Re: [PATCH][RSDL-mm 0/7] RSDL cpu scheduler for 2.6.21-rc3-mm2

[Mike Galbraith <efault@gmx.de>]

On Tue, 2007-03-13 at 21:06 +1100, Con Kolivas wrote:
> On Tuesday 13 March 2007 20:39, Ingo Molnar wrote:
> > * Mike Galbraith <efault@gmx.de> wrote:
> > > I just retested with the encoders at nice 0, and the x/gforce combo is
> > > terrible. [...]
> >
> > ok. So nice levels had nothing to do with it - it's some other
> > regression somewhere. How does the vanilla scheduler cope with the
> > exactly same workload? I.e. could you describe the 'delta' difference in
> > behavior - because the delta is what we are interested in mostly, the
> > 'absolute' behavior alone is not sufficient. Something like:
> >
> >  - on scheduler foo, under this workload, the CPU hogs steal 70% CPU
> >    time and the resulting desktop experience is 'choppy': mouse pointer
> >    is laggy and audio skips.
> >
> >  - on scheduler bar, under this workload, the CPU hogs are at 40%
> >    CPU time and the desktop experience is smooth.
> >
> > things like that - we really need to be able to see the delta.
> 
> I only find a slowdown, no choppiness, no audio stutter (it would be extremely 
> hard to make audio stutter in this design without i/o starvation or something 
> along those lines). The number difference in cpu percentage I've already 
> given on the previous email. The graphics driver does feature in this test 
> case as well so others' mileage may vary. Mike said it was terrible.

My first test run with lame at nice 0 was truly horrid, but has _not_
repeated, so disregard that as an anomaly.  For the most part, it is as
you say, things just get slower with load, any load.  I definitely am
seeing lurchiness which is not present in mainline.  No audio problems
with either kernel.

> > > [...]  Funny thing though, x/gforce isn't as badly affected with a
> > > kernel build.  Any build is quite noticable, but even at -j8, the
> > > effect doen't seem to be (very brief test warning applies) as bad as
> > > with only the two encoders running.  That seems quite odd.
> >
> > likewise, how does the RSDL kernel build behavior compare to the vanilla
> > scheduler's behavior? (what happens in one that doesnt happen in the
> > other, etc.)
> 
> Kernel compiles seem similar till the jobs get above about 3 where rsdl gets 
> slower but still smooth. Audio is basically unaffected either way.

It seems to be a plain linear slowdown.  The lurchiness I'm experiencing
varies in intensity, and is impossible to quantify.  I see neither
lurchiness nor slowdown in mainline through -j8.

> Don't forget all the rest of the cases people have posted.

Absolutely, all test results count.

	-Mike