Skip to content

Latest commit

 

History

History
814 lines (574 loc) · 22.2 KB

slides.md

File metadata and controls

814 lines (574 loc) · 22.2 KB
Raw
title
Modern JVM Garbage Collection

Modern JVM
Garbage Collection


Java User Group Switzerland
2020-12-01

Puzzle ITC

Puzzle ITC

What to Expect

  • Some GC theory
  • New GCs in OpenJDK HotSpot
  • How to choose and tune GC in OpenJDK HotSpot
  • GC in containers

Garbage Collector Theory

What Does a GC Do?

  • Tracks every object in JVM Heap
  • Removes unused objects

    ⇒ Easy?

GC Design Factors

  • Program throughput
  • GC throughput
  • Heap overhead
  • Pause times
  • Pause frequency
  • Pause distribution
  • Allocation performance
  • Compaction
 

GC Design Factors

  • Concurrency
  • Scaling
  • Tuning
  • Warmup time
  • Page release
  • Portability
  • Compatibility

    ⇒ Very hard, lots of tradeoffs

References: 1

Generational Hypothesis

  • Empirical observation
  • (Weak): Most objects die young
  • Strong: The older the object, the less chance it has to die
  • Notable exception: LRU caches

References: 1

Throughput and Latency

Throughput
% of time not spent in garbage collection
Latency
Responsiveness of an application, affected by GC pauses

References: 1

STW, Parallel and Concurrent

STW (Stop-The-World)
All app threads are stopped during GC
Concurrent
All app threads are running during GC
Parallel
GC uses multiple threads in STW and/or concurrent phases

OpenJDK HotSpot Garbage Collectors

JVM Ergonomics

  • JVM ergonomics selects GC if none specified 1
  • Serial GC if 1 CPU or < 1.75 GiB RAM 2
  • OpenJDK < 8u191 not fully aware of containers 3
  • GCs ergonomics auto-tune low-level params

OpenJDK 8 Available GCs

GC Option Comment
Serial -XX:+UseSerialGC default
Parallel -XX:+UseParallelGC default
CMS -XX:+UseConcMarkSweepGC
G1 -XX:+UseG1GC
Shenandoah -XX:+UseShenandoahGC non-mainline backport

References: 1 2

OpenJDK 11 Available GCs

GC Option Comment
Serial -XX:+UseSerialGC default
Parallel -XX:+UseParallelGC
CMS -XX:+UseConcMarkSweepGC deprecated
G1 -XX:+UseG1GC default
Shenandoah -XX:+UseShenandoahGC non-mainline backport
ZGC -XX:+UseZGC exp., Linux x86_64 only
Epsilon -XX:+UseEpsilonGC experimental

References: 1 2 3 4

Epsilon GC

  • No-op gc, does not collect garbage at all
  • Useful for:
    • Measurements
    • Short living processes
    • Garbage-free applications

References: 1 2

GC Concurrency

Young Generation Old Generation
Serial Copy Mark Compact
Parallel Copy Mark Compact
CMS Copy Conc Mark Conc Sweep
G1 Copy Conc Mark Compact
Shenandoah — Conc Mark Conc Compact
ZGC — Conc Mark Conc Compact

  Stop-The-World   Concurrent with application

References: 1

Concurrent GC Phases

Concurrent GC Phases

References: 1

Concurrent GCs

  • All expensive GC phases run concurrent
  • Coordination between GC & app through barriers
  • STW pauses for root set scan and cleanup
  • Pauses are short and predictable
  • Better scalability for large heaps

Barriers

  • Machine code injected by JIT compiler
  • Additional metadata needed for coordination
  • Throughput reduction
    • Predictable
    • Can be offset with more resources

Shenandoah Garbage Collector

Shenandoah Garbage Collector

  • Developed by Red Hat
  • Named after Shenandoah national park
  • Originally based on G1 GC
  • Concurrent marking with SATB like G1

Shenandoah Garbage Collector

  • Pause times independent of heap and live-set size
  • Single generation, multiple regions
  • Multiple heuristics and failure modes

Shenandoah Garbage Collector

  • Concurrent compaction:
    • v1: Brooks pointers, read and write barriers
    • v2: on-heap forwarding pointers, load barriers
  • Barrier loop optimizations

References: 1 2 3 4 5

On-Heap Forwarding Pointers

Shenandoah On-Heap Forwarding Pointers

References: 1 2

Load (Reference) Barriers

Extra code when object reference is loaded from heap:

Object obj2 = obj.field1;  // Loading an object reference from heap
// Load barrier needed here

Object obj3 = obj2;        // No barrier, not a load from heap
obj.doSomething();         // No barrier, not a load from heap
int i = obj.field2;        // No barrier, not an object reference

Optimized for common case

References: 1

Shenandoah Load Barrier

Pseudocode:

load_reference_barrier(addr):
  if in_evac_phase() and in_collection_set(addr) and !is_forwarded(addr):
    new_addr = copy_object(addr)
    if cas_fwd_pointer(addr, new_addr):
      return new_addr
    else:
      return get_fwd_pointer(addr)  # Another thread copied object

References: 1 2

Shenandoah JDK Support

  • Available in OpenJDK, except Oracle builds
  • Fixes and updates backported to LTS JDKs
JDK Support Status
8 LTS Ready for production
11 LTS Ready for production
12-14 STS Experimental, discontinued
15 STS Ready for production

References: 1 2

Z Garbage Collector

Z Garbage Collector

  • Developed by Oracle, initially proprietary
  • Inspired by patented Azul C4 GC
  • Pause times independent of heap and live-set size
  • Single generation, multiple regions

Z Garbage Collector

  • Concurrent marking and compaction:
    • Load barriers and colored pointers
      • 64 bit only
      • no compressed oops
    • off-heap forward pointers

Z Garbage Collector

  • Failure modes not documented
  • Windows and MacOS support in JDK 14+

References: 1 2 3 4

Colored Pointers

ZGC Colored Pointers

Object address size was changed from 42 to 44 bits in JDK 13 1

Heap Multi-Mapping on x86_64

ZGC Heap Multi-Mapping

References: 1 2

ZGC Load Barrier

Pseudo code:

load_reference_barrier(addr):
  if color(addr) is bad:
    return slow_path(addr)  # mark/relocate/remap, depends on gc phase

References: 1 2

ZGC JDK Support

  • Available in OpenJDK, including Oracle builds
  • No backports ⇒ upgrade for fixes & new features
JDK Support Status
11 LTS Experimental, discontinued
12-14 STS Experimental, discontinued
15 STS Production ready

References: 1

Choosing and Tuning GCs

When to use which GC?

Criteria GC
Heap ≤ 100 MiB Serial
Single CPU, long pauses ok Serial
Maximum throughput, long pauses ok Parallel
Minimum latency, reduced throughput ok Shenandoah or ZGC
Minimum latency, JDK 8 or 11 LTS Shenandoah
Large heap, long pauses not ok Shenandoah or ZGC
Slow hardware, long pauses not ok Shenandoah (or ZGC?)
Balanced / otherwise G1

References: 1 2

Garbage Collector Design Goals

GC Throughput Pause time
Serial 99% —
Parallel 99% —
G1 90% 200 ms
Shenandoah 85% 10 ms
ZGC 85% 10 ms
Epsilon 100% 0 ms

List actual values: -XX:+UnlockDiagnosticVMOptions -XX:+PrintFlagsFinal References: 1 2 3

Generic GC Tuning

  • Set max heap size based on live-set and alloc rate
  • For Parallel and G1 and with specific requirement:
    • Set max pause-time or throughput goal
  • Leave other options on default
  • If goal isn't met tune according to next slides
  • If goal still isn't met tune according to references

References: 1 2

Serial GC Tuning

  • Throughput goal not met:
    • Increase max heap size
  • Pauses too long:
    • Choose different GC

References: 1

Parallel GC Tuning

  • Throughput goal not met:
    • Increase max heap size
    • Remove or raise max pause-time goal
  • Pauses too long:
    • Set or lower max pause-time goal

References: 1

G1 GC Tuning

  • Throughput too low:
    • Increase max heap size
    • Raise max pause-time goal
  • Pauses too long:
    • Lower max pause-time goal
    • Disable string deduplication

References: 1 2

ZGC Tuning

  • Throughput too low:
    • Decrease number of GC threads
  • Latency too high / allocation failures:
    • Increase max heap size
    • Increase GC threads

References: 1

Shenandoah Tuning

  • Latency too high / allocation failures:
    • Increase max heap size
    • Change/tune heuristics to run GC sooner
    • Increase allocator thread pacing delay

References: 1

Monitoring GCs

  • GCs provide detailed metrics in logs 1
  • GCs provide partial metrics via MBeans 2 3
  • Each GC reports metrics differently
  • Compare with GC logs to find what is reported

Monitoring GCs

  • Use tool like GCViewer to interpret logs 1
  • Export metrics to Prometheus with Micrometer 2
    • Supports concurrent GCs in 1.6+ 3
    • Reports all GC phases as pauses
  • Compare with non-GC metrics 4

Java GCs in Containers

JVM Ergonomics in Containers

  • Same rules apply in containers since 8u191 1
    • Serial GC if 1 CPU or < 1792 MiB RAM
  • Red Hat/fabric8 images use run-java.sh 2
    • Defaults to Parallel GC on JVM < 10
    • Sets max heap size by default

Heap Size

  • Optimize performance 1:
    • Set min heap size = max heap size
    • Enable -XX:+AlwaysPreTouch
  • Optimize cost:
    • Select GC with page release
      • G1 (JDK 12+), Shenandoah or ZGC 2 3
    • Don't set min heap size

GraalVM Native Image

  • Community Edition: Serial GC
  • Enterprise Edition: Serial GC, G1 GC (Linux only)
  • Implemented in Java
  • Native image not as efficient as JIT code

References: 1 2 3

Finalization

Finalization is deprecated. Instead use:

  • Explicit clean-up method
  • AutoClosable
  • PhantomReference and ReferenceQueue

References: 1

What's Next

  • Check references in slides
  • Look at some benchmarks 1 2 3
  • Verify and tune GC in your apps
  • Collect GC metrics in your apps
  • Read more about GC Theory 4