RFR: 8295476: Split G1 cost per byte predictor on gc phase

Albert Mingkun Yang ayang at openjdk.org
Mon Oct 31 15:17:29 UTC 2022


On Tue, 25 Oct 2022 20:00:32 GMT, Thomas Schatzl <tschatzl at openjdk.org> wrote:

> Hi,
> 
>   please review this change to the cost-per-byte predictor to be split on gc phase (young-only/mixed) instead of during marking/outside of marking.
> 
> The reason is that the distinction between young-only/mixed of these values is higher than marking/outside of marking, although the latter degrades somewhat to the former in case of continuous concurrent marking where both are almost equivalent anyway. I believe that has been the original reason to do the marking/not-marking distinction.
> 
> I also removed the implied 10% extra cost for copies during marking (i.e. during the young-only phase). I could not find any particular reason why copying would be more expensive then with the current code base.
> 
> The distinction in general makes sense as the data copied during young-only and mixed is different (only short-lived data vs. short+long-lived data).
> 
> Testing: gha, examining predictions with some benchmarks
> 
> Thanks,
>   Thomas

Marked as reviewed by ayang (Reviewer).

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

PR: https://git.openjdk.org/jdk/pull/10862


More information about the hotspot-gc-dev mailing list