Some testing by [~huntch] suggests that it would be useful to reclaim soft references more aggressively; particularly it seems that the value should somehow be related to memory pressure in the old gen.

This leads to better overall pause times, and the work being distributed more evenly.

Since concurrent mark typically indicates that memory pressure is high(er) as we start it in response to old gen getting full, the suggestion is to apply a different SoftRefLRUMbPerSec (or so) value during that time.

JRockit divided the value by (I think) by 2.5 for the concurrent mark soft reference policy.

However other heuristics could be interesting.