Re: hackbench regression with kernel 2.6.32-rc1

[Zhang, Yanmin]

On Tue, 2009-10-27 at 15:42 +0100, Mike Galbraith wrote:
> On Tue, 2009-10-27 at 16:03 +0800, Zhang, Yanmin wrote:
> > On Fri, 2009-10-16 at 13:06 +0200, Mike Galbraith wrote:
> > > On Tue, 2009-10-13 at 11:12 +0800, Zhang, Yanmin wrote:
> > > 
> > > > NEXT_BUDDY has no help on volanoMark and tbench.
> > > 
> > > Can you try the patch below please?  It does tries to preserve buddy
> > > affinity where possible, and mitigates over-preemption by strengthening
> > > buddies a bit.  It improves vmark here by ~7%.
> > I ran some benchmarks against 2.6.32-rc1+Peter_2_patches+below_patch.
> > Below result is against 2.6.32-rc1.
> > hackbench result has about 10% improvement on stoakley (2*4 cores) and
> > tigerton (4*4 cores).
> > tbench still has about 5% regression on ïstoakley and tigerton.
> > VolanoMark has 33% regression on tigerton, but has 2% improvement on stoakley.
> > 
> > I also ran the benchmarks against the latest tips/master and got the similiar
> > results like above testing.
> > ï
> > The testing against tips on Nehalem machine didn't show much improvement/regression.
> 
> Thanks for the testing.  Your results suggest that I should revive the
> mark buddies whether you use them or not idea.
> 
> 	-Mike
I'm investigating 5% tbench regression on Nehalem machine. perf_counter shows
select_task_rq_fair consumes about 5% cpu time with 2.6.32-rc1 while it consumes
less than 0.5% with 2.6.31.

Patch c88d5910890 has comments to explain it, but I still can't understand why
to add complicated balance logic when selecting task rq.

I will check which section in function ïselect_task_rq_fair consumes so much time.

Yanmin


--
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/