RFR: Load balance remembered set scanning [v3]

[Kelvin Nilsen]

On Mon, 1 Aug 2022 20:31:29 GMT, Y. Srinivas Ramakrishna <ysr at openjdk.org> wrote:

> Are any specific or general performance numbers available to share for this change? Thanks!

In a 20-minute Extremem workload comparing mainline against this branch, the average time for remembered-set scanning was reduced from 28,634 microseconds to 21,810 microseconds, an improvement of 17.55%.  This was out of n = 1893 scans for the balanced branch and n = 1925 for the mainline.  I don't have a full analysis of why n is different between the two workloads.  I have seen previously that unbalanced scanning of remembered set sometimes has much more significant impact on GC completion times, and can result in allocation failures and degeneration.  On this workload, I don't see evidence of this.  But shorter GC times perhaps means less floating garbage, and that might be one reason we needed fewer total GC cycles.

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

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