Currently G1 groups all of the regions to be expected to be evacuated (minimal # of regions) into a single remset group to be evacuated in the first mixed gc.

The problem is that this selection is done at the remark pause, so there is incomplete information about how long it takes to actually evacuate this group.

This causes unexpectedly long pauses; another related issue is that in some cases this huge group will also be ordered (in the cleanup pause) after the first mixed gc.

Since this group is so large, it is likely that it will accumulate lots of additional remembered set entries during that time, causing pause times misses too.

There are two related issues here: the size of that group, and that it can be reordered. The latter is expected, and by decreasing the size that issue may not appear any more.