On 7/10/07, Bryan Cantrill <bmc at eng.sun.com> wrote:
>
>
> On Tue, Jul 10, 2007 at 11:49:48AM -0600, Bruce Shaw wrote:
> > >The kstat facility has long been in need of a serious overhaul; this
> > >presents a good opportunity to do this work.
> >
> >
> > I want in on this.  I use kstat to derive CPU information for net-snmp.
> > If we're opening the API, I've got some suggestions.
>
> Excellent.  So do I. ;)  To me, this would be an incredibly valuable
> project, and it would allow us to get the benefits of a "CPUfs" besides.
> And I think this might be one of those instances where we want to at least
> perform the thought experiment of breaking the API (especially the
> in-kernel
> API, which is unspeakably broken in that it exports far too much
> implementation).  Not that we necessarily need to break the API in order
> to move kstat forward, just that this is a situation where we want to
> at least understand what we might be able to do if we were to start with
> a clean(er) sheet of paper...


I have been thinking about this, and one of the things I don't understand,
is how the Scheduler selects, which CMT VCPU, CORE, or socket to schedule
processes or threads. Could someone give me a quick explanation, and how it
relates to kstat? (Or a reference book/link where I can find the info)

Looking down the road, I can also see a further silicon layer, "core
clusters" that share a databus, w/ the databuses tied together with an
onchip xbar. (This may or may not come to fruition, but as we scale core
counts, we will probably see (cc-)numa designs on a single piece of silicon.

I ask because I am thinking about various workloads, will have different
requirements. Some may perform best with core/cache affinity, and others
that would best be utilized by "striping" processes/threads across
cores/sockets. How do we account for this?

--brian

P.S. - What I seek is a stable interface to get this system information via
the commandline. (IE: scriptable). I understand that this may require the
API and the knobs to be hammered out first.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: 
<http://mail.opensolaris.org/pipermail/observability-discuss/attachments/20070710/059eac0e/attachment.html>

Reply via email to