On Mon, 20 Jul 2020 17:05:25 +0200 Marc Roos wrote: >> I'm a bit suspicious about some of the speedup figures quoted, and >> whether rspamd was tested against an optimized and similarly >> parameterized SA. It's very easy to make SA look bad. > > I agree. I have even asked on the mailing list how many test rspamd > does and how I can configure it to do just one test. Both questions > were left unanswered.
rspamd is threaded so it should process a single message much faster than SA. Whether it has better throughput under heavy load is a different matter. If it's still true that most cpu time is spent on evaluating rule regexes then the throughput should be similar, with SA getting a modest boost from sa-compile and rspamd getting a modest boost from shortcircuiting out some subrules. It wouldn't surprise me if the factor of 6 simply came from scanning the test corpus one at a time on a 6 core cpu.
