RFR: 8336707: Contention of ForkJoinPool grows when stealing works

[Doug Lea]

On Wed, 16 Oct 2024 16:38:13 GMT, Viktor Klang <vklang at openjdk.org> wrote:

>> This addresses tendencies in previous update to increase fencing, scanning, and signalling that can increase contention, and slow down performance especially on ARM platforms. It also uses more ARM-friendly constructions to reduce overhead (leading to several changes that all of the same form),
>
> src/java.base/share/classes/java/util/concurrent/ForkJoinPool.java line 1969:
> 
>> 1967:             int phase = w.phase, r = w.stackPred;     // seed from registerWorker
>> 1968:             int cfg = w.config, nsteals = 0, src = -1;
>> 1969:             for (;;) {
> 
> Could the non-labeled version introduce different GC safepoints which could have adverse impacts? 🤔

Yes, one of many uncertainties and tradeoffs...

> src/java.base/share/classes/java/util/concurrent/ForkJoinPool.java line 2279:
> 
>> 2277:                                 if (U.compareAndSetReference(a, k, t, null)) {
>> 2278:                                     q.base = b + 1;
>> 2279:                                     w.source = j;
> 
> Might be worth commenting the volatile-write piggybacking which allows to get away with not calling updateBase above? 🤔

Thanks; done

> src/java.base/share/classes/java/util/concurrent/ForkJoinPool.java line 2433:
> 
>> 2431:                             if (U.compareAndSetReference(a, k, t, null)) {
>> 2432:                                 q.base = nb;
>> 2433:                                 w.source = j;
> 
> Might be worth commenting the volatile-write piggybacking which allows to get away with not calling updateBase above? 🤔

done

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

PR Review Comment: https://git.openjdk.org/jdk/pull/21507#discussion_r1806452514
PR Review Comment: https://git.openjdk.org/jdk/pull/21507#discussion_r1806454588
PR Review Comment: https://git.openjdk.org/jdk/pull/21507#discussion_r1806454936