G1 estimates the size (cardinality) of combined remembered sets to determine whether a given set of regions can be evacuated within a certain time.

The current way of having a single predictor predicting the percentage of the number of remembered set entries of a given region to be added to the total when adding that given region is totally inadequate.

While it works fairly well for young gen regions, the estimation for old gen regions is completely wrong.

E.g. having determined that a given young gen will yield ~1000 cards to actually scan, and an old gen region having 42k regions, G1 estimates that of those 42k (distinct entries) only 108 will be added to the 1000.
This does not make sense at  all, and ultimately causes G1 to underestimate the time taking this region into the collection set by a big margin.

(From printing the estimations from a specjvm2008 xml.validation run at 2GB, 15ms pause time goal)

Try to improve the estimation.