RFR: Variable steps in adaptive heuristics

Roman Kennke rkennke at redhat.com
Tue May 16 07:49:56 UTC 2017


Am 15.05.2017 um 21:24 schrieb Aleksey Shipilev:
> Hi,
>
> Current adaptive heuristics makes only +1 or -1 steps for free threshold,
> depending on conditions.
>
> This is not good for several reasons:
>  a) Full GC means something going completely awry, and we need to step back hard
> to alleviate this. Stepping one percent is not going to cut it, because we will
> probably run into another Full GC;
>  b) Cancelled GC means something had changed so much, that we blew the free
> threshold completely -- like a workload phase change -- and we should step back
> harder to see what is going on.
>  c) When the initial guess for free threshold is wrong, it would take a while to
> arrive to most optimal one with small steps.
>
> The idea is to make variable steps, depending on distance from target and/or
> exceptional circumstances:
>   http://cr.openjdk.java.net/~shade/shenandoah/adaptive-varsteps/webrev.01/
>
> Testing: hotspot_gc_shenandoah, specjbb, specjvm
>
> Thanks,
> -Aleksey
>
>
Ok, yes!




More information about the shenandoah-dev mailing list