--
On Thu, 25 Oct 2007, Gregory Haskins wrote:

quoted text
> Oh crap.  I just realized this is an older version of the patch..mustv'e
> forgot to refresh...grr.  Ill send out the refreshed one.

Been there, done that ;-)

quoted text
>
> But anyway, I digress.
>
> On Thu, 2007-10-25 at 11:27 -0400, Steven Rostedt wrote:
>
> >
> > The cpu_priority and the cp->lock will be aboslutely horrible for
> > cacheline bouncing.
>
> In the form presented here in this email, perhaps.  I think you will see
> some significant improvements in the refreshed version.  The big change
> is that the global lock is gone.
>
> >   Ironically, this will kill performance for the very
> > machines this code is to help with.  The larger the number of CPUs you
> > have the more cacheline bouncing this code will create.
>
> Don't forget:  The same is precisely true for the current -rt2
> algorithm.  For instance, the -rt2 algorithm aside from being linear in
> general, scales cacheline bouncing linearly as well.  Each cpu is going
> to trash rq->rt.highest_prio and then we will walk them for each scan.

Yep, -rt2 (and -rt3) are both horrible too. That's why I'm working on a
sched-domain version now to handle that.

quoted text
>
> The fact is, you can't maintain a global dynamic policy without bouncing
> cachelines, period.  But hopefully we can minimize it, and I just want
> to see the fastest code here.

Well, we need a balance between cacheline hits, and where to pull from.
As performance goes, the two are inverse perportional. The secret is to
find a nice middle ground.

quoted text
>
> >
> > I still don't see the benefit from the cpupri code.
>
> I still owe you timing data, but at this juncture I think I can beat
> linear (especially as we start throwing in big-iron) ;)  I originally
> got involved in this scheduler rework from observations of poor scaling
> on our 8/16-ways, so I want it to scale as much as you ;)  If this alg
> doesn't pan out, that's cool.  But I think it will in the end.  Linear
> algs in the fast path just make my skin crawl.  Perhaps it will still be
> the best design in the end, but I am not giving up that easy until I
> prove it to myself.

It's only linear in respect to the size of the domain or cpus allowed. Any
8/16 way boxes worried about cacheline bouncing need to reign in on the
cpus allowed mask. Otherwise, we would simply have bouncing anyway with
the tasks themselves.

-- Steve

-
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/