Hi Vincent, On Tue, Nov 25, 2014 at 05:15:09PM +0100, Vincent Bernat wrote: > ??? 25 novembre 2014 16:46 +0100, Emeric Brun <[email protected]> : > > > Indeed on haproxy.com appliances, we made the choice to dedicate several > > cores/CPUs for NICs interrupts (manual bind, no usage of irqbalance) and > > to dedicate one other for the haproxy process with nbproc=1. > > On a related topic, did you get a chance to look at the new > SO_INCOMING_CPU (which should land in 3.18)? From my understanding, this > doesn't apply to HAProxy architecture since it assumes that incoming > connections will be handed to workers. > > See: > http://lwn.net/Articles/619862/
Interesting, thanks for the link, I hadn't noticed it yet. It might be interesting if/when we implement multithreading, so that each epoll works primarily on sockets from the same CPU. Definitely something to test at some point :-) Cheers, Willy
