CFV: New Project: ZGC

Per Liden per.liden at
Wed Oct 25 19:45:23 UTC 2017

I hereby propose the creation of the ZGC Project with myself (Per Liden) 
as the Lead and the HotSpot Group as the sponsoring Group.

In accordance with the OpenJDK guidelines [1], this project will provide 
a home for the continued development of the Z Garbage Collector, also 
known as ZGC. ZGC is a new garbage collector optimized for low latency 
and very large heaps. We've developed ZGC internally at Oracle so far, 
and we're now open-sourcing it so as to broaden the base of both 
contributors and users.

ZGC has been designed with the following goals in mind:
* Handle multi-terabyte heaps
* GC pause times not exceeding 10ms
* No more than 15% application throughput reduction compared to using G1

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. We are 
however currently able to meet or exceed these goals on some well-known 
industry standard benchmarks.

At a glance, ZGC is a concurrent, currently single-generation, 
region-based, incrementally compacting collector. Stop-The-World phases 
are limited to root scanning, meaning GC pause times do not increase 
with the heap- or live-set size.

While there is still work to do, the design and implementation is 
reasonably mature and stable. ZGC today executes the following GC 
tasks/phases concurrently:
* Marking
* Reference processing (java.lang.ref.*)
* Relocation set selection
* Relocation/Compaction

And we're actively working on making the remaining GC tasks/phases 
concurrent. These are:
* Weak root processing (StringTable, JNIWeakGlobalRefs)
* Class unloading

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. To name a few:

* It allows us to reclaim and reuse memory during the 
relocation/compaction phase, before pointers pointing into the 
reclaimed/reused regions have been fixed. This helps keep the general 
heap overhead down. It also means that there is no need to implement a 
separate mark-compact algorithm to handle "Full GC".

* It allows us to have relatively few and simple GC barriers. This helps 
keep the runtime overhead down. It also means that it's easier to 
implement, optimize and maintain the GC barrier code in our interpreter 
and JIT compilers.

* We currently store marking and relocation related information in the 
colored pointers. However, the versatile nature of this scheme allows us 
to store any type of information (as long as we can fit it into the 
pointer) and let the load barrier take any action it wants to based on 
that information. We believe this will lay the foundation for many 
future features. To pick one example, in a heterogeneous memory 
environment, this could be used to track heap access patterns to guide 
GC relocation decisions to move rarely used objects to "cold storage".

Much of the remaining work involves addressing latency issues in non-GC 
subsystems in HotSpot, such as being able to concurrently unlink stale 
entries in the StringTable. We hope and expect to see a fair bit of 
collaboration with people working on other garbage collectors in areas 
where we have a common interest.

Some of the work coming out of the ZGC project has already been seen, 
either in the form of general improvements, or because a feature has 
found use cases outside of ZGC, such as:
* Atomics re-write
* GC Barrier API
* Thread local handshakes

I (Per Liden) am a member of the HotSpot GC team at Oracle, and have 
been working on JRockit and HotSpot projects for the past 8 years. I'm 
the initial author of ZGC, but many people have made significant 
contributions since then.

Special thanks to Stefan Karlsson, who has been working with me on ZGC 
since the very early phases of this project.

The initial Reviewers and Committers will be (based on people who have 
contributed to ZGC development within Oracle so far):

* Stefan Karlsson (Reviewer)
* Erik Österlund (Committer)
* Mikael Gerdin (Committer)
* Kim Barret (Committer)
* Nils Eliasson (Committer)
* Rickard Bäckman (Committer)
* Roland Westrelin (Committer)
* Coleen Philimore (Committer)
* Robin Ehn (Committer)
* Gerard Ziemski (Committer)

The initial source of this project will be based on a clone of a JDK 10 
repository, plus the latest ZGC patch set. Changes from the JDK 10 
parent will be synced into ZGC periodically. Change review policy will 
be determined by the Lead and a consensus of Reviewers. Review is 
expected to be relaxed initially, but made more strict as we get closer 
to integration.

The project will host at least the following mailing list:

* zgc-dev for developers

Votes are due by 23:59 CET on Wednesday, November 8, 2017.

Only current OpenJDK Members [1] are eligible to vote on this motion. 
Votes must be cast in the open on the discuss list. Replying to this 
message is sufficient if your mail program honors the Reply-To header.

For Lazy Consensus voting instructions, see [2].

Per Liden


More information about the announce mailing list