RFR: Compact heuristics adjustments

[Aleksey Shipilev]

On 11/15/2018 05:19 PM, Roman Kennke wrote:
> Out of curiousity: what is the practical consequence of this? Even
> compacter heap at the expense of even more CPU burn?

More prompt uncommit seems have no detectable overhead under load. Lower garbage threshold makes
apps have ~10% denser heap at expense of little more evacuation work.

The most effect would be changing allocation threshold, as it mostly defines how frequently GC runs
under load. Current value of 10% seems to work fine for the tests I tried. But this is where the
min-free-threshold rule comes in: it was possible to make allocation threshold "too high", and make
GC OOM.

-Aleksey