New Algorithms in Java
0 likes89 views
The document discusses various Java garbage collection (GC) algorithms, including G1, Shenandoah, and ZGC, highlighting their functionalities, benefits, and tuning methods. It emphasizes the need for low-pause collectors suitable for diverse applications and addresses the trade-offs involved in achieving low pause times. The document also mentions future developments in GC technology within the Java ecosystem.
New Algorithms in Java
- 1. Krystian Zybała @k_zybala (Modern) New? GC algorithms in Java
- 2. # whoami • Java Principal Engineer / Java Performance Engineer • Specialization : JVM, Apache Kafka, Cassandra • Hobby : JVM Performance, Reactive Systems • Workshops, Consultations Twitter: @k_zybala Blog: https://kzybala.pl Mail: kontakt@kzybala.pl
- 4. Let's hold our hands and wait for the next GC cycle
- 5. What is GC?
- 6. GC is your friend 🤗
- 8. Previous GC’s algorithms • SerialGC • ParallelGC • CMS - not available in current JDKs builds • G1GC
- 11. The most of young objects are much more likely to die than old objects
- 12. Eden Survivor 1 Survivor 2 Tenured New object Old generation Young generation
- 13. G1
- 14. G1 • Low-pause collector - First papers date back to 2004 • Supported since JDK 7u4 (April 2012) • Replacement for CMS • Low pauses valued more than max throughput • Designed to be really ease to tune • Tuning based on max Stop-The-World pause ( -XX:MaxGCPauseMillis=X) Default: 250ms
- 15. G1 • Default GC since Java 9 • Use regions instead generations • Dynamic size of region • Parallel Full GC in Java10(jdk-10-ea+34)
- 16. G1 Concurrent Marking Concurrent liveness summary Init Marking Remark Clean up GC Thread Application Thread Copy or Space reclamation
- 18. Ergonomic
- 19. bool UseG1GC = false {product} {default} bool UseSerialGC = true {product} {ergonomic} http://hg.openjdk.java.net/jdk/jdk10/ fi le/b09e56145e11/src/hotspot/share/runtime/os.cpp#l1606
- 21. Why we need something new?
- 23. GC Optimized for Best for Pause Multi- threaded Number of cores SerialGC Memory Footprint Single core, small heaps ( micro-service ? ) Yes No 1 ParallelGC Throughput Batch job Yes Yes 2+ G1GC Throughput / Latency Balance request-response db integractions Yes Yes 2+* ZGC Low latency request-response db integractions Yes Yes 2+* ShenandoahGC Low latency request-response db integractions Yes Yes 2+*
- 24. How does it work?
- 25. How does it work? • Requires meta data storing in object • Requires JIT support - Load GC barrier or Store GC barrier
- 26. Mutator problems Headers x = 1 y = 2 z = 1 Object A Object B
- 27. Mutator problems Headers x = 1 y = 2 z = 1 Headers x = 1 y = 2 z = 1 from-space to-space Object A Object B Object A Object B Object C
- 28. Mutator problems Headers x = 1 y = 4 z = 1 Headers x = 3 y = 2 z = 1 from-space to-space Application thread 10 Application thread 3 Object A Object B Object A Object B
- 29. Shenandoah
- 30. Shenandoah • Available in AdoptOpenJDK builds • NOT available in Oracle’s OpenJDK builds • Present in RedHat OpenJDK 8,11+ • Low pause GC • Concurrent Compaction, Single Generation
- 31. Shenandoah • Available for x86 32-bit and 64-bit • ARM ports available • Linux, MacOS, Windows • Region Based
- 32. Shenandoah 1.0 * Maintains a weak to-space invariant * Load barrier * Store barrier
- 34. Shenandoah Mark Class Field1 Field2 Mark Class Field1 Field2 Mark Class Field1 Field2 Mark Class Field1 Field2 Fwd Ptr Fwd Ptr 1.0 Object A Object B Object C Object A Object B Object C
- 35. Shenandoah Problems * More memory need ( due to the forward pointer) worst case 50% common case 5-10% * Complicated GC barriers ( Store and Load)
- 36. Shenandoah 2.0 * Maintains a strong to-space invariant * Load barrier (LRB - Load Reference Barrier)
- 38. Shenandoah 2.0 Class Field1 Field2 Mark Class Field1 Field2 Fwd Ptr Mark Class Field1 Field2 From Space To Space Object A Object B Object C
- 39. Shenandoah Concurrent Mark Concurrent Evacuation Init Mark Final Mark Init U-R Concurrent Update References GC Thread Application Thread Final U-R
- 41. Shenandoah Marking + Evacuating + Update References
- 43. Shenandoah Mark Class Field1 Field2 Evacuation from-space to-space Mark Class Field1 Field2 Application Thread copy Object A Object B Object B’
- 44. Shenandoah Mark Class Field1 Field2 Evacuation Mark Class Field1 Field2 from-space to-space Mark Class Field1 Field2 GC Thread copy Application Thread copy Object A Object B Object B’’ Object B’
- 45. Shenandoah Evacuation Mark Class Field1 Field2 from-space to-space Mark Class Field1 Field2 Class Field1 Field2 Fwd Ptr GC Thread copy Application Thread copy Object A Object B Object B’’ Object B’
- 46. Shenandoah Evacuation from-space to-space Mark Class Field1 Field2 Class Field1 Field2 Fwd Ptr Object A Object B Object B’
- 47. Shenandoah Update references from-space to-space Mark Class Field1 Field2 Class Field1 Field2 Fwd Ptr Object A Object B
- 48. Tuning
- 49. -Xms -Xmx
- 50. Shenandoah Tuning • -XX:+AlwaysPreTouch • -XX:+UseLargePages • -XX:+UseNUMA (not supported yet) • -XX:-UseBiasedLocking • -XX:+DisableExplicitGC
- 51. ZGC
- 52. ZGC • Sub-millisecond max pause times (less than 10ms) (less than 1ms) • Pauses are O(1) • Pause times do not increase with the heap, live-set or root-set size • Handle heaps ranging from a 8MB to 16TB in size • ZGC was initially introduced as an experimental feature in JDK 11, and was declared Production Ready in JDK 15.
- 53. ZGC • Only for Linux x86 64-bit • compressed class pointers • no support compress OOPs • Region based • ZPages • Small ( 2MiB - object size up to 256KiB) • Medium ( 32MiB - object size up to 4MiB) • Large ( 4+ MiB - object size > 4MiB)
- 54. ZGC • Concurrent • Region-based • Compacting • NUMA-aware • Using colored pointers • Using load barriers
- 55. ZGC Uses colored pointers. Color indicates GC metadata Object address - 44 bits 16 bit 4 color bits 64 bits pointer address
- 56. ZGC Multi-Mapping ∂ ∂ 001<addr> Virtual memory Physical memory 010<addr> 100<addr>
- 57. ZGC ZGC uses a load GC barrier Object someObject = person.name; <load_barrier> // String n = person.name; mov 0x10(%rax), %rbx // bad color test %rbx, (0x16)%r15 // If yes, enter slow path jnz slow_path
- 58. ZGC Concurrent Mark Reference processing Unload classes Free memory pages Prepare relocation Mark start Mark End Relocate Concurrent Relocate Concurrent Remap GC Thread Application Thread
- 59. ZGC Concurrent Mark Concurrent Reference processing Unload classes Free memory pages Prepare relocation Concurrent Relocate Remap Concurrent Mark …
- 60. ZGC
- 61. ZGC Relocation Forwarding table A -> A1 B -> B1 C -> C1
- 62. ZGC Relocation Forwarding table A -> A1 B -> B1 C -> C1 D -> D1 Application thread copy
- 63. ZGC Relocation Forwarding table A -> A1 B -> B1 C -> C1 D -> D1 Application thread copy GC thread copy
- 64. ZGC Relocation Forwarding table A -> A1 B -> B1 C -> C1 D -> D1 E -> E1 Application thread copy GC thread copy
- 65. ZGC Remapping
- 66. ZGC Future changed * Generational * Sub-millisecond max pause time * Graal JIT support
- 67. Tuning
- 68. -XMX<Size>
- 69. ZGC Tuning • -XX:ZAllocationSpikeTolerance=factor • -XX:ZCollectionInterval=seconds • -XX:ZFragmentationLimit=percent • -XX:ZUncommitDelay
- 71. Java 19
- 72. Project Loom
- 74. JEP 376 ZGC Concurrent Thread-Stack Processing
- 75. Concurrent Thread-Stack Processing ▪Remove thread-stack processing from ZGC safepoints. ▪Make stack processing lazy, cooperative, concurrent, and incremental. ▪Remove all other per-thread root processing from ZGC safepoints. ▪Provide a mechanism by which other HotSpot subsystems can lazily process stacks.
- 76. Conclusions
- 77. Conclusions * Still JDK8 ? - Shenandoah * JDK 11+ ? - ZGC or Shenandoah * Nothing is for free - you need more resources for achieve low pause
- 78. Is G1 dead?
- 79. JEP-423 Region Pinning for G1
- 80. JEP-423 • No stalling of threads due to JNI critical regions. • No additional latency to start a garbage collection due to JNI critical regions. • No regressions in GC pause times when no JNI critical regions are active. • Minimal regressions in GC pause times when JNI critical regions are active.
- 81. Over 30 major performance improvements
- 83. QA
- 84. Thank you!