Re: [PATCH 7/6][RFC] sched: unify load_balance{,_newidle}()

[Mike Galbraith <efault@gmx.de>]

On Thu, 2009-12-24 at 10:29 +0100, Peter Zijlstra wrote:
> On Thu, 2009-12-24 at 05:43 +0100, Mike Galbraith wrote:
> > On Wed, 2009-12-23 at 16:13 +0100, Peter Zijlstra wrote:
> > > load_balance() and load_balance_newidle() look remarkably similar, one
> > > key point they differ in is the condition on when to active balance.
> > > 
> > > So split out that logic into a separate function.
> > > 
> > > One side effect is that previously load_balance_newidle() used to fail and
> > > return -1 under these conditions, whereas now it doesn't. I've not yet fully
> > > figured out the whole -1 return case for either load_balance{,_newidle}().
> > > 
> > > It also differs in that sd->cache_nice_tries is now added on the 
> > > CPU_NEWLY_IDLE case.
> > 
> > Unification Looks like a good idea, less being more and all that.  I
> > suspect that last bit is why newidle effectiveness has been heavily
> > impacted.  x264 ultrafast testcase is whimpering pathetically again ;-)
> 
> That could be easily verified by setting cache_nice_tries to 0.

Was bad wild guess.

> However, I would suspect need_active_balance(.idle = CPU_NEWLY_IDLE) to
> always fail on your machine, since I don't think you've got all that
> power savings muck enabled.

Nope. Guess I need to poke need_active_balance() with a stick.

> Is that with just this patch applied or also with the next one? I
> worried more about the next one.

With both... all 8 actually, but popping last two cures idle woes.

> If just this one, that funny -1 return value thing might have played a
> role, since that seems to trigger the:
> 
>  if (pulled_task) {
>    this_rq->idle_stamp = 0;
>    break;
>  }
> 
> logic in idle_balance()
> 
> Which didn't make any sense to me, since it didn't move any task, so why
> pretend it did...

Heh, pretending is unlikely to help.

	-Mike