Re: [Patch v4 02/22] sched/cache: Limit the scan number of CPUs when calculating task occupancy

["Chen, Yu C" <yu.c.chen@intel.com>]

Hi Gengkun,

On 4/9/2026 9:17 PM, Luo Gengkun wrote:
> 
> 
> On 2026/4/2 5:52, Tim Chen wrote:
>> From: Chen Yu <yu.c.chen@intel.com>

[ ... ]

> 
> To address the issue of scanning overhead, there is a more targeted 
> approach: only scanning the CPUs actually accessed by the process, and 
> evicting these CPUs when they remain unaccessed for a specific period of 
> time.
> 

Thanks for the patch. This approach looks quite sensible to me, and

[ ... ]

> 
>          /*
>           * The update to mm->sc_stat should not be reordered
> @@ -1582,6 +1584,7 @@ void account_mm_sched(struct rq *rq, struct 
> task_struct *p, s64 delta_exec)
>                  pcpu_sched->runtime += delta_exec;
>                  rq->cpu_runtime += delta_exec;
>                  epoch = rq->cpu_epoch;
> +               cpumask_set_cpu(cpu_of(rq), &mm->sc_stat.visited_cpus);

I would refer a check before writing to avoid c2c overhead:
	if (!cpumask_test_cpu(cpu_of(rq), &mm->sc_stat.visited_cpus))
		cpumask_set_cpu(cpu_of(rq), &mm->sc_stat.visited_cpus);

>          }
> 
> 
>          /*
> @@ -1724,7 +1727,10 @@ static void task_cache_work(struct callback_head 
> *work)
>                  return;
> 
> 
>          scoped_guard (cpus_read_lock) {
> -               get_scan_cpumasks(cpus, p);
> +               if (unlikely(llc_epoch_visited_timeout == 0))
> +                       get_scan_cpumasks(cpus, p);
> +               else
> +                       cpumask_and(cpus, cpu_online_mask, &mm- 
>  >sc_stat.visited_cpus);
> 

I believe this patch will bring significant benefits(reduce jitter)
whenthe thread count is low, since it skips many CPUs. The only concern
is the frequency of reading and writing the "global" 
mm->sc_stat.visited_cpus
from multiple threads. However, if Peter's sbm can be leveraged in the 
future,
this  overhead should be further reduced. Anyway, let me run tests on our
machines to seethe results. Much appreciated!

thanks,
Chenyu