Sample runs shows >10 MB/s write I/O with -m sanity on large machine. Those writes seem to originate from UsePerfData subsystem. We should do -XX:-UsePerfData, since jcstress does not need it anyway.