Re: Announce - Staircase Deadline cpu scheduler v0.42

[Nick Piggin]

On Thu, Apr 19, 2007 at 07:40:04PM +1000, Con Kolivas wrote:
> On Thursday 19 April 2007 13:22, Nick Piggin wrote:
> > On Thu, Apr 19, 2007 at 12:12:14PM +1000, Con Kolivas wrote:
> > > Version 0.42
> > >
> > > http://ck.kolivas.org/patches/staircase-deadline/2.6.21-rc7-sd-0.42.patch
> >
> > OK, I run some tests later today...
> 
> 
> Thank you very much.

lmbench numbers are roughly comparable to mainline (lmbench seems to be
a bit erratic, but there isn't the obvious drop that cfs has).

Didn't worry about hackbench ;)

kbuild:
2.6.21-rc7
508.87user 32.47system 2:17.82elapsed 392%CPU
509.05user 32.25system 2:17.84elapsed 392%CPU
508.75user 32.26system 2:17.83elapsed 392%CPU
508.63user 32.17system 2:17.88elapsed 392%CPU
509.01user 32.26system 2:17.90elapsed 392%CPU
509.08user 32.20system 2:17.95elapsed 392%CPU

2.6.21-rc7-sd42
512.78user 31.99system 2:18.41elapsed 393%CPU
512.55user 31.90system 2:18.57elapsed 392%CPU
513.05user 31.78system 2:18.48elapsed 393%CPU
512.46user 32.06system 2:18.63elapsed 392%CPU
512.78user 31.81system 2:18.49elapsed 393%CPU
512.41user 32.08system 2:18.70elapsed 392%CPU

sd42 is doing about 745 context switches per second here, and perfomance is
slightly below mainline. But it isn't doing badly.

507.87user 32.53system 2:17.50elapsed 392%CPU
508.47user 32.40system 2:17.56elapsed 393%CPU
508.59user 32.27system 2:17.53elapsed 393%CPU 

A few runs with rr_interval at 100 show that ctxsw numbers drop to 587, and
performance is up to slightly above mainline.

With the results I've got so far with all scedulers (actually I didn't try
nicksched with a small timeslice, but I'm sure it would give the expected
result)... I'd say 5ms might be too small a timeslice. Even 15ms will hurt
some people I think.

Although we could arguably tolerate this kind of regression, my box only
has 1MB caches, and kbuild is naturally context switching at over 500 per
second anyway. On something with bigger caches and less context switchy /
more cache sensitive workloads, the regression could be quite a bit worse.

(not directed at anyone in particular, but food for thought)
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/