Re: [PATCH 1/2] sched/fair: Use weight of SD_NUMA domain in find_busiest_group

[Mel Gorman <mgorman@techsingularity.net>]

On Tue, Dec 21, 2021 at 11:53:50AM +0100, Vincent Guittot wrote:
> On Fri, 10 Dec 2021 at 10:33, Mel Gorman <mgorman@techsingularity.net> wrote:
> >
> > find_busiest_group uses the child domain's group weight instead of
> > the sched_domain's weight that has SD_NUMA set when calculating the
> > allowed imbalance between NUMA nodes. This is wrong and inconsistent
> > with find_idlest_group.
> 
> I agree that find_busiest_group and find_idlest_group should be
> consistent and use the same parameters but I wonder if sched_domain's
> weight is the right one to use instead of the target group's weight.
> 

Ok

> IIRC, the goal of adjust_numa_imbalance is to keep some threads on the
> same node as long as we consider that there is no performance impact
> because of sharing  resources as they can even take advantage of
> locality if they interact.

Yes.

> So we consider that tasks will not be
> impacted by sharing resources if they use less than 25% of the CPUs of
> a node. If we use the sd->span_weight instead, we consider that we can
> pack threads in the same node as long as it uses less than 25% of the
> CPUs in all nodes.
> 

I assume you mean the target group weight instead of the node. The
primary resource we are concerned with is memory bandwidth and it's a
guess because we do not know for sure where memory channels are or how
they are configured in this context and it may or may not be correlated
with groups. I think using the group instead would deserve a series on
its own after settling on an imbalance number when there are multiple
LLCs per node.

-- 
Mel Gorman
SUSE Labs