Phase-based Adaptive Recompilation in a JVM

Transcription

1 Phase-based Adaptive Recompilation in a JVM Dayong Gu Clark Verbrugge Sable Research Group, School of Computer Science McGill University, Montréal, Canada {dgu1, clump}@cs.mcgill.ca April 7, 2008 Sable Group, McGill University 1/33

2 Outline 1 Motivation 2 Hardware Information Analysis 3 Adaptive Recompilation 4 Conclusions and Future Work Sable Group, McGill University 2/33

3 Motivation Java Programs JVM (Virtual Hardware) Sable Group, McGill University 3/33

4 Motivation Java Programs JVM (Virtual Hardware) Real Hardware Sable Group, McGill University 4/33

5 Motivation Hardware Impact The impact of hardware on program behaviour can be significant Strong correlation exists: hardware performance program behaviour Hardware Event Counters Hardware counters widely exist in modern processors Accessible from software libs: PAPI, PMAPI,PCL,... Sable Group, McGill University 5/33

6 Motivation Hardware Impact The impact of hardware on program behaviour can be significant Strong correlation exists: hardware performance program behaviour Hardware Event Counters Hardware counters widely exist in modern processors Accessible from software libs: PAPI, PMAPI,PCL,... Use hardware event data to improve adaptive optimizations in JVM Sable Group, McGill University 5/33

7 Program Phases There are different types of program phases Sable Group, McGill University 6/33

8 Program Phases There are different types of program phases Flat Phases refer to the contiguous intervals where show a stable, flat performance on a type of sampled, profiling data Periodic Phases long term repetitions in execution, showing unstable performance but sharing similar patterns Sable Group, McGill University 6/33

9 Program Phases There are different types of program phases Flat Phases refer to the contiguous intervals where show a stable, flat performance on a type of sampled, profiling data Periodic Phases long term repetitions in execution, showing unstable performance but sharing similar patterns Detect this long term periodic phases from hardware data and employ the phase information in adaptive optimizations Sable Group, McGill University 6/33

10 Contributions Highlight the hardware impact on Java program execution Develop online pattern creation algorithm to represent hardware event Detect long term periodic phases from hardware patterns Implement an adaptive recomplication strategy using phase information Sable Group, McGill University 7/33

11 Hardware Event Data Example Hardware Event Time Sable Group, McGill University 8/33

12 Recurrent Phases Hardware Event Time Sable Group, McGill University 9/33

13 Hardware Pattern Hardware Event Time Sable Group, McGill University 10/33

14 Hardware Pattern Hardware Event Time Sable Group, McGill University 11/33

15 Hardware Pattern Hardware Event Time Sable Group, McGill University 12/33

16 Hardware Pattern Hardware Event Time Sable Group, McGill University 13/33

17 Hardware Pattern Hardware Event Time Sable Group, McGill University 14/33

18 Hardware Pattern Hardware Event Time Sable Group, McGill University 15/33

19 Pattern Building Algorithm (Simplified) Get HW data D Get the variation: D Vs former Data Yes Large enough? No Begin a new pattern building Continue the current pattern building; Add info. about D Conditions: 1. Small variations in a row 2. Pattern reaches the max length 3. Meet a more important variation (stop in the middle) Pattern Finish? Yes Report pattern No Sable Group, McGill University 16/33

20 Pattern Building Example (1) Hardware Data Sable Group, McGill University 17/33

21 Pattern Building Example (1) Hardware Data (2) Variation Sable Group, McGill University 17/33

22 Pattern Building Example (1) Hardware Data (3) Level Lv3 Lv2 Lv1 Lv (2) Variation Sable Group, McGill University 17/33

23 Pattern Building Example (1) Hardware Data (3) Level Lv3 Lv2 Lv1 Lv (2) Variation (4) Pattern Lv3 Lv2 Lv1 Lv Sable Group, McGill University 17/33

24 Pattern Analysis Patterns are stored and analyzed The number of occurrences determines the hotness of a pattern The hottest pattern is used to represent the current program phase The phase information is used to archieve a better adaptive hot method recompilation strategy in Jikes RVM Sable Group, McGill University 18/33

25 Recompilation in Jikes RVM Get method samples Compute the Past time in a method Opt. Level Opt0 Opt1 Opt2 Opt3 Estimate the compilation cost C i to optimization level i Samples Recompile to level i, if (SpeedupRate i Past) > C i Sable Group, McGill University 19/33

26 Recompilation in Jikes RVM Get method samples Compute the Past time in a method Opt. Level Opt0 Opt1 Opt2 Opt3 Estimate the compilation cost C i to optimization level i Samples Recompile to level i, if (SpeedupRate i Past) > C i Assume Future Past Sable Group, McGill University 19/33

27 Optimization Opportunity 4 3 Opt0 Opt1 Opt2 Opt3 Opt. Level Samples Sable Group, McGill University 20/33

28 Optimization Opportunity 4 3 Opt. Level Samples Sable Group, McGill University 21/33

29 Optimization Opportunity 4 3 Opt. Level Samples Sable Group, McGill University 22/33

30 Optimization Opportunity Recompilation is not free, cannot always be aggressive Sable Group, McGill University 23/33

31 Optimization Opportunity Recompilation is not free, cannot always be aggressive Our approach: Program Sampled Methods Hardware Event Recompilation State Behaviour Aggressiveness New The Beginners No recurrence Low of patterns Young Important methods Recurrences High of patterns Mature Optimized methods Less fresh patterns Moderate Less important methods More old patterns Rejuvenated New important More fresh High methods patterns again Sable Group, McGill University 23/33

32 Optimization Opportunity Recompilation is not free, cannot always be aggressive Our approach: Program Sampled Methods Hardware Event Recompilation State Behaviour Aggressiveness New The Beginners No recurrence Low of patterns Young Important methods Recurrences High of patterns Mature Optimized methods Less fresh patterns Moderate Less important methods More old patterns Rejuvenated New important More fresh High methods patterns again Future Past Fixed aggressiveness Adaptive aggressiveness Sable Group, McGill University 23/33

33 Optimization Opportunity Optimize code to higher levels earlier Possibly save recompilation overhead for intermediate levels Save unnecessary recompilation for the beginners Sable Group, McGill University 24/33

34 Original 4 3 Opt0 Opt1 Opt2 Opt3 Opt. Level Samples Sable Group, McGill University 25/33

35 Online Optimization 4 3 Opt. Level Samples Sable Group, McGill University 26/33

36 Offline Head Space Study 4 3 Opt. Level Samples Sable Group, McGill University 27/33

37 Experimental Set-up Benchmarks: SPECjvm98 suite Dacapo benchmarks: antlr, bloat, fop, pmd, xalan Soot and PseudoJbb Test on Athlon 1.4G, 1GB memory, Debian Linux kernel Average of the middle 11 in 15 runs Sable Group, McGill University 28/33

38 Recompilation Results Normalized execution time comp db jack javac jess mpeg mtrt rt antlr bloat fop pmd xalan soot pjbb Avg Orig Online Offline Whole Execution Time Reduction Average Up to (incl. all overhead) Offline Use training runs 8.7% 21% Online Use HW pattern info. 4.4% 18% Sable Group, McGill University 29/33

39 Whole Overhead Benchmark Overhead (%) Benchmark Overhead (%) compress 2.02 antlr 2.12 db 1.39 bloat 1.65 jack 1.71 fop 1.69 javac 1.13 pmd 1.70 jess 0.49 xalan 1.07 mpegaudio 1.76 soot 1.85 mtrt 0.82 PseudoJbb 0.77 raytrace 1.30 Average 1.43 The overhead includes all sources from hardware monitoring, pattern construction, information analysis, and building control events to adaptive engine Sable Group, McGill University 30/33

40 Conclusions Understanding repetitive program behaviour and exploring phases in program execution is important We implemented a technique for determining program phases from hardware data We applied the phase information in adaptive recomplication Hardware information can be used in a wide range of areas Runtime profiling, selecting GC points Program understanding, system reconfiguration, instruction/data relocation and prefetch... Sable Group, McGill University 31/33

41 Future Work Test other hardware events/combinations Use offline analysis results for repeatable executions Attach hardware variation with software structures Advanced static analysis can be helpful Develop other adaptive applications Sable Group, McGill University 32/33

42 Thank you! Questions? Sable Group, McGill University 33/33