|Oracle9i Database Performance Planning
Release 2 (9.2)
Part Number A96532-01
Good system performance begins with design and continues throughout the life of your system. Carefully consider performance issues during the initial design phase, and it will be easier to tune your system during production.
This chapter contains the following sections:
System performance has become increasingly important as computer systems get larger and more complex and as the Internet plays a bigger role in business applications. In order to accommodate this, Oracle Corporation has designed a new performance methodology. It is based on years of Oracle designing and performance experience, and it explains clear and simple activities that can dramatically improve system performance.
Performance strategies vary in their effectiveness, and systems with different purposes, such as operational systems and decision support systems, require different performance skills. This book examines the considerations that any database designer, administrator, or performance expert should focus their efforts on.
System performance is designed and built into a system. It does not just happen. Performance problems are usually the result of contention for, or exhaustion of, some system resource. When a system resource is exhausted, the system is unable to scale to higher levels of performance. This new performance methodology is based on careful planning and design of the database, to prevent system resources from becoming exhausted and causing down-time. By eliminating resource conflicts, systems can be made scalable to the levels required by the business.
With the availability of relatively inexpensive, high-powered processors, memory, and disk drives, there is a temptation to buy more system resources to improve performance. In many situations, new CPUs, memory, or more disk drives can indeed provide an immediate performance improvement. However, any performance increases achieved by adding hardware should be considered a short-term relief to an immediate problem. If the demand and load rates on the application continue to grow, then the chance that you will face the same problem in the near future is very likely.
In other situations, additional hardware does not improve the system's performance at all. Poorly designed systems perform poorly no matter how much extra hardware is allocated. Before purchasing additional hardware, make sure that there is no serialization or single threading going on within the application. Long-term, it is generally more valuable to increase the efficiency of your application in terms of the number of physical resources used for each business transaction.
The word scalability is used in many contexts in development environments. The following section provides an explanation of scalability that is aimed at application designers and performance specialists.
Scalability is a system's ability to process more workload, with a proportional increase in system resource usage. In other words, in a scalable system, if you double the workload, then the system would use twice as many system resources. This sounds obvious, but due to conflicts within the system, the resource usage might exceed twice the original workload.
Examples of bad scalability due to resource conflicts include the following:
An application is said to be unscalable if it exhausts a system resource to the point where no more throughput is possible when it's workload is increased. Such applications result in fixed throughputs and poor response times.
Examples of resource exhaustion include the following:
This means that application designers must create a design that uses the same resources, regardless of user populations and data volumes, and does not put loads on the system resources beyond their limits.
Applications that are accessible through the Internet have more complex performance and availability requirements. Some applications are designed and written only for Internet use, but even typical back-office applications, such as a general ledger application, might require some or all data to be available online.
Characteristics of Internet age applications include the following:
Figure 1-1 illustrates the classic Internet/e-business and demand growth curve, with demand growing at an increasing rate. Applications must scale with the increase of workload and also when additional hardware is added to support increasing demand. Design errors can cause the implementation to reach its maximum, regardless of additional hardware resources or re-design efforts.
Internet applications are challenged by very short development timeframes with limited time for testing and evaluation. However, bad design generally means that at some point in the future, the system will need to be re-architected or re-implemented. If an application with known architectural and implementation limitations is deployed on the Internet, and if the workload exceeds the anticipated demand, then there is real chance of failure in the future. From a business perspective, poor performance can mean a loss of customers. If Web users do not get a response in seven seconds, then the user's attention could be lost forever.
In many cases, the cost of re-designing a system with the associated downtime costs in migrating to new implementations exceeds the costs of properly building the original system. The moral of the story is simple: design and implement with scalability in mind from the start.
When building applications, designers and architects should aim for as close to perfect scalability as possible. This is sometimes called linear scalability, where system throughput is directly proportional to the number of CPUs.
In real life, linear scalability is impossible for reasons beyond a designer's control. However, making the application design and implementation as scalable as possible should ensure that current and future performance objectives can be achieved through expansion of hardware components and the evolution of CPU technology.
The application has the biggest impact on scalability. For example:
However, the design is not the only problem. The physical implementation of the application can be the weak link. For example:
Bad capacity planning of all hardware components is becoming less of a problem as relative hardware prices decrease. However, too much capacity can mask scalability problems as the workload is increased on a system.
All software components have scalability and resource usage limitations. This applies to application servers, database servers, and operating systems. Application design should not place demands on the software beyond what it can handle.
Hardware is not perfectly scalable. Most multiprocessor machines can get close to linear scaling with a finite number of CPUs, but after a certain point each additional CPU can increase performance overall, but not proportionately. There might come a time when an additional CPU offers no increase in performance, or even degrades performance. This behavior is very closely linked to the workload and the operating system setup.
There are two main parts to a system's architecture:
Today's designers and architects are responsible for sizing and capacity planning of hardware at each tier in a multitier environment. It is the architect's responsibility to achieve a balanced design. This is analogous to a bridge designer who must consider all the various payload and structural requirements for the bridge. A bridge is only as strong as its weakest component. As a result, a bridge is designed in balance, such that all components reach their design limits simultaneously.
The main hardware components are the following:
There can be one or more CPUs, and they can vary in processing power from simple CPUs found in hand-held devices to high-powered server CPUs. Sizing of other hardware components is usually a multiple of the CPUs on the system.
Database and application servers require considerable amounts of memory to cache data and avoid time-consuming disk access.
The I/O subsystem can vary between the hard disk on a client PC and high performance disk arrays. Disk arrays can perform thousands of I/Os each second and provide availability through redundancy in terms of multiple I/O paths and hot pluggable mirrored disks.
All computers in a system are connected to a network, from a modem line to a high speed internal LAN. The primary concerns with network specifications are bandwidth (volume) and latency (speed).
Oracle9i Database Performance Tuning Guide and Reference for more information on tuning these resources
The same way computers have common hardware components, applications have common functional components. By dividing software development into functional components, it is possible to comprehend the application design and architecture better. Some components of the system are performed by existing software bought to accelerate application implementation or to avoid re-development of common components.
The difference between software components and hardware components is that while hardware components only perform one task, a piece of software can perform the roles of various software components. For example, a disk drive only stores and retrieves data, but a client program can manage the user interface and perform business logic.
Most applications involve the following components:
This component is the most visible to application users. This includes the following functions:
This component implements core business rules that are central to the application function. Errors made in this component could be very costly to repair. This component is implemented by a mixture of declarative and procedural approaches. An example of a declarative activity is defining unique and foreign keys. An example of procedure-based logic is implementing a discounting strategy.
Common functions of this component include the following:
This component is implemented in all pieces of software. However, there are some requests and resources that can be influenced by the application design and some that cannot.
In a multiuser application, most resource allocation by user requests are handled by the database server or the operating system. However, in a large application where the number of users and their usage pattern is unknown or growing rapidly, the system architect must be proactive to ensure that no single software component becomes overloaded and unstable.
Common functions of this component include the following:
This component is largely the responsibility of the database server and the operating system.
Common functions of this component include the following:
Configuring the initial system architecture is a largely iterative process. Architects must satisfy the system requirements within budget and schedule constraints. If the system requires interactive users transacting business or making decisions based on the contents of a database, then user requirements drive the architecture. If there are few interactive users on the system, then the architecture is process-driven.
Examples of interactive user applications:
Examples of process-driven applications:
In many ways, process-driven applications are easier to design than multiuser applications because the user interface element is eliminated. However, because the objectives are process-oriented, architects not accustomed to dealing with large data volumes and different success factors can become confused. Process-driven applications draw from the skills sets used in both user-based applications and data warehousing. Therefore, this book focuses on evolving system architectures for interactive users.
Generating a system architecture is not a deterministic process. It requires careful consideration of business requirements, technology choices, existing infrastructure and systems, and actual physical resources, such as budget and manpower.
The following questions should stimulate thought on architecture, though they are not a definitive guide to system architecture. These questions demonstrate how business requirements can influence the architecture, ease of implementation, and overall performance and availability of a system. For example:
Most applications fall into one of the following categories:
For this type of application, there is usually one user. The focus of the application design is to make the single user as productive as possible by providing good response time, yet make the application require minimal administration. Users of these applications rarely interfere with each other and have minimal resource conflicts.
For this type of application, the users are limited by the number of employees in the corporation actually transacting business through the system. Therefore, the number of users is predictable. However, delivering a reliable service is crucial to the business. The users will be using a shared resource, so design efforts must address response time under heavy system load, escalation of resource for each session usage, and room for future growth.
For this type of application, extra engineering effort is required to ensure that no system component exceeds its design limits. This would create a bottleneck that brings the system to a halt and becomes unstable. These applications require complex load balancing, stateless application servers, and efficient database connection management. In addition, statistics and governors should be used to ensure that the user gets some feedback if their requests cannot be satisfied due to system overload.
The choices of user interface range from a simple Web browser to a custom client program.
The distance between users influences how the application is engineered to cope with network latencies. The location also affects which times of the day are busy, when it is impossible to perform batch or system maintenance functions.
Network speed affects the amount of data and the conversational nature of the user interface with the application and database servers. A highly conversational user interface can communicate with back-end servers on every key stroke or field level validation. A less conversational interface works on a screen-sent and a screen-received model. On a slow network, it is impossible to get good data entry speeds with a highly conversational user interface.
The amount of data queried online influences all aspects of the design, from table and index design to the presentation layers. Design efforts must ensure that user response time is not a function of the size of the database. If the application is largely read only, then replication and data distribution to local caches in the application servers become a viable option. This also reduces workload on the core transactional server.
Consideration of the user type is important. If the user is an executive who requires accurate information to make split second decisions, then user response time cannot be compromised. Other types of users, such as users performing data entry activities, might not need such a high level of performance.
This is mandatory for today's Internet applications where trade is conducted 24 hours a day. However, corporate systems that run in a single time zone might be able to tolerate after-hours downtime. This after-hours downtime can be used to run batch processes or to perform system administration. In this case, it might be more economic not to run a fully-available system.
It is important to determine if transactions need to be executed within the user response time, or if they can they be queued for asynchronous execution.
The following are secondary questions, which can also influence the design, but really have more impact on budget and ease of implementation. For example:
This influences the sizing of the database server machine. On systems with a very large database, it might be necessary to have a bigger machine than dictated by the workload. This is because the administration overhead with large databases is largely a function of the database size. As tables and indexes grow, it takes proportionately more CPUs to allow table reorganizations and index builds to complete in an acceptable time limit.
This section describes design decisions that are involved in building applications.
Applications are no different than any other designed and engineered product. Well-designed structures, machines, and tools are usually reliable, easy to use and maintain, and simple in concept. In the most general terms, if the design looks right, then it probably is right. This principle should always be kept in mind when building applications.
Consider some of the following design issues:
Data modeling is important to successful relational application design. This should be done in a way that quickly represents the business practices. Chances are, there will be heated debates about the correct data model. The important thing is to apply greatest modeling efforts to those entities affected by the most frequent business transactions. In the modeling phase, there is a great temptation to spend too much time modeling the non-core data elements, which results in increased development lead times. Use of modeling tools can then rapidly generate schema definitions and can be useful when a fast prototype is required.
Table design is largely a compromise between flexibility and performance of core transactions. To keep the database flexible and able to accommodate unforeseen workloads, the table design should be very similar to the data model, and it should be normalized to at least 3rd normal form. However, certain core transactions required by users can require selective denormalization for performance purposes.
Examples of this technique include storing tables pre-joined, the addition of derived columns, and aggregate values. Oracle provides numerous options for storage of aggregates and pre-joined data by clustering and materialized view functions. These features allow a simpler table design to be adopted initially.
Again, focus and resources should be spent on the business critical tables, so that good performance can be achieved. For non-critical tables, shortcuts in design can be adopted to enable a more rapid application development. If, however, in prototyping and testing a non-core table becomes a performance problem, then remedial design effort should be applied immediately.
Index design is also a largely iterative process, based on the SQL generated by application designers. However, it is possible to make a sensible start by building indexes that enforce primary key constraints and indexes on known access patterns, such as a person's name. As the application evolves and testing is performed on realistic sizes of data, certain queries will need performance improvements for which building a better index is a good solution. The following list of indexing design ideas should be considered when building a new index:
One of the easiest ways to speed up a query is to reduce the number of logical I/Os by eliminating a table access from the execution plan. This can be done by appending to the index all columns referenced by the query. These columns are the select list columns and any required join or sort columns. This technique is particularly useful in speeding up online applications response times when time-consuming I/Os are reduced. This is best applied when testing the application with properly sized data for the first time.
The most aggressive form of this technique is to build an index-organized table (IOT). However, you must be careful that the increased leaf size of an IOT does not undermine the efforts to reduce I/O.
There are several index types available, and each index has benefits for certain situations. The following list gives performance ideas associated with each index type.
These are the standard index type, and they are excellent for primary key and highly-selective indexes. Used as concatenated indexes, B-tree indexes can be used to retrieve data sorted by the index columns.
These are suitable for low cardinality data. Through compression techniques, they can generate a large number of rowids with minimal I/O. Combining bitmap indexes on non-selective columns allows efficient
OR operations with a great number of rowids with minimal I/O. Bitmap indexes are particularly efficient in queries with
COUNT(), because the query can be satisfied within the index.
These indexes allow access through a B-tree on a value derived from a function on the base data. Function-based indexes have some limitations with regards to the use of nulls, and they require that you have the cost-based optimizer enabled.
Function-based indexes are particularly useful when querying on composite columns to produce a derived result or to overcome limitations in the way data is stored in the database. An example of this is querying for line items in an order exceeding a certain value derived from (sales price - discount) x quantity, where these were columns in the table. Another example is to apply the
UPPER function to the data to allow case-insensitive searches.
Partitioning a global index allows partition pruning to take place within an index access, which results in reduced I/Os. By definition of good range or list partitioning, fast index scans of the correct index partitions can result in very fast query times.
These are designed to eliminate index hot spots on insert applications. These indexes are excellent for insert performance, but they are limited in that they cannot be used for index range scans.
Building and maintaining an index structure can be expensive, and it can consume resources such as disk space, CPU, and I/O capacity. Designers must ensure that the benefits of any index outweigh the negatives of index maintenance.
Use this simple estimation guide for the cost of index maintenance: Each index maintained by an
UPDATE of the indexed keys requires about three times as much resource as the actual DML operation on the table. What this means is that if you
INSERT into a table with three indexes, then it will be approximately 10 times slower than an
INSERT into a table with no indexes. For DML, and particularly for
INSERT-heavy applications, the index design should be seriously reviewed, which might require a compromise between the query and
Use of sequences, or timestamps, to generate key values that are indexed themselves can lead to database hotspot problems, which affect response time and throughput. This is usually the result of a monotonically growing key that results in a right-growing index. To avoid this problem, try to generate keys that insert over the full range of the index. This results in a well-balanced index that is more scalable and space efficient. You can achieve this by using a reverse key index or using a cycling sequence to prefix and sequence values.
Designers should be flexible in defining any rules for index building. Depending on your circumstances, use one of the following two ways to order the keys in an index:
Views can speed up and simplify application design. A simple view definition can mask data model complexity from the programmers whose priorities are to retrieve, display, collect, and store data.
However, while views provide clean programming interfaces, they can cause sub-optimal, resource-intensive queries. The worst type of view use is when a view references other views, and when they are joined in queries. In many cases, developers can satisfy the query directly from the table without using a view. Usually, because of their inherent properties, views make it difficult for the optimizer to generate the optimal execution plan.
In the design and architecture phase of any system development, care should be taken to ensure that the application developers understand SQL execution efficiency. To do this, the development environment must support the following characteristics:
Connecting to the database is an expensive operation that is highly unscalable. Therefore, the number of concurrent connections to the database should be minimized as much as possible. A simple system, where a user connects at application initialization, is ideal. However, in a Web-based or multitiered application, where application servers are used to multiplex database connections to users, this can be difficult. With these types of applications, design efforts should ensure that database connections are pooled and are not reestablished for each user request.
Maintaining user connections is equally important to minimizing the parsing activity on the system. Parsing is the process of interpreting a SQL statement and creating an execution plan for it. This process has many phases, including syntax checking, security checking, execution plan generation, and loading shared structures into the shared pool. There are two types of parse operations:
Because parsing should be minimized as much as possible, application developers should design their applications to parse SQL statements once and execute them many times. This is done through cursors. Experienced SQL programmers should be familiar with the concept of opening and re-executing cursors.
Application developers must also ensure that SQL statements are shared within the shared pool. To do this, bind variables to represent the parts of the query that change from execution to execution. If this is not done, then the SQL statement is likely to be parsed once and never re-used by other users. To ensure that SQL is shared, use bind variables and do not use string literals with SQL statements. For example:
Statement with string literals:
Statement with bind variables:
The following example shows the results of some tests on a simple OLTP application:
Test #Users Supported No Parsing all statements 270 Soft Parsing all statements 150 Hard Parsing all statements 60 Re-Connecting for each Transaction 30
These tests were performed on a four-CPU machine. The differences increase as the number of CPUs on the system increase.
The choice of development environment and programming language is largely a function of the skills available in the development team and architectural decisions made when specifying the application. There are, however, some simple performance management rules that can lead to scalable, high-performance applications.
The programming model can vary between HTML generation and calling the windowing system directly. The development method should focus on response time of the user interface code. If HTML or Java is being sent over a network, then try to minimize network volume and interactions.
Interpreted languages, such as Java and PL/SQL, are ideal to encode business logic. They are fully portable, which makes upgrading logic relatively easy. Both languages are syntactically rich to allow code that is easy to read and interpret. If business logic requires complex mathematical functions, then a compiled binary language might be needed. The business logic code can be on the client machine, the application server, and the database server. However, the application server is the most common location for business logic.
Most of this is not affected by the programming language, but tools and 4th generation languages that mask database connection and cursor management might use inefficient mechanisms. When evaluating these tools and environments, check their database connection model and their use of cursors and bind variables.
Most of this is not affected by the programming language.
DECODEcase statements are very often candidates for optimization, as are statements with a large amount of
ORpredicates or set operators, such as
The most common examples of candidates for local caching include the following:
DUALcan account for over 60% of the workload on a database.
The design implications of using a cache should be considered. For example, if a user is connected at midnight and the date is cached, then the date value he has becomes invalid.
The two biggest challenges in application development today are the increased use of Java to replace compiled C or C++ applications, and increased use of object-oriented techniques, influencing the schema design.
Java provides better portability of code and availability to programmers. However, there are a number of performance implications associated with Java. Because Java is an interpreted language, it is slower at executing similar logic than compiled languages such as C. As a result, resource usage of client machines increases. This requires more powerful CPUs to be applied in the client or middle-tier machines and greater care from programmers to produce efficient code.
Because Java is an object-oriented language, it encourages insulation of data access into classes not performing the business logic. As a result, programmers might invoke methods without knowledge of the efficiency of the data access method being used. This tends to result in database access that is very minimal and uses the simplest and crudest interfaces to the database.
With this type of software design, queries do not always include all the
WHERE predicates to be efficient, and row filtering is performed in the Java program. This is very inefficient. In addition, for DML operations, and especially for
INSERTs are performed, making use of the array interface impossible. In some cases, this is made more inefficient by procedure calls. More resources are used moving the data to and from the database than in the actual database calls.
In general, it is best to place data access calls next to the business logic to achieve the best overall transaction design.
The acceptance of object-orientation at a programming level has led to the creation of object-oriented databases within the Oracle Server. This has manifested itself in many ways, from storing object structures within
BLOBs and only using the database effectively as an indexed card file to the use of the Oracle object relational features.
If you adopt an object-oriented approach to schema design, then make sure that you do not lose the flexibility of the relational storage model. In many cases, the object-oriented approach to schema design ends up in a heavily denormalized data structure that requires considerable maintenance and
REF pointers associated with objects. Often, these designs represent a step backward to the hierarchical and network database designs that were replaced with the relational storage method.
In summary, if you are storing your data in your database for the long-term and you anticipate a degree of ad hoc queries or application development on the same schema, then you will probably find that the relational storage method gives the best performance and flexibility.
You could experience errors in your sizing estimates when dealing with variable length data if you work with a poor sample set. Also, as data volumes grow, your key lengths could grow considerably, altering your assumptions for column sizes.
When the system becomes operational it becomes harder to predict database growth, especially that of indexes. Tables grow over time, and indexes are subject to the individual behavior of the application in terms of key generation, insertion pattern, and deletion of rows. The worst case is where you insert using an ascending key and then delete most rows from the left-hand side but not all the rows. This leaves gaps and wasted space. If you have index use like this make sure that you know how to use the online index rebuild facility.
Most good DBAs monitor space allocation for each object and look for objects that could grow out of control. A good understanding of the application can highlight objects that could grow rapidly or unpredictably. This is a crucial part of both performance and availability planning for any system. When implementing the production database, the design should attempt to ensure that minimal space management takes place when interactive users are using the application. This applies for all data, temp, and rollback segments.
Estimation of workloads for capacity planning and testing purposes is often described as a black art. When considering the number of variables involved it is easy to see why this process is largely impossible to get precisely correct. However, designers need to specify machines with CPUs, memory, and disk drives, and eventually roll out an application. There are a number of techniques used for sizing, and each technique has merit. When sizing, it is best to use at least two methods to validate your decision-making process and provide supporting documentation.
This is an entirely empirical approach where an existing system of similar characteristics and known performance is used as a basis system. The specification of this system is then modified by the sizing specialist according to the known differences. This approach has merit in that it correlates with an existing system, but it provides little assistance when dealing with the differences.
This approach is used in nearly all large engineering disciplines when preparing the cost of an engineering project be it a large building, a ship, a bridge, or an oil rig. If the reference system is an order of magnitude different in size from the anticipated system, then some of the components could have exceeded their design limits.
The benchmarking process is both resource and time consuming, and it might not get the correct results. By simulating in a benchmark an application in early development or prototype form, there is a danger of measuring something that has no resemblance to the actual production system. This sounds strange, but over the many years of benchmarking customer applications with the database development organization, we have yet to see good correlation between the benchmark application and the actual production system. This is mainly due to the number of application inefficiencies introduced in the development process.
However, benchmarks have been used successfully to size systems to an acceptable level of accuracy. In particular, benchmarks are very good at determining the actual I/O requirements and testing recovery processes when a system is fully loaded.
Benchmarks by their nature stress all system components to their limits. As all components are being stressed be prepared to see all errors in application design and implementation manifest themselves while benchmarking. Benchmarks also test database, operating system, and hardware components. Because most benchmarks are performed in a rush, expect setbacks and problems when a system component fails. Benchmarking is a stressful activity, and it takes considerable experience to get the most out of a benchmarking exercise.
Modeling the application can range from complex mathematical modeling exercises to the classic simple calculations performed on the back of an envelope. Both methods have merit, with one attempting to be very precise and the other making gross estimates. The down side of both methods is that they do not allow for implementation errors and inefficiencies.
The estimation and sizing process is an imprecise science. However, by investigating the process, some intelligent estimates can be made. The whole estimation process makes no allowances for application inefficiencies introduced by writing bad SQL, poor index design, or poor cursor management. A good sizing engineer builds in margin for application inefficiencies. A good performance engineer discovers the inefficiencies and makes the estimates look realistic. The process of discovering the application inefficiencies is described in the Oracle performance method.
The testing process mainly consists of functional and stability testing. At some point in the process, performance testing is performed.
The following list describes some simple rules for performance testing an application. If correctly documented, this provides important information for the production application and the capacity planning process after the application has gone live.
All testing must be done with fully populated tables. The test database should contain data representative of the production system in terms of data volume and cardinality between tables. All the production indexes should be built and the schema statistics should be populated correctly.
All testing should be performed with the optimizer mode that will be used in production. All Oracle research and development effort is focused upon the cost-based optimizer, and therefore Oracle Corporation recommends the use of the cost-based optimizer.
A single user on an idle or lightly used system should be tested for acceptable performance. If a single user cannot get acceptable performance under ideal conditions, it is impossible there will be good performance under multiple users where resources are shared.
Obtain an execution plan for each SQL statement, and some metrics should be obtained for at least one execution of the statement. This process should be used to validate that a good execution plan is being obtained by the optimizer and the relative cost of the SQL statement is understood in terms of CPU time and physical I/Os. This process assists in identifying the heavy use transactions that will require the most tuning and performance work in the future.
This process is difficult to perform accurately, because user workload and profiles might not be fully quantified. However, transactions performing DML statements should be tested to ensure that there are no locking conflicts or serialization problems.
It is important to test with a configuration as close to the production system as possible. This is particularly important with respect to network latencies, I/O sub-system bandwidth and processor type and speed. A failure to do this could result in an incorrect analysis of potential performance problems.
When benchmarking, it is important to measure the performance under steady state conditions. Each benchmark run should have a ramp-up phase, where users are connected to the application and gradually start performing work on the application. This process allows for frequently cached data to be initialized into the cache and single execution operations, such as parsing, to be completed prior to the steady state condition. Likewise, at the end of a benchmark run, there should be a ramp-down period, where resources are freed from the system and users cease work and disconnect.
This section describes design decisions involved deploying applications.
When new applications are rolled out, two strategies are commonly adopted:
Both approaches have merits and disadvantages. The Big Bang approach relies on good testing of the application at the required scale, but has the advantage of minimal data conversion and synchronization with the old system, because it is simply switched off. The Trickle approach allows debugging of scalability issues as the workload increases, but might mean that data needs to be migrated to and from legacy systems as the transition takes place.
It is hard to recommend one approach over the other, because each method has associated risks that could lead to system outages as the transition takes place. Certainly, the Trickle approach allows profiling of real users as they are introduced to the new application and allows the system to be reconfigured only affecting the migrated users. This approach affects the work of the early adopters, but limits the load on support services. This means that unscheduled outages only affect a small percentage of the user population.
The decision on how to roll out a new application is specific to each business. The approach adopted will have its own unique pressures and stresses. The more testing and knowledge derived from the testing process, the more you will realize what is best for the rollout.
To assist in the rollout process, build a list of tasks that, if performed correctly, increase the chance of good performance in production and, if there is a problem, enable rapid debugging of the application. For example:
MAXLOGHISTORYto values higher than what you anticipate for the rollout. This results in more disk space usage and bigger control files, but saves time later should these need extension in an emergency.
Oracle9i Database Performance Tuning Guide and Reference for guidance on setting minimal parameters in initial instance configuration
DELETErates should be created with either automatic segment space management or multiple freelists and an increased setting of
INITRANS. To avoid contention of rollback segments, either automatic undo management should be used or multiple rollback segments should be created to support the required user population.
Oracle9i Database Administrator's Guide for more information on using automatic undo management and on managing free space with automatic segment space management
WHEREclause predicates are sent as string literals. If the precompilers are used to develop the application, then make sure that the parameters
RELEASE_CURSORhave been reset from the default values prior to precompiling the application.