Re: [PATCH] sched: balance_cpu to consider other cpus in its group as target of (pinned) task migration

[Peter Zijlstra <peterz@infradead.org>]

On Mon, 2012-06-04 at 17:11 +0530, Srivatsa Vaddagiri wrote:
> * Peter Zijlstra <peterz@infradead.org> [2012-06-04 11:00:54]:
> 
> > > Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
> > 
> > Did vatsa write this patch?
> 
> I wrote the first version of the patch which Prashanth took, tested,
> fixed a bug and is finally publishing it. So yes,
> 
> > If so, you forgot a From header, if not, wtf!?
> 
> it is missing the From header.
> 
> > OK, so previously we only pulled to ourselves,
> 
> That't not entirely true isn't it i.e this_cpu need not equal
> smp_processor_id even before this change.

You forgot to finish that, I presume you were thinking of nohz idle
balancing? True, but in that case the target was at least idle.

> > now you make cpu x move
> > from cpu y to cpu z. This changes the dynamic of the load-balancer, not
> > a single word on that and its impact/ramifications.
> 
> The other possibility is for the right sibling cpus to do load balance
> in the same domain (noting that it needs to pull a task from another
> sched_group to itself and ignoring balance_cpu). That seemed like a more
> invasive change than this patch. We'd be happy to try any other approach
> you have in mind.

I'm not saying the approach is bad, I'm just saying the patch is bad.
Mostly because there's a distinct lack of information on things.

There's nothing to indicate you've considered stuff, found this the best
solution because of other stuff etc... thus I think its the first thing
that came to mind without due consideration.

I don't like unconsidered poking at the load-balancer.