Re: tbench regression with 2.6.33-rc1

From: Lin Ming
Date: Tue Jan 12 2010 - 04:36:21 EST

On Tue, 2010-01-12 at 16:54 +0800, Peter Zijlstra wrote:
> On Tue, 2010-01-12 at 09:09 +0800, Lin Ming wrote:
> > I test this patch applied to 2.6.33-rc3, but no help on tbench
> > regression.
> How stable is this regression on your machines? That is, can you bisect
> it? Me and Mike seem to have no luck bisecting it, ending up at totally
> unrelated patches and the like :/

I tried bisect weeks ago, but the bisect result is not stable.

Then I tried another way to reproduce this regression.
I picked up most scheduler related patches that merged into 2.6.33-rc1,
and applied them to 2.6.32, then the ~4% regression was reproduced.

But again, the bisect result among these scheduler related patches was
not stable either. The regression seems not caused by one single patch,
but the accumulation of some patches.

You may try the attached scheduler patches to reproduce the regression.

git checkout v2.6.32
git am sched-2.6.33-rc1.patch

Lin Ming

Attachment: sched-2.6.33-rc1.patch.tar.bz2
Description: application/bzip-compressed-tar