RFR (S): Parallelize code cache scan better

[Zhengyu Gu]

Good to me.

-Zhengyu

On 04/27/2017 01:00 PM, Aleksey Shipilev wrote:
> Hi,
>
> Turns out, our parallel code cache iterator is not that parallel: it processes
> entire code heaps in one stride. In our current mode, there are only three code
> heaps. So, in some workloads, e.g. Compiler.sunflow, code cache scan for the
> largest code heap for initial evacuation dominates the final mark pause time.
>
> The fix is to parallelize iterator better:
>   http://cr.openjdk.java.net/~shade/shenandoah/code-cache-parallel/webrev.01/
>
> Before:
>
>  Final Mark Pauses (N) = 0.35 s (a = 6797 us) (n =    51)
>      (lvls, us = 4219, 5508, 6680, 7168, 12895)
>
>     Initial Evacuation = 0.28 s (a = 5499 us) (n =    51)
>      (lvls, us = 3516, 4570, 5625, 5996, 9967)
>
> After:
>
>  Final Mark Pauses (N) = 0.26 s (a = 5261 us) (n =    49)
>      (lvls, us = 3574, 3926, 4375, 5547, 14308)
>
>     Initial Evacuation = 0.16 s (a = 3183 us) (n =    49)
>      (lvls, us = 2324, 2617, 2949, 3262, 5801)
>
> Testing: hotspot_gc_shenandoah, jcstress -m quick
>
> Thanks,
> -Aleksey
>