On large heaps (100G+), even with maximum region size, there are many regions.

Since we currently have this 1:1 mapping of region and remembered set, this means that the total remembered set gets very large just by the fact of having many regions and many inter-region pointers.

Remembered set management has a large overhead then, both in size and time (RS scan, scrub, ....). One idea to reduce this overhead is to slightly decouple remembered sets from regions, i.e. have one remembered set spanning multiple regions.

This can decrease the number of inter-remembered set pointers significantly.

The only disadvantage is that this decreases the flexibility during evacuation slightly: you can only evacuate all regions with a particular remembered set at once. However, at these large heap sizes, the corresponding machine the VM runs on is typically sufficiently powerful to handle multiple regions in a reasonable time.

I.e. the evacuation granularity, if used judiciously, would be sufficiently small for keeping time goals.