Reverse Engineering Performance Models for ComponentBased Software Systems

1
Reverse EngineeringPerformance Models for
Component-BasedSoftware Systems

• 2008-04-22
• SOSoRNet Workshop

Klaus Krogmann (krogmann_at_ipd.uka.de)
2
Motivation
3
Performance Prediction
?
How to get such models?

• Performance modelof a component-basedsoftware
  architecture

• Performance data
• Execution time
• Throughput
• Resource utilisation

4
Why Creating Models? (1)
Sizing / Scalability / Relocation
vs.
Changes in usage profile
vs.
Extensions of legacy software systems
Performance predictions at design time
5
Why Creating Models? (2)
Analytical solvable Performance jumps
Critical contention levels
Simulation Faster than real execution
6
Approach

• Process and Models

7
Reverse Engineering Component Behaviour
I
(Java)Source Code
8
Example File Share
BusinessLogic
compressed true
compressed false
Callcompress cfile.length fileInput.length
CallcheckFile if(compressed) file.length
fileInput.lengthelse file.length
compress.out.length
checkFile.out true
checkFile.out false
InternalStoreFile HDD file.length 2
25CPU file.length 5.34
Legend Control flow Dataflow Resource
consumption
9
Process for II
10
Genetic Algorithm

• Learns on monitoring data from running
  application
• Data argument (call / return) properties
  (primitive types, length of collection types, )
• Calling frequencies
• Pre-configured by static analysis results
• Gene selection (e.g. constants, if-then-else)
• Initial chromosomes
• Usually (non-stochastic) functional dependency
  exists

11
Machine Learning

• Not limited to one approach
• Genetic Algorithms (GA)
• Support Vector Machines (SVM)
• Hill-Climbing / simulated annealing
• Regression (Splines / linear)
• Stochastic approximation
• Greedy optimization

12
Conclusion

• Approach summed up

13
Conclusion

• Combination of static and dynamic source code
  analysis for reverse engineering
• Machine learning applied to reverse engineer a
  parameterised behavioural specification of
  component services
• Focus of target model
• Performance prediction
• Component-based software architectures

14
BACKUP
15
What is to be learned?

• Loops
• Break conditions
• Loop number
• Branches
• Branching conditions
• Probabilities
• Call of component-external services
• Frequency / probability
• Conditions
• Call parameters (x, y)
• Return value of provided service (long)
• Based on (input dimensions)
• Service input parameters (of described service)
  (a, b)
• Return parameters (of component-external calls)
  (float)