Parallel distributes card scanning work on fixed chunks of memory, sizing them on a fixed number of cards (128, see PSCardTable::scavenge_contents_parallel), calling out a work unit of 64k.

With configurable card sizes this work unit automatically adjusts.

Investigate if there is a sizable impact on performance by that, and whether it would be useful to base the work unit on (area) size, not number of cards.

G1 should not be affected by changes to card size in that respect because it uses a fraction of heap region size.