Date: Tue, 20 May 2014 19:41:13 +0400 From: Slawa Olhovchenkov <slw@zxy.spb.ru> To: Adrian Chadd <adrian@freebsd.org> Cc: "svn-src-head@freebsd.org" <svn-src-head@freebsd.org>, "svn-src-all@freebsd.org" <svn-src-all@freebsd.org>, "src-committers@freebsd.org" <src-committers@freebsd.org> Subject: Re: svn commit: r265792 - head/sys/kern Message-ID: <20140520154113.GA23318@zxy.spb.ru> In-Reply-To: <CAJ-Vmo=x-JkHSVDT4REWQ_g=vb45T%2BTfrpZ8bc28NNbPd-ZYsg@mail.gmail.com> References: <201405100053.s4A0rbF9080571@svn.freebsd.org> <20140511083114.GA53503@zxy.spb.ru> <CAJ-Vmo=x-JkHSVDT4REWQ_g=vb45T%2BTfrpZ8bc28NNbPd-ZYsg@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sun, May 11, 2014 at 08:32:34AM -0700, Adrian Chadd wrote: > On 11 May 2014 01:31, Slawa Olhovchenkov <slw@zxy.spb.ru> wrote: > > On Sat, May 10, 2014 at 12:53:37AM +0000, Adrian Chadd wrote: > > > >> Author: adrian > >> Date: Sat May 10 00:53:36 2014 > >> New Revision: 265792 > >> URL: http://svnweb.freebsd.org/changeset/base/265792 > >> > >> Log: > >> Add in support to optionally pin the swi threads. > >> > >> Under enough load, the swi's can actually be preempted and migrated > >> to other currently free cores. When doing RSS experiments, this lead > >> to the per-CPU TCP timers not lining up any more with the RX CPU said > >> flows were ending up on, leading to increased lock contention. > >> > >> Since there was a little pushback on flipping them on by default, > >> I've left the default at "don't pin." > >> > >> The other less obvious problem here is that the default swi > >> is also the same as the destination swi for CPU #0. So if one > >> pins the swi on CPU #0, there's no default floating swi. > >> > >> A nice future project would be to create a separate swi for > >> the "default" floating swi, as well as per-CPU swis that are > >> (optionally) pinned. > > > > MFC planed? > > I have 10.0 box with aprox. 16Gbit TCP at peak. > > I've no plans to MFC it at the present. > > By itself it shouldn't do very much. The rest of RSS stack and driver > changes have to go in before it'll matter. > > (But if you try it on 10.0 and it changes things, by all means let me know.) I am try on 10.0, but not sure about significant improvement (may be 10%). For current CPU (E5-2650 v2 @ 2.60GHz) hwpmc don't working (1. after collect some data `pmcstat -R sample.out -G out.txt` don't decode any; 2. kldunload hwpmc do kernel crash) and I can't collect detailed profile information.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20140520154113.GA23318>