Re: [PATCH 1/2] sched/fair: Consider SD_NUMA when selecting the most idle group to schedule on

[Mel Gorman]

On Tue, Feb 13, 2018 at 02:04:45PM +0100, Peter Zijlstra wrote:
> On Tue, Feb 13, 2018 at 11:35:48AM +0000, Mel Gorman wrote:
> > > However, if we have numa balancing enabled, that will counteract
> > > the normal spreading across nodes, so in that regard it makes sense, but
> > > the above code is not conditional on numa balancing.
> > > 
> > 
> > It's not conditional on NUMA balancing because one case where it mattered
> > was a fork-intensive workload driven by shell scripts. In that case, the
> > workload benefits from preferring a local node without any involvement from
> > NUMA balancing. I could make it conditional on it but it's not strictly
> > related to automatic NUMA balancing, it's about being less eager about
> > starting new children on remote nodes.
> 
> Yeah, I suppose. And you're right, there's no real winning this. It's
> all tea-leaves and entrails.
> 

That is my new favourite description of this portion of the scheduler :D

> In any case, I think I prefer the kill sync early variant and you were
> going to ammend some comments. Can you respin and resend all these
> patches (can do in a single series)?

No problem. I had it prepared already and am just waiting for one result
before I push send.

Thanks.

-- 
Mel Gorman
SUSE Labs