Skip site navigation (1)Skip section navigation (2)
Date:      Mon, 9 Oct 2006 00:55:24 +0200
From:      "Attilio Rao" <attilio@freebsd.org>
To:        "John-Mark Gurney" <gurney_j@resnet.uoregon.edu>,  "David Xu" <davidxu@freebsd.org>, freebsd-current@freebsd.org,  "Kip Macy" <kmacy@fsmware.com>, "Ivan Voras" <ivoras@fer.hr>
Subject:   Re: [PATCH] MAXCPU alterable in kernel config - needs testers
Message-ID:  <3bbf2fe10610081555r67265368sf7f12edbf35bff0d@mail.gmail.com>
In-Reply-To: <20061008225150.GK793@funkthat.com>
References:  <2fd864e0610080423q7ba6bdeal656a223e662a5d@mail.gmail.com> <20061008135031.G83537@demos.bsdclusters.com> <4529667D.8070108@fer.hr> <200610090634.31297.davidxu@freebsd.org> <20061008225150.GK793@funkthat.com>

next in thread | previous in thread | raw e-mail | index | archive | help
2006/10/9, John-Mark Gurney <gurney_j@resnet.uoregon.edu>:
> David Xu wrote this message on Mon, Oct 09, 2006 at 06:34 +0800:
> > On Monday 09 October 2006 04:58, Ivan Voras wrote:
> > > Kip Macy wrote:
> > > > It will only cover the single chip Niagara 2 boxes.
> > >
> > > Oh right, they'll doing multi chips in Niagara 2 :) Go Sun :)
> > >
> > > Still, single T2 chips should be more common, so I'd guess it will pay
> > > to optimize for that case.
> > >
> > > (For the rest of the audience: Niagara 1 has 32 logical CPUs and
> > > supports only one physical CPU/socket; Niagara 2 will have 64 logical
> > > CPUs and support > 1 CPUs/sockets; so a 2 socket Niagara 2 box will have
> > > 128 logical processors! Cue SciFi music...)
> > >
> > > Any word on how will they handle migration of threads across sockets (or
> > > will it be OS's job)? Judging from T1 architecture, I think such event
> > > would create a very large performance penalty, but I'm not an expert.
> > > __________
> >
> > The current 4BSD scheduler does not handle large number of cores  very  well,
> > also the single sched_lock will be a bottleneck for such a configuration.
>
> Bad enough that Kip had to reduce HZ down to 100... since sched_lock
> ends up serializing ALL cpus when scheduling needs to happen..

How would you see a sched_lock decomposition (and, if it is possible,
how many locks it could be decomposed in?)

Attilio


-- 
Peace can only be achieved by understanding - A. Einstein



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?3bbf2fe10610081555r67265368sf7f12edbf35bff0d>