RFR: Specialize ShenandoahGenerationType to avoid generational check penalty on fast paths [v3]
Aleksey Shipilev
shade at openjdk.org
Tue Apr 18 18:12:16 UTC 2023
On Tue, 18 Apr 2023 17:36:12 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:
>> The target for this PR is to resolve the performance TODO at `ShenandoahMark::mark_through_ref`.
>> The same can probably be used in native barrier code too, if we decide to specialize those with `GENERATION`.
>>
>> With sample HyperAlloc test that is configured to stress the marking code:
>>
>>
>> $ java -XX:+UseShenandoahGC -XX:+UnlockDiagnosticVMOptions -XX:ShenandoahGCMode=passive -XX:-ShenandoahDegeneratedGC -Xlog:gc+stats -Xms1g -Xmx1g -XX:+AlwaysPreTouch -XX:ParallelGCThreads=1 -jar HyperAlloc-1.0.jar -d 30 -s 800 -t 1
>>
>>
>> We have consistent improvements in mark times for about 2.5%, look at average times below:
>>
>>
>> # Baseline
>>
>> [31.009s][info][gc,stats] Pause Full GC (G) = 25.546 s (a = 532201 us) (n = 48) (lvls, us = 203125, 539062, 542969, 544922, 559177)
>> [31.009s][info][gc,stats] Pause Full GC (N) = 25.543 s (a = 532136 us) (n = 48) (lvls, us = 203125, 539062, 542969, 544922, 559061)
>> [31.009s][info][gc,stats] Mark = 10.076 s (a = 209922 us) (n = 48) (lvls, us = 74609, 212891, 212891, 212891, 215692)
>>
>> [31.065s][info][gc,stats] Pause Full GC (G) = 25.618 s (a = 533715 us) (n = 48) (lvls, us = 230469, 541016, 542969, 544922, 556460)
>> [31.065s][info][gc,stats] Pause Full GC (N) = 25.615 s (a = 533650 us) (n = 48) (lvls, us = 230469, 539062, 542969, 544922, 556338)
>> [31.065s][info][gc,stats] Mark = 10.090 s (a = 210202 us) (n = 48) (lvls, us = 83398, 212891, 212891, 214844, 217067)
>>
>> [30.975s][info][gc,stats] Pause Full GC (G) = 25.613 s (a = 533606 us) (n = 48) (lvls, us = 201172, 539062, 544922, 546875, 558696)
>> [30.975s][info][gc,stats] Pause Full GC (N) = 25.611 s (a = 533566 us) (n = 48) (lvls, us = 201172, 539062, 544922, 546875, 558576)
>> [30.975s][info][gc,stats] Mark = 10.152 s (a = 211491 us) (n = 48) (lvls, us = 73633, 214844, 214844, 214844, 218065)
>>
>> # Patched
>>
>> [30.607s][info][gc,stats] Pause Full GC (G) = 25.224 s (a = 525492 us) (n = 48) (lvls, us = 216797, 531250, 537109, 539062, 541888)
>> [30.607s][info][gc,stats] Pause Full GC (N) = 25.220 s (a = 525411 us) (n = 48) (lvls, us = 216797, 531250, 537109, 539062, 541788)
>> [30.607s][info][gc,stats] Mark = 9.841 s (a = 205013 us) (n = 48) (lvls, us = 77734, 207031, 208984, 208984, 211176)
>>
>> [30.647s][info][gc,stats] Pause Full GC (G) = 25.293 s (a = 526937 us) (n = 48) (lvls, us = 287109, 531250, 535156, 539062, 545759)
>> [30.647s][info][gc,stats] Pause Full GC (N) = 25.290 s (a = 526868 us) (n = 48) (lvls, us = 287109, 531250, 535156, 539062, 545660)
>> [30.647s][info][gc,stats] Mark = 9.833 s (a = 204856 us) (n = 48) (lvls, us = 101562, 207031, 207031, 207031, 212495)
>>
>> [30.640s][info][gc,stats] Pause Full GC (G) = 25.193 s (a = 524860 us) (n = 48) (lvls, us = 210938, 531250, 535156, 537109, 540524)
>> [30.640s][info][gc,stats] Pause Full GC (N) = 25.190 s (a = 524792 us) (n = 48) (lvls, us = 210938, 531250, 535156, 537109, 540420)
>> [30.640s][info][gc,stats] Mark = 9.830 s (a = 204788 us) (n = 48) (lvls, us = 76367, 207031, 207031, 208984, 210008)
>
> Aleksey Shipilev has updated the pull request with a new target base due to a merge or a rebase. The incremental webrev excludes the unrelated changes brought in by the merge/rebase. The pull request contains ten additional commits since the last revision:
>
> - Merge branch 'master' into gentype-non-generational
> - Handle more omissions
> - Add a performance blurb
> - Merge branch 'master' into gentype-non-generational
> - More work
> - Micro-optimize in_generation
> - Touchups
> - Cleanups
> - Initial work
This now passes `hotspot_gc_shenandoah` on Linux x86_64 for me, so I am integrating.
-------------
PR Comment: https://git.openjdk.org/shenandoah/pull/254#issuecomment-1513595762
More information about the shenandoah-dev
mailing list