RFR: 8338737: Shenandoah: Reset marking bitmaps after the cycle [v3]

[Xiaolong Peng]

On Fri, 18 Oct 2024 08:22:56 GMT, Xiaolong Peng <xpeng at openjdk.org> wrote:

>> All the shenandoah passed, still waiting for our test farm to verify performance.
>> 
>> Reset marking bitmaps after collection cycle, for GenShen only do this for young generation, also choose not do this for Degen and full GC since both are running at safepoint, we should leave safepoint as ASAP.
>> 
>> I have run same workload for 30s with shenandoah in generational mode and classic mode, average average time of concurrent reset dropped significantly(283us to 109us for GenShen, 276us to 167us for classic shenandoah. 
>> 
>> Classic shenandoah baseline:
>> 
>> [30.973s][info][gc,stats    ] Concurrent Reset               =    0.148 s (a =      276 us) (n =   536) (lvls, us =      141,      205,      240,      291,      694)
>> [30.973s][info][gc,stats    ] Pause Init Mark (G)            =    0.136 s (a =      253 us) (n =   536) (lvls, us =      119,      242,      254,      264,      926)
>> [30.973s][info][gc,stats    ] Pause Init Mark (N)            =    0.047 s (a =       87 us) (n =   536) (lvls, us =       62,       84,       88,       91,      110)
>> ...
>> 
>> 
>> Classic shenandoah reset bitmaps after cycle:
>> 
>> [30.967s][info][gc,stats    ] Concurrent Reset               =    0.109 s (a =      167 us) (n =   652) (lvls, us =       48,      113,      139,      160,      888)
>> [30.967s][info][gc,stats    ] Concurrent Reset After Collect =    0.085 s (a =      132 us) (n =   648) (lvls, us =      107,      121,      125,      131,      532)
>> [30.967s][info][gc,stats    ] Pause Init Mark (G)            =    0.189 s (a =      289 us) (n =   652) (lvls, us =      117,      260,      277,      291,     2297)
>> [30.967s][info][gc,stats    ] Pause Init Mark (N)            =    0.058 s (a =       89 us) (n =   652) (lvls, us =       62,       85,       89,       94,      143)
>> ...
>> 
>> 
>> GenShen baseline
>> 
>> [31.008s][info][gc,stats    ] Concurrent Reset               =    0.107 s (a =      283 us) (n =   379) (lvls, us =      143,      225,      283,      330,      753)
>> [31.008s][info][gc,stats    ] Pause Init Mark (G)            =    0.098 s (a =      259 us) (n =   379) (lvls, us =      104,      227,      262,      305,     1033)
>> [31.008s][info][gc,stats    ] Pause Init Mark (N)            =    0.034 s (a =       90 us) (n =   379) (lvls, us =       67,       81,       89,       99,      130)
>> ...
>> 
>> 
>> GenShen reset bitmaps after cycle
>> 
>> 
>> [30.977s][info][gc,stats    ] Concurrent Reset               =    0.050 s (a =      109 us) (n =   462) (l...
>
> Xiaolong Peng has updated the pull request incrementally with one additional commit since the last revision:
> 
>   set need_bitmap_reset flag when change affiliation of heap region from OLD_GENERATION

My change also has stale mark bits issue I believe, gitfarm Genshen test fails sometimes due to crash caused by stale bitmap marks.

Amended one fix in this PR, it should solve the issue: https://github.com/openjdk/shenandoah/pull/516/commits/d4fddac5a1e7321ff3b4441bd5cd096e949cd56a#diff-7b87cde2e13d8769dc21cf6f7ae45c92942776fcb44275339182b7fa5c290ce0

I have run TestGCOldWithShenandoah.java#generational 100 times w/o crash caused by stale bitmap marks last night.

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

PR Comment: https://git.openjdk.org/shenandoah/pull/516#issuecomment-2422731057