I expect that as memory usage increases, the resulting increase in GC pressure causes the GC to run more often. Most GCs that I know of are designed to have some memory free for the young generation (for fast, cheap allocations) and some memory free in the old generation (so objects can be moved from the young generation to the old generation). Maybe the JVM detects that it's approaching the limits of physical memory and starts GCing more often to avoid running out of space?

EDIT: From rereading the original post it seems more likely that it was approaching the heap size limit set when running the JVM and not a physical memory limit. I would expect the same principle applies, though.