Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

PerOpteryx

545 views

Published on

Automatically Improve Software Architecture Models for Performance, Reliability, and Costs using Evolutionary Algorithms

Published in: Technology, Business
  • Be the first to comment

  • Be the first to like this

PerOpteryx

  1. 1. PerOpteryx<br />Automatically Improve Software Architecture Modelsfor Performance, Reliability, and Costs using Evolutionary Algorithms<br />Anne Martens Karlsruhe Institute of Technology (KIT), GermanyHeiko Koziolek, Steffen Becker, Ralf Reussner<br />WOSP / SIPEW 2010<br />
  2. 2. Software Performance Engineering<br />Anne Martens<br />C1 <br /> 3 sec<br />A<br />Transform<br />Solve<br />Change component and deployment<br />2.5 sec<br />C2 <br />A<br />Solve<br />Transform<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  3. 3. Not only Performance!<br />Anne Martens<br />C1 <br />3 sec<br />p(fail) 0.01%$ 5700 <br />A<br />Transform<br />Solve<br />Change component and deployment<br />2.5 sec<br />p(fail) 0.02%<br />$ 12000 <br />C2 <br />A<br />Solve<br />Transform<br />Optimise multiple criteria at once<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  4. 4. Multicriteria Optimisation<br />Anne Martens<br />ArchitecturalCandidate<br />A<br />Time<br />5s<br />$40K<br />Costs<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  5. 5. Multicriteria Optimisation<br />Anne Martens<br />Generated & Evaluated<br />A<br />Time<br />5s<br />$40K<br />Costs<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  6. 6. Multicriteria Optimisation<br />Anne Martens<br />A<br />Time<br />5s<br />B<br />3s<br />C<br />2s<br />Pareto-optimal<br />$40K<br />$33K<br />$20K<br />Costs<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  7. 7. Related Work: Quality Optimisation<br />Anne Martens<br />Rule-based approaches: Single quality only<br />Parsons2008, Cortellessa2009, PerformanceBooster (Xu&Woodside2008),ArchE (McGregor2007)<br />Multicriteria evaluation: No improvement<br />Bondarev2007, Grunske2007<br />Optimisation: Limited degrees of freedom<br />ArcheOpteryx (Aleti2009), Canfora2005,Kavimandan2009, Sassy (Menascé2010)<br />Missing: Flexible multicriteriaoptimisationatthe design level<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  8. 8. PerOpteryx Approach<br />Anne Martens<br />Flexible degreesoffreedom<br />Multiplequalities<br />Multi-criteriaoptimization<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  9. 9. Degrees of Freedom<br />Anne Martens<br />Design decision that can still be made<br />C<br />Degree of freedom<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  10. 10. Types of Degrees of Freedom in CBSE<br />Anne Martens<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  11. 11. Instances of Degrees of Freedom<br />Anne Martens<br />Each component<br />Allocation<br />Componentselection for D<br />Each server<br />Processor speed<br />Component selection<br />Search alternatives<br />Component selection for C<br />Allocation of D<br />C<br />D<br />Processor speed of server 1<br />Allocation of C<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  12. 12. Search Problem<br />Anne Martens<br />evaluate<br />Response in 2.5 s<br />P(failure) 0.02%<br />Cost $6000<br />transform<br />evaluate<br />initial model<br />candidate model<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  13. 13. SearchImplementation<br />Anne Martens<br />NSGA-II<br />[Deb2002]<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  14. 14. Quality evaluation<br />Anne Martens<br />PalladioComponentModel<br />[Becker2007]<br />PCM2LQN<br />[Koziolek2008]<br />PCM2Cost<br />[Martens2010]<br />PCM2DTMC<br />[Brosch2009]<br /><br />Cost<br />Reliability<br />Performance<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  15. 15. CaseStudywith PerOpteryx (1/2)<br />Anne Martens<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  16. 16. CaseStudywith PerOpteryx (1/2)<br />1235 candidates<br />58 Pareto optimal<br />8h running time<br />Anne Martens<br />Componentallocation<br />Processingrates<br />Componentselection<br />Response Time<br />POFOD<br />Costs<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  17. 17. Case Study with PerOpteryx (2/2)<br />Anne Martens<br />RT: 1.34 s<br />POFOD: 5.2E-4<br />Cost: 69.83<br />Onlyfour, but fasterserversDifferent Webserver<br />RT: 2.2 s<br />POFOD: 6E-4<br />Cost: 98<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  18. 18. Future Work<br />Anne Martens<br />?<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  19. 19. Conclusions<br />Anne Martens<br />AutomatedArchitectureImprovement<br />Flexible degreesoffreedom<br />Multiplequalities<br />Multi-criteriaOptimization<br />http://sdqweb.ipd.kit.edu/wiki/PerOpteryx<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  20. 20. Conclusions<br />Anne Martens<br />Multi-criteriaOptimization<br />PerOpteryx<br />Benefits<br />Performance, Reliability, Costs<br />Implemented& Validated<br />Automatedimprovement<br />Performance istrade-off<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  21. 21. PerOpteryx Implementation (1/3)<br />Anne Martens<br />Degrees of Freedom:<br />ProcessingRate<br />ComponentSelection<br />ComponentAllocation<br />?<br />?<br />e.g., 1.5 – 3 GHz<br />Motivation – Related Work – Approach – Case Study – Future Work –Conclusion<br />
  22. 22. SearchProcess<br />Anne Martens<br />3 sec<br />p(fail) 0.01%5700 $<br />Change alongdegrees<br />2.8 sec<br />p(fail) 0.02%<br />6000 $<br />Change alongdegrees<br />2.5 sec<br />p(fail) 0.02%<br />12000 $<br />
  23. 23. Not only Performance!<br />Anne Martens<br />C1 <br /> 3 sec<br />A<br />Transform<br />Solve<br />Change component and deployment<br /> 2.5 sec<br />C2 <br />A<br />Solve<br />Transform<br />

×