On Thu, Feb 20, 2014 at 12:09:04AM +0200, Alexander Motin wrote: > On 19.02.2014 23:44, Slawa Olhovchenkov wrote: > > On Wed, Feb 19, 2014 at 11:04:49PM +0200, Alexander Motin wrote: > > > >> On 19.02.2014 22:04, Adrian Chadd wrote: > >>> On 19 February 2014 11:59, Alexander Motin <mav_at_freebsd.org> wrote: > >>> > >>>>> So if we're moving towards supporting (among others) a pcbgroup / RSS > >>>>> hash style work load distribution across CPUs to minimise > >>>>> per-connection lock contention, we really don't want the scheduler to > >>>>> decide it can schedule things on other CPUs under enough pressure. > >>>>> That'll just make things worse. > >>> > >>>> True, though it is also not obvious that putting second thread on CPU run > >>>> queue is better then executing it right now on another core. > >>> > >>> Well, it depends if you're trying to optimise for "run all runnable > >>> tasks as quickly as possible" or "run all runnable tasks in contexts > >>> that minimise lock contention." > >>> > >>> The former sounds great as long as there's no real lock contention > >>> going on. But as you add more chances for contention (something like > >>> "100,000 concurrent TCP flows") then you may end up having your TCP > >>> timer firing stuff interfere with more TXing or RXing on the same > >>> connection. > >> > >> 100K TCP flows probably means 100K locks. That means that chance of lock > >> collision on each of them is effectively zero. More realistic it could > > > > What about 100K/N_cpu*PPS timer's queue locks for remove/insert TCP > > timeouts callbacks? > > I am not sure what this formula means, but yes, per-CPU callout locks > can much more likely be congested. They are only per-CPU, not per-flow. 100K TCP flows distributed between CPU (100K/N_cpu). every TCP flow several times per seconds touch his callout (*PPS)Received on Wed Feb 19 2014 - 22:07:21 UTC
This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:47 UTC