon and Microsoft provide instructions about the customisation of their cloud
Hadoop clusters to run Spark. In the case of Google instead there is a publicly available third-party script
allowing for easy deployment of Spark clusters on Compute Engine [490]. Databricks [491] - the software
company founded by Spark developers - offers automated deployment and management of Spark clusters
on Amazon AWS [192]. Unfortunately, pricing is not clearly stated on the company website.
Among the above mentioned Hadoop distributions, all but Hortonworks [454] also comprise Apache Spark
as part of their software stack.
Apache Spark guarantees reliability exploiting information about the lineage of RDDs. In the event of node
failures the framework identifies missing partitions and schedules re-computation only of those.
Dependencies in the lineage are used to obtain a graph of required transformations and this is followed up
to the closest available ancestor.
D.2.3.3. Spark Quality metrics and models
The key metrics to assess quality of Spark jobs are, as for Hadoop, throughput and compliance with
deadlines [492].
Being a rather young and poorly documented project, no performance models are available yet. However,
in literature it has been observed that the shuffle phase is a bottleneck and some works investigate the
optimisation of specific parts of the framework, as in [493], [494] even if some recent works have shown
Copyright © 2015, DICE consortium – All rights reserved
84
Deliverable 1.1. State of the art analysis
how business intelligence workloads are mainly CPU-bounded and little improvement can be achieved if
network and I/O in general are improved [495].
D.3.
Streaming
D.3.1
Introduction
Stream processing is a technology designed to collect, integrate, analyse and visualise data streams or
sensor data with a high ratio of event throughput versus number of queries, all in real-time as the data is
being produced.
Stream processing solutions are designed to handle high volume in real time without disrupting the activity
of existing sources and enterprise systems and with a scalable, highly available and fault-tolerant
architecture [496], [497].
Use cases and application domains:







Telecom: Security breaches, network outages, bandwidth allocation...
Financial services: anti-fraud, operational risks, order routing, pricing...
Retail: shrinkage, stock outs, offers, pricing...
Manufacturing: preventive maintenance, quality assurance, supply chain optimisation...
Transportation: driver monitoring, predictive maintenance, routes, pricing...
Web: Application failures, operational issues, personalised content
Research: Digital signal processing, databases, operating systems and networks, complex event
processing, machine learning...
D.3.2
Stream processing architecture
Streaming architectures are typically organised in three layers [498]:
1. Data collection: A large scale data collection layer (Sensors, social media monitoring, financial
market data, business transactions, machine-to-machine data etc.).
2. Real-time stream processing.
3. Data output: Redirect the processed data to other sub-systems for further processing.
Data collection and data output usually is performed through a highly efficient message broker such as
Kafka [21] or RabbitMQ [499]. Figure 15 shows a typical streaming processing solution [500].
Figure 15. Architecture of a typical streaming processing solution [500].
Copyright © 2015, DICE consortium – All rights reserved
85
Deliverable 1.1. State of the art analysis
D.3.3
Public cloud offerings
Table 21 lists some of the most important public and commercial cloud offerings with some streaming
components integrated.
Table 21: Public and commercial cloud offerings for stream processing.
Product
Microsoft HD Insight
[501]
AWS Kinesis [504]
Databricks [491]
Components
Free
trial
Pricing
One
month
Hadoop + Yarn + Spark +
Storm [502]
Difficult to
calculate [503]
Kinesis API + Integration
with AWS [192] and Storm
Depends on the
amount of records No
streamed [505]
Black boxed
components available
through APIs
Spark
Not publicly
available
One
month
-
Kafka + Spark + HBase [455] -
-
Based on Apache
components. More
private cloud oriented
Dataflow SDK + Google
Cloud Platform (Kafka,
RabbitMQ [499], HBase
[462], …)
Two
months
Beta
Cloudera [506]
Google Cloud Dataflow
[27]
Observations
Difficult to
calculate
Based on Apache
components
As a conclusion, most public clouds offerings are based on open source software components, mostly
provided by the Apache foundation, while others provide custom APIs and SDKs for the streaming
process.
D.3.4
Open source solutions.
A detailed view of some of the most important streaming solutions is described below:
D.3.4.1. Apache Storm
Apache Storm [19], [507] is a distributed real-time computation system for processing large volumes of
high-velocity data.
Features:





Fast: one million 100 byte messages per second per node.
Scalable.
Fault tolerant: workers automatically restarted.
Reliable: guarantees every message is processed.
Easy to operate: ready for production.
Figure 16 shows a typical Storm architecture [508].
Copyright © 2015, DICE consortium – All rights reserved
86
Deliverable 1.1. State of the art analysis
Figure 16. Typical Storm architecture [508].
D.3.4.2. Apache Spark Streaming
Apache Spark Streaming [509] is a batch Big Data processor that provides a module for streaming
processing.
Features:



API: supports Java [99], Scala [433] and Python [128].
Fault tolerance: recovers both lost work and operator state.
Integration: Allows reuse the same code for batch processing.
Figure 17 shows a typical Spark architecture [508].
Figure 17. Typical Spark architecture [508].
Copyright © 2015, DICE consortium – All rights reserved
87
Deliverable 1.1. State of the art analysis
D.3.4.3. Comparison between Storm and Spark
For what concerns streaming analysis, the main difference between Storm and Spark resides in the different
processing models implemented. Storm processes the incoming pieces of information (tuples) one at a time
and this leads to a lower latency but also to a lower throughput compared to Spark which is based on the
mini-batch analysis paradigm. In a nutshell, a mini-batch is a collection of tuples arrived at the systems
during a (short) time window. This approach raises the latency to the order of seconds.
Moreover, in Storm each tuple is acknowledged back to the source implementing the at-least-once fault
tolerance mechanism. This means that Storm cannot avoid considering twice the same tuple while
recovering from a fault. Spark, instead, tracks the data at batch level guaranteeing to analyse each batch
exactly once. From the programming languages point of view, Storm is developed in Clojure [352], while
Spark in Java and Scala [433]. Storm natively supports a larger set of languages but many third-party
projects aiming at extending Spark-supported languages are available. Storm is undoubtedly more widely
adopted in real-life deployments, whereas only one company officially uses Spark in production. Finally,
although both solutions can run on their own and on Mesos clusters [15], Spark also supports YARN [18]
natively.
This short comparison, summarised in the Table 22, is inevitably partial and circumstantial. The
performance, for example, greatly depends on the version and configuration of the tools. Moreover,
recently Storm introduced Trident [417], which is a solution for the mini-batch processing that implements
many features that are typical of Spark.
Table 22: Storm vs Spark comparison.
Storm [19], [507]
Spark [411]
Stream type
Stream processing (one record at time)
Mini-batching
Latency
Less than a second
Seconds
Fault tolerance mechanism
At least once (may have duplicated)
Exactly once
Implementation language
Clojure
Scala and Java
Directly supported languages Java, Scala, Clojure, Python, Ruby Java, Scala, Python
[103]
Production deployments
Many
Only one
Resource Management
Mesos
YARN, Mesos
Throughput
10k records/s/node
400k records/s/node
D.3.4.4. Apache Samza
Apache Samza [510] is a distributed stream processing framework. It makes use of Apache Kafka [21] and
Apache Hadoop YARN [18].
Features:




Simple API: provides a very simple callback-based ‘process message’ API comparable to
MapReduce.
Managed state: Snapshots and restoration of streaming processors’ state.
Fault tolerance: When a cluster fails, tasks are transparently migrated.
Durability: No messages are ever lost.
Copyright © 2015, DICE consortium – All rights reserved
88
Deliverable 1.1. State of the art analysis



Scalability: Partitioned and distributed.
Pluggable.
Processor isolation: Hadoop’s security model and resource isolation through Linux CGroups.
Figure 18 shows a typical Samza architecture [508].
Figure 18. Typical Samza architecture [508].
D.3.4.5. Other solutions
Other open source solutions not as mature as the above ones:







Hadoop streaming [511].
Apache S4 [512].
Streamparse [513].
Stratio [514].
Summingbird [515].
Facebook Scribe [516].
LinkedIn Pinot [517].
Comparison matrix of Streaming solutions discussed in detail in the sections D.3.4.1-D.3.4.4 is presented in
the Table 23.
Copyright © 2015, DICE consortium – All rights reserved
89
Deliverable 1.1. State of the art analysis
Table 23: Comparison of streaming processing frameworks [508], [518]-[520].
Technology
Apache Storm
Apache Spark
Apache Samza
Delivery semantics
at-most-once
at-least-once
exactly-once (with
Trident)
exactly-once
at-least-once
State management
stateless, stateful (with
Trident)
stateful
stateful
Latency
sub-second
seconds
sub-second
Language support
Java, Clojure and any non Scala, Java, Python
JVM language
Java, Scala, only JVM
languages
Maturity
***
**
**
Ease of use
**
***
*
Originates from
Twitter
AMPLab
LinkedIn
Companies using it
Groupon, Twitter,
Yahoo!, Spotify, Yelp
Alibaba Taobao, Amazon, LinkedIn, Intuit,
Autodesk, Ebay, Yandex MobileAware, Fortscale
Stream source (S)
spouts
receivers
consumers
Stream primitive (P)
tuple
Dstream
message
transformations windows
operations
tasks
Stream computation (C) bolts
D.3.5



D.3.6
Message queues
Apache Kafka [21]
Apache ActiveMQ/Apollo [521]
RabbitMQ [499]
General characteristics
D.3.6.1. Key monitoring metrics
 Latency, as the difference between the time message enter the system until it contribute to the final
results
 24x7 streaming
 I/O troughput
 Maximum delivery time
D.3.6.2. Main quality assurance challenges
Real-time processing is the next large step in Big Data analysis. Main difficulties have to be with the ease
of configuring and deploying testing environments and make them ready to be deployed in production
without breaking the streaming flow.
D.3.6.3. Reliability
Some streaming processors guarantee that each message will be fully processed. This is the case of Storm
[522].
Streaming processors are fault-tolerant: when workers die, they will be automatically restarted. If a node
dies, the worker will be restarted on another node [523].
Copyright © 2015, DICE consortium – All rights reserved
90
Deliverable 1.1. State of the art analysis
D.3.6.4. Scalability
Streaming processing solutions are inherently parallel and run across a cluster of machines. Different parts
of the processors can be scaled individually by tweaking their parallelism even on the fly [524], [525].
D.3.6.5. Efficiency
Streaming processing solutions can process more than 100 000 messages per second per node.
D.3.6.6. Privacy
Usually privacy and data protection is provided by other collateral components. For example, Apache
Ranger delivers a comprehensive approach to security for a Hadoop cluster. It provides central security
policy administration across the core enterprise security requirements of authorisation, audit and data
protection [526], [527].
D.4.
NoSQL
Traditionally, the Relational DataBase Management Systems (RDBMS) were the main type of databases in
use in the application. With the growth of the level of distribution of both storage and computation, and
with the ever growing amount of data, the NoSQL (Not only SQL) database model became more suitable
for certain data-intensive use cases. For example, Facebook, Google or LinkedIn support millions of
connection in each seconds. The data does not exhibit the strong structural and relational requirements, but
instead requires a quick and simple scaling up to handle increase of the data sources and the growth of the
users accessing the data.
NoSQL databases are data storage solution suitable for building very large and scalable web application
that have to deal with analysing and moving huge data sets. NoSQL refers to those distributed databases
that avoid the rigidity of the relational model in favor of other kind of data models. The reason to move to a
NoSQL database arises from a mismatch between the needs of modern DIAs compared to the following
characteristics of relational databases:
1. Rigidity of the data model. In a relational data model, data have a precise structure and all those
belonging to a certain relation have to share exactly the same structure. This means that, for
instance, it is difficult to represent, as part of an e-commerce application, situations in which we
have a catalog of heterogeneous products all having a size, but where the way to represent such
size can vary from by height, width and depth for one item and radius for others.
2. Difficulty of handling large quantities of data. Relational databases have been designed and built to
handle data that stay within a few Gigabytes. When considering Terabytes and Petabytes, which
are the targets for DIAs, they fail and should be replaced by simpler systems that make less strict
assumptions about the ACID properties of data.
The main kinds of not relational data models offered by NoSQL are the following:



Key value: each data item is identified by a unique key that is used to retrieve it. The actual data
item is stored as a sequence of bytes and is opaque to the NoSQL database.
Document-based: this model is an evolution of the key-value one. Each data item is still univocally
identified by a key, but the internal structure of the data is known by the database and it is not fixed
and equal for all data items.
Column-based: these databases owe their name to the data model proposed by Google BigTable
paper [528]. Data are stored inside structures named Columns, which, in turn, are contained inside
Column Families, and are indexed by Keys.
Copyright © 2015, DICE consortium – All rights reserved
91
Deliverable 1.1. State of the art analysis

Graph-based: in this case the data model follows the graph theory and allows adjacent nodes to be
directly connected one with the other. Such kind of data model is often adopted to represent highly
interconnected structures such as social networks.
The data models described above are quite different from each other, but generally aim at simplifying the
organisation of data to make them more suitable for distribution. Distribution and replication are in fact the
two mechanisms to enable scalability of the database with respect to the growth of the quantity of data to
be handled. More specifically, replication of data on different nodes increases tolerance to faults and
network partitions. Distribution of data allows the processing work to be split among different nodes so that
it becomes more manageable while the size of the whole data set grows. Distribution of data occurs by
partitioning them horizontally or vertically. Horizontal partitioning, also called sharding, means that
different data items can reside in different nodes of a database cluster. Vice versa, vertical partitioning
applies to columns-based databases and means that data distribution occurs by column.
In the field of distributed databases the CAP theorem [529] state that it is impossible for a distributed
database system to achieve at the same time the three properties of Consistency, Availability and Partition
tolerance, but just two of them can be guaranteed simultaneously. While classical centralised relational
databases aim at guarantee the consistency and availability properties, NoSQL databases tend to privilege
availability and partition tolerance. More specifically, they typically fulfill the BASE requirements (basic
availability, soft state, eventually consistent).
D.4.1
High level architecture of a NoSQL database
A typical architecture for a distributed NoSQL database is the one shown in Figure 19. The Figure
highlights the presence of different distributed nodes each one managing its local storage. Such nodes are
typically coordinated by a master node that receives the requests from clients and propagates them to the
others. Other nodes are used to manage the life-cycle of the others.
Figure 19. Architecture of a distributed NoSQL. This Figure refers to the HBASE architecture.
D.4.2
List of public cloud offerings available for this technology
Table 24 lists some of the DaaS (Database as a Service) available on the market. While their quality e.g., in
terms of number of entities written or retrieved in the time unit, or of possible parallel requests handled is
of crucial importance for an application developer [529], still such information is not available to users of
such services.
Copyright © 2015, DICE consortium – All rights reserved
92
Deliverable 1.1. State of the art analysis
Table 24: Public cloud offerings for NoSQL (Database as a Service).
Name
Data model
Cost
Amazon
DynamoDB
[531]
Key-value
and
Document
Free usage tier.
The prices below depend on the region.
Provisioned Throughput Capacity
 Write Throughput: less than $0.01/hour for every 10 units of Write Capacity
(approx. 36,000 writes/hour);
 Read Throughput: less than $0.01/hour for every 50 units of Read Capacity
(approx. 180,000 strongly consistent reads/hour, or 360,000 eventually
consistent reads/hour).
Indexed Data Storage
 First 25 GB stored/month is free;
 less than $0.4/GB/month thereafter
Reserved Capacity
 depends on the time; less than $0.0150 hour for 100 Write/Read Capacity Units
Data Transfer IN: free
Data Transfer OUT:
 first 1GB/month free of charge
 Up to 10 TB/month: $0.09/GB
 lower price for more transfer
Amazon
SimpleDB
[532]
Columnbased
Free usage tier
Machine Utilisation
 First 25 Amazon SimpleDB Machine Hours consumed per month are free;
 $0.2/Amazon SimpleDB Machine Hour consumed thereafter
Data Transfer IN: free
Data Transfer OUT:
 first 1GB/month free of charge
 Up to 10 TB/month: $0.25/GB
 lower price for more transfer
Structured Data Storage
 first 1GB/month free of charge
 less than $0.4/GB/month thereafter
Redis Labs
[533]
Key-value
There are available 10 types of virtual machines from different providers, for each of
them certain plan costs are available depending on the features provided.
Google Cloud Columnbased
Datastore
[534]
Azure
DocumentDB
[535]
Document
There are two types of costs: free cota/day, respectively prices if one exceeds the quotas
Stored Data:
 free 1GB total limit;
 after $0.18/GB/month
Read Operations:
 free 50;
 after $0.06/100k operations
Write Operations:
 free 50k;
 after $0.06/100k operations
Small Operations:
 free 50k;
 after free
It is billed based on the number of collections contained in a database account. There are
three types of collections S1, S2 and S3:
 S1: SSD Storage 10GB; Requests Units 250/sec.; Scale out limits up to 100;
SLA 99.95%; $0.034/hr (~$25/mo)
 S2: SSD Storage 10GB; Requests Units 1000/sec.; Scale out limits up to 100;
SLA 99.95%;$0.067/hr (~$50/mo)
 S3: SSD Storage 10GB; Requests Units 2500/sec.; Scale out limits up to 100;
SLA 99.95%;$0.134/hr (~$100/mo).

Copyright © 2015, DICE consortium – All rights reserved
93
Deliverable 1.1. State of the art analysis
Azure Redis
Cache [536]
Key-value
Azure Storage Multiple
[537]
It gives the ability to use a secure open source Redis cache, managed by Microsoft, to
build highly scalable and responsive applications by providing you super-fast access to
your data.
It is offered in two tiers:
 basic - a single cache node (ideal for development/test and non-critical
workloads);
 standard - replicated cache in a two-node Primary/Secondary configuration;
automatic replication between the two node; high-availability SLA.
There are six configuration types,two of them being, for example:
 C0: cache size 250MB; basic $0.022/hr (~$16/mo); standard $0.055/hr
(~$41/mo)
 C6: cache size 53 GB; basic $0.84/hr (~$625/mo); standard $2.10/hr
(~$1,562/mo)
Azure has four types of storage, each coming with two options standard, respectively
premium. We detail here the prices for the standard storage.
 block blobs, suitable for streaming and storing documents, videos, pictures,
backups, and other unstructured text or binary data. For example, for the first 1
TB / Month, prices are starting from $0.024/GB for LRS (Locally redundant
storage - maintaining three copies of your data) and gets to $0.061/GB for RAGRS (Read access geo-redundant storage - replicating the data to a secondary
geographic location, and also providing read access to the data in the secondary
location). In this category one can buy storage capacity up to Over 5,000 TB /
Month
 page blobs and disks, suitable for random read and write operations (VHD
images). For example, for the first 1 TB / Month, prices are starting from
$0.05/GB for LRS and gets to $0.12/GB for RA-GRS. In this category one can
buy storage capacity up to over 5,000 TB / Month.
 tables and queues; tables offer NoSQL storage for unstructured and semistructured data (suitable for web applications, address books, and other user
data). Queues provide a reliable messaging solution for your apps. For example,
for the first 1 TB / Month, prices are starting from $0.07/GB for LRS and gets
to $0.12/GB for RA-GRS. In this category one can buy storage capacity up to
over 5,000 TB / Month.
 files (preview): suitable for sharing the files between applications running in the
virtual machines using familiar Windows APIs or file REST API. For LRS
option, the price is $0.04/GB.
Their cost depends on how much one stores, the volume of storage transactions out data
transfer, which data redundancy option is chosen.
MongoLab
[538]
Document
There exist three types of predefined databases for various needs:
1. Sandbox - for development and prototyping, free
2. Shared - for small datasets and light workloads, $15 and more
3. Dedicated - large datasets and demanding workloads, $180 and more
There exists also other plans, the machines can be chosen to be hosted by different
providers (Amazon, Google, Microsoft Azure):
 dedicated cluster plans
o standard, e.g. RAM 68GB SSD 700GB $3520/month
o high storage, e.g. RAM 68GB SSD 1TB $3790/month
o high performance, e.g. RAM 61GB SSD 640GB $5890/month
 dedicated single-node plans, e.g. RAM 68GB SSD 700GB $2045/month
Rackspace
[539]
Redis – Key- Redis. There are three options for using Redis at Rackspace: Managed Cloud, Redis as a
value
Service and Private Cloud. For example, using Redis as a Service, one should pay from
$59 for 500MB up to $7499 for 100GB.
MongoDB. The same three usage options as above. For example, using MongoDB as a
MongoDB - Service, one should pay from $19 for 1GB/month and customized price for storage in the
Document
interval 100GB - 10PB.
Copyright © 2015, DICE consortium – All rights reserved
94
Deliverable 1.1. State of the art analysis
D.4.3
Open source solutions
List of open source solutions for adoption of NoSQL technology in private clouds is presented in the Table
25:
Table 25: Open source solutions for NoSQL.
Name
Website
Apache HBase http://hbase.apache.org/
Data model
Chef
Cookbook
Column-based Version 0.1.0
Cassandra
https://cassandra.apache.org/
Column-based Version 0.2.4
Hypertable
http://hypertable.org/
Column-based Version 0.3.2
Accumulo
http://accumulo.apache.org/
Key-value
Cloudata
https://github.com/gruter/cloudata
Column-based Not found
MongoDB
http://www.mongodb.org/
Document
Version 0.16.2
CouchBase
Server
http://www.couchbase.com/
Document
Version 1.3.1
CouchDB
http://couchdb.apache.org/
Document
Version 2.5.2
RethinkDB
http://www.rethinkdb.com
Document
Version 0.1.0
SequoiaDB
http://www.sequoiadb.com/en/index.php?p=index&j=2
Document
Not found
RavenDB
https://github.com/ravendb/ravendb
Document
Not found
Riak
http://riak.basho.com/
Key value
Version 3.1.0
Redis
http://redis.io/
Key value
Version 3.0.4
Aerospike
http://www.aerospike.com/
Key-value
In-memory
Version 0.0.12
LevelDB
https://github.com/google/leveldb
Key value
Version 1.18.0
Berkeley DB
http://www.oracle.com/technetwork/database/databasetechnologies/berkeleydb/
Key value
Not found
Oracle
NoSQL
Database
http://www.oracle.com/technetwork/database/databasetechnologies/nosqldb/overview/index.html
Key value
Not found
Dynomite
https://github.com/moonpolysoft/dynomite/wiki/dynomite- Key value
framework
Version 0.2.2
HamsterDB
http://hamsterdb.com/
Not found
Key value
Not found
Table 25 lists in column ‘Chef Cookbook’ the version supported by Chef [49] recipes.
D.4.4
Quality Assurance
D.4.4.1. Key monitoring metrics
There is not much work in the literature aiming at assessing the performance of NoSQL. Yahoo has defined
in 2010 a benchmarking framework by identifying a set of workloads useful to analyse the performance of
Copyright © 2015, DICE consortium – All rights reserved
95
Deliverable 1.1. State of the art analysis
NoSQL [540]. A very recent application of such benchmarks is reported in [541]. Another work going in
this direction is described in [529]. In general, the metrics being considered are the following:



Throughput by workload: number of CRUD (Create, Read, Update and Delete) operations
performed on the database in the time unit, considering different types of workloads.
Latency by workload.
Support to parallel connections: that is, ability of the database to support various client connections
without cutting them or showing unforeseen failures.
D.4.4.2. Main quality assurance challenges.
As already mentioned, the Brewer's CAP Theorem [529] states that in a distributed system it is impossible
to have at the same time availability, consistency and partition tolerance satisfied at the same level. NoSQL
databases tend to privilege availability and partition tolerance.
While most NoSQL databases and, in particular, the column-based ones have been developed to support
high scalability with respect to the growth of data, they delegate to the developer most of the activities
related to the optimisation of CRUD operations. For instance, if a developer wants to extract from a
NoSQL the data items fulfilling a certain complex condition, most likely, this developer will have to extract
and filter all data in the database in order to find the ones that are interesting to him/her. If such extraction
and filtering operations are programmed in a non-optimised way, all the advantages offered by NoSQL in
terms of scalability may be lost at the application level because of the execution of many operations on
such data.
D.4.4.3. Reliability
NoSQL stores use data sharding and replication by design. Several configuration options allow database
administrators to tune between performance and reliability, such as: number of replica (shards), number of
writes to consider a write operation successful, number of reads to consider a read operation successful etc.
D.4.4.4. Efficiency
Unlike relational database management systems that were designed initially as monolithic systems, NoSQL
stores are designed upfront as distributed systems. They support scale-out approach; they allow ease
addition or removal of nodes in/from the storage nodes cluster.
D.4.4.5. Safety
Database NoSQL have been built for offering high performance and for being available also in the presence
of network partitioning. However, they have not been developed for meeting hard-deadlines.
D.4.4.6. Privacy
NoSQL has not been designed with security as a priority. Many NoSQL products allow and even
recommend the use of a ‘trusted environment’ assuming that only trusted machines can access the
database's TCP ports. SSL connections are another measure to be enabled in order to offer better protection
at network layer.
At the application layer, products such as Riak [542] and MongoDB [543] already include support for
users’ authentication and authorisation, similarly to relational database management systems. To conclude,
security is a shared responsibility between level of protection offered by the product and the deployment of
the product within an organisation and it safe to say that the latest versions of NoSQL software are no lesssecure than their relational systems.
Copyright © 2015, DICE consortium – All rights reserved
96
Deliverable 1.1. State of the art analysis
D.4.5
Models
We are not aware of general meta-models focusing on NoSQL. In [544] a meta-model specifically
developed to support migration between column-based NoSQL of different vendors is defined.
D.4.5.1. QoS prediction models.
QoS research for NoSQL databases focuses on performance in terms of response times and throughput,
availability and scalability and consistency guarantees. Most studies that address performance, availability
and scalability are benchmarking studies with limited work in modelling, e.g. [545] and utilisation of
models in feedback loops, e.g. [546]. Research into predicting consistency guarantees, i.e., the probability
of retrieving stale data from quorum-like NoSQL databases, has been studied using Monte Carlo simulation
in [547].
D.5.
Software-Defined Networking
Big Data systems are usually massively parallel and require at some stage of the computation to transfer
partial results from one node to another (e.g. the copy shuffle phase of Hadoop and Spark). Data transfers
could become the bottleneck of the computation [548] and Software-Defined Networking (SDN) solutions
are widely adopted within data centres nowadays to configure the underlying networking infrastructure to
cope with such issues.
SDN is a new network paradigm based on the decoupling of the Control Plane from the Data Plane to make
computer networks more programmable [549]. SDN provides a global vision of all network nodes state
inside a logically centralised controller. Now programmers have the opportunity to easily create
applications that before had to be taught and developed over multiple nodes in a distributed way. All the
nodes had to be configured one by one in order to obtain the desired global network behaviour. It is much
easier reasoning over a complete network graph rather than on single nodes.
In the past, network devices were closed, proprietary, vertically-integrated systems with the drawback that
third-party programmes could not easily manage them. In addition the configuration interfaces varied from
vendor to vendor and even from device to device of the same vendor.
Through the creation of a software-based application, operators can control in an easier way the
infrastructure, having the ability to customise, optimise and deploy new services without having to upgrade
the underlying hardware.
D.5.1
Architecture
SDN architecture consists of three main layers: 1) Data plane, 2) Control plane and 3) Application layer.
Data plane includes the elements of the network infrastructure (physical and virtual switches). These
devices implement the forwarding behaviour dictated by the controller, which installs forwarding rules
through an abstract interface.
The Network Operating System (NOS) lies in the control plane. NOS is the intermediate layer between the
underlying network and the application layer. It gives a consistent abstract view of the network global state
and offers an interface for controlling the network to the application layer. The NOS acts like a standard
computer OS and abstracts the resources of the whole network to the applications executed on top of it.
The application layer is placed at the top of the SDN stack. It contains applications that provide network
services. The above-defined abstractions allow third-party developers to easily deploy new services in
different heterogeneous networks (data centres, WAN, mobile networks, etc.).
To allow the communication between the layers, well-defined interfaces are used:
Copyright © 2015, DICE consortium – All rights reserved
97
Deliverable 1.1. State of the art analysis



Southbound API: for the communication between the controller and the network infrastructure (e.g.
OpenFlow API).
Northbound API: for the communication between the network applications and the controller.
EastWest API: In the case of a multi-controller-based architecture, this interface manages
interactions between the various controllers.
Table 26 reports the most widely adopted open source SDN solutions that support OpenFlow [550], the
most used protocol for managing network devices. Two examples of SDN for WANs inter-data center are
B4 [551] and SWAN [552], deployed by Google and Microsoft respectively.
Table 26: Open source SDN solutions supporting OpenFlow.
Name
Description
OpenDaylight
[553]
A collaborative open source project hosted by The Linux Foundation. Its goal is to accelerate
the adoption of SDNs and create a solid foundation for Network Functions Virtualisation
(NFV).
OpenContrail
[554]
An Apache 2.0-licensed project that is built using standards-based protocols and provides all
the necessary components for network virtualisation–SDN controller, virtual router, analytics
engine, and published northbound APIs. It is the open source version of Juniper Networks
Contrail controller
Floodlight
[555]
The Floodlight Open SDN Controller is an enterprise-class, Apache-licensed, Java-based
OpenFlow Controller. It is supported by a community of developers including a number of
engineers from Big Switch Networks.
Ryu [556]
A component-based SDN framework. Ryu provides software components with well defined
API that make it easy for developers to create new network management and control
applications.
FlowVisor
FlowVisor is an experimental SDN controller that enables network virtualisation by dividing
a physical network into multiple logical networks.
Error!
eference
source not
found.
ONOS [558]
ONOS is the Open Network Operating System developed by the Open Networking Lab
(ON.Lab) from Stanford and UC Berkeley. Although ONOS is now aimed at carrier
networks, ON.Lab intends it as an alternative to the vendor-driven OpenDaylight initiative
and hopes to extend it to enterprises in the future.
D.5.2
Quality Assurance in Software Defined Networking
Traffic Engineering (TE) is an important mechanism to optimise the performance of a data network at both
traffic and resource level. SDN allows for centralised visibility of global network topology and status, thus
providing great incentive for new TE techniques. SDN TE mechanisms mainly focus on four areas: flow
management, fault tolerance, topology update and traffic analysis [559]. Flow management deals with the
process of installation of new forwarding rules: when a flow does not match any rule in the switch, its first
packet is sent to the controller that will install new forwarding rules. This process can introduces delay and
significant overhead if a significant number of new flows arrive at switches. Some solutions exploit
wildcards rules to reduce the control plane load [560]-[562], others deploy multiple controllers
architecture to obtain load-balancing [562]-[569].
Fault tolerance is a mandatory requirement to ensure network reliability: networks should detect and react
to a failure transparently and without affecting user experience. Fault tolerance is needed for both Data
Copyright © 2015, DICE consortium – All rights reserved
98
Deliverable 1.1. State of the art analysis
Plane and Control Plane. For the data plane some solutions – such as [570] and [571] – are based on the
restoration approach (backup path rules are installed reactively by the controller), while others such as
[572] and [573] have a proactive approach (protection paths are pre-computed and pre-installed, so no
additional switch-controller signalling is required). Control plane reliability is handled by replicating the
controller with a primary-backup approach [574] or by means of a distributed cluster of controllers [575],
[576].
Topology update mechanisms are related to planned network policies changes: updates must guarantee
consistency so that each individual packet or flow is not handled by a conflicting mix of old and new
policy.
Some approaches modifies packet’s header fields with an ID of the policy and temporarily keeping both the
new and old flow rules [577], [578], others are time-based [579], [580].
Finally traffic analysis tools aim to supply instruments to monitor traffic, collect statistics at different
aggregation levels, check network invariants and debug programming errors.
Some approaches are active and based on polling from the controller [581], [582]. Others are passive and
based on the analysis of control messages [583]. Some others exploit dedicated servers for monitoring
traffic and triggering updates to the controller [584].
D.5.3
Current Issues
Networks consist of a wide variety of devices, and each time a new network function has to be developed,
network administrators must express policies through a tedious box-by-box configuration, dealing with a
multitude of protocols and vendor-specific interfaces. In the next paragraphs is explained why SDN
simplifies the network management and how is possible to avoid the usage of low level APIs.
D.5.4
Controllers architecture and low levels APIs
In SDN, the presence of a unique logically centralised controller simplifies and redefines how to manage
networks, letting the applications to be network aware.
The controller is in charge of both providing an abstract view of the network topology to the application
layer and of managing and communicating the applications requirements down to the dataplane layer [585].
OpenFlow [550], API reference of SDN, is a standard open interface which allows the communication
between the controller and the dataplane layer (southbound API). The controller uses the OpenFlow
messages to install rules, query traffic statistics and learn the network topology.
While SDN makes it possible to programme the network, it does not make it easy. Since OpenFlow is a
low level APIs, a programmer is forced to write application in a low level language by handling each single
packet field in each single network device.
D.5.5
Languages and network policies
Openflow low level API makes difficult also the definition of policies, since a programme must both take
into account the shared rule-table space and manipulate single packet fields.
Several high level languages (Pyretic [585], Merlin [586], etc.) were born to encourage programmers to
focus on how to specify a network policy at a high level of abstraction, rather than how to implement it
using low-level OpenFlow mechanism.
Copyright © 2015, DICE consortium – All rights reserved
99
Deliverable 1.1. State of the art analysis
One of the disadvantages of programming using OpenFlow is the non-modularity. Today, controllers allow
programming only monolithic controller applications bringing to an increasing complexity in debugging
and testing. Thanks to the presence of the northbound programming interface, it is possible to define a
module for each control application. Each policy can be developed independently and the interactivity
between them is orchestrated by runtime software, a compiler that lies in the controller. Each policy is
compiled and transformed in specific low-level forwarding rules to be installed in the switches, avoiding
each kind of collision between them.
High level languages allows to define policies operating on an abstract view of the network: for example
multiple underlying switches can be abstracted as a single derived virtual switch or, alternatively, one
underlying switch can be represented as multiple derived virtual switches. The compiler will produce lowlevel rule according to the actual topology.
D.5.6
SDNs in DICE
The integration of SDN technologies into Big Data applications allows a more accurate control of network
resources than traditional switching technologies. This can be incorporated at design time by specifying
requirements for the Data-Intensive Application and at deployment time reserving the necessary network
resources for the specific data flows in the different phases of the processing, ensuring predictable quality
levels expected from the network.
D.6.
Cloud-based blob storage
D.6.1
Ceph
Ceph is an open source distributed storage technology for efficient handling of Big Data requirements
[587]. Ceph presents object, block, and file storage from a single distributed computer cluster and supports
data replication for fault tolerance. Ceph’s object storage system allows users to mount Ceph as a thinly
provisioned block device, automatically stripes and replicates the data across the cluster. Ceph’s distributed
block storage distributes the storage requirements of VMs across a large number of storage devices as
opposed to centralised Storage Area Network (SAN) and Network-Attached Storage (NAS) solutions. This
way, each VM disk data is stored across many (at least 2) different smaller storage devices called 'replicas'.
If one replica is lost, the VM’s disk will not be corrupted or the data lost. Ceph’s file system and Ceph
metadata server cluster provide a service that maps the directories and file names of the file system to
objects stored within expandable Reliable Autonomic Distributed Object Store (RADOS) [588] clusters
ensuring high performance and avoidance of heavy loads. This makes Ceph a relevant and apt solution for
the handling of large data sets. Ceph supports a RESTful API that is compatible with the data access model
of the Amazon S3 API.
D.6.1.1. Typical architecture of CEPH
Ceph clusters are made up of machines running the following types of services:




Monitors (MON): these maintain a map of the current state of the Ceph cluster.
Object Storage Devices (OSD): these store the data.
Metadata server (MDS): these maintain metadata on the placement of files in the cephfs filing
system.
Auxiliary services, such as RADOS gateways to S3 and SWIFT.
An example of a Ceph configuration can be found in Figure 20 [588].
Copyright © 2015, DICE consortium – All rights reserved
100
Deliverable 1.1. State of the art analysis
Figure 20. Ceph integration example [589].
D.6.1.2. Cloud offerings using CEPH
Flexiant Cloud Orchestration (FCO) [180] is a cloud orchestration software solution and allows service
providers to manage their own public, private or hybrid cloud solution. FCO supports the use of CEPH as a
storage solution for the node data storage, however pricing and quality will depend on the service
provider.
OpenStack [188] is a cloud operating system that allows control of compute, storage, and networking
resources throughout a datacenter. OpenStack supports the use of CEPH as a storage solution, however
pricing and quality will depend on the datacenter operator [590]. An example of a public cloud offering
using OpenStack and Ceph is Dreamhost.
Dreamhost is a public cloud company that uses OpenStack and Ceph for their cloud and storage solution.
Currently Dreamhost offers multiple options for its storage solution. These can be found in the Table 27
[591].
Table 27: Dreamhost pricing storage [591].
Storage Included
40 GB
200 GB
1,024 GB
2,048 GB
20,480 GB
Monthly Price
$0.95
$4.50
$19.95
$34.95
$299.95
Effective Price/GB
2.38¢
2.25¢
1.95¢
1.71¢
1.46¢
D.6.1.3. Open Source solutions
Ceph is an open source project [592] and can be easily adapted to any cloud environment.
D.6.1.4.
Quality assurance
D.6.1.4.1. Key monitoring metrics
 Available Disk Space.
 IOWait.
 Read Operations per second.
 Write Operations per second.
Copyright © 2015, DICE consortium – All rights reserved
101
Deliverable 1.1. State of the art analysis



Network Throughput.
Cluster health.
OSD status.
D.6.1.4.2. Quality assurance challenges for CEPH literature.
Distributed block storage is a newer technology than centralised storage. Arguably, it is more difficult to
administer, and carries the risks of being less well tested in mainstream production environments. In
particular, tuning it for performance can be more difficult. However, if deployed properly it offers
enhanced scalability, can be very economical and can perform as well as a centralised storage solution.
CEPH does this by using CRUSH, a data distribution function. CRUSH enforces data replica separation for
improved data safety [593].
D.6.1.4.3. Reliability and high-availability support for CEPH
Ceph’s foundation is the Reliable Autonomic Distributed Object Store (RADOS), which provides
applications with object, block, and file system storage in a single unified storage cluster. Ceph’s RADOS
provides scalability that can handle thousands of client hosts or KVMs accessing Petabytes to Exabytes of
data.
D.6.1.4.4. Scalability and performance support for CEPH
To provide scalability Ceph uses the CRUSH algorithm [594]. This algorithm helps liberate storage
clusters from the scalability and performance limitations imposed by centralised data table mapping as it
replicates and re-balances data within the cluster dynamically and delivers high-performance and infinite
scalability.
D.6.1.4.5. Privacy and data protection with CEPH
Data security is typically Cloud provider-dependent. However, Ceph does offer some data security by
default. Ceph provides a cephx authentication system, which authenticates users operating Ceph clients.
With Ceph a user contacts a monitor that is used to authenticate them and distribute keys. The monitor
returns an authentication data structure that contains a session key for use in obtaining Ceph services. This
key is self encrypted and only the user can request services from the Ceph monitors.
The client then uses the session key to request its desired services from the monitor, and the monitor
provides the client with a ticket that will authenticate the client to the OSDs that actually handle data. Ceph
monitors and OSDs share a secret, so the client can use the ticket provided by the monitor with any OSD or
metadata server in the cluster. This form of authentication will prevent attackers with access to the
communications medium from either creating bogus messages under another user’s identity or altering
another user’s legitimate messages, as long as the user’s secret key is not divulged before it expires [595].
The users can grant or revoke access to objects or buckets for other users in an S3 compatible way.
D.6.2
Amazon Simple Service Storage (Amazon S3)
D.6.2.1. Overview
Amazon S3 (Simple Storage Service) [596] is an online file storage web service offered by Amazon Web
Services (AWS) [192]. Amazon provides storage through web services interfaces as well as through the use
of their APIs (REST, SOAP) [596].
The main benefits of S3 and why it has grown in popularity is due to its low storage cost, redundancy and
unlimited storage capacity. S3 is typically used to back data off site, storage for large amounts of data
generated or even hosting static websites.
Copyright © 2015, DICE consortium – All rights reserved
102
Deliverable 1.1. State of the art analysis
S3 stores data as objects within ‘buckets.’ These buckets are unique to an AWS account and are identified
by a unique, user-assigned key. Objects are stored with buckets and can store as many objects as required.
They can either be accessed using the web interface or API to write, read and delete objects. These Objects
can be up to 5 Terabytes in size, however within a single PUT the object can only be a maximum of 5GB.
Data stored within S3 is stored on redundant servers in multiple data centres and can be publicly accessible
if required. One example in which this is used is for the hosting of static websites.
D.6.2.2. Public cloud offerings of S3
AWS is the only public cloud that offers this technology. A full breakdown of pricing can be found in
Table 28 and Table 29 [597] (note these prices are the US Standard pricing and may change depending on
AWS region used).
Table 28: AWS Storage Pricing US Standard [597].
First 1 TB / month
Next 49 TB / month
Next 450 TB / month
Next 500 TB / month
Next 4000 TB / month
Over 5000 TB / month
Standard Storage
3.00¢ per GB
2.95¢ per GB
2.90¢ per GB
2.85¢ per GB
2.80¢ per GB
2.75¢ per GB
Reduced Redundancy Storage
2.40¢ per GB
2.36¢ per GB
2.32¢ per GB
2.28¢ per GB
2.24¢ per GB
2.20¢ per GB
Table 29: Data Transfer Pricing US Standard [597].
Pricing
Data Transfer IN To Amazon S3
All data transfer in
Data Transfer OUT From Amazon S3 To
Amazon EC2 in the Northern Virginia Region
Another AWS Region
Amazon CloudFront
Data Transfer OUT From Amazon S3 To Internet
First 1 GB / month
Up to 10 TB / month
Next 40 TB / month
Next 100 TB / month
0.00¢ per GB
0.00¢ per GB
2.00¢ per GB
0.00¢ per GB
0.00¢ per GB
9.00¢ per GB
8.50¢ per GB
7.00¢ per GB
D.6.2.3. Open source solutions with S3
With AWS services there are no open source options as they must all be used from within an AWS
account. However AWS allows S3 to be used with an on-site solution but only with the use of an AWS
storage gateway VM [598]. An example of this can be found in within section detailing on site usage with
S3.
Other solutions such as Ceph provide S3 compatibility by offering a S3 compatible API, while native
installation of S3 is currently not available.
D.6.2.4. Quality assurance.
D.6.2.4.1. Key S3 monitoring metrics
 Size of all objects present in bucket(s).
 Number of objects present in bucket(s).
 Get transfer speed.
 Push transfer speed.
Copyright © 2015, DICE consortium – All rights reserved
103
Deliverable 1.1. State of the art analysis
D.6.2.4.2. Main quality assurance challenges.
AWS conduct continuous research and development of S3, but this research is not made public. External
research is conducted. It looks at the S3 solution as a black box approach and focuses on pricing, transfer
speeds and multi cloud reliability.
D.6.2.4.3. Configuration options for reliability and high-availability with S3
Amazon S3 Standard Storage is designed to achieve 99.999999999% durability [599]. To achieve this level
of durability S3 redundantly stores objects on multiple devices across multiple facilities in an Amazon S3
Region. The service is designed to sustain concurrent device failures by quickly detecting and repairing any
lost redundancy. S3 also verifies the integrity of data using checksums.
Reduced Redundancy Storage is a storage option within S3 that enables customers to reduce their costs by
storing data at lower levels of redundancy than S3’s standard storage offering. With RRS, data is replicated
fewer times so the durability offered is only 99.99%. However, both are backed by Amazon S3’s Service
Level Agreement.
D.6.2.4.4. Scalability and performance for S3
The total volume of data and number of objects that can be stored is unlimited. Individual Amazon S3
objects can range in size from 1 byte to 5 Terabytes however the largest object that can be uploaded in a
single PUT is 5 Gigabytes.
D.6.2.4.5. Privacy and data protection within S3
Within S3 a user can choose to make their data private or publicly accessible. Amazon state they will only
track usage for billing purposes, however if required by law this data can be made available to relevant
authorities. It is up-to the end user to encrypt their data before storing within S3.
D.7.
In-Memory Analytics
D.7.1
Introduction
In-memory analytics is a computing paradigm aimed at processing large volumes of data in main memory,
while minimising usage of disk I/O. Often, in-memory databases rely on column based representation and
compression techniques to optimise processing and memory usage. Thanks to the constant decrease of
DRAM costs, this approach is now mainstream in databases for analytical workloads, which can then enjoy
vastly shortened response times compared to the recent past. As a result, business analytics is now in
strong demand and several companies are developing in-memory database-as-a-service offerings in the
cloud.
In-memory analytics offerings rely on in-memory database technologies. Since the offerings in this area are
mostly commercial, we focus on describing a specific commercial implementation, i.e., the SAP HANA
platform.
D.7.2
Diagram showing a typical architecture of this technology.
N/A, this is a single box, there is no architecture. In case of high-availability architecture, standard layouts
exist, e.g., [600].
D.7.3
Public cloud offerings available for this technology.
In-memory DBMS are mostly commercial and include, among others, SAP HANA, ScaleOut and Actian
Matrix that are all available on Amazon AWS. In-memory extensions are also features in Oracle and
Microsoft SQL databases.
Copyright © 2015, DICE consortium – All rights reserved
104
Deliverable 1.1. State of the art analysis
D.7.4
Open source solutions for adoption of this technology in private clouds.
Open source in-memory DBMS include Apache Derby, HyperSQL (HSQLDB) and SQLite.
While Apache Spark is not an in-memory DBMS, it is an open source solution to perform some in-memory
analytics.
D.7.5
Quality assurance
D.7.5.1. Key monitoring metrics.
Typical metrics include CPU Utilisation, query response times, threading level, thread affinity, mean
memory consumption, peak memory consumption.
D.7.5.2. Main quality assurance challenges.
In-memory databases are a new type of DBMS that is not well understood in terms of quality assurance.
While benchmarking studies are available that prove that in-memory databases can be orders of magnitude
faster than traditional databases, the definition of techniques for optimal deployment, configuration and
resource management of these DBMS is still an open research problem.
As noted in [601], a significant problem for quality assurance of in-memory databases is the ability to
capture time-varying threading levels that are used internally to these databases to process queries
efficiently. At present, there are limited studies available on this aspect of performance.
Another important problem is the characterisation of peak memory consumption, since this affect memory
sizing. The goal is to minimise the impact of disk spill-off when memory consumption exceeds the
available physical memory. Understanding peak memory requirements is simple testing a reference
workload, but it is generally hard to predict on unobserved configurations.
Characterising and assigning threads affinities is also important, due to cache locality. As the threading
level of a query is dynamically increased, there can be in NUMA architectures substantial latencies for jobs
that reference data stored in a cache residing on a different socket.
D.7.5.3. Reliability
Some in-memory DMBS come with high-availability (HA) capabilities, such as backup & restore, disaster
recovery and reliable architectures with stand-by servers and standard layouts to replicate data in order to
minimise the changes of data loss.
D.7.5.4. Efficiency
Admission control allows to limit the maximum number of queries that can simultaneously execute on an
in-memory database system. This is the primary control known to limit excessive resource utilisation.
D.7.5.5. Privacy
In-memory DMBS and related database-as-a-service products typically offer multi-tenancy. Multi-tenancy
implies privacy and data protections, which can be enforced at various granularity levels, from separate
index servers to separate tables, up to separating individual table rows.
D.7.6
Models
D.7.6.1. Meta-models.
To the best of our knowledge, no meta-models explicitly captures the properties of in-memory databases.
Such an extension, in order to feed design-time reasoning, would require at least an explicit
characterisation in the meta-model of memory consumption and parallelism levels for each query type,
Copyright © 2015, DICE consortium – All rights reserved
105
Deliverable 1.1. State of the art analysis
which does not appear in principle too complex. It would be instead more complex to derive an
optimisation tool to decide the deployment and configuration of such systems.
D.7.6.2. Quality of Service prediction models.
A number of works have explored the problem of performance modelling of in-memory databases, such as
[601] that defines a queueing network model to predict query response times and CPU utilisations. Similar
works, in the arena of general DBMS, have adopted machine learning approach to predict response times
that appear partly applicable to in-memory DBMS [602], [603].
Copyright © 2015, DICE consortium – All rights reserved
106
Deliverable 1.1. State of the art analysis
Conclusion
In this deliverable, we have provided a comprehensive overview of the state-of-the-art in several areas
related to the DICE project including, but not limited to: DevOps, Model Driven Engineering (MDE),
Quality Engineering Methods, architecture styles and technologies for Big Data. One of the main
takeaways that emerge from this deliverable is the breadth, heterogeneity and complexity of the challenges
that exist in building high-quality data-intensive applications:

First, we noticed a generalised lack of extensions to annotate, model or test the behaviour of
complex software systems in terms of data usage, especially in the domain of MDE. On the one
hand this is challenging, since it implies that DICE has to build a foundation in MDE to describe
Data-Intensive Applications. On the other hand, having a “first mover” advantage like this
increases the opportunities for impact and exploitation.

Next, we observed that the technology stack of Big Data is very heterogeneous, and this poses the
question on how a 3-years research project like DICE could produce a useful contribution. In fact,
Big Data is so new and diverse that it does not seem always possible to come up with abstractions
(e.g., UML annotations or forma models) that can be reused for more than 1-2 implementations of
a given technology. It seems therefore more interested for DICE to focus on modelling a few
technologies of clear impact and diffusion (e.g., MapReduce, Spark and Storm) rather than
supporting a plethora of diverse technologies for the sake of comprehensiveness. Extensibility of
the approach to other technologies should still be assessed and, where possible, supported via
appropriate guidelines.

Another major lesson that was learned by the state of the art review is the disruptive growth of
DevOps in recent months, which is attractive large parts of the industry towards changing their
product offering and the way they address service delivery problems. This realisation has led us to
elaborate the positioning of DICE at the crossroad between Model-Driven Engineering and
DevOps, as we discuss in Deliverable D1.2 “Requirements specification”.

The investigation in Chapter D has led us to assess and compare the commercial offerings and open
source implementations of some popular Big data technologies. After performing this review, the
consortium has defined an initial software stack for DICE to support in the different areas of
interest of the project, such as Apache Hadoop, Apache Spark, Apache Storm, Amazon S3, Apache
Cassandra, and Cloudera Oryx 2. The analysis in this deliverable has led us to believe that such
technology stack is sufficiently broad to cover commons classes of data-intensive applications.

Lastly, it should be remarked that the analysis performed in this deliverable has driven the
definition of deliverable D1.2, where we report an extensive set of requirements for the DICE
project and an overview of the general technical approach.
In conclusion, Big Data is a diverse, heterogeneous area where model-driven software engineering has a lot
of potential, but where language, tool and formal model support is still lacking. This opens exciting
possibilities for the DICE project to make an impact in this area, by defining a novel UML profile and tools
for building Data-Intensive Applications with quality guarantees
Copyright © 2015, DICE consortium – All rights reserved
107
Deliverable 1.1. State of the art analysis
References
[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]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
NIST ‘Big Data Interoperability Framework: Volume 1, Definitions’. Draft Release. National Institute of Standards and
Technology. April 2015.
NIST ‘Big Data Interoperability Framework: Volume 2, Big Data Taxonomies’. Draft Release. National Institute of
Standards and Technology. April 2015.
‘The Digital Universe of Opportunities: Rich Data and the Increasing Value of the Internet of Things’ Vernon Turner,
John F. Gantz, David Reinsel, and Stephen Minton, Report from IDC for EMC April 2014.
Big Data Value Association http://www.bdva.eu/
‘Exploring Data-Driven Innovation as a New Source of Growth – mapping the policy issues raised by ‘Big Data’’,
Report from OECD 18 June 2013.
Applying assumptions from the McKinsey Global Institute report ‘Big Data: The next frontier for innovation,
competition, and productivity’, June 2011, to the European healthcare sector.
‘Big Data Vendor Revenue and Market Forecast 2013-2017’, article, Wikibon, February 2014.
Big Data Market by Types (Hardware; Software; Services; BDaaS - HaaS; Analytics; Visualisation as Service); By
Software (Hadoop, Big Data Analytics and Databases, System Software (IMDB, IMC): Worldwide Forecasts & Analysis
(2013 – 2018), available online at: www.marketsandmarkets.com, August 2013.
‘Worldwide Big Data Technology and Services 2013–2017 Forecast’, report, IDC, December 2013.
‘Big and open data in Europe - A growth engine or a missed opportunity?’ Sonia Buchholtz, Maciej Bukowski,
Aleksander Śniegocki (Warsaw Institute for Economic Studies), report commissioned by demosEUROPA, 2014.
Big Data Value calculation based on http://www.eskillslandscape.eu/ict-workforce-in-europe/
‘The European Data Market’, Gabriella Catteneo, IDC, presentation given at the NESSI summit in Brussels on 27 May
2014, available online at: http://www.nessi-europe.eu/?Page=nessi_summit_2014
IDC European Vertical Markets Survey, October 2013. http://www.idc.com/getdoc.jsp?containerId=M07W
Apache Hadoop https://hadoop.apache.org/
Hindman, B., Konwinski, A., Zaharia, M., Ghodsi, A., Joseph, A. D., Katz, Shenker, S., Stoica, I. ‘Mesos: A Platform
for Fine-Grained Resource Sharing in the Data Center’. In NSDI. March 2011.
Marz, N. (2013). Big Data: Principles and best practices of scalable real-time data systems. O’Reilly Media.
A repository dedicated to the Lambda Architecture (LA), http://lambda-architecture.net/
Hadoop YARN, http://hadoop.apache.org/docs/stable/hadoop-yarn/hadoop-yarn-site/YARN.html, accessed on
08/05/2015.
Apache Storm http://storm.apache.org/
Apache Zookeeper https://zookeeper.apache.org/
Apache Kafka http://kafka.apache.org/
Fan, W. and Bifet, A. (2013). Mining Big Data: current status, and forecast to the future. ACM SIGKDD Explorations
Newsletter, 14(2):1–5.
Robak, S., Franczyk, B., and Robak, M. (2013). Applying Big Data and linked data concepts in supply chains
management. In Computer Science and Information Systems (FedCSIS), 2013 Federated Conference on, pages 1215–
1221. IEEE.
Nicolas Bär, ‘Investigating Lambda Architecture’, Master Thesis, August 2014, University of Zurich.
SRBench benchmark http://www.w3.org/wiki/SRBench
DEBS Grand Challenge 2014 http://lsds.doc.ic.ac.uk/projects/SEEP/DEBS-GC14
Google Cloud Dataflow, https://cloud.google.com/dataflow/
Amazon Kinesis, http://docs.aws.amazon.com/kinesis/latest/dev/key-concepts.html
Amazon Elastic MapReduce, http://aws.amazon.com/elasticmapreduce/, accessed on 08/05/2015.
http://lambdoop.com/
Oryx 2 http://oryxproject.github.io/oryx/
Oryx 1 https://github.com/cloudera/oryx
Hadoop Distributed File System (HDFS) http://hadoop.apache.org/docs/r1.2.1/hdfs_design.html
http://spark.apache.org/docs/latest/mllib-collaborative-filtering.html
http://arxiv.org/abs/1203.6402
http://radar.oreilly.com/2014/07/questioning-the-lambda-architecture.html
R. C. Fernandez, P. Pietzuch, J. Koshy, J. Kreps, D. Lin, N. Narkhede, J. Rao, C. Riccomini, and G. Wang. Liquid:
Unifying Nearline and Offline Big Data Integration. In CIDR, 2015.
http://blog.acolyer.org/2015/02/04/liquid-unifying-nearline-and-offline-big-data-integration/
Loukides, Mike (2012-06-07). ‘What is DevOps?’
Loukides, Mike (2011-05-03) ‘Software Velocity 2011 - retrospective’
FP7 Research and Innovation http://ec.europa.eu/research/fp7/index_en.cfm
BTO Research Technical Briefing ‘[email protected]’, May 2015
Copyright © 2015, DICE consortium – All rights reserved
108
Deliverable 1.1. State of the art analysis
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
[64]
[65]
[66]
[67]
[68]
[69]
[70]
[71]
[72]
[73]
[74]
[75]
[76]
[77]
[78]
[79]
[80]
[81]
[82]
[83]
[84]
[85]
[86]
[87]
Ant http://wiki.mxunit.org/display/default/Continuous+Integration+--+ANT++The+Glue+between+Jenkins+and+your+Tests
Capistrano http://capistranorb.com/
RPM Package Manager http://www.rpm.org/
Cobbler http://cobbler.github.io/
Crowbar https://crowbar.github.io/
Puppet https://puppetlabs.com/puppet/what-is-puppet
Chef https://www.chef.io/chef/
Nagios https://www.nagios.org/
Sensu https://sensuapp.org/
Software Engineering Institute - Blog ‘Continuous Integration and DevOps’
Matt Weinberger, DevOps explained: A philosophy of speed, not momentum
Gartner. ‘Research Note‘
Hüttermann, Michael (2012). ‘DevOps for Developers’. Apress.
PuppetLabs, 2014. ‘State of DevOps Report ‘
Model-Driven Architecture Specification and Standardisation, The Object Management Group (OMG) http://www.omg.org/mda/
OMG Object Management Group http://www.omg.org/
Model-Driven Engineering Reference Guide, An Enterprise Architect’s Perspective http://www.theenterprisearchitect.eu/blog/2009/01/15/mde-model-driven-engineering-reference-guide/
MODAClouds EU http://www.modaclouds.eu/
MODAClouds Documentation Portal, D4.2.1, MODACloudML Version 1 - http://www.modaclouds.eu/wpcontent/uploads/2012/09/MODAClouds_D4.2.1_MODACloudMLDevelopmentInitialVersion.pdf
Reuse and Migration of legacy applications to Interoperable Cloud Services - EU REMICS, Methodology http://www.remics.eu/publicdeliverables
ARTIST EU Project Homepage, Resource Gallery - http://www.artist-project.eu/vision
JUNIPER EU Project Homepage, Resource Gallery - http://www.juniper-project.org/
CORBA http://www.corba.org/
Architecture-Driven Modernisation Taskforce, Will Ulrich, Tactical Strategy Group http://www.omg.org/news/meetings/workshops/ADM_2005_Proceedings_FINAL/T-1_Ulrich.pdf
Unified Modelling Language: Infrastructure, 2011. Version 2.4.1, OMG document: formal/2011-08-05.
OMG: UML Profile for MARTE: Modelling and Analysis of Real-Time and Embedded Systems, June 2011. Version
1.1, OMG document: formal/2011-06-02.
Bernardi, S., Merseguer, J., and Petriu, D.C.: A dependability profile within MARTE. Software and Systems Modelling,
10(3):313–336, 2011.
J2EE Java Enterprise Edition SDK http://www.oracle.com/technetwork/java/javaee/download-141771.html
.NET Framework http://www.microsoft.com/net
Object Constraint Language, 2014. OMG document: formal/2014-02-03, v2.4.
Avizienis, A., Laprie, J.C., Randell, B. and Landwehr, C.: Basic concepts and taxonomy of dependable and secure
computing. IEEE Trans. on Dependable and Secure Computing, 01(1):11–33, 2004.
Leveson, N.G.: Safeware. Addison-Wesley, USA, 1995.
Bernardi, S., Merseguer, J. and Petriu, D.C.: Model-driven Dependability Assessment of Software Systems. Springer,
2013.
Basin, D., Doser, J., and Lodderstedt, T.: Model Driven Security: From UML Models to Access Control Infrastructures.
ACM Transactions on Software Engineering and Methodology, 15(1): 39-91, 2006.
Ferraiolo, D.F., Sandhu, R., Gavrila, S., Kuhn, D.R., and Chandramouli, R.: Proposed NIST standard for role-based
access control. ACM Trans. on Information Systems Security (TISSEC) 4(3):224–274, 2001.
Rodríguez, R., Merseguer, J. and Bernardi, S.: Modelling Security of Critical Infrastructures: A Survivability
Assessment. The Computer Journal 2014. DOI :10.1093/comjnl/bxu096.
Jürjens, J: Secure systems development with UML. Springer 2005, ISBN 978-3-540-00701-2.
CARiSMA, https://www-secse.cs.tu-dortmund.de/carisma/index.shtml.
XMI XML Meta-data Interchange http://www.omg.org/spec/XMI/
Enterprise Architect https://www.sparxsystems.es/
Papyrus https://www.eclipse.org/papyrus/
Eclipse Public License 1.0 https://www.eclipse.org/legal/epl-v10.html
Eclipse Papyrus WIKI http://wiki.eclipse.org/Papyrus_User_Guide
UML Activity, Sequence, Profile Diagrams https://www.eclipse.org/papyrus/usersTutorials/usersTutorialsIndex.php
UML Diagrams (and Models) with Papyrus http://lowcoupling.com/post/47802411601/uml-diagrams-and-models-withpapyrus
Copyright © 2015, DICE consortium – All rights reserved
109
Deliverable 1.1. State of the art analysis
[88]
[89]
[90]
[91]
[92]
[93]
[94]
[95]
[96]
[97]
[98]
[99]
[100]
[101]
[102]
[103]
[104]
[105]
[106]
[107]
[108]
[109]
[110]
[111]
[112]
[113]
[114]
[115]
[116]
[117]
[118]
[119]
[120]
[121]
[122]
[123]
[124]
[125]
[126]
[127]
[128]
[129]
[130]
[131]
[132]
[133]
[134]
[135]
[136]
[137]
[138]
[139]
[140]
[141]
[142]
[143]
[144]
[145]
[146]
UML Sequence Diagrams http://lowcoupling.com/post/47844944042/uml-sequence-diagrams
UML State Machines http://lowcoupling.com/post/47845261252/uml-state-machines
UML Use case http://www.cs.fsu.edu/~baker/swe1/restricted/notes/tutorial_papyrus/tutorial_papyrus.html
UML Timing diagram http://www.eclipse.org/papyrus/project-info/new_and_noteworthy_0.10.0.php
Papyrus for Real-Time Embedded Systems, Charles Rivet https://www.youtube.com/watch?v=kfD7Ejxo6dQ
Modelio https://www.modelio.org/
Modeliosoft https://www.modeliosoft.com/
BPMN standard http://www.bpmn.org/
http://rd.softeam.com/prototypes/martedesigneralfaformodelio2x
MOSKitt https://moskitt.gva.es/redmine
ArgoUML http://sourceforge.net/projects/argouml/
Java programming language http://www.oracle.com/technetwork/java/index.html
C++ programming language https://isocpp.org/
C# programming language https://msdn.microsoft.com/en-us/library/67ef8sbd.aspx
PHP https://www.php.net/
Ruby programming language https://www.ruby-lang.org/ru/
Delphi programming language http://www.delphibasics.co.uk/
StarUML version 1. FreeSoftware http://staruml.sourceforge.net/v1/index.php
StarUML version 2. Commercial http://staruml.io/
GNU GPL http://www.gnu.org/licenses/gpl-3.0.en.html
UML Designer http://www.umldesigner.org/
UML2 https://wiki.eclipse.org/MDT-UML2
https://www.eclipse.org/sirius/
MARTE Designer http://marketplace.obeonetwork.com/module/marte
http://www.nomagic.com/products/magicdraw.html
SySML http://sysml.org/
UPDM http://www.omg.org/spec/UPDM/
XML Schema http://www.w3.org/standards/xml/schema
WSDL – Web Service Description Language http://www.w3.org/TR/wsdl
BluAge http://www.bluage.com/
http://www-03.ibm.com/software/products/en/ratisoftarch
Sparx Systems http://www.sparxsystems.com/
CEA http://www.cea.fr/english-portal
Atos http://atos.net/en-us/home.html
http://www.citma.gva.es/ (the website is in Spanish)
Tigris.org http://www.tigris.org/
Obeo http://www.obeo.fr/en/
No Magic, Inc. http://www.nomagic.com/
IBM http://www.ibm.com/uk/en/
ActionScript http://help.adobe.com/livedocs/specs/actionscript/3/wwhelp/wwhimpl/js/html/wwhelp.htm
Python https://www.python.org/
Visual Basic https://msdn.microsoft.com/en-us/vstudio/ms788229.aspx
VB .NET https://msdn.microsoft.com/en-us/vstudio/hh388573
DDL https://msdn.microsoft.com/en-us/library/ff848799.aspx
Ada http://www.adaic.org/
VHDL http://www.eda.org/twiki/bin/view.cgi/P1076/WebHome
Verilog http://www.verilog.com/
BPEL http://docs.oasis-open.org/wsbpel/2.0/OS/wsbpel-v2.0-OS.html
Visual Studio https://www.visualstudio.com/
AndroMDA http://www.andromda.org/
NetBeans https://netbeans.org/
EA User Guide http://www.sparxsystems.com.au/bin/EAUserGuide.pdf
Modelio User Manual https://www.modelio.org/documentation/user-manuals.html
Papyrus User Manual http://wiki.eclipse.org/Papyrus_User_Guide
MOSKitt User Manual www.moskitt.org/eng/med/print/manuales/
ArgoUML Documentation http://argouml-users.net/index.php?title=Documentation
UML Designer Documentation http://www.umldesigner.org/ref-doc/usage.html
http://www.nomagic.com/files/manuals/MagicDraw%20UserManual.pdf
Enterprise Architect Forum http://www.sparxsystems.com.au/cgi-bin/yabb/YaBB.cgi
Copyright © 2015, DICE consortium – All rights reserved
110
Deliverable 1.1. State of the art analysis
[147]
[148]
[149]
[150]
[151]
[152]
[153]
[154]
[155]
[156]
[157]
[158]
[159]
[160]
[161]
[162]
[163]
[164]
[165]
[166]
[167]
[168]
[169]
[170]
[171]
[172]
[173]
[174]
[175]
[176]
[177]
[178]
[179]
[180]
[181]
[182]
[183]
[184]
[185]
[186]
[187]
[188]
[189]
[190]
[191]
[192]
[193]
[194]
[195]
[196]
[197]
[198]
[199]
Modelio Forum https://www.modelio.org/forum/index/6-general.html
MOSKitt forum https://joinup.ec.europa.eu/software/gvcase-db/forum/all
ArgoUML Forum http://argouml-users.net/forum/
UML Designer Forum http://www.obeonetwork.com/group/uml-designer
https://community.nomagic.com/magicdraw-f22.html
Eclipse Mars project https://projects.eclipse.org/releases/mars
J. Oldevik, A. Solberg , Ø. Haugen, and B. Møller-Pedersen Evaluation Framework for Model-Driven Product Line
Engineering Tools,, Software Product Line, 2006, pp 589-618
S. Becker, H. Koziolek, and R. Reussner. The Palladio component model for model-driven performance prediction.
Journal of Systems and Software,82(1), 3-22. 2009
D. Varró, and A. Balogh. The model transformation language of the VIATRA2 framework. Science of Computer
Programming, 68(3), 214-234. 2007
Model Editing VIATRA2 http://wiki.eclipse.org/VIATRA2/GettingStarted/Model_Editing
VIATRA2 http://www.eclipse.org/viatra/
UML Model transformation tool http://sourceforge.net/projects/umt-qvt/
UPUPA http://sealabtools.di.univaq.it/tools.php
fUML http://www.omg.org/spec/FUML/
Meta-Object Facility Query-Views-Transformations Specification - http://www.omg.org/spec/QVT/
AMMA - ATLAS Model Management Architecture https://wiki.eclipse.org/AMMA
Atlas-Transformation Language Specification - https://eclipse.org/atl/
Derek Palma and Thomas Spatzier, Topology and Orchestration Specification for Cloud Applications Version 1.0,
OASIS standard, November 2013 http://docs.oasis-open.org/tosca/TOSCA/v1.0/os/TOSCA-v1.0-os.html
Open Source CONvention OSCON 2015 http://www.oscon.com/open-source-2015
TOSCA-TC SubCommittee (SC) on interoperability https://www.oasisopen.org/committees/tc_home.php?wg_abbrev=tosca-interop
YAML http://yaml.org/
SaltStack ALT http://saltstack.com/
Ansible automation tool for application deployment http://www.ansible.com/home
Vagrant – tool for configuring development environments https://www.vagrantup.com/
http://en.wikipedia.org/wiki/Orchestration_%28computing%29
http://www.ubuntu.com/cloud/tools/juju
https://jujucharms.com/docs/stable/authors-charm-components
https://jujucharms.com/u/asanjar/spark/trusty
http://getcloudify.org/
http://getcloudify.org/2012/03/20/big-data-in-the-cloud-using-cloudify.html
http://alien4cloud.github.io/index.html
https://brooklyn.incubator.apache.org/
https://www.oasis-open.org/committees/tc_home.php?wg_abbrev=camp#technical
Flexiant Cloud Orchestrator: https://www.flexiant.com/flexiant-cloud-orchestrator/
Rundeck: http://rundeck.org/
CAMF: https://projects.eclipse.org/proposals/cloud-application-management-framework
CELAR: https://github.com/CELAR/c-Eclipse
Open-TOSCA Winery: https://projects.eclipse.org/projects/soa.winery
Open-TOSCA VinoThek: http://install.opentosca.org/
AGPL – GNU Affero General Public License https://gnu.org/licenses/agpl-3.0.en.html
Apache license 2.0 https://www.apache.org/licenses/
OpenStack – Open source software for creating public and private clouds https://www.openstack.org/
Apache CloudStack – open source cloud computing https://cloudstack.apache.org/
FCO Blueprints https://www.flexiant.com/2013/07/11/what-are-cloud-blueprints/
FCO Triggers https://www.flexiant.com/2014/11/13/flexiant-trigger-plugin-technology/
Amazon Web Services (AWS) http://aws.amazon.com/
SoftLayer http://www.softlayer.com/
VMware vSphere https://www.vmware.com/uk/products/vsphere
VMware vCloudAir http://vcloud.vmware.com/uk/explore-vcloud-air/what-is-vcloud-air
Apache Java Multi-Cloud Toolkit https://jclouds.apache.org/
Flexiant Hypervisors comparison https://www.flexiant.com/2014/02/12/hypervisor-comparison-kvm-xen-vmware-hyperv/
MIKELANDGELO http://www.mikelangelo-project.eu/
Docker https://www.docker.com/
Copyright © 2015, DICE consortium – All rights reserved
111
Deliverable 1.1. State of the art analysis
[200]
[201]
[202]
[203]
[204]
[205]
[206]
[207]
[208]
[209]
[210]
[211]
[212]
[213]
[214]
[215]
[216]
[217]
[218]
[219]
[220]
[221]
[222]
[223]
[224]
[225]
[226]
[227]
[228]
[229]
[230]
[231]
[232]
[233]
[234]
[235]
[236]
[237]
[238]
[239]
[240]
[241]
[242]
[243]
[244]
[245]
[246]
[247]
LXC https://linuxcontainers.org/
IPMI http://www.intel.com/content/www/us/en/servers/ipmi/ipmi-home.html
OSV http://osv.io/
Martin Fowler, Contionuous Integration, May 2006, http://www.martinfowler.com/articles/continuousIntegration.html
http://en.wikipedia.org/wiki/Continuous_integration
http://www.yegor256.com/2014/10/08/continuous-integration-is-dead.html
http://confluence.public.thoughtworks.org/display/CC/CI+Feature+Matrix
TeamCity https://www.jetbrains.com/teamcity/
XCode https://developer.apple.com/xcode/
Jenkins https://jenkins-ci.org/
Hudson http://hudson-ci.org/
http://en.wikipedia.org/wiki/Jenkins_%28software%29
https://www.atlassian.com/software/bamboo
JIRA issue tracking service https://www.atlassian.com/software/jira
Stash https://www.atlassian.com/software/stash
http://www.thoughtworks.com/products/go-continuous-delivery
Mingle http://www.thoughtworks.com/mingle/
https://github.com/Strider-CD/strider
Node.js https://nodejs.org/
http://buildbot.net/
https://circleci.com/
http://en.wikipedia.org/wiki/Xvfb
MIT license http://opensource.org/licenses/MIT
BSD license https://en.wikipedia.org/wiki/BSD_licenses
Software versioning https://en.wikipedia.org/wiki/Software_versioning
Revision control https://en.wikipedia.org/wiki/Revision_control
CVS - Concurrent Versions System https://en.wikipedia.org/wiki/Concurrent_Versions_System
http://subversion.apache.org/
http://git-scm.com/
SHA-1 https://en.wikipedia.org/wiki/SHA-1
Loeliger, Jon and McCullough Matthew, Version Control with Git, Second Edition, O’Reilly Media, Inc, August 2012
https://mercurial.selenic.com/
http://www.openoffice.org/
Versions Maven Plugin: http://www.mojohaus.org/versions-maven-plugin/
Python-versioneer: https://github.com/warner/python-versioneer
Dist::Zilla: http://dzil.org/index.html
Perl programming language https://www.perl.org/
Casale, G., Ardagna, D., Artac, M., Barbier, F., Di Nitto, E., Henry, H, Iuhasz, G., Joubert, C., Merseguer, J., Ion
Monteanu, V., Perez, J. F., Petcu, D., Rossi, M., Sheridan, C., Spais I., Vladusic, D., DICE: Quality-Driven
Development of Data-Intensive Cloud Applications, In the proceedings of the MISE 2015 workshop
List of Web Services standards
https://en.wikipedia.org/wiki/List_of_web_service_specifications#Web_Service_Standards_Listings
Lazowska, E.D., Zahorjan, J., Graham, S.G., and Sevcik, K.C.: Quantitative system performance: computer system
analysis using queueing network models. Prentice-Hall, Inc. 1984, ISBN: 0-13-746975-6.
JOHNSON, B. W. 1989. Design and Analysis of Fault-Tolerant Digital Systems. Addison-Wesley
C. Baier, J-P. Katoen, 2008. Principles of Model Checking, MIT Press
C.A. Furia, D. Mandrioli, A. Morzenti, M. Rossi, 2012. Modelling Time in Computing, Springer.
Dwyer, M. B., Avrunin, G. S., and Corbett, J. C. 1999. Patterns in property specifications for finite-state verification.
Proceedings of the 1999 International Conference on Software Engineering (ICSE)
S. Konrad and B. H. C. Cheng. Real-time specification patterns. Proceedings of the 2005 International Conference on
Software Engineering (ICSE), pages 372-381
L. Grunske. Specification patterns for probabilistic quality properties. In Proceedings of the 2008 International
Conference on Software Engineering (ICSE), pages 31-40
D. Bianculli, C. Ghezzi, C. Pautasso, and P. Senti, 2012. Specification patterns from research to industry: A case study in
service-based applications. Proceedings of the 2012 International Conference on Software Engineering (ICSE), pages
968-976
Jain, R.: The Art of Computer Systems Performance Analysis: Techniques for Experimental Design, Measurement,
Simulation, and Modelling. Wiley, Wiley professional computing series, 1991, ISBN: 978-0-471-50336-1.
Copyright © 2015, DICE consortium – All rights reserved
112
Deliverable 1.1. State of the art analysis
[248]
[249]
[250]
[251]
[252]
[253]
[254]
[255]
[256]
[257]
[258]
[259]
[260]
[261]
[262]
[263]
[264]
[265]
[266]
[267]
[268]
[269]
[270]
[271]
[272]
[273]
[274]
[275]
[276]
Nicol, D.M., Sanders, W.H., Trivedi, K.S.: Model-Based Evaluation: From Dependability to Security. IEEE Trans.
Dependable Sec. Comput. 1(1): 48-65 (2004).
Trivedi, K.S.: Probability and Statistics with Reliability, Queueing, and Computer Science Applications. 2nd Edition,
Wiley, 2001, ISBN: 978-0-471-33341-8,
U.S. Nuclear Regulatory Commission: Fault Tree Handbook. NUREG-0492, January 1981.
Dugan, B.J., Bavuso, S.J., and Boyd M.A.: Dynamic fault-tree models for fault tolerant computer systems. IEEE
Transactions on Reliability, 41(3):363-373, 1992.
Bolch, G., Greiner, S., de Meer, H. and Trivedi, K.S.: Queueing Networks and Markov Chains: Modelling and
Performance Evaluation with Computer Science Applications. Wiley-Interscience, New York, NY, USA (1998).
Haverkort, B.R., Marie, R., Rubino, G. and Trivedi, K.S.: Performability Modelling: Techniques and Tools. Wiley &
Sons, 2001.
Stewart, W.J.: Introduction to the Numerical Solution of Markov Chains. Princeton University Press, 1994, ISBN-13:
978-0691036991.
Muppala, J., Sathaye, A., Howe, R., and Trivedi, K.S.: Dependability Modelling of a Heterogeneous VAXcluster System
Using Stochastic Reward Nets. Hardware and Software Fault Tolerance in Parallel Computing Systems, D. Avresky
(ed.), pp. 33-59, Ellis Horwood Ltd., 1992.
Sahner, R.A., Trivedi, K.S., and Puliafito, A.: Performance and Reliability Analysis of Computer Systems: An ExampleBased Approach Using the SHARPE Software Package. Kluwer Academic Publishers, 1996.
Buchholz, P.: Exact and Ordinary Lumpability in Finite Markov Chains. Journal of Applied Probability, 31(1),pp. 59-75,
1994.
Bobbio, A. and Trivedi, K.S.: An Aggregation Technique for the Transient Analysis of Stiff Markov Chains. IEEE Trans.
Computers 35(9): 803-814 (1986).
Miner, A.S., and Ciardo, G.: Efficient Reachability Set Generation and Storage Using Decision Diagrams. ICATPN
1999: 6-25.
Malhotra, M., Muppala, J.K., and Trivedi, K.T.: Stiffness-tolerant methods for transient analysis of stiff Markov chains.
Microelectronics and Reliability, vol.34, pp. 1825-1841, 1994.
Molloy, M.K. On the Integration of Delay and Throughput Measures in Distributed Processing Models. PhD thesis,
UCLA, Los Angeles (CA), 1981.
Kelling, C.: Conventional and Fast Simulation Techniques for Stochastic Petri Nets. Forschungsbericht des Fachbereichs
Informatik, Heft 25, Technische Universität Berlin, Berlin, 1996.
Rodríguez, R.J., Júlvez, J., and Merseguer, J.: On the Performance Estimation and Resource Optimisation in Process
Petri Nets. IEEE T. Systems, Man, and Cybernetics: Systems 43(6): 1385-1398 (2013).
Ajmone Marsan, M., Balbo, G., Conte, G., Donatelli, S., and Franceschinis, G.: Modelling with Generalised Stochastic
Petri Nets, Wiley & Sons, 1994, ISBN:0471930598.
Sanders, W.H., and Meyer, J.F.: Stochastic Activity Networks: Formal Definitions and Concepts. European Educational
Forum: School on Formal Methods and Performance Analysis 2000: 315-343.
Chiola, G., Dutheillet, C., Franceschinis, G., and Haddad, S.: Stochastic Well-Formed Colored Nets and Symmetric
Modelling Applications. IEEE Trans. Computers 42(11): 1343-1360 (1993).
Dugan, B.J., Trivedi, K.S., Geist, R., and Nicola, V.F.: Extended Stochastic Petri Nets: Applications and Analysis.
Performance 1984: 507-519.
Lindemann, C.: Performance Modelling with Deterministic and Stochastic Petri Nets. SIGMETRICS Performance
Evaluation Review 26(2): 3 (1998).
Choi, H., Kulkarni, V.K., and Trivedi, K.S.: Markov Regenerative Stochastic Petri Nets. Perform. Eval. 20(1-3): 337357 (1994).
Puliafito, A., Scarpa, M., and Trivedi, K.S.: Petri Nets with k Simultaneously Enabled Generally Distributed Timed
Transitions. Perform. Eval. 32(1): 1-34 (1998).
Baskett, F., Chandy, K.M., Muntz, R.R., and Palacios, F.G.: Open, closed, and mixed networks of queues with different
classes of customers. J. ACM 22(2), 248–260 (1975).
Franks, G., Al-Omari T., Woodside, C.M., Das, O., and Derisavi, S.: Enhanced modelling and solution of layered
queueing networks. IEEE Trans Softw. Eng 35(2):148–161, 2009.
Omari, T., Franks, G., Woodside C.M., and Pan A.: Efficient performance models for layered server systems with
replicated servers and parallel behaviour. J Syst Softw 80(4):510–527, 2007.
Tribastone, M.: A fluid model for layered queueing networks. IEEE Trans Softw Eng 39(6):744–756, 2013.
Pérez J.F., and Casale G. Assessing sla compliance from Palladio component models. In: Proceedings of the 2013 15th
International Symposium on Symbolic and Numeric Algorithms for Scientific Computing, SYNASC ’13, Timisoara,
Romania, pp 409–416, 2013.
Bacigalupo, D., van Hemert, J., Chen, X., Usmani, A., Chester, A., He, L. Dillenberger, D., Wills, G., Gilbert, L., and
Jarvis, S.: Managing dynamic enterprise and urgent workloads on clouds using layered queueing and historical
performance models. Simul. Model Prac. Theory 19:1479–1495, 2011.
Copyright © 2015, DICE consortium – All rights reserved
113
Deliverable 1.1. State of the art analysis
[277]
[278]
[279]
[280]
[281]
[282]
[283]
[284]
[285]
[286]
[287]
[288]
[289]
[290]
[291]
[292]
[293]
[294]
[295]
[296]
[297]
[298]
[299]
[300]
[301]
[302]
[303]
[304]
[305]
[306]
[307]
[308]
[309]
[310]
[311]
[312]
Nance, R.R.: A History of Discrete Event Simulation Programming Languages. SIGPLAN Not., 28(3), March 1993, pp.
149-175.
Banks J, Carson J.S., Nelson, B.L., and Nicol D.M.: Discrete Event System Simulation, Upper Saddle River, N.J.,
Prentice-Hall, 2000.
Shelder, G.S.: Regenerative Stochastic Simulation. Boston, Prentice-Hall, 1993.
Gillespie, D.T.: Approximate accelerated stochastic simulation of chemically reacting systems. Journal of Chemical
Physics, 115(4), 2001.
Heidelberger, P.: Fast Simulation of Rare Events in Queueing and Reliability Models. ACM Trans. Model. Comput.
Simul., Jan. 1995, 5(1), pp.43-85.
University of Hamburg: Petri Net tool database. Accessed: 2015-06-02. URL: http://www.informatik.unihamburg.de/TGI/PetriNets/tools/db.html.
Charalambous, A.: Extension of PIPE2 to Support Coloured Generalised Stochastic Petri Nets. Imperial College,
Tech.Rep. 2014.
Jensen, K.: Coloured Petri Nets. Basic Concepts, Analysis Methods and Practical Use. Basic Concepts. EATCS
Monographs on Theoretical Computer Science, vol. 1. Springer, Germany (1992).
GreatSPN2.0 http://www.di.unito.it/~greatspn/index.html
TimeNET https://www.tu-ilmenau.de/sse/timenet/general-information/petri-nets/
PIPE http://pipe2.sourceforge.net/
PeabraiN http://webdiis.unizar.es/GISED/?q=tool/peabrain
JMT http://jmt.sf.net
LINE https://line-solver.sourceforge.net
LQNS http://www.sce.carleton.ca/rads/lqns/
Clarke, E. M. (2008). The birth of model checking. 25 Years of Model Checking, 1-26.
Krstić, S. (2014). Quantitative properties of software systems: specification, verification, and synthesis. Companion
Proceedings of the 36th International Conference on Software Engineering. ACM.
Kwiatkowska, M. (2013). Advances in quantitative verification for ubiquitous computing. Theoretical Aspects of
Computing–ICTAC 2013, 42-58.
Basin, D., Caronni, G., Ereth, S., Harvan, M., Klaedtke, F., & Mantel, H. (2014). Scalable Offline Monitoring. Runtime
Verification. Springer International Publishing.
Camilli, M. (2014). Formal verification problems in a Big Data world: towards a mighty synergy. Companion
Proceedings of the 36th International Conference on Software Engineering. ACM.
Yang, F., Su, W., Zhu, H., & Li, Q. (2010). Formalising MapReduce with csp. Engineering of Computer Based Systems
(ECBS), 2010 17th IEEE International Conference and Workshops on. IEEE.
Hoare, C. A. R. (1985). Communicating sequential processes (Vol. 178). Englewood Cliffs: Prentice-hall.
Refinement, F. (2003). FDR2 User Manual. Formal Systems (Europe) Ltd.
Clarke, Edmund M, Orna Grumberg, and Doron Peled. Model checking. MIT press, 1999.
Di Noia, T., Mongiello, M., & Di Sciascio, E. A computational model for MapReduce job flow.
Uppaal. http://uppaal.org/
Barbierato, E., Gribaudo, M., & Iacono, M. (2013). Modelling apache hive based applications in Big Data architectures.
Proceedings of the 7th International Conference on Performance Evaluation Methodologies and Tools. ICST (Institute
for Computer Sciences, Social-Informatics and Telecommunications Engineering).
Fisher, M. (2011). An Introduction to Practical Formal Methods Using Temporal Logic. John Wiley & Sons.
Dwyer, M. B., Avrunin, G. S., & Corbett, J. C. (1999). Patterns in property specifications for finite-state verification.
Software Engineering, 1999. Proceedings of the 1999 International Conference on. IEEE.
Ciapessoni, E., Mirandola, P., Coen-Porisini, A., Mandrioli, D., & Morzenti, A. (1999). From formal models to formally
based methods: an industrial experience. ACM Transactions on Software Engineering and Methodology (TOSEM), 8(1),
79-113.
Baresi, L., Blohm, G., Kolovos, D. S., Matragkas, N., Motta, A., Paige, R. F., Radjenovik A. & Rossi, M. (2015).
Formal verification and validation of embedded systems: the UML-based MADES approach. Software & Systems
Modelling, 1-21.
Bianculli, D., Ghezzi, C., Pautasso, C., & Senti, P. (2012). Specification patterns from research to industry: a case study
in service-based applications. Proceedings of the 34th International Conference on Software Engineering. IEEE Press.
The SPIN Model checker. http://spinroot.com/
NuSMV: A New Symbolic Model Checker. http://nusmv.fbk.eu/
The PRISM probabilistic model checker. http://www.prismmodelchecker.org/
Bersani, M. M., Rossi, M., & Pietro, P. S. (2013). A tool for deciding the satisfiability of continuous-time metric
temporal logic. In Temporal Representation and Reasoning (TIME), 2013 20th International Symposium on(pp. 99-106).
IEEE
Copyright © 2015, DICE consortium – All rights reserved
114
Deliverable 1.1. State of the art analysis
[313]
[314]
[315]
[316]
[317]
[318]
[319]
[320]
[321]
[322]
[323]
[324]
[325]
[326]
[327]
[328]
[329]
[330]
[331]
[332]
[333]
[334]
[335]
[336]
[337]
[338]
[339]
[340]
[341]
[342]
[343]
[344]
[345]
[346]
[347]
[348]
De Moura, L., & Bjørner, N. (2008). Z3: An efficient SMT solver. Tools and Algorithms for the Construction and
Analysis of Systems, 337-340.
The Zot bounded model/satisfiability checker. http://zot.googlecode.com
CorrettoUML https://github.com/deib-polimi/Corretto
Jackson, D. (2001). Lightweight formal methods. FME 2001: Formal Methods for Increasing Software Productivity, 1-1.
Ghilardi, Silvio et al. ‘Towards SMT model checking of array-based systems.’ Automated Reasoning (2008): 67-82.
Ledru Y., Idani A. and Richier J.L.. Validation of a security policy by the test of its formal B specification - a case study.
FormaliSE: FME Workshop on Formal Methods in Software Engineering. ICSE 2015.
Lodderstedt T., Basin D., and Doser J. SecureUML: A UML-based modelling language for model-driven security. The
Unified Modelling Language (2002): 426-441.
Abrial, J. (2005).The B-Book: Assigning programmes to meanings. Cambridge University Press.
The nuXmv model checker. https://nuxmv.fbk.eu/
The Z3 Theorem Prover. https://github.com/Z3Prover
Parsons T., Murphy, J. ‘Detecting Performance Antipatterns in Component Based Enterprise Systems’, Journal of Object
Technology, vol. 7, no. 3, March - April 2008, pp. 55-90.
Cortellessa V., ‘Performance antipatterns: state of the art and future perspectives’, 10th European Workshop, EPEW
2013, Venice, Italy, September 16-17, 2013. Proceedings, pp. 1-6. DOI: 10.1007/978-3-642-40725-3_1
Catia Trubiani, Anne Koziolek, Vittorio Cortellessa, Ralf Reussner, Guilt-based handling of software performance
antipatterns in palladio architectural models, Journal of Systems and Software, Volume 95, September 2014, Pages 141165, ISSN 0164-1212.
Wert et. al, ‘Supporting swift reaction: automatically uncovering performance problems by systematic experiments’.
Proceedings of the 2013 International Conference on Software Engineering, pp. 552-561 .
Wert et. al., ‘Automatic detection of performance anti-patterns in inter-component communications’. Proceedings of the
10th international ACM Sigsoft conference on Quality of software architectures, 2014, pp. 3-12 .
Trubiani et. al, ‘Exploring Synergies between Bottleneck Analysis and Performance Antipatterns’. Proceedings of the 5th
ACM/SPEC international conference on Performance engineering, 2014. Pages 75-86 .
Jeroslow, R. G. The polynomial hierarchy and a simple model for competitive analysis. Mathematical Programming, 32,
146–164, 1985.
Colson, B., Marcotte, P., Savard G. An overview of bilevel optimisation. Annals of Operations Research, 153, 235–256,
2007.
El-Ghazali, T. Metaheuristics: From Design to Implementation. Wiley Publishing, 2009.
Paul Baker et al. Model-Driven Testing, 183 pages, Springer, 2008.
OMG http://utp.omg.org
Arilo Dias Neto et al. A Survey on Model-based Testing Approaches: A Systematic Review, WEASELTech’07, ACM,
2007.
Diwakar Krishnamurthy, Jerome A. Rolia, Shikharesh Majumdar. A Synthetic Workload Generation Technique for
Stress Testing Session-Based Systems. IEEE Trans. Software Eng. 32(11): 868-882 (2006).
Gabriel Iuhasz et al. MODAClouds Deliverable D6.5.3: Runtime Environment Final Release, March 2015. (Available at
http://www.modaclouds.eu )
André van Hoorn, Christian Vögele, Eike Schulz, Wilhelm Hasselbring, Helmut Krcmar. Automatic Extraction of
Probabilistic Workload Specifications for Load Testing Session-Based Application Systems, in Proc. of Valuetools 2014.
Ningfang Mi, Giuliano Casale, Lucy Cherkasova, Evgenia Smirni. Injecting realistic burstiness to a traditional clientserver benchmark. Proceedings of the 6th international conference on Autonomic computing, 149-158, 2009.
Giuliano Casale, Amir Kalbasi, Diwakar Krishnamurthy, Jerry Rolia. BURN: Enabling Workload Burstiness in
Customised Service Benchmarks. IEEE Trans. Software Eng. 38(4): 778-793 (2012).
Vahid Garousi, Lionel C. Briand and Yvan Labiche. Traffic-aware Stress Testing of Distributed Systems Based on UML
Models, Proc. of ICSE 2006.
YCSB. https://github.com/brianfrankcooper/YCSB
Transaction Processing Council Benchmarks. http://www.tpc.org
Tilmann Rabl, Meikel Poess, Hans-Arno Jacobsen, Patrick O'Neil, and Elizabeth O'Neil. 2013. Variations of the star
schema benchmark to test the effects of data skew on query performance. In Proceedings of the 4th ACM/SPEC
International Conference on Performance Engineering (ICPE '13), ACM, New York, NY, USA, 361-372.
http://www.cs.brandeis.edu/~linearroad/
H. Herodotou, H. Lim, G. Luo, N. Borisov, L. Dong, F. B. Cetin, and S. Babu, ‘Starfish: A Self-Tuning System for Big
Data Analytics,’ Proc. 5th Biennial Conf. on Innovative Data Syst. Res. (CIDR ’11) (Asilomar, CA, 2011), pp. 261–272.
https://community.progress.com/community_groups/technology_partner/w/technologypartner/401.sonic-testharness.aspx
IBM Hursley, Performance Harness for Java Message Service, 2005. http://www.alphaworks.ibm.com/tech/perfharness
ActiveMQ, JMeter performance test, 2006 http://activemq.apache.org/jmeter-performance-tests.html
Copyright © 2015, DICE consortium – All rights reserved
115
Deliverable 1.1. State of the art analysis
[349]
[350]
[351]
[352]
[353]
[354]
[355]
[356]
[357]
[358]
[359]
[360]
[361]
[362]
[363]
[364]
[365]
[366]
[367]
[368]
[369]
[370]
[371]
[372]
[373]
[374]
[375]
[376]
[377]
[378]
[379]
[380]
[381]
[382]
[383]
[384]
[385]
[386]
[387]
[388]
[389]
[390]
[391]
[392]
[393]
[394]
[395]
[396]
[397]
[398]
[399]
JBoss, JBoss JMS Messaging Performance Framework, 2006 http://docs.jboss.org/jbossmessaging/docs/guide1.0.1.SP5/html_single/
Grinder: http://grinder.sourceforge.net/
Jython http://www.jython.org/
Clojure http://clojure.org/
Jmeter: http://jmeter.apache.org/
Markov4Jmeter: https://www.se.informatik.uni-kiel.de/en/research/projects/markov4jmeter
Selenium: http://www.seleniumhq.org/
MDload: https://github.com/imperial-modaclouds/modaclouds-mdload
Chaos Monkey https://github.com/Netflix/SimianArmy
AWS Autoscaling http://aws.amazon.com/autoscaling/
Microsoft Azure http://azure.microsoft.com/en-gb/
WazMonkey https://github.com/smarx/WazMonkey
Hadoop toolkit, https://code.google.com/p/hadoop-toolkit/wiki/HadoopPerformanceMonitoring
SequenceIQ, http://sequenceiq.com/
http://blog.sequenceiq.com/blog/2014/10/07/hadoop-monitoring/
Elasticsearch, https://www.elastic.co
Kibana, https://www.elastic.co/products/kibana
Logstash, http://logstash.net/
Collectd https://collectd.org/
Hadoop Vaidya, http://hadoop.apache.org/docs/r1.2.1/vaidya.html
Ganglia, http://ganglia.info
RRDtool http://oss.oetiker.ch/rrdtool/
Apache Ambari, https://ambari.apache.org/
Apache Chukwa, https://chukwa.apache.org/
DATASTAX, http://www.datastax.com
OpsCentre, http://www.datastax.com/what-we-offer/products-services/datastax-opscenter
Apache Cassandra http://cassandra.apache.org/
MMS MongoDB, https://www.mongodb.com/cloud
Server Density, https://www.serverdensity.com/plugins/mongodb
Manage Engine, https://www.manageengine.com/
Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P. and Witten, I. H. The WEKA Data Mining Software: An
Update SIGKDD Explor. Newsl., ACM, 2009, Vol. 11(1), pp. 10-18
http://weka.sourceforge.net/packageMetaData/distributedWekaBase/index.html
Bifet, A., Holmes, G., Kirkby, R. and Pfahringer, B. MOA: Massive Online Analysis J. Mach. Learn. Res., JMLR.org,
2010, Vol. 11, pp. 1601-1604
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R.,
Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M. and Duchesnay, E. Scikit-learn:
Machine Learning in Python, Journal of Machine Learning Research, 2011, Vol. 12, pp. 2825-2830
Cython C-Extension for Python http://cython.org/
http://www.numpy.org/
http://docs.scipy.org/doc/numpy/index.html
http://matplotlib.org/
Berthold, M. R., Cebron, N., Dill, F., Gabriel, T. R., Kötter, T., Meinl, T., Ohl, P., Sieb, C., Thiel, K. and Wiswedel, B.
KNIME: The Konstanz Information Miner Studies in Classification, Data Analysis, and Knowledge Organisation (GfKL
2007) Springer, 2007
LIBSVM https://www.csie.ntu.edu.tw/~cjlin/libsvm/
R programming language http://www.r-project.org/
Achtert, E., Kriegel, H.-P., Schubert, E. and Zimek, A. Interactive Data Mining with 3D-parallel-coordinate-trees
Proceedings of the 2013 ACM SIGMOD International Conference on Management of Data ACM, 2013, pp. 1009-1012
Fortran http://www.fortran.com/
http://cran.r-project.org/web/packages/available_packages_by_name.html
Wolfram Research, I. Mathematica 2015
Gdeisat, M. and Lilley, F. Matlab by Example: Programming Basics Elsevier Science Publishers B. V., 2013
https://github.com/shadanan/HadoopLink
http://www.mathworks.com/help/matlab/large-files-and-big-data.html
Heaton, J. Programming Neural Networks with Encog 2 in Java Heaton Research, Inc., 2010
http://www.heatonresearch.com/wiki/Encog_Performance
GPGPU https://en.wikipedia.org/wiki/General-purpose_computing_on_graphics_processing_units
Copyright © 2015, DICE consortium – All rights reserved
116
Deliverable 1.1. State of the art analysis
[400]
[401]
[402]
[403]
[404]
[405]
[406]
[407]
[408]
[409]
[410]
[411]
[412]
[413]
[414]
[415]
[416]
[417]
[418]
[419]
[420]
[421]
[422]
[423]
[424]
[425]
[426]
[427]
[428]
[429]
[430]
[431]
[432]
[433]
[434]
[435]
[436]
[437]
[438]
A. Coates, B. Huval, T. Wang, D. J. Wu, A. Y. Ng, B. Catanzaro, Deep learning with COTS HPC Systems, Proceedings
of the 30th International Conference on Machine Learning, Atlanta, USA, 2013
H. Schwenk, A. Rousseau, M Attik, Large, Pruned or Continuous Space Language Models on a GPU for Statistical
Machine Translation, Proceedings of the NAACL-HLT 2012 Workshop: Will We Ever Really Replace the N-gram
Model? On the Future of Language Modelling for HLT, June 2012, Montreal, Canada
Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R., Guadarrama, S. and Darrell, T. Caffe:
Convolutional Architecture for Fast Feature Embedding Proceedings of the ACM International Conference on
Multimedia ACM, 2014, pp. 675-678
Collobert, R., Kavukcuoglu, K. and Farabet, C. Torch7: A Matlab-like Environment for Machine Learning BigLearn,
NIPS Workshop 2011
Bastien, F., Lamblin, P., Pascanu, R., Bergstra, J., Goodfellow, I. J., Bergeron, A., Bouchard, N. and Bengio, Y. Theano:
new features and speed improvements Deep Learning and Unsupervised Feature Learning NIPS 2012 Workshop, 2012
https://code.google.com/p/cuda-convnet2/
http://deeplearning4j.org/gettingstarted.html
Ashari, A., Tatikonda, S., Boehm, M., Reinwald, B., Campbell, K., Keenleyside, J. and Sadayappan, P. On Optimising
Machine Learning Workloads via Kernel Fusion Proceedings of the 20th ACM SIGPLAN Symposium on Principles and
Practice of Parallel Programming ACM, 2015, pp. 173-182
http://devblogs.nvidia.com/parallelforall/digits-deep-learning-gpu-training-system/
H2O http://0xdata.com/product/
Franklin, M., Gonzalez, J., Jordan, M. I., Pan, X., Smith, V., Sparks, E., Talwalkar, A., Venkataraman, S. and Zaharia,
M. MLlib 2013
Apache Spark https://spark.apache.org/
Zaharia, M., Chowdhury, M., Franklin, M. J., Shenker, S. and Stoica, I. Spark: Cluster Computing with Working Sets
Proceedings of the 2Nd USENIX Conference on Hot Topics in Cloud Computing USENIX Association, 2010, pp. 10-10
Owen, S., Anil, R., Dunning, T. and Friedman, E. Mahout in Action Manning Publications Co., 2011
https://flink.apache.org/
jubat.us
Hido, S., Tokui, S. and Oda, S. Jubatus: An Open Source Platform for Distributed Online Machine Learning NIPS 2013
Workshop on Big Learning, Lake Tahoe
Jain, A. and Nalya, A. Learning Storm Packt Publishing, 2014 inproceedings (Jia14)
Shi, Q., Petterson, J., Dror, G., Langford, J., Smola, A. and Vishwanathan, S. Hash Kernels for Structured Data J. Mach.
Learn. Res., JMLR.org, 2009, Vol. 10, pp. 2615-2637
Microsoft Research http://research.microsoft.com/en-us/
Agarwal, A., Chapelle, O., Dudik, M. and Langford, J. A Reliable Effective Terascale Linear Learning System CoRR,
2011, Vol. abs/1110.4198
Rosen, J., Polyzotis, N., Borkar, V. R., Bu, Y., Carey, M. J., Weimer, M., Condie, T. and Ramakrishnan, R. Iterative
MapReduce for Large Scale Machine Learning CoRR, 2013, Vol. abs/1303.3517
Petuum http://petuum.github.io/
GraphLab Create https://dato.com/products/create/
Deeplearning4j http://deeplearning4j.org/
MLBase http://www.mlbase.org/
FinkML https://ci.apache.org/projects/flink/flink-docs-master/libs/ml/
NIMBLE http://users.cis.fiu.edu/~lzhen001/activities/KDD2011Program/docs/p334.pdf
SystemML http://researcher.watson.ibm.com/researcher/files/us-ytian/systemML.pdf
Amazon EC2 http://aws.amazon.com/ec2/
Google GCE https://cloud.google.com/compute/
Julia parallel computing http://julia.readthedocs.org/en/latest/manual/parallel-computing/
Julia programming language http://julialang.org/
Scala programming language http://www.scala-lang.org/
http://ellogon.org/
van Hoorn et. al., ‘Kieker: a framework for application performance monitoring and dynamic software analysis’.
Proceedings of the 3rd ACM/SPEC International Conference on Performance Engineering, 2014. Pages 247-248.
Perez et. al., ‘Towards a DevOps Approach for Software Quality Engineering’. Proceedings of the 2015 Workshop on
Challenges in Performance Methods for Software Development, 2015. Pages 5-10
Calinescu et.al., ‘Adaptive Model Learning for Continual Verification of Non-Functional Properties’. Proceedings of the
5th ACM/SPEC international conference on Performance engineering, 2014. Pages 87-98.
Bauer, A., Leucker, M., Schallhart, C.: Runtime Verification for LTL and TLTL. ACM Trans. Softw. Eng. Methodol,
20.4. September 2011.
Copyright © 2015, DICE consortium – All rights reserved
117
Deliverable 1.1. State of the art analysis
[439]
[440]
[441]
[442]
[443]
[444]
[445]
[446]
[447]
[448]
[449]
[450]
[451]
[452]
[453]
[454]
[455]
[456]
[457]
[458]
[459]
[460]
[461]
[462]
[463]
[464]
[465]
[466]
[467]
[468]
[469]
[470]
[471]
[472]
[473]
[474]
[475]
[476]
[477]
Pnueli 1977. The temporal logic of programmes. In Symposium on the Foundations of Computer Science (FOCS). IEEE
Computer Society Press, Providence, Rhode Island, 46–57.
Havelund and Rosu 2002. Synthesising Monitors for Safety Properties. In Tools and Algorithms for Construction and
Analysis of Systems (TACAS). Lecture Notes in Computer Science, vol. 2280. 342–356.
Hakansson, J., Jonsson, B., and Lundqvist, O. 2003. Generating online test oracles from temporal logic specifications.
Journal on Software Tools for Technology Transfer (STTT) 4, 4, 456–471.
Thati, P. and Rosu, G. 2005. Monitoring algorithms for metric temporal logic specifications. Electronic Notes in
Theoretical Computer Science (ENTCS) 113, 145–162.
Oded Maler, Dejan Nickovic 2004: Monitoring Temporal Properties of Continuous Signals. Proceedings of
FORMATS/FTRTFT: 152-166
Alur, R., Henzinger, T. A.: A Really Temporal Logic. J. ACM 41.4, pp. 181-203. ACM (1994)
Kuhtz, L., Finkbeiner: LTL path checking is efficiently parallelisable. In: Proc of ICALP 2009. LNCS, vol. 5556, pp.
235–246. Springer (2009)
Markey, N., Raskin, J.: Model Checking Restricted Sets of Timed Paths. CONCUR 2004. LNCS, vol. 3170, pp. 432-447.
Springer (2004).
Rosu, G., Chen, F.: Semantics and algorithms for parametric monitoring. Logical Methods in Computer Science 8(1)
(2012)
Medhat, R., Joshi, Y., Bonakdarpour, B., Fischmeister, S.: Parallelised runtime verification of first-order LTL
specifications (2014), technical report
Bauer, A., Falcone, Y. : Decentralised LTL monitoring. In: Proc of FM 2012. LNCS, vol. 7436, pp. 85–100. Springer
(2012)
Barre, B., Klein, M., Soucy-Boivin, M., Ollivier, P.A., Hallé, S.: MapReduce for parallel trace validation of LTL
properties. In: Proc. of RV 2012. LNCS, vol. 7687, pp. 184–198. Springer (2012)
Bianculli, D., Ghezzi, C., Krstic , S.: Trace checking of metric temporal logic with aggregating modalities using
MapReduce. In: Proc. of SEFM 2014. LNCS, vol. 8702, pp. 144–158. Springer (2014)
Bersani, M.M, Bianculli D., Ghezzi, C., Krstic, S., San Pietro, P.]: Lazy Semantics for MTL: An Application to Trace
Checking with MapReduce . Submitted to ATVA 2015
Apache Tez, http://tez.apache.org, accessed on 08/05/2015.
Hortonworks Data Platform, http://hortonworks.com/hdp/, accessed on 08/05/2015.
Cloudera CDH, http://www.cloudera.com/content/cloudera/en/products-and-services/cdh.html, accessed on 08/05/2015.
MapR, https://www.mapr.com/products/mapr-distribution-including-apache-hadoop, accessed on 08/05/2015.
IBM BigInsights, http://www-01.ibm.com/software/data/infosphere/hadoop/enterprise.html, accessed on 08/05/2015.
Big R https://www01.ibm.com/support/knowledgecenter/SSPT3X_3.0.0/com.ibm.swg.im.infosphere.biginsights.analyze.doc/doc/t_analyze
_data_bigr.html
Big Sheets http://www-01.ibm.com/software/ebusiness/jstart/bigsheets/
Hive https://hive.apache.org/
Pig https://pig.apache.org/
HBase http://hbase.apache.org/
SQL-on-Hadoop https://www.mapr.com/why-hadoop/sql-hadoop/sql-hadoop-details
HDP cookbooks, rackerlabs/hdp-cookbooks, accessed on 08/05/2015.
Cloudera cookbook, RiotGamesCookbooks/cloudera-cookbook, accessed on 08/05/2015.
MapR cookbooks, boorad/mapr_cookbooks, accessed on 08/05/2015.
Amazon EC2 pricing http://aws.amazon.com/elasticmapreduce/pricing/
EMRFS http://docs.aws.amazon.com/ElasticMapReduce/latest/DeveloperGuide/emr-fs.html
Microsoft HDInsight, http://azure.microsoft.com/en-gb/services/hdinsight/, accessed on 08/05/2015.
Azure Blob http://justazure.com/azure-blob-storage-part-one-introduction/
https://www.flexiant.com/2012/12/03/application-provisioning/
Swift http://docs.openstack.org/developer/swift/
Villalpando, Bautista, April, and Abran. ‘Performance analysis model for Big Data applications in cloud computing.’
Journal of Cloud Computing 3.1 (2014): 1–20.
Verma, Cherkasova, and Campbell. ‘Profiling and evaluating hardware choices for MapReduce environments: An
application-aware approach.’ Performance Evaluation 79 (2014): 328–344.
Barbierato, Gribaudo, and Iacono. A Performance Modelling Language For Big Data Architectures. ECMS 2013: 511–
517.
Herodotou. ‘Hadoop performance models.’ arXiv preprint arXiv:1106.0940 (2011).
Zhang, Rajasekaran, Duan, Wood, and Zhu. ‘Minimising Interference and Maximising Progress for Hadoop Virtual
Machines.’
Copyright © 2015, DICE consortium – All rights reserved
118
Deliverable 1.1. State of the art analysis
[478]
[479]
[480]
[481]
[482]
[483]
[484]
[485]
[486]
[487]
[488]
[489]
[490]
[491]
[492]
[493]
[494]
[495]
[496]
[497]
[498]
[499]
[500]
[501]
[502]
[503]
[504]
[505]
[506]
[507]
[508]
[509]
[510]
[511]
[512]
[513]
[514]
[515]
[516]
[517]
Verma, Cherkasova, and Campbell. ‘ARIA: automatic resource inference and allocation for MapReduce environments.’
Proceedings of the 8th ACM International Conference on Autonomic Computing 14 Jun. 2011: 235-244.
Zhang, Cherkasova, and Thau Loo. ‘Parameterisable benchmarking framework for designing a MapReduce performance
model.’ Concurrency and Computation: Practice and Experience 26.12 (2014): 2005–2026.
Malekimajd, Rizzi, Ardagna, Ciavotta, Passacantando, and Moghavar. ‘Optimal Capacity Allocation for executing
MapReduce Jobs in Cloud Systems.’ SYNASC 2014: 385–392.
Yang and Sun. ‘An analytical performance model of MapReduce.’ Cloud Computing and Intelligence Systems (CCIS),
2011 IEEE International Conference on 15 Sep. 2011: 306–310.
Lin, Meng, Xu, and Wang. ‘A practical performance model for Hadoop MapReduce.’ Cluster Computing Workshops
(CLUSTER WORKSHOPS), 2012 IEEE International Conference on 24 Sep. 2012: 231–239.
Lin, Zhang, Wierman, and Tan. ‘Joint optimisation of overlapping phases in MapReduce.’ Performance Evaluation
70.10 (2013): 720–735.
Tan, Wang, Yu, and Zhang. ‘Non-work-conserving effects in MapReduce: diffusion limit and criticality.’ The 2014 ACM
international conference on Measurement and modelling of computer systems 16 Jun. 2014: 181–192.
Bardhan and Menascé. ‘Queueing network models to predict the completion time of the map phase of MapReduce jobs.’
Proceedings of the Computer Measurement Group International Conference 3 Dec. 2012.
Vianna, Comarela, Pontes, J. Almeida, V. Almeida, Wilkinson, Kuno, and Dayal. ‘Analytical performance models for
MapReduce workloads.’ International Journal of Parallel Programming 41.4 (2013): 495–525.
Aguilera-Mendoza and Llorente-Quesada. Modelling and simulation of Hadoop Distributed File System in a cluster of
workstations. Model and Data Engineering 2013: 1–12.
Castiglione, Gribaudo, Iacono, and Palmieri. ‘Exploiting mean field analysis to model performances of Big Data
architectures.’ Future Generation Comp. Syst. 37 (2014): 203–211.
Zaharia, Chowdhury, Das, Dave, Ma, McCauley, Franklin, Shenker, Stoica. Resilient Distributed Datasets: A FaultTolerant Abstraction for In-Memory Cluster Computing. Berkeley University Technical Report 2012.
https://www.cs.berkeley.edu/~matei/papers/2012/nsdi_spark.pdf
Spark GCE, sigmoidanalytics/spark_gce, accessed on 07/05/2015.
Databricks Inc., https://databricks.com, accessed on 07/05/2015.
Ousterhout, Panda, Rosen, Venkataraman, Xin, Ratnasamy, Shenker, Stoica. The Case for Tiny Tasks in Compute
Clusters. https://www.eecs.berkeley.edu/~apanda/papers/tiny-tasks.pdf
Chowdhury. Performance and Scalability of Broadcast in Spark. Berkeley University Technical Report 2010.
http://www.cs.berkeley.edu/~agearh/cs267.sp10/files/mosharaf-spark-bc-report-spring10.pdf
Davidson, and Or. ‘Optimising Shuffle Performance in Spark.’ University of California, Berkeley — Department of
Electrical Engineering and Computer Sciences, Tech. Rep (2013).
Ousterhout, Rasti, Ratnasamy, Shenker, Chun. Making Sense of Performance in Data Analytics Frameworks.
https://www.eecs.berkeley.edu/~keo/publications/nsdi15-final147.pdf
Whänner K. Sep 10 2014 Real-Time Stream Processing as Game Changer in a Big Data World with Hadoop and data
Warehouse. May 7 2015 http://www.infoq.com/articles/stream-processing-hadoop
http://venturebeat.com/2014/03/19/without-stream-processing-theres-no-big-data-and-no-internet-of-things/
http://www.happiestminds.com/sites/all/themes/happiestminds/pdf/ANALYTICS_ON_BIG_FAST_DATA_USING_A_
REALTIME_STREAM_DATA_PROCESSING_ARCHITECTURE.pdf
RabbitMQ https://www.rabbitmq.com/
http://blog.infochimps.com/wp-content/uploads/2012/10/StormKafka.png
http://azure.microsoft.com/es-es/pricing/details/hdinsight/
http://azure.microsoft.com/es-es/documentation/articles/hdinsight-component-versioning/
http://azure.microsoft.com/es-es/pricing/calculator/
http://aws.amazon.com/es/kinesis/
http://aws.amazon.com/kinesis/pricing/
http://www.cloudera.com/content/cloudera/en/products-and-services.html
http://hortonworks.com/hadoop/storm/
http://dzone.com/articles/streaming-big-data-storm-spark
http://spark.apache.org/streaming/
http://samza.apache.org/
http://hadoop.apache.org/docs/r1.2.1/streaming.html
http://incubator.apache.org/s4/
https://github.com/Parsely/streamparse
http://docs.stratio.com/modules/streaming-cep-engine/development/
https://github.com/twitter/summingbird
https://github.com/facebookarchive/scribe
https://github.com/linkedin/pinot
Copyright © 2015, DICE consortium – All rights reserved
119
Deliverable 1.1. State of the art analysis
[518]
[519]
[520]
[521]
[522]
[523]
[524]
[525]
[526]
[527]
[528]
[529]
[530]
[531]
[532]
[533]
[534]
[535]
[536]
[537]
[538]
[539]
[540]
[541]
[542]
[543]
[544]
[545]
[546]
[547]
[548]
[549]
[550]
[551]
[552]
[553]
[554]
[555]
[556]
[557]
http://lambda-architecture.net/components/2014-06-30-speed-components/
https://samza.apache.org/learn/documentation/0.9/comparisons/storm.html
http://samza.apache.org/learn/documentation/0.7.0/comparisons/spark-streaming.html
ActiveMQ http://activemq.apache.org/
https://storm.apache.org/documentation/Guaranteeing-message-processing.html
https://storm.apache.org/documentation/Fault-tolerance.html
https://storm.apache.org/about/scalable.html
https://labs.spotify.com/2015/01/05/how-spotify-scales-apache-storm/
http://hortonworks.com/hadoop/ranger/
https://github.com/apache/storm/blob/master/SECURITY.md
Fay Chang, Jeffrey Dean, Sanjay Ghemawat, Wilson C. Hsieh, Deborah A. Wallach, Mike Burrows, Tushar Chandra,
Andrew Fikes, and Robert E. Gruber. Bigtable: A Distributed Storage System for Structured Data. OSDI'06: Seventh
Symposium on Operating System Design and Implementation, Seattle, WA, November, 2006.
Brewer, Eric. Towards Robust Distributed Systems. http://www.cs.berkley.edu/~brewer/cs262b-2004/PODC-keynote.pdf
M. Scavuzzo, E. Di Nitto, D. Ardagna, Building Data-Intensive Applications Exploiting Data as a Service Systems:
Experiences and Challenges. Politecnico di Milano. Technical Report n. 2014.10.
Amazon DynamoDB https://aws.amazon.com/dynamodb/
Amazon SimpleDB https://aws.amazon.com/simpledb/
Amazon Redis Labs https://redislabs.com/
Google Cloud Datastore https://cloud.google.com/datastore/
Azure DocumentDB http://azure.microsoft.com/en-us/services/documentdb/
Azure Redis Cache http://azure.microsoft.com/en-us/services/cache/
Azure Storage http://azure.microsoft.com/en-us/services/storage/
MongoLab https://mongolab.com/company/
Rackspace http://www.rackspace.com/data/managed-nosql
Brian F. Cooper, Adam Silberstein, Erwin Tam, Raghu Ramakrishnan, and Russell Sears. 2010. Benchmarking cloud
serving systems with YCSB. In Proceedings of the 1st ACM symposium on Cloud computing (SoCC '10). ACM, New
York, NY, USA, 143-154.
End Point. Benchmarking Top NoSQL Databases Apache Cassandra, Couchbase, HBase, and MongoDB. Revised: May
27, 2015. http://www.datastax.com/wp-content/themes/datastax-2014-08/files/NoSQL_Benchmarks_EndPoint.pdf
http://basho.com/fundamentals-of-nosql-security/
http://docs.mongodb.org/manual/security/
M. Scavuzzo, E. Di Nitto, S. Ceri, Interoperable Data Migration between NoSQL Columnar Databases, Enterprise
Distributed Object Computing Conference Workshops and Demonstrations (EDOCW), 2014 IEEE 18th International.
R. Osman and P. Piazzolla (Sept 2014). Modelling Replication in NoSQL Datastores. 11th International Conference on
Quantitative Evaluation of SysTems (QEST 2014), Sept 8–10, 2014, Florence, Italy.
C. Stewart, A. Chakrabarti, and R. Griffith. Zoolander: Efficiently meeting very strict, low-latency slos. In 10th ICAC,
2013.
P. Bailis, S. Venkataraman, M. J. Franklin, J. M. Hellerstein, and I. Stoica. Probabilistically bounded staleness for
practical partial quorums. PVLDB, 5(8):776{787, 2012.
Aaron Davidson, Andrew Or. Optimising Shuffle Performance in Spark. UC Berkeley technical report.
http://www.cs.berkeley.edu/~kubitron/courses/cs262a-F13/projects/reports/project16_report.pdf. Last accessed 19th May
2015.
N. Feamster, J. Rexford, E. Zegura. The Road to SDN: An Intellectual History of Programmable Networks
Nick McKeown, Tom Anderson, Hari Balakrishnan, Guru Parulkar, Larry Peterson, Jennifer Rexford, Scott Shenker and
Jonathan Turner. OpenFlow: Enabling Innovation in Campus Networks
Sushant Jain, Alok Kumar, Subhasree Mandal, Joon Ong, Leon Poutievski, Arjun Singh, Subbaiah Venkata, Jim
Wanderer, Junlan Zhou, Min Zhu, Jonathan Zolla, Urs Hölzle, Stephen Stuart and Amin Vahdat. B4: Experience with a
Globally-Deployed Software Defined WAN
Chi-Yao Hong, Srikanth Kandula, Ratul Mahajan, Ming Zhang, Vijay Gill, Mohan Nanduri, Roger Wattenhofer.
Achieving High Utilisation with Software-Driven WAN
https://www.openstack.org/assets/presentation-media/OpenDaylightOpenStackIcehouseSummit.pdf
http://www.opencontrail.org/
http://www.projectfloodlight.org/floodlight/
http://osrg.github.io/ryu/
Rob Sherwood , Glen Gibb , Kok-Kiong Yap , Guido Appenzeller, Martin Casado, Nick McKeown , Guru Parulkar.
FlowVisor: A Network Virtualisation Layer, Available online at:
http://archive.openflow.org/downloads/technicalreports/openflow-tr-2009-1-flowvisor.pdf
Copyright © 2015, DICE consortium – All rights reserved
120
Deliverable 1.1. State of the art analysis
[558]
[559]
[560]
[561]
[562]
[563]
[564]
[565]
[566]
[567]
[568]
[569]
[570]
[571]
[572]
[573]
[574]
[575]
[576]
[577]
[578]
[579]
[580]
[581]
[582]
[583]
[584]
[585]
Pankaj Berde , Matteo Gerola, Jonathan Hart, Yuta Higuchi, Masayoshi Kobayashi, Toshio Koide, Bob Lantz, Brian
O’Connor, Pavlin Radoslavov, William Snow, Guru Parulkar. ONOS: Towards an Open, Distributed SDN OS, Proc. of
HotSDN’14. Available online from: http://www-cs-students.stanford.edu/~rlantz/papers/onos-hotsdn.pdf
Ian F. Akyildiz, Ahyoung Lee, Pu Wang, Min Luo and Wu Chou. A roadmap for traffic engineering in SDN-OpenFlow
networks, Computer Networks, Volume 71, 4 October 2014, Pages 1–30
A.R. Curtis, J.C. Mogul, J. Tourrilhes, P. Yalagandula, P. Sharma, S. Banerjee, Devoflow: scaling flow management for
high performance networks, Proc. of SIGCOMM'11. Available online from:
http://conferences.sigcomm.org/sigcomm/2011/papers/sigcomm/p254.pdf
J.C. Mogul, J. Tourrilhes, P. Yalagandula, P. Sharma, A.R. Curtis, S. Banerjee, Devoflow: cost-effective flow
management for high performance enterprise networks, Proc. of Hotnets'10. Available online from:
http://conferences.sigcomm.org/hotnets/2010/papers/a1-mogul.pdf
M. Yu, J. Rexford, M.J. Freedman, J. Wang, Scalable flow-based networking with DIFANE, Proc. of SIGCOMM'10.
Available online from: http://www-bcf.usc.edu/~minlanyu/writeup/sigcomm10.pdf
A. Tootoonchian, Y. Ganjali, Hyperflow: a distributed control plane for openflow. Available online from:
http://www.cse.iitd.ac.in/~siy107537/csl374/a5/files/Tootoonchian.pdf
T. Koponen, M. Casado, N. Gude, J. Stribling, L. Poutievski, M. Zhu, R. Ramanathan, Y. Iwata, H. Inoue, T. Hama, et
al., Onix: a distributed control platform for large-scale production networks. Available online from:
https://www.usenix.org/legacy/event/osdi10/tech/full_papers/Koponen.pdf
Y. Hu, W. Wang, X. Gong, X. Que, S. Cheng, Balanceflow: controller load balancing for openflow networks, Proc. of
2nd International Conference on Cloud Computing and Intelligent Systems (CCIS), 2012 IEEE
S. Hassas Yeganeh, Y. Ganjali, Kandoo: a framework for efficient and scalable offloading of control applications. Proc.
of HotSDN'12. Available online from: http://conferences.sigcomm.org/sigcomm/2012/paper/hotsdn/p19.pdf
E. Ng, Maestro: A System For Scalable Openflow Control. Available online from:
http://www.cs.rice.edu/~eugeneng/papers/TR10-11.pdf
D. Erickson, The Beacon Openflow Controller, Proc. of HotSDN'13. Available online from:
http://yuba.stanford.edu/~derickso/docs/hotsdn15-erickson.pdf
A. Tootoonchian, S. Gorbunov, Y. Ganjali, M. Casado, R. Sherwood, On controller performance in software-defined
networks. Available online from: https://www.usenix.org/system/files/conference/hot-ice12/hotice12-final33_0.pdf
S. Sharma, D. Staessens, D. Colle, M. Pickavet, P. Demeester, Enabling fast failure recovery in openflow networks, Proc
of DRCN, 2011.
D. Staessens, S. Sharma, D. Colle, M. Pickavet, P. Demeester, Software defined networking: Meeting carrier grade
requirements, Proc. LANMAN, 2011.
S. Sharma, D. Staessens, D. Colle, M. Pickavet, P. Demeester, Openflow: meeting carrier-grade recovery requirements,
Computer Communications 36(6): 656-665 (2013)
A. Sgambelluri, A. Giorgetti, F. Cugini, F. Paolucci, P. Castoldi, Openflow-based segment protection in ethernet
networks, Optical Communications and Networking, IEEE/OSA Journal of (Volume:5 , Issue: 9), 2013.
P. Fonseca, R. Bennesby, E. Mota, A. Passito, A replication component for resilient openflow-based networking, NOMS
2012: 933-939.
M. Luo, Y. Tian, Q. Li, J. Wang, W. Chou, Sox – a generalised and extensible smart network openflow controller, First
SDN World Summit, 2012.
M. Luo, Y. Zeng, J. Li, An adaptive multi-path computation framework for centrally controlled networks, Journal of
Computer Networks, Elsevier, February 2015.
M. Reitblatt, N. Foster, J. Rexford, D. Walker, Consistent updates for software-defined networks: Change you can
believe in!, Proc. Hotnets, 2011.
M. Reitblatt, N. Foster, J. Rexford, C. Schlesinger, D. Walker, Abstractions for network update. Proc. of ACM
SIGCOMM 2012.
T. Mizrahi, Y. Moses, Time-based updates in software defined networks, SDNRG, IETF Meeting, Berlin, July 2013.
T. Mizrahi, Y. Moses, Time-based Updates in Openflow: A Proposed Extension to the Openflow Protocol, CCIT Report
#835, July 2013, EE Pub No. 1792, Technion, Israel.
A. Tootoonchian, M. Ghobadi, Y. Ganjali, Opentm: traffic matrix estimator for openflow networks, PAM'10 Proceedings
of the 11th international conference on Passive and active measurement.
M. Yu, L. Jose, R. Miao, Software defined traffic measurement with opensketch, Proc of NSDI 2013.
C. Yu, C. Lumezanu, Y. Zhang, V. Singh, G. Jiang, H.V. Madhyastha, Flowsense: monitoring network utilisation with
zero measurement cost, PAM 2013.
T. Benson, A. Anand, A. Akella, M. Zhang, Microte: fine grained traffic engineering for data centers, Proc of Co-Next
2011.
Joshua Reich, Christopher Monsanto, et al. Modular SDN Programming with Pyretic. ;login: issue: October 2013,
Volume 38, Number 5.
Copyright © 2015, DICE consortium – All rights reserved
121
Deliverable 1.1. State of the art analysis
[586]
[587]
[588]
[589]
[590]
[591]
[592]
[593]
[594]
[595]
[596]
[597]
[598]
[599]
[600]
[601]
[602]
[603]
Robert Soulé, Shrutarshi Basu, Parisa Jalili Marandi, Fernando Pedone, Robert Kleinberg, Emin Gün Sirer, and Nate
Foster. Merlin: A Language for Provisioning Network Resources. Proc. of CONEXT 2014.
http://ceph.com/
RADOS https://ceph.com/dev-notes/the-rados-distributed-object-store/
http://docs.flexiant.com/display/DOCS/Ceph+Support
https://www.openstack.org/summit/openstack-summit-hong-kong-2013/session-videos/presentation/ceph-the-de-factostorage-backend-for-openstackd
https://www.dreamhost.com/cloud/computing/
https://github.com/ceph/ceph
http://www.ssrc.ucsc.edu/Papers/weil-osdi06.pdf
Ceph: A Scalable, High-Performance Distributed File System Sage A. Weil Scott A. Brandt Ethan L. Miller Darrell D. E.
Long (http://ceph.com/docs/master/rados/operations/crush-map/)
https://ceph.com/docs/v0.79/rados/operations/auth-intro/
http://aws.amazon.com/s3/
http://aws.amazon.com/s3/pricing/
http://aws.amazon.com/storagegateway/
http://aws.amazon.com/s3/faqs/
Analysis of SAP HANA High Availability Capabilities. http://www.oracle.com/technetwork/database/availability/saphana-ha-analysis-cwp-1959003.pdf
K.Molka et al. Memory-Aware Sizing for In-Memory Databases. In Proc. NOMS 2014, 10 pages, IEEE Press.
Jennie Duggan, Ugur Çetintemel, Olga Papaemmanouil, Eli Upfal. Performance prediction for concurrent database
workloads. Proceedings of SIGMOD 2011, 337-348.
A. J. Elmore, S. Das, A. Pucher, D. Agrawal, A. El Abbadi, and X. Yan, ‘Characterizing tenant behaviour for placement
and crisis mitigation in multitenant DBMSs,’ in Proceedings of SIGMOD 2013, pp. 517– 528, ACM Press.
Copyright © 2015, DICE consortium – All rights reserved
122
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement