E135 EAServer Goes to the Races: Performance Testing, Tuning, and Tweaking for Speed and Endurance

1
E135 EAServer Goes to the Races Performance
Testing, Tuning, and Tweaking for Speed and
Endurance

• Karen Hartley
• Nomad Development, Inc.
• Khartley_at_NomadDev.com

2
Agenda

• Examining Performance
• What are the best tools?
• Limitations of Component Generation
• Streamlining the Architecture
• Boosting Performance

3
Performance What is it?

• Performance
• In system testing, this generally refers to the
  ability of a system to meet the requirements
  defined prior to testing.
• Systems are usually said to exhibit good or poor
  performance, depending on what criteria are used
  to measure that performance.
• Performance can be expressed in terms of
• Concurrent users
• Transactions per day
• Average response time
• Any other measure of speed

4
The Starting Gate

• Why are we here? How did we get here?
• EAServer applications are in place
• Component-based design
• 3 tier?
• Based upon a framework?

Business Logic
Presentation
Data Access
5
The Starting Gate

• Why are we here? How did we get here?
• EAServer applications are in place
• gtUsage is growing
• gtPerformance may be suffering
• EAServer limitations?
• Design Issues?

Usage
Performance
6
The Starting Gate

• Why are we here? Why should we spend the time
  and money to tweak and tune?
• Customers who encounter a smooth, error-free
  online experience are significantly more likely
  to engage in repeat transactions
• Lower-than-average performance means that they
  will find a site with higher performance
• By pinpointing performance bottlenecks and
  differentiating between software-and
  hardware-based problems, testing and monitoring
  solutions allow the highest possible productivity
  from existing equipment, bandwidth and personnel

7
Examining Performance - Tools

• What are we looking for?
• Problem Components
• When performance problems are found, you need the
  tools to quickly identify the component(s) at
  fault in order to implement a timely and
  effective remedy.
• Architecture/Design Issues
• Settings and Standards that could be modified?
• How do pooling, caching, and component types
  change testing results?

8
Testing Tools Load Testing

• Load Testing will demonstrate the effects of
  increasing the number of users and therefore the
  number of components instantiated/activated
  within a system
• What are your options?
• Mercury Interactive LoadRunner, Astra LoadTest
• Rational - Quantify
• Segue - SilkPerformer
• Empirix e-Load
• Compuware QA Center Performance Edition
• RadView Webload Resource Manager
• Cyrano - OpenSTA
• Others

9
Testing Tools Load Testing

• Lets look at some examples of what these tools
  can do for you
• Benefits?
• Pitfalls?
• Limitations?

10
Testing with EAServers Monitors

• Build your own component test bed
• You can then use EAServers monitoring tools to
  examine and tune your component performance
• Changes can be made and tested concerning
• Connection Caches
• Instance Pools
• Property Settings
• For Example
• 1) One pool, more StateFull instances -gt Jaguar
  crashes at about 2000 instances (-gt 2000
  threads)2) More pools, one StateLess instance
  per pool-gt Jaguar crashes at about 200 instances
  (-gt 200 threads)

11
Testing What can you learn?

• Memory utilization Check for memory leaks on
  the application processes and JVM memory usage if
  using a system designed around Java
• CPU High CPU can be an indicator of a
  well-designed system working very hard or a
  poorly designed system trying to work but being
  limited by other problems.
• Connection pool Watch the pool of connections
  to the database and verify they are being used
  correctly (released, reused).
• Queue Watch the queue. When requests start to
  queue, you have hit a bottleneck.
• Queue wait time A queue is acceptable, as long
  as the wait time in the queue is not excessive.
• Disk IO Verify that disk access is fast.

12
Architecture and Design Issues Streamlining for
Better Performance

• Now that weve got some information on
  performance, lets look at areas for improvement
• First Lets take a look at your architecture
• How many are using frameworks?
• May have gotten into the mode of creating a
  Component for each
• Interface Object
• Business Object
• Data Access Object

13
Framework Architecture Component Overload?

• Select a customer from a list, pull up one of
  their orders
• Component for the customer list (interface)
• Component for the customer business object
• Component for the actual data access to the
  customer
• Component for the data access of the order
• Component for the orders business object
• Component for the display of the order
  (interface)
• Misc other framework/support components
• What are the ramifications? How can this be
  simplified?

14
Component Overload Reversal

• Consider changing components to NVOs
• Do all of your objects inherit from COMPONENT
  ancestors?
• Lets talk through an example Interface objects
  
• Many are called up a via a director or factory,
  returning their output to a JSP
• Change their ancestor to an NVO may greatly
  reduce components, no change in app structure or
  code (other than call to instantiate)

15
Component Overload Reversal

• Consider denormalizing the Business and Data
  Access Layers
• How often does the business object merely pass
  along a get_data call to the data access layer?
• Consider combining the access of simple data into
  the business object
• If the business object instantiates multiple data
  access components, consider combining those
  components into one
• Can other simplifications or mergers be made?

16
Component Overload Reversal Why?

• Lets look at what happens when a component
  instance is loaded to see why reduction helps
• When a component is loaded by the PBVM, the PB
  classes are loaded from the PBDs specified in the
  com.sybase.jaguar.component.pb.librarylist
  property.
• Within the PBVM, a class group loader is
  responsible for accessing the PBDs and loading
  the appropriate PB class definitions to execute
  the business logic within a component. Those
  definitions consist of
• The virtual function table for the class
• The list of the prototypes for the functions and
  events defined in the class
• The p-codes for each routine defined in the class

17
Component Overload Reversal Why?

• Component instance loading continued
• gtThe symbol tables, which contain information
  about the variables of the class including shared
  variables, instance variables, and local
  variables
• The instance image, which is the blueprint for
  creating instances of a PowerBuilder class
• The constant pools, which contain literal values
  used explicitly within the PowerScript code as
  well as function names, variable names, and
  argument names

18
Component Overload Reversal Why?

• In addition, a unique instance of a class group
  loader is assigned to each distinct occurrence of
  a library list.
• This means that multiple instances of a single
  component will share the resources loaded for
  that component's definition because each instance
  of a component shares the same library list.
• Different components require different class
  group loaders.
• The com.sybase.jaguar.component.pb.librarylist
  property of those components will differ - the
  resolution of the in the property will point to
  different subdirectories within the Repository

19
Component Overload What else can be done?

• Sharing Class Group Loaders
• The instantiation of a class group loader is tied
  to the uniqueness of a component's library list
• Modify the com.sybase.jaguar.component.pb.libraryl
  ist property of the affected components. (This
  technique may help noticeably with component
  frameworks or components that share many common
  objects, such as 'helper' NVOs.)
• Include all PBLs containing the NVOs that are
  sharing resources in a single, combined library
  list.
• Place all of the component NVOs into the same PBL
  and target, or
• Modify the library lists of each component target
  to essentially be the concatenation of all the
  PBL(s) used to implement each individual
  component.

20
Component Overload What else can be done?

• Sharing Class Group Loaders continued
• Deploy each component normally from the PB
  project painter
• Ensure that the exact same deployment options in
  terms of PBL consolidation are specified on the
  Libraries tab of the EAServer Component Generator
  (deployment project properties) dialog for each
  project used to deploy these components.

21
Component Overload What else can be done?

• Sharing Class Group Loaders continued
• After deployment use Jaguar Manager to modify the
  com.sybase.jaguar.component.pb.librarylist
  property in the properties file of each component
  to be the exact same list by referencing a fully
  qualified path to the constituent PBDs
• Ensure that the designator is explicitly
  converted to a full path.
• Point to the PBDs within the Repository for one
  of the deployments, or
• Copy the PBDs elsewhere. If you do not qualify
  the path to the PBDs, the Windows PATH will be
  searched for those files.

22
Component Overload What else can be done?

• Sharing Class Group Loaders continued
• Use Jaguar Manager to refresh the components for
  which you made property file modifications.
• Backup the changes to the components' properties
  files, because each new deployment from the PB
  development environment will overwrite your
  changes to the com.sybase.jaguar.component.pb.libr
  arylist property.

23
Component Overload What else can be done?

• Sharing Class Group Loaders continued
• Use Jaguar Manager to refresh the components for
  which you made property file modifications.
• Backup the changes to the components' properties
  files, because each new deployment from the PB
  development environment will overwrite your
  changes to the com.sybase.jaguar.component.pb.libr
  arylist property.

24
What else can be tuned?

• Connection Caches
• As you test, do you see an increasing number of
  connections?
• This may point to a problem in the way you are
  handling transactional behavior.
• EAServer appears to think those connections are
  still part of an outstanding transaction
• A new request will always give you a new
  connection.
• This combined with the default setting of
  JAG_CM_FORCE for PowerBuilder's
  GetConnectionOption, could result in a larger
  number of active connections than allowed by the
  cache.

25
What else can be tuned?

• Connection Caches Continued
• Examine the transactional semantics of your
  components
• Make sure that you are disconnecting every
  connection at the appropriate point in a
  component's lifecycle

26
What else can be tuned?

• Instance Limitations
• EAServer 4.0 allows for a limit to be placed on
  the number of active instances of a component.
• You can also specify a maximum wait time (time to
  wait for an instance to become available).
• If the maximum number is hit, the constructor
  throws an exception
• The client or factory should catch and can retry.

27
Examples, Questions, and Additional Information

• Questions?
• For a copy of the latest version of these slides,
  demos, etc. please visit
• www.NomadDev.com