RFR: Purge batched matrix cleanup

[Roman Kennke]

Am 13.10.2017 um 17:36 schrieb Aleksey Shipilev:
> http://cr.openjdk.java.net/~shade/shenandoah/matrix-no-batched/webrev.01/
>
> Before asynchronous region recycling, we had faced problems with matrix cleanups: it takes a while,
> and we could not accept it for the pause. There, we had to do deferred, batched, parallel Matrix
> cleanup [1] to alleviate STW costs. Now, when region recycling and matrix cleanups are handled in
> the concurrent phase, we care about this much less.
>
> Current code handles an interesting complication: we cannot add the batched-cleanup regions to the
> free set, for the fear we would clean up matrix for the regions that are used by freeset for
> allocation, thus breaking the matrix. On closer inspection, the same thing happens when allocation
> paths *assist* with recycling some of the trash regions into empty ones! Which is rare case, but it
> is nevertheless a bug.
>
> This can be mitigated by acquiring the heap lock for the batched matrix cleanup, but it would
> potentially block allocators for hundreds of milliseconds, which defeats the purpose.
>
> My suggestion is to ditch the batched matrix cleanup code, and leverage async recycling for doing
> the right thing. Allocators would normally assist with matrix cleanup, if async recycling is late.
> The experiments show this adds around 100us latency on allocation path with 32K regions (which is
> above our target anyhow), and it is negligible with the 4K target.
>
> Testing: hotspot_gc_shenandoah
>
> Thanks,
> -Aleksey
>
> [1] http://mail.openjdk.java.net/pipermail/shenandoah-dev/2017-May/002299.html
>
I think that makes sense. Patch looks good.


Roman