I would imagine a heuristic to keep enough of the filter in L2 cache to avoid pipeline stalls might be useful. Sort of a double bloom, but weighted for common lookups.