Re: [PATCH -v4 0/5] x86,smp: make ticket spinlock proportional backoff w/ auto tuning

[Mike Galbraith <bitbucket@online.de>]

On Sat, 2013-01-26 at 13:05 +0100, Ingo Molnar wrote: 
> * Mike Galbraith <bitbucket@online.de> wrote:
> 
> > On Fri, 2013-01-25 at 14:05 -0500, Rik van Riel wrote:
> > 
> > > The performance issue observed with AIM7 is still a mystery.
> > 
> > Hm.  AIM7 mystery _may_ be the same crud I see on a 4 node 40 
> > core box. Stock scheduler knobs are too preempt happy, produce 
> > unstable results. I twiddle them as below to stabilize 
> > results.
> > 
> > I'm testing a load balancing series from Alex Shi with AIM7 
> > and whatnot, added your series on top of it and retested.  
> > What I see is improvement.
> 
> Since we want both series that's good news, agreed?

Yes, both look good.

BTW, the balance and powersaving low end improvements in the Alex set
are not due to any select_idle_sibling() misbehavior as one might
suspect when looking at the set, it's innocent (this time), it's the
balancing policies themselves.  I don't see where it comes from, gotta
be some turbo thingy methinks, but who cares, better is better.

-Mike