The intersection problem they are solving is pretty trivial. I don't know why you would even consider using a bloom filter for it, that's not exactly the best domain for it.
Also, I would have just used the C++ set_intersection function. It seems unlikely that the 2x speedup matters since they already got it down from 7 seconds to 6ms.
With the work we were performing, the 2x improvement may not have been necessary, but it was useful. We could run everything on 1 box and keep up with roughly 30-50% system utilization continuously. Had we used the STL version, we would have needed another machine, and may not have been able to use mmaps to share data between cores.
Also, I would have just used the C++ set_intersection function. It seems unlikely that the 2x speedup matters since they already got it down from 7 seconds to 6ms.