Date: Mon, 8 Aug 2016 19:18:56 +0700 From: Eugene Grosbein <eugen@grosbein.net> To: Sepherosa Ziehau <sepherosa@gmail.com> Cc: "freebsd-net@freebsd.org" <freebsd-net@freebsd.org> Subject: Re: 40Gbps http client benchmark Message-ID: <57A878B0.9040004@grosbein.net> In-Reply-To: <CAMOc5cyHsLZ93R9jQncBgReii0nCTnW0ZoBFbA4RkHpTcwHcnw@mail.gmail.com> References: <57A62668.7020309@grosbein.net> <CAMOc5cyHsLZ93R9jQncBgReii0nCTnW0ZoBFbA4RkHpTcwHcnw@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On 08.08.2016 08:57, Sepherosa Ziehau wrote: > We have an optimized wrk here: > https://github.com/sepherosa/wrk Thank you, I'll try it. > It greatly reduces the # of kqueue syscalls and avoids unnecessary > setsockopt etc. BTW, how many concurrent connections and threads are > you testing w/? Did you reduce the MSL on your client machines? > Default local port range probably should be ok, but it still worth > checking. I use default port ranges and get best results with command like this: cpuset -l 0-7 wrk -t8 -c60 -d60s http://x.x.x.x/index This way, I get 4.38GB out of 4x10G lagg (lagghash l4), that is about 36Gbps. It seems, uneven lagg load balancing prevents it from reaching higher numbers. I tried to increase connection number from -c60 to large values but then unpatched wrk overhelms CPU with syscalls.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?57A878B0.9040004>