RFR: Load balance remembered set scanning

[Kelvin Nilsen]

On Wed, 27 Jul 2022 19:38:05 GMT, Kelvin Nilsen <kdnilsen at openjdk.org> wrote:

> This branch divides remembered set scanning into smaller units of work so that multiple cores can more effectively share the workload between them.  The benefit is to reduce concurrent scan remembered set times and to increase the parallelism of this phase.

The reason to "group chunks" is to reduce the need for synchronized coordination between worker threads.  Dividing work into entire ShenandoahHeapRegions seems to work fairly well for some workloads.  The idea is to start out each remembered set scan effort with work assignments that are a "large number" chunks.  Then, as we get closer to the end of the total scanning effort, we began dividing the worker into smaller and smaller assignments.  This way, a thread that gets an early "dense" work assignment might only complete that single heavy assignment while other worker threads can each complete a large number of lighter worker assignments.

-------------

PR: https://git.openjdk.org/shenandoah/pull/153