Re: hackbench regression with kernel 2.6.32-rc1

[Mike Galbraith <efault@gmx.de>]

On Thu, 2009-10-29 at 08:50 +0800, Zhang, Yanmin wrote:
> On Wed, 2009-10-28 at 15:22 +0100, Mike Galbraith wrote:
> > On Wed, 2009-10-28 at 17:29 +0800, Zhang, Yanmin wrote:
> > > 	-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 select_task_rq_fair consumes so much time.
> > 
> > Turn off SD_WAKE_BALANCE as it was called in rc1.  See commit 182a85f.
> I run testing against 2.6.32-rc1 which already includes the patch.

Duh, I checked the wrong tree.

SD_PREFER_LOCAL is still on in rc1 though (double checks;), so you'll go
through the power saving code until you reach a domain containing both
waker's cpu and wakee's previous cpu even if that code already found
that a higher domain wasn't overloaded.  Looks to me like that block
wants a want_sd && qualifier. 

Even it you turn SD_PREFER_LOCAL off, you can still hit the overhead if
SD_POWERSAVINGS_BALANCE is set, so I'd make sure both are off and see if
that's the source (likely, since the rest is already off).

	-Mike