RFR: 8344665: Refactor PartialArrayState allocation for reuse
Kim Barrett
kbarrett at openjdk.org
Thu Nov 21 00:02:16 UTC 2024
On Wed, 20 Nov 2024 23:40:41 GMT, Kim Barrett <kbarrett at openjdk.org> wrote:
> This change splits the existing PartialArrayStateAllocator class into an
> allocator class and a manager class. The allocator class is per worker
> thread. The manager class provides the memory management context for a
> group of allocators.
>
> This change is in preparation for some other refactorings around partial array
> state handling. That work is intended to make it easier for various
> collections to adopt the use of that mechanism for chunking the processing of
> large objArrays.
>
> The new implementation for the memory management context is based on the
> existing one, with an Arena per worker, now controlled by the manager object.
> Potential improvements to that can be explored in the future. Some ideas
> include improvements to the Arena API or a single thread-safe Arena variant
> (trading slower arena allocation (which is the slow path) for less memory
> usage).
>
> G1 has a single manager, reused by each young/mixed GC. Associated state
> allocators are nested in the per-worker structures, so deleted at the end of
> the collection. The manager is reset at the end of the collection to allow the
> memory to be recycled. It is planned that the STW full collector will also use
> this manager when converted to use PartialArrayState. So it will be reused by
> all STW collections.
>
> ParallelGC has a single manager, reused by each young collection. Because the
> per-worker promotion managers are never destroyed, their nested state
> allocators are never destroyed. So the manager is not reset, instead leaving
> previously allocated states in the allocator free lists for use by the next
> collection. This means the full collector won't be able to use the same
> manager object as the young collectors.
>
> Testing: mach5 tier1-5
I'm not sure why skara initially labelled this as `hotspot`. Re-labelled as `hotspot-gc`. [later] Oh, it was because of
the jfr test.
-------------
PR Comment: https://git.openjdk.org/jdk/pull/22287#issuecomment-2489782134
More information about the hotspot-gc-dev
mailing list