Reuse your threads, and undef as much as possible within the thread, when the thread run finishes. This way, your memory use should level out eventually, as each thread reuses it's own memory space. Repeatedly running threads async, is similar to that create-destroy cycle which tends to gain memory in OO programs. Reusing objects fixes it.

Without testing, I would say that each threads memory use will be the maximum of it's most memory-intensive run. But if you undef everything, then on it's next run, it should reuse it's memory.