RFR: 8337511: Implement JEP 404: Generational Shenandoah (Experimental) [v5]

[William Kemper]

On Wed, 30 Oct 2024 02:09:46 GMT, Liang Mao <lmao at openjdk.org> wrote:

>> We disabled tiered compilation to force everything to compile through C2 to get more consistent results.
>
> @earthling-amzn ，thanks for providing your options. Looks like disabling pacing would help the score. 
> 
> -XX:+UseShenandoahGC -XX:+UnlockExperimentalVMOptions -XX:+UnlockDiagnosticVMOptions -XX:-TieredCompilation -XX:-ShenandoahPacing -XX:+AlwaysPreTouch -XX:+DisableExplicitGC
> 
> I guess 4 cores may not be a target configuration for pauseless GC since generational pauseless GC requires at least 2 concurrent GC threads that already occupy half of CPU resource. I did a rough test on 8 cores and 16 cores according to your options. GenShen doesn't have obvious advantages with 8 cores and Shenandoah seems to outperform genshen significantly with 16 cores. Would you please help verify the result in your environment?
> 
> 8 cores -Xmx8g -Xms8g:
> 
> GenShen:
> RUN RESULT: hbIR (max attempted) = 9640, hbIR (settled) = 8050, max-jOPS = 8098, critical-jOPS = 6132
> RUN RESULT: hbIR (max attempted) = 8944, hbIR (settled) = 8574, max-jOPS = 7781, critical-jOPS = 6130
> 
> Shenandoah:
> RUN RESULT: hbIR (max attempted) = 9640, hbIR (settled) = 8050, max-jOPS = 7616, critical-jOPS = 6194
> RUN RESULT: hbIR (max attempted) = 9640, hbIR (settled) = 8050, max-jOPS = 7712, critical-jOPS = 6262
> 
> 16 cores -Xmx20g -Xms20g:
> 
> GenShen:
> RUN RESULT: hbIR (max attempted) = 19881, hbIR (settled) = 16584, max-jOPS = 17694, critical-jOPS = 13274
> RUN RESULT: hbIR (max attempted) = 19881, hbIR (settled) = 18724, max-jOPS = 17694, critical-jOPS = 13445
> Shenandoah:
> RUN RESULT: hbIR (max attempted) = 23838, hbIR (settled) = 20446, max-jOPS = 18594, critical-jOPS = 15441
> RUN RESULT: hbIR (max attempted) = 20138, hbIR (settled) = 19989, max-jOPS = 18728, critical-jOPS = 14967

@mmyxym - I disabled Shenandoah's pacer, and find that the generational mode out performs the non-generational mode:

Shen: RUN RESULT: hbIR (max attempted) = 16584, hbIR (settled) = 13837, max-jOPS = 11609, critical-jOPS = 11062
Shen: RUN RESULT: hbIR (max attempted) = 16584, hbIR (settled) = 13837, max-jOPS = 11609, critical-jOPS = 10425

Gen:  RUN RESULT: hbIR (max attempted) = 16584, hbIR (settled) = 14144, max-jOPS = 13267, critical-jOPS = 12087
Gen:  RUN RESULT: hbIR (max attempted) = 16584, hbIR (settled) = 13837, max-jOPS = 13267, critical-jOPS = 12151

The shared options for these runs were:

"-Xmx8g -Xms8g -XX:ActiveProcessorCount=8 -XX:+UseShenandoahGC -XX:+UnlockExperimentalVMOptions -XX:MetaspaceSize=1g -XX:-ShenandoahPacing

This is an aarch64 host with 16 cores and 64G of memory.

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

PR Comment: https://git.openjdk.org/jdk/pull/21273#issuecomment-2448245419