Yes, that is how it works, except (as you implicitly note) that large heaps in single processes can cause problems; allowing incremental collection per heap would flatten the latency profile further.