RFR: "Connected" should select bounded number of least connected regions

[Zhengyu Gu]

Okay.

-Zhengyu

On 08/08/2017 09:12 AM, Aleksey Shipilev wrote:
> Have been looking into why "connected" heuristics timeout our testing with GCOld test. It now takes
> around 500 seconds to run, while the timeout is set to 480s. Other heuristics, including
> LRU/Generational finish in under 120s.
>
> The reason for that seems to be a naive cset selection in "connected" that could possibly sweep up
> lots of regions, because each of those regions individually is lightly connected. So it seems better
> to make "connected" do the similar thing LRU/Generational is doing: cap the number of regions to
> consider. This requires sorting by connectedness to select the best candidates. It also makes sense
> to lock the allocation threshold to the region cap, as LRU/Gen does.
>
> See:
>  http://cr.openjdk.java.net/~shade/shenandoah/partial-connected-sync/webrev.01/
>
> With this change, the running time for GCOld under "connected" falls to ~120s, no testing timeouts
> observed. The pure allocation tests seem to tell that "connected" now performs on par with LRU/Gen,
> which is also a good sign.
>
> Testing: hotspot_gc_shenandoah
>
> Thanks,
> -Aleksey
>