RFR: Improve adaptive heuristics

[Zhengyu Gu]

Looks good.

-Zhengyu

On 02/08/2017 10:09 AM, Roman Kennke wrote:
> While watching Shenandoah in the visualizer, I noticed that my double-
> marking change caused something else:
>
> - the cset size would go up and down now, fairly chaotically. it used
> to be fairly evenly sized
> - more worrying: it would almost always leave a lot of heap headroom
> unused
>
> Both is caused by the same thing: the collection-set threshold is fixed
> and is likely not optimal for the current situation of the program.
>
> Running SPECjvm shows that the actual collection set size varies
> greatly with the different benchmarks (between 0 and >50), and so
> should the cset-threshold (ideally, the threshold lies close to the avg
> cset size).
>
> So I changed that: I keep the last N cset-sizes in a TruncatedNumSeq
> and use the decaying-avg as the current cset-threshold. This way, I get
> much better heap usage & more even cset sizes. The 'downside' is that
> GC phases are not *exactly* evenly spaced anymore, but still very close
> to, and most importantly, they don't tend to clump together.
>
> Bottom-line: it's all inter-connected.
>
> Tested with SPECjvm and SPECjbb.
>
> Ok to push?
>
> http://cr.openjdk.java.net/~rkennke/adaptive2/webrev.00/
>
> Roman
>