Hi Cyril, On Tue, Oct 09, 2012 at 11:30:59PM +0200, Cyril Bonté wrote: > Le 09/10/2012 23:21, Cyril Bonté a écrit : > >Hi Willy, > > > >Le 09/10/2012 22:53, Willy Tarreau a écrit : > >>both patches merged, thank you Cyril! > > > >Less than 10 minutes to merge them, amazing :-)
probably that you sent the mail just before I opened my box :-) > >I also noticed that "tune.sslcachesize" was undocumented. Possible indeed. BTW I think we should rename it "tune.ssl.cache-size" to avoid reintroducing the naming nightmare we have already known in the past. > >As a side note, when ssl was announced, I gave it a try and was > >surprised by the performances which were great for a first pre-release, > >but I didn't have time to test it further. > >Tonight, I've replayed some bench with the last git commits but I didn't > >find anymore this feeling of performance : around 70 requests/sec > >whereas a similar configuration with stunnel brings 500 requests/sec. > >I'll try to investigate this next days (currently, I can't say if I made > >a mistake in the configuration or in the tests). > > Oh btw, I've just made a new test with haproxy-1.5-dev12 and the > performance is here again : around 500 req/s with "nbproc 1" and 1200 > req/s with "nbproc 8". OK. Its possible that we did something wrong in one version. I know that Emeric does a lot of performance tests, but it is always possible that I break something each time I touch the lower layers. If you observe anything suspect again, do not hesitate to report it. The difference you're talking about could also be caused by too low a cache size, causing constant rekeying. Cheers, Willy
