time matters
play

Time Matters Minimizing Garbage Collection Overhead with Minimal - PowerPoint PPT Presentation

Time Matters Minimizing Garbage Collection Overhead with Minimal Effort Gnther Blaschek Philipp Lengauer 2015-11-05 Configuration Hell OpenJDK 8 681 VM parameters (1300+ including tracing and debugging flags) ~ 10 205 configurations


  1. Time Matters Minimizing Garbage Collection Overhead with Minimal Effort Günther Blaschek Philipp Lengauer 2015-11-05

  2. Configuration Hell OpenJDK 8 681 VM parameters (1300+ including tracing and debugging flags) ~ 10 205 configurations assuming only boolean parameters 2 2015-11-05 Johannes Kepler University Linz

  3. Garbage Collection 37 GC parameters – 7.2 * 10 35 configurations 3 2015-11-05 Johannes Kepler University Linz

  4. Previous Approach Throughput 2 3 4 5 4 4 4 4 4 3 NewRatio - - - - + + + + + + AdaptiveBoundary 1024 1024 1024 1024 1024 512 2048 4096 3072 2048 PLABSize Hill Climbing (ParamILS) The Taming of the Shrew: Increasing Performance by Automatic Parameter Tuning for Java Garbage Collectors , Lengauer et. al., International Conference on Performance Engineering, 2014 (ICPE'14) 4 2015-11-05 Johannes Kepler University Linz

  5. Previous Results 100 – 400 configurations tested 5 2015-11-05 Johannes Kepler University Linz

  6. Goal 100% GC ratio 1) How good can we get? ?% 2) How can we achieve the same / better results faster? 3) How can we determine “important” parameters? 0% Effort 6 2015-11-05 Johannes Kepler University Linz

  7. Brute Force 14879 configurations tested 7 2015-11-05 Johannes Kepler University Linz

  8. Parameter Relevance Work with “Top 10” Parameters 8 2015-11-05 Johannes Kepler University Linz

  9. Signficant Parameters 1 … AdaptiveSizeDecrementScaleFactor 6 … AdaptiveSizePolicyWeight 2 … MaxTenuringThreshold 7 … MinHeapFreeRatio 2304 configurations 3 … MaxHeapFreeRatio 8 … YoungPLABSize 4 … UseAdaptiveGCBoundary 9 … UseAdaptiveSizePolicyWithSystemGC 5 … SurvivorPadding 10 … NewRatio 9 2015-11-05 Johannes Kepler University Linz

  10. Experiment Sequence 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 10 9 1 2 3 4 5 6 7 9 8 10 1 Select best value one after the other 1 2 3 4 1 5 6 7 10 9 8 1 → Linear number of experiments (22) “Pseudo Brute Force” 1 1 ... 1 1 10 9 8 7 1 6 5 4 3 2 1 Try all permutations of parameters 10 2015-11-05 Johannes Kepler University Linz

  11. Omitting Parameters Try all combinations of omissions (Pseudo brute force) 11 2015-11-05 Johannes Kepler University Linz

  12. Final Result 12 2015-11-05 Johannes Kepler University Linz

  13. Recommendation # Parameter default alt. 1 4 5, 6 AdaptiveSizeDecrementScaleFactor 2 15 1 MaxTenuringThreshold 3 70 50 MaxHeapFreeRatio 4 0 1 UseAdaptiveGCBoundary 5 3 1 SurvivorPadding 6 10 50 AdaptiveSizePolicyWeight 7 40 20, 25 MinHeapFreeRatio 8 4096 1024 YoungPLABSize 9 0 1 UseAdaptiveSizePolicyWithSystemGC 10 2 1 NewRatio 13 2015-11-05 Johannes Kepler University Linz

  14. Q&A ? 14 2015-11-05 Johannes Kepler University Linz

Recommend


More recommend