JEP 333: ZGC: A Scalable Low-Latency Garbage Collector (Experimental)

AuthorsPer Liden, Stefan Karlsson
OwnerPer Liden
TypeFeature
ScopeImplementation
StatusClosed / Delivered
Release11
Componenthotspot / gc
Discussionhotspot dash gc dash dev at openjdk dot java dot net
EffortL
DurationL
BlocksJEP 351: ZGC: Uncommit Unused Memory (Experimental)
JEP 364: ZGC on macOS (Experimental)
JEP 365: ZGC on Windows (Experimental)
DependsJEP 312: Thread-Local Handshakes
JEP 304: Garbage Collector Interface
Relates toJEP 377: ZGC: A Scalable Low-Latency Garbage Collector (Production)
Reviewed byMikael Vidstedt, Stefan Karlsson
Endorsed byMikael Vidstedt
Created2018/02/13 09:58
Updated2020/03/13 08:31
Issue8197831

Summary

The Z Garbage Collector, also known as ZGC, is a scalable low-latency garbage collector.

Goals

We have strong ambitions to meet these goals for a large set of relevant workloads. At the same time, we want to acknowledge that we don't see these goals as hard requirements for every conceivable workload.

Non-Goals

It is not a goal to provide working implementations for platforms other than Linux/x64. Support for additional platforms can be added later, if there is enough demand.

Motivation

Garbage collection is one of Java's main strengths. However, when garbage collection pauses become too long they start to affect application response times negatively. By removing or drastically reducing the length of GC pauses, we'll make Java a more attractive platform for an even wider set of applications.

Furthermore, the amount of memory available in modern systems continues to grow. Users and application developers expect the JVM to be equipped to take full advantage of this memory in an efficient manner, and without long GC pause times.

Description

At a glance, ZGC is a concurrent, single-generation, region-based, NUMA-aware, compacting collector. Stop-the-world phases are limited to root scanning, so GC pause times do not increase with the size of the heap or the live set.

A core design principle/choice in ZGC is the use of load barriers in combination with colored object pointers (i.e., colored oops). This is what enables ZGC to do concurrent operations, such as object relocation, while Java application threads are running. From a Java thread's perspective, the act of loading a reference field in a Java object is subject to a load barrier. In addition to an object address, a colored object pointer contains information used by the load barrier to determine if some action needs to be taken before allowing a Java thread to use the pointer. For example, the object might have been relocated, in which case the load barrier will detect the situation and take appropriate action.

Compared to alternative techniques, we believe the colored-pointers scheme offers some very attractive properties. In particular:

Performance

Regular performance measurements have been done using SPECjbb® 2015 [1]. Performance is looking good, both from a throughput and latency point of view. Below are typical benchmark scores (in percent, normalized against ZGC's max-jOPS), comparing ZGC and G1, in composite mode using a 128G heap.

(Higher is better)

ZGC
       max-jOPS: 100%
  critical-jOPS: 76.1%

G1
       max-jOPS: 91.2%
  critical-jOPS: 54.7%

Below are typical GC pause times from the same benchmark. ZGC manages to stay well below the 10ms goal. Note that exact numbers can vary (both up and down, but not significantly) depending on the exact machine and setup used.

(Lower is better)

ZGC
                avg: 1.091ms (+/-0.215ms)
    95th percentile: 1.380ms
    99th percentile: 1.512ms
  99.9th percentile: 1.663ms
 99.99th percentile: 1.681ms
                max: 1.681ms

G1
                avg: 156.806ms (+/-71.126ms)
    95th percentile: 316.672ms
    99th percentile: 428.095ms
  99.9th percentile: 543.846ms
 99.99th percentile: 543.846ms
                max: 543.846ms

Ad-hoc performance measurements have also been done on various other SPEC® benchmarks and internal workloads. In general, ZGC manages to maintain single-digit millisecond pause times.

[1] SPECjbb® 2015 is a registered trademark of the Standard Performance Evaluation Corporation (spec.org). The actual results are not represented as compliant because the SUT may not meet SPEC's requirements for general availability.

Limitations

The initial experimental version of ZGC will not have support for class unloading. The ClassUnloading and ClassUnloadingWithConcurrentMark options will be disabled by default. Enabling them will have no effect.

Also, ZGC will initially not have support for JVMCI (i.e. Graal). An error message will be printed if the EnableJVMCI option is enabled.

These limitations will be addressed at a later stage in this project.

Building and Invoking

By convention, experimental features in the JVM are disabled by default by the build system. ZGC, being an experimental feature, will therefore not be present in a JDK build unless explicitly enabled at compile-time using the configure option --with-jvm-features=zgc.

(ZGC will be present in all Linux/x64 JDK builds produced by Oracle)

Experimental features in the JVM also need to be explicitly unlocked at run-time. To enable/use ZGC, the following JVM options will therefore be needed: -XX:+UnlockExperimentalVMOptions -XX:+UseZGC.

Please see the ZGC Project Wiki for more information on how to setup and tune ZGC.

Alternatives

Testing

Most of our existing functional and stress tests are collector agnostic and can be reused as-is. Additional tests targeting properties and functions specific to ZGC will be added.

Dependencies