for coherence (16g 50% read 50% write) and ycsb hbase workload (100g heap), to avoid to-space exhausted for mixed gc, we have to push G1NewSizePercent down. The default is 5, which can lead to too many young regions, or not enough old regions added to the CSet, so that we run into to-space exhausted.
On the other hand, pushing G1NewSizePercent to 1 does not give the best performance. For coherence, 2 gives best throughput w/o to-space exhausted. For hbase ycsb, 3 gives best throughput.

As we discussed, there might be a way to consider free space available when constructing cset mixed gc.This includes lowering G1NewSizePercent enough to avoid to-space exhausted, as well as adding old regions to cset