Re: [PATCH RESEND] sched: prefer an idle cpu vs an idle sibling for BALANCE_WAKE

[Mike Galbraith <umgwanakikbuti@gmail.com>]

On Wed, 2015-06-03 at 16:34 -0400, Josef Bacik wrote:
> On 06/03/2015 01:43 PM, Mike Galbraith wrote:

> > There are also other loads like your server where waking to an idle cpu
> > dominates all else, pgbench is one of those.  In that case, you've got a
> > 1:N waker/wakee relationship, and what matters above ALL else is when
> > the mother of all work (the single server thread) wants a CPU, it had
> > better get it NOW, else the load stalls.  Likewise, 'mom' being
> > preempted hurts truckloads.  Perhaps your server has a similar thing
> > going on, keeping wakees the hell away from the waker rules all.
> >
> 
> Yeah our server has two waker threads (one per numa node) and then the N 
> number of wakee threads.  I'll run tbench and pgbench with the new 
> patches and see if there's a degredation.  Thanks,

If you look for wake_wide(), it could perhaps be used to select wider
search for only the right flavor load component when BALANCE_WAKE is
set.  That would let the cache lovers in your box continue to perform
while improving the 1:N component.  That wider search still needs to
become cheaper though, low hanging fruit being to stop searching when
you find load = 0.. but you may meet the energy efficient folks, who
iirc want to make it even more expensive.

wake_wide() inadvertently helped another sore spot btw - a gaggle of
pretty light tasks being awakened from an interrupt source tended to
cluster around that source, preventing such loads from being all they
can be in a very similar manner.  Xen (shudder;) showed that nicely in
older kernels, due to the way its weird dom0 gizmo works.

	-Mike