2 minutes is cute. If you stress a CMS setup hard enough that the young generation is completely full, it will allocate directly in the old generation. This of

2 minutes is cute. If you stress a CMS setup hard enough that the young generation is completely full, it will allocate directly in the old generation. This of course screws the full gc heuristic totally, up to the point where the GC is started too late and you fully run out of memory. At which point the JVM drops down to a single threaded oldschool serial GC as last line of defense. On a 96GiB heap, that thing can take hours; all stuck 100% on a single cpu with even signal handling suspended. Fun times. That said, for heaps above 32ish GiB, we still go with our tuned CMS settings and overcommit one or two additional memory modules. It's a lot cheaper than the time it takes trying to tune in G1 on a large heap with a lot of gc pressure.