RFR: 8325553: Parallel: Use per-marker cache for marking stats during Full GC

[Stefan Johansson]

On Tue, 13 Feb 2024 09:39:54 GMT, Albert Mingkun Yang <ayang at openjdk.org> wrote:

> > have you tried the new SystemGC tests
> 
> Yes, it's perf-neutral. The reason is that the obj-size (int-array) used is too large, ~1k bytes. As a result, most per-region live-words updates are done by a single gc-worker: no contention. (If I reduce the obj-size to ~10 bytes, I can see large improvement.)
> 
Maybe adding such a scenario to the test once they are in would make sense. 

> > To avoid duplicating the logic
> 
> The duplication will hopefully be removed after #17822

Ok, yes I saw that there was only a single other user of `add_object(...)`. To me the logic is more connected with the `ParallelCompactData` class and is a better fit there, but you have looked more at this code recently and if you have a plan on how to get rid of the duplication after this other change I'm ok with dealing with it there.

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

PR Comment: https://git.openjdk.org/jdk/pull/17788#issuecomment-1941001381