Re: [PATCH] sched/fair: Adjust the allowed NUMA imbalance when SD_NUMA spans multiple LLCs

[Valentin Schneider]

On 28/10/21 14:03, Mel Gorman wrote:
> Commit 7d2b5dd0bcc4 ("sched/numa: Allow a floating imbalance between NUMA
> nodes") allowed an imbalance between NUMA nodes such that communicating
> tasks would not be pulled apart by the load balancer. This works fine when
> there is a 1:1 relationship between LLC and node but can be suboptimal
> for multiple LLCs if independent tasks prematurely use CPUs sharing cache.
>
> Zen* has multiple LLCs per node with local memory channels and due to
> the allowed imbalance, it's far harder to tune some workloads to run
> optimally than it is on hardware that has 1 LLC per node. This patch
> adjusts the imbalance on multi-LLC machines to allow an imbalance up to
> the point where LLCs should be balanced between nodes.
>

I've run out of brain juice for today and didn't get to decipher the logic
you're implementing, but for now I do have a comment on the topology
detection side of things (see inline).

> --- a/kernel/sched/topology.c
> +++ b/kernel/sched/topology.c
> @@ -644,6 +644,7 @@ static void destroy_sched_domains(struct sched_domain *sd)
>  DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
>  DEFINE_PER_CPU(int, sd_llc_size);
>  DEFINE_PER_CPU(int, sd_llc_id);
> +DEFINE_PER_CPU(int, sd_numaimb_shift);
>  DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
>  DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
>  DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
> @@ -672,6 +673,20 @@ static void update_top_cache_domain(int cpu)
>       sd = lowest_flag_domain(cpu, SD_NUMA);
>       rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
>
> +	/*
> +	 * Save the threshold where an imbalance is allowed between SD_NUMA
> +	 * domains. If LLC spans the entire node, then imbalances are allowed
> +	 * until 25% of the domain is active. Otherwise, allow an imbalance
> +	 * up to the point where LLCs between NUMA nodes should be balanced
> +	 * to maximise cache and memory bandwidth utilisation.
> +	 */
> +	if (sd) {
> +		if (sd->span_weight == size)
> +			per_cpu(sd_numaimb_shift, cpu) = 2;
> +		else
> +			per_cpu(sd_numaimb_shift, cpu) = max(2, ilog2(sd->span_weight / size * num_online_nodes()));
> +	}
> +

So nodes are covered by the NODE topology level which *doesn't* have
SD_NUMA set. I always get confused on how MC/DIE/NODE is supposed to look
on those sub-NUMA clustering thingies, but either way consider:

  NUMA-20 [              ]
  NODE    [      ][      ]
  DIE     [      ][      ]
  MC      [  ][  ][  ][  ]

NODE level gets degenerated, update_top_cache_domain() is invoked with:

  NUMA-20 [              ]
  DIE     [      ][      ]
  MC      [  ][  ][  ][  ]

That lowest_flag_domain(cpu, SD_NUMA) will span the entire system.

Conversely, with this topology where node == LLC:

  NUMA-20 [              ]
  NODE    [      ][      ]
  DIE     [      ][      ]
  MC      [      ][      ]

You get

  NUMA-20 [              ]
  MC      [      ][      ]

lowest_flag_domain(cpu, SD_NUMA)->span_weight > size, even though LLC =
node.

Long story short, I think you want to use sd->child here - that *should*
point to a domain that spans exactly one node (it's gonna be NODE, or some
other domain that has the same span because NODE was degenerated).

>       sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
>       rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
>