[RFC PATCH v6 3/4] scheduler: scan idle cpu in cluster for tasks within one LLC

[Dietmar Eggemann]

On 20/04/2021 02:18, Barry Song wrote:

[...]

> @@ -5786,11 +5786,12 @@ static void record_wakee(struct task_struct *p)
>   * whatever is irrelevant, spread criteria is apparent partner count exceeds
>   * socket size.
>   */
> -static int wake_wide(struct task_struct *p)
> +static int wake_wide(struct task_struct *p, int cluster)
>  {
>  	unsigned int master = current->wakee_flips;
>  	unsigned int slave = p->wakee_flips;
> -	int factor = __this_cpu_read(sd_llc_size);
> +	int factor = cluster ? __this_cpu_read(sd_cluster_size) :
> +		__this_cpu_read(sd_llc_size);

I don't see that the wake_wide() change has any effect here. None of the
sched domains has SD_BALANCE_WAKE set so a wakeup (WF_TTWU) can never
end up in the slow path.
Have you seen a diff when running your `lmbench stream` workload in what
wake_wide() returns when you use `sd cluster size` instead of `sd llc
size` as factor?

I guess for you,  wakeups are now subdivided into faster (cluster = 4
CPUs) and fast (llc = 24 CPUs) via sis(), not into fast (sis()) and slow
(find_idlest_cpu()).

>  
>  	if (master < slave)
>  		swap(master, slave);

[...]

> @@ -6745,6 +6748,12 @@ static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
>  	int want_affine = 0;
>  	/* SD_flags and WF_flags share the first nibble */
>  	int sd_flag = wake_flags & 0xF;
> +	/*
> +	 * if cpu and prev_cpu share LLC, consider cluster sibling rather
> +	 * than llc. this is typically true while tasks are bound within
> +	 * one numa
> +	 */
> +	int cluster = sched_cluster_active() && cpus_share_cache(cpu, prev_cpu, 0);

So you changed from scanning cluster before LLC to scan either cluster
or LLC.

And this is based on whether `this_cpu` and `prev_cpu` are sharing LLC
or not. So you only see an effect when running the workload with
`numactl -N X ...`.

>  
>  	if (wake_flags & WF_TTWU) {
>  		record_wakee(p);
> @@ -6756,7 +6765,7 @@ static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
>  			new_cpu = prev_cpu;
>  		}
>  
> -		want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
> +		want_affine = !wake_wide(p, cluster) && cpumask_test_cpu(cpu, p->cpus_ptr);
>  	}
>  
>  	rcu_read_lock();
> @@ -6768,7 +6777,7 @@ static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
>  		if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
>  		    cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
>  			if (cpu != prev_cpu)
> -				new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
> +				new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync, cluster);
>  
>  			sd = NULL; /* Prefer wake_affine over balance flags */
>  			break;
> @@ -6785,7 +6794,7 @@ static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
>  		new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
>  	} else if (wake_flags & WF_TTWU) { /* XXX always ? */
>  		/* Fast path */
> -		new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
> +		new_cpu = select_idle_sibling(p, prev_cpu, new_cpu, cluster);
>  
>  		if (want_affine)
>  			current->recent_used_cpu = cpu;

[...]