Architecting Object Applications for High Performance

Architecting Object Applications for High Performance
Architecting Object Applications for High
Performance with Relational Databases
Shailesh Agarwal1
Christopher Keene2
1.0 Abstract
Arthur M. Keller3
for building object-oriented applications that
access relational databases.
This paper presents an approach for architecting OO applications for high performance with
relational databases. The key ideas of this
approach are:
This approach enables organizations to derive
the benefits of OO technology while leveraging their investments in relational technology.
Developers building such object-relational
applications face some difficult problems such
as (i) Mapping the objects from the application
model to the relational schema in the database,
(ii) Managing the locking and transactions to
ensure data integrity, and (iii) Optimizing with
consideration of the performance characteristics of relational databases. Each of these problems present several interesting issues worthy
of discussion (see for example [Keller93]).
However, this paper focuses primarily on the
performance issues for such object-relational
applications and presents our experiences with
architecting high performance object-relational applications.
2.0 Introduction
3.0 Object-Relational Mediators
Object-oriented software development is rapidly becoming the leading approach for building flexible, scalable software systems in
client/server environments. Additionally, over
the past decade, relational technology has
matured and has been widely adopted for managing corporate data. Relational databases
have now become the standard data stores for
on-line transaction processing (OLTP) applications. These two trends are motivating the need
An object-relational application provides an
object-oriented interface to relational data. In
such applications the application object model
is mapped to a relational schema in the underlying database. Typically such applications
build or use a mediator for transforming object
operations to relational database calls and
vice-versa [Wiederhold92].
• Optimize business object mapping: tune the
mapping the between business objects and
relational tables to leverage relational technology.
• Perform client object management: use a
client-side object manager to minimize
database traffic.
FIGURE 1. Object/relational mediator
1. Author’s address: Persistence Software, 1700 S
Amphlett Blvd., Suite 250 San Mateo, CA 94402, or
[email protected]
2. Author’s address: Persistence Software
3. Author’s address: Stanford University, Computer Science Department, Stanford, CA 94305-2140 or
[email protected]
High Performance Object/Relational Applications
August 10, 1995
This document was created with FrameMaker 4.0.4
As shown in Figure 1, the mediator maps
objects to relational tuples and also ensure data
integrity with appropriate locking and transaction management. Tuples in a relational database are presented as instances of a class and
any updates on those instances are translated
into updates of the corresponding tuples in the
database. Additionally, the mediator could also
provide a client-side object cache for enhancing performance. In a sense, such object-relational mediators enable an OODBMS interface
to a legacy RBDMS.
completed without changing the object model
at all. The resulting application ran properly,
but where inheritance had assisted clustering
in the object database, it imposed a high performance penalty in the relational database.
Through our experiences with companies
implementing large scale object systems, we
have developed an approach to architecting
object applications for high performance.
Persistence, an object/relational environment
from Persistence Software, is a mediator which
creates and manages C++ objects linked to
relational databases. Examples of high performance systems built with Persistence include:
Our experience has shown that object-relational mapping and high performance can go
together, provided proper consideration is
given to object mapping and object management issues. Developers can optimize the mapping between a set of interrelated business
objects and a relational schema to leverage
relational technology. Further developers can
use a client-side object cache to enhance application performance.
• AT&T materials logistics: the LOGIC system supports over 300 warehouses worldwide to manage all of AT&T’s network
systems inventory. The application, which
took 50 developers over two years to build,
is the largest object-oriented application
ever deployed by AT&T.
• CBIS cellular billing: the Project 2000 system built by Cincinnati Bell Information
Systems manages cellular billing for Cincinnati Bell and Sprint. The system, which
uses Persistence teamed with an Oracle
database and Tuxedo, processes over 1 million transactions daily.
Other examples include the Federal Express
Ground Operations Command and Control
System, and the Conrail Locomotive Maintenance Scheduler. We have found that there are
a number of “ground rules” for achieving high
performance from object/relational applications. Developers who do not follow these
rules risk building systems which do not scale
well in operation.
In one memorable example, a team ported an
application from an object database to a relational database using Persistence. The port was
High Performance Object/Relational Applications
August 10, 1995
4.0 Achieving Performance
The basic approaches for optimizing object
access to relational data are similar to the
objectives for tuning the relational database
itself. The two approaches are:
• Maximize Server Performance: Applications should be written to ensure that the
queries and updates posed to the server can
be processed efficiently (e.g., minimize the
use of joins).
• Minimize Server Requests: Applications
should be written to minimize the total
number of queries and updates sent to the
relational data server (e.g., send all the
changes to a record at once, rather than
sending an update as each attribute is
Architecting applications with these two
approaches will enable them to achieve high
performance. The following sections discuss
these approaches in greater detail.
5.0 Server Performance
• Inheritance to relational rows or joins
The key objective of this approach is to pose
only those queries and updates that can be processed efficiently by the server. There are
potentially three ways to achieve this objective:
5.1 Mapping Classes
• Choose the appropriate object to relational
• Use query capabilities of relational databases.
• Take advantage of special performance features of relational databases.
Mapping efficiency between object classes and
relational tables is the most critical factor for
achieving high performance. “Pure” object
models sometimes map poorly to relational
structures, for example by requiring many
joins to construct simple objects. Such a mapping can significantly deteriorate application
performance. Hence, starting with an object
model and attempting to map to a relational
schema can sometimes lead to poor performance. On the other hand Entity-Relational
(E-R) models convert naturally to object models, mapping entities to objects and relationships to associations. E-R models can then be
optimized through selective denormalization
based on expected usage. The advantage of
this approach is that such denormalization is
well understand while optimizing object models is not. The best way to achieve this efficiency is to use a normalized relational schema
(E-R model) as the basis for the corresponding
object model.
Developers will find that the simplest mapping
between objects and normalized table typically
provides the best performance. Object-relational mapping requires the mapping of the
following three major modeling concepts:
• Class to relational rows
• Relationships to foreign keys
High Performance Object/Relational Applications
August 10, 1995
The most efficient mapping is most often the
direct mapping of a class to a relational table.
A more complex mapping can lead to performance penalties. Classes which represent a
join between several tables will be expensive
to construct and potentially impossible to
update [Keller85]. Hence, in general, developers should map each class to a single table;
more complex objects can always be built up
out of simple, “table-based” classes.
Two exceptions to this rule include creating
projection objects to minimize network traffic
and view objects for decision support.
• “Projection objects”: for tables with many
columns, it may be efficient to map a projection of the table to a “projection class”,
and retrieve a full row only when it is
needed. For example, a customer row may
contain 50 columns, but only the name and
phone number are needed most of the time.
By creating a class which maps to just the
needed columns, the developer can avoid
having to pass unnecessary information
across the network. Such a class must contain the primary key columns at a minimum.
• “View objects”: for decision support applications, it is often useful to create a view
table which represents a database join, and
then map this view to an object class. This
allows developers to take full advantage of
relational algebra to hide the physical data
model from the object application
5.2 Mapping Relationships
Relationships between objects map to foreign
keys between rows. How each mapping is
implemented, however, has a significant
impact on the performance and flexibility of
the resulting application. Developers have several choices for mapping object relationships
to relational tables:
• Distinct table: In this approach the relationship is represented as a distinct table in the
database. This approach provides the most
modelling flexibility since it makes adding
and removing relationships transparent to
other tables. However, this approach can be
expensive if the relationship is frequently
requirement for speed versus flexibility, different relationship mapping techniques are appropriate.
FIGURE 3. Relationship mapping
Mapping Approach
- Efficient
- Flexible
- Inefficient
- Requires
- Efficient
- Flexible
- Efficient
Relationship Type
FIGURE 2. Distinct table mapping
Object Model
Relationship: Customer
has-an Account
RDB Mapping
• CustId (PK)
Key Table
• CustId (FK)
• AcctId (FK)
• AcctId (PK)
• Embedded foreign key: This is the most
common approach for one-one and onemany relationships. In this case, for a given
class the primary key of a related class is
embedded in the class itself. This results in
a performance characteristic better than the
“distinct table” approach but worse than the
“embedded class” approach.
• Embedded classes: In this case, two classes
are merged into a single table. Such merging improves performance at the cost of
extensibility and violation of normalization.
In a one-many relationship, this approach
requires a fixed length array of the “many”
class to be embedded into the “one” class
As shown in the following table, depending on
the cardinality of the relationship and the
High Performance Object/Relational Applications
August 10, 1995
5.3 Mapping Inheritance
As with relationships, how object inheritance
is mapped to the database can have a profound
impact on performance. Here, particular attention to expected query paths plays an important role on the mapping choice.
A parent class can be either abstract or concrete. An abstract parent class is one which
will never be instantiated on its own, and consequently does not have any physical data
associated with it (e.g., no corresponding
table). A concrete parent class is one which
can exist on its own, and typically will have its
own associated table.
Developers also have choices about how to
map inheritance associations (for example, the
customer class inherits from the person class)
into relational tables.
• Vertical Partitioning: In this type of mapping each object model class maps to a corresponding table. The classic objectrelational mapping “mistake” is to create
1. Theoretically, an alternative is to embed the “one”
class in the “many” class. This option is not supported
by Persistence because of resulting update anomalies.
deep inheritance hierarchies which require
multi-way joins to retrieve basic object
• Horizontal Partitioning: In this case, only
leaf classes are mapped to tables and
include all of their inherited attributes. This
approach may give improved performance
since only one table needs to be accessed
for instances of a given leaf class.
with concrete parent classes must perform
expensive joins to retrieve instances.
As shown in the table below, depending on
whether the inherited superclass is abstract or
concrete and the requirement for speed versus
flexibility, different inheritance mapping techniques are appropriate.
FIGURE 5. Inheritance mapping
Mapping Approach
FIGURE 4. Horizontal partitioning
Object Model
Inheritance Type
Inheritance: Customer
and Employee Inherit
From Person
- Efficient
- Inefficient
- Efficient
- Inefficient
- Efficient
- Efficient
Abstract parent
Concrete parent
RDB Mapping
• PersonId (PK)
• CustName
• PersonId (PK)
• EmpName
• Typed Partitioning: Another way to handle
inheritance is to map all classes in an inheritance tree to a single table, using a type column to distinguish between subclasses. This
enables the retrieval of objects from multiple classes in a single query, but violates
Queries made at the parent class level will be
implemented very differently depending on the
inheritance mapping. If the parent class is concrete, a parent class query can be efficiently
handled by querying the table corresponding to
the parent class. If the parent class is abstract,
the query must be run against each child table,
a potentially time consuming operation.
On the other hand, retrieving a single instance
of a child class is much faster if all the parent
classes are abstract, as all necessary columns
are located in the child table. Child classes
High Performance Object/Relational Applications
August 10, 1995
5.4 Database Query Support
Developers can gain significant performance
benefits by using the sophisticated querying
capabilities available with relational databases
for ad-hoc access of objects. For the initial
select of a set of objects, ad-hoc access via the
relational query mechanism using indexes is
significantly faster than navigational access.
Once the data has been retrieved from the database then in-memory navigational queries
allow efficient use of the cached object information. In order to take advantage of this
query support, appropriate indexes should be
built on columns that are to be used for class
and relationship queries. In general, indexes to
be built include all primary and foreign keys in
the database schema.
5.5 Database Optimizations
Developers of object applications should make
sure that their applications take advantage of
any special features that the underlying database provides for optimizing performance.
Some examples of such special features are:
• Stored procedures: Implementing object
methods to create, read, update, and delete
objects through stored procedures can speed
performance by 20 percent or more.
• Array and bulk interfaces: Many databases
provide array interfaces where updates on a
collection of tuples can be sent in a batch. A
batched operation is significantly faster than
the corresponding operation done on a tuple
by tuple basis.
• Asynchronous queries: Some databases also
support non-blocking queries where the
database server returns control immediately
after receiving the query request from the
application and the application can then poll
the database server for the results of query
at a later time. This approach allows the
application to optimize its response time
and also use multi-threaded capabilities to
perform background processing of the
query results.
6.0 Minimize Server Requests
The previous section described how to optimize performance for an object application by
choosing the appropriate object-relational
mapping and by taking advantage of the features of the underlying database. This section
discusses how to enhance performance by
using a smart cache on the client side. Intelligent caching of data on the client can significantly reduce the database traffic for an
application and provide orders of magnitude
performance enhancements [Cattell94].
The basic approach to optimizing client object
management is to reduce number of database
accesses by using the client side cache as much
as possible. This provides a significant performance benefit since: (i) In-memory access is
High Performance Object/Relational Applications
August 10, 1995
three orders of magnitude faster than disk
access, and (ii) Costly network traffic is
6.1 Client-side cache
The basic model for client object management
is to cache data instances read from the database, register their primary key values, and
respond to queries based on this cached data.
As tuples are retrieved from the database, they
are converted to objects and “knit” together
according to the object model mapping to form
a network of in-memory objects.
FIGURE 6. Object cache
Client Object Manager
- Object model built by "knitting" separate queries
- Objects can be traversed in memory
- Objects model flushed upon commit
Cached data can also be changed repeatedly
without accessing the database, holding the
database update until the transaction is committed. This prevents sending multiple updates
for the same object to the database. In the case
that the transaction is aborted, no changes need
to be sent to the database at all. Client-side
caching is also critical for ensuring object data
integrity [Keller93].
6.2 Optimizing object navigation
Using such an in-memory network, queries
which follow foreign key pointers (navigations
queries) can be performed fully in the cache
once the basic object model information has
been retrieved [Keller95]. There is however an
issue of how to tell if a query can be satisfied
fully from cache data. This is related to the
well-known problem of “phantoms” [Gray93]
in databases. There are no simple solutions to
the problem in the general case. However,
often for specific applications it is possible to
assert that the cache contains all the data necessary for the query. In such cases, using the
cache for responding to the query can provide
significant performance benefits without compromising the validity of the response.
6.3 Concurrency and cache
Efficient cache management can also speed
transaction throughput, particularly by making use of techniques such as optimistic concurrency. In an optimistic transaction model no
database locks are placed for cached data.
Cache updates are made assuming that the
underlying data has not changed. Before sending the cache updates to the database, a check
is performed to ensure that the data update is
valid. The basic premise of this scheme is that
only a small fraction of the updates will be
found to be invalid. Such a scheme enables
high concurrency without data inconsistencies
and lost updates.
Another approach for cache synchronization is
to use a notification mechanism which notifies
applications of changes in the database that are
of interest to those applications. This feature is
not commonly available in commercial relational databases.
7.0 Benchmarking Results
Based on our experience, a well-constructed
object application should be somewhat faster
than a native relational database application:
for the “cold” database operations (create,
read, update and delete) the application should
be within 5 to 10 percent of native database
access; while for “warm” operations performed by querying or navigating the object
High Performance Object/Relational Applications
August 10, 1995
cache, the application should be roughly three
orders of magnitude faster than database
The actual performance is greatly dependent
on the degree to which the application can take
advantage of data stored in the object cache.
As [Nag95] reported, “Persistence offers many
of the performance benefits of an OODBMS
while retaining the reliability and portability of
the underlying relational database.”
8.0 Conclusions
Developers can achieve high performance
building object systems linked to relational
databases, provided they make the appropriate
trade-offs. Object models must be modified to
take into account the strengths and weaknesses
of the underlying datastore. Particular care
should be given to the complexity of the query
required to return a instances of commonly
used classes - joins can be deadly in these situations.
Here is a summary of the design guidelines for
achieving high performance in object applications linked to relational data:
1. Classes: keep it simple. Map classes to
tables where possible, except for special
cases such as projection classes to minimize
network traffic or view classes for decision
2. Relationships: keep it fast. Embed one-toone relationships into a single table, use foreign keys to navigate one-to-many relationships, avoid many-to-many relationships if
3. Inheritance: balance retrieval speed against
query efficiency. In general, horizontal
inheritance is the best choice as it minimizes retrieval speeds for child classes. For
special situations, vertical or typed inheritance provides faster query responses.
4. Object caching: reuse retrieved data. Caching object data allows multiple updates and
is critical to ensuring object integrity.
5. Object navigation: perform client-side relationship queries. Knitting retrieved objects
into an in-memory object model provides
orders of magnitude performance benefits
for data-intensive applications.
By following these guidelines and using the
appropriate tools, developers can achieve the
performance required for deploying large scale
object systems.
9.0 For More Information
Object Applications With Relational Databases,” ACM SIGMOD Proceedings, 1993.
[Keller95] Keller, A. M., and Basu, J., “A
Predicate-based Caching Scheme for ClientServer Database Architectures,” to appear in
VLDB Journal, 1995.
[Nag95] Nag, B., and Zhao, Y., “Implementing
the 007 Benchmark on Persistence,” Master’s
Thesis, Computer Sciences Department, University of Wisconsin-Madison, 1995.
[Wiederhold92] Wiederhold, G., Mediators in
the Architecture of Future Information Systems, IEEE Computer, March 1992.
For further information on Persistence Software, please contact [email protected] or
10.0 References
[Barsalou91] Barsalou, T., Siambela, N.,
Keller, A. M., and Wiederhold, G., Updating
Relational Databases through Object-Based
Views, ACM SIGMOD Proceedings, Denver,
CO, May 1991.
[Cattell94] Cattell, R., Object Data Management: Object-Oriented and Extended Relational Database Systems, Addison-Wesley,
Menlo Park, California, revised edition, 1994.
[Gray93] Gray, J., and Reuter, A., Transaction
Processing: Concepts And Techniques, Morgan Kaufman, San Mateo, California, 1993.
[Keller85] Keller, A. M., “Updating Relational
Databases through Views,” Ph.D. Dissertation,
Department of Computer Science, Stanford
University, February, 1985.
[Keller93] Keller, A. M., Agarwal, S. and
Jensen, R., “Enabling The Integration of
High Performance Object/Relational Applications
August 10, 1995
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