Hotspot gc

1,606 views
1,391 views

Published on

0 Comments
4 Likes
Statistics
Notes
  • Be the first to comment

No Downloads
Views
Total views
1,606
On SlideShare
0
From Embeds
0
Number of Embeds
4
Actions
Shares
0
Downloads
38
Comments
0
Likes
4
Embeds 0
No embeds

No notes for slide

Hotspot gc

  1. 1. Hotspot GC Tuning Michał Warecki
  2. 2. Generational space
  3. 3. G1 Heap (Dynamic) Eden Survivor Old Humongous Unused
  4. 4. Know your goal ● Different applications require different GC behavior – ● No one-size-fits-all solution regarding GC and performance User aspects: – – ● Throughput Pauses Engineering aspects: – Footprint – Scalability
  5. 5. GC Benchmarks ● Hardware – Production like ● ● Data – Production like in volumes and veracity ● ● Why? Phantom bottlenecks Why? Phantom bottlenecks i.e. caches Tests – No micro-benchmarks – Production like tests
  6. 6. Tuning for throughput ● Increase heap size – ● Reduce overall need for GC Let objects die in young generation – – GC in old generation is more expensive than in young generation prevent promotion of medium lifetime objects into old generation
  7. 7. Let objects die in young generation ● Increase young generation size – ● Only limited by need for old generation size Keep objects in survivor space – Increase survivors space – Raise occupancy threshold – Raise age threshold ● ● ● pro: prevents promotion of medium lifetime objects con: needlessly copies around long lifetime objects Use parallel young GC – increases throughput, if >>2 CPUs available
  8. 8. Tuning for minimal pause time ● Use parallel GC (parallel young and parallel compact) – ● Reduces pause time, if >>2 CPUs available Use concurrent GC (CMS) – – ● pro: mostly concurrent con: fragmentation + more expensive young GC Try out "G1" – Designed to limit pause time and frequency
  9. 9. Tuning CMS ● Avoid stop-the-world pauses – Reduce duration of full GC – Avoid full GC altogether
  10. 10. Prevent fallback to stop-the-world GC ● Increase heap size – ● Start CMS early, i.e. lower occupancy threshold – ● Defers the problems ("night time GC") Reduces throughput because GC runs practically all the time Increase young generation size – Avoids fragmentation in the first place
  11. 11. Hotspot GC “packs” ● -XX:+UseSerialGC ● -XX:+UseParallelGC ● -XX:+UseParallelOldGC ● -XX:+UseConcMarkSweepGC ● -XX:+UseG1GC
  12. 12. Serial GC (STW) ● -XX:+UseSerialGC – ● Single threaded young and old generation -XX:+UseParallelGC – Single threaded old generation
  13. 13. Parallel GC (STW) ● -XX:+UseParallelGC – ● -XX:UseParallelOldGC – ● Parallel young generation Parallel young and old generation -XX:+UseConcMarkSweepGC – Parallel young generation
  14. 14. (Mostly) Concurrent GC ● -XX:+UseConcMarkSweepGC – Mostly concurrent old generation – Young gen is still full STW
  15. 15. G1 GC ● -XX:UseG1GC – Serial, Parallel and Concurrent
  16. 16. Scavenger Phases ● Scavenge ● Reference ● SoftReference ● WeakReference ● FinalReference ● PhantomReference ● JNI Weak Reference ● String Table
  17. 17. CMS phases (1/2)
  18. 18. CMS phases (2/2) ● Initial mark ● Concurrent mark ● Precleaning ● Remark ● Sweep ● Reset data structures
  19. 19. GC Monitoring JVisualVM
  20. 20. GC Monitoring Java Mission Control
  21. 21. GC Monitoring Logs -verbose:gc -Xloggc:gclogs.txt -XX:+PrintGCDetails -XX:+PrintGCDateStamps -XX:+PrintTenuringDistribution -XX:+PrintGCApplicationConcurrentTime -XX:+PrintGCApplicationStoppedTime
  22. 22. GC logs ParNew
  23. 23. GC logs ParNew
  24. 24. GC logs CMS
  25. 25. GC logs CMS
  26. 26. GC Logs G1 (wow!) 2013-09-29T13:17:21.424+0200: 10.978: [GC pause (young) (to-space overflow), 0.01560900 secs] [Parallel Time: 9.9 ms] [GC Worker Start (ms): 10978.3 10978.4 10978.4 10978.4 10978.5 10978.6 10978.6 10978.6 Avg: 10978.5, Min: 10978.3, Max: 10978.6, Diff: 0.3] [Ext Root Scanning (ms): 2.3 2.0 2.2 2.2 2.0 1.9 1.6 2.4 Avg: 2.1, Min: 1.6, Max: 2.4, Diff: 0.8] [Update RS (ms): 0.0 0.2 0.0 0.0 0.0 0.0 0.0 0.0 Avg: 0.0, Min: 0.0, Max: 0.2, Diff: 0.2] [Processed Buffers : 0 26 0 0 0 5 0 0 Sum: 31, Avg: 3, Min: 0, Max: 26, Diff: 26] [Scan RS (ms): 0.2 0.3 0.0 0.2 0.1 0.3 0.0 0.0 Avg: 0.1, Min: 0.0, Max: 0.3, Diff: 0.2] [Object Copy (ms): 5.9 5.9 7.3 5.9 6.1 7.1 6.6 6.8 Avg: 6.5, Min: 5.9, Max: 7.3, Diff: 1.4] [Termination (ms): 1.2 1.2 0.0 1.2 1.2 0.0 1.2 0.0 Avg: 0.7, Min: 0.0, Max: 1.2, Diff: 1.2] [Termination Attempts : 1 2 1 2 2 1 1 1 Sum: 11, Avg: 1, Min: 1, Max: 2, Diff: 1] [GC Worker End (ms): 10988.0 10988.0 10987.9 10988.0 10988.0 10987.9 10988.0 10988.0 Avg: 10988.0, Min: 10987.9, Max: 10988.0, Diff: 0.1] [GC Worker (ms): 9.6 9.6 9.6 9.5 9.5 9.4 9.4 9.3 Avg: 9.5, Min: 9.3, Max: 9.6, Diff: 0.3] [GC Worker Other (ms): 0.3 0.3 0.4 0.4 0.5 0.5 0.5 0.6 Avg: 0.4, Min: 0.3, Max: 0.6, Diff: 0.3] [Clear CT: 0.7 ms] [Other: 5.0 ms] [Choose CSet: 0.1 ms] [Ref Proc: 0.9 ms] [Ref Enq: 0.0 ms] [Free CSet: 1.2 ms] [Eden: 400M(400M)->0B(377M) Survivors: 0B->23M Heap: 476M(500M)->157M(500M)] [Times: user=0.07 sys=0.02, real=0.01 secs]
  27. 27. GC Logs G1 ● Ext Root Scanning ● Update RS (Remembered Set) ● Scan RS ● Object Copy ● Termination ● GC Worker End ● GC Worker ● GC Worker Other ● Clear CT (Card table) ● Other (Choose CSet – Collection Set, Ref Proc, Ref Enq, Free CSet)
  28. 28. G1 Remembered Sets Ref: http://www.infoq.com/articles/tuning-tips-G1-GC
  29. 29. Useful options -Xmn -Xmx -Xms -XX:SurvivorRatio=X -XX:MaxTenuringThreshold=XX -XX:+UseConcMarkSweepGC -XX:+CMSScavengeBeforeRemark -XX:+CMSParallelRemarkEnabled -XX:CMSInitiatingOccupancyFraction=XX -XX:+UseCMSInitiatingOccupancyOnly -XX:ParallelGCThreads=X -XX:+UseGCTaskAffinity -XX:+BindGCTaskThreadsToCPUs -XX:ParGCCardsPerStrideChunk=XXXXX -XX:+UseCondCardMark
  30. 30. Thanks! Questions?

×