On 9/29/2025 9:44 PM, Peter Zijlstra wrote:
> On Sat, Aug 09, 2025 at 01:02:18PM +0800, Chen Yu wrote:
>> From: Tim Chen <tim.c.chen@linux.intel.com>
>>
>> Switching a process's preferred LLC generates lots of task
>> migrations across LLCs. To avoid frequent switches
>> of home LLC, implement the following policy:
>>
>> 1. Require a 2x occ change threshold to switch preferred LLC
>> 2. Don't discard preferred LLC for a task
>>
>> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
>> ---
>> kernel/sched/fair.c | 24 ++++++++++++++++--------
>> 1 file changed, 16 insertions(+), 8 deletions(-)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index a7be5c5ecba3..9e3c6f0eb934 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -1175,6 +1175,14 @@ static s64 update_curr_se(struct rq *rq, struct sched_entity *curr)
>> #define EPOCH_PERIOD (HZ/100) /* 10 ms */
>> #define EPOCH_OLD 5 /* 50 ms */
>>
>> +static int llc_id(int cpu)
>> +{
>> + if (cpu < 0)
>> + return -1;
>> +
>> + return per_cpu(sd_llc_id, cpu);
>> +}
>> +
>> void mm_init_sched(struct mm_struct *mm, struct mm_sched __percpu *_pcpu_sched)
>> {
>> unsigned long epoch;
>> @@ -1307,6 +1315,7 @@ static void __no_profile task_cache_work(struct callback_head *work)
>> struct task_struct *p = current;
>> struct mm_struct *mm = p->mm;
>> unsigned long m_a_occ = 0;
>> + unsigned long last_m_a_occ = 0;
>> int cpu, m_a_cpu = -1;
>> cpumask_var_t cpus;
>>
>> @@ -1345,11 +1354,13 @@ static void __no_profile task_cache_work(struct callback_head *work)
>> per_cpu(sd_llc_id, i), occ, m_occ, m_cpu, nr);
>> }
>>
>> - a_occ /= nr;
>> + // a_occ /= nr;
>
> This seems broken.
>
The original thought was that not dividing the nr might help in
an asymmetric LLC scenario.
If there are 2 threads in the 4-CPU LLC and 3 threads in the
8-CPU LLC, it might be better to choose the one with 3 threads.
But it wouldn't be the case if division is done by the nr.
And the NUMA balancing fault statistic behaves similarly:
it compares the total number of faults per node rather than
the average number of faults per CPU, IIUC.
>> if (a_occ > m_a_occ) {
>> m_a_occ = a_occ;
>> m_a_cpu = m_cpu;
>> }
>> + if (llc_id(cpu) == llc_id(mm->mm_sched_cpu))
>> + last_m_a_occ = a_occ;
>
> Not 'last', 'curr' perhaps?
>
OK.
>>
>> trace_printk("(%d) a_occ: %ld m_a_occ: %ld\n",
>> per_cpu(sd_llc_id, cpu), a_occ, m_a_occ);
>> @@ -1363,13 +1374,10 @@ static void __no_profile task_cache_work(struct callback_head *work)
>> }
>> }
>>
>> - /*
>> - * If the max average cache occupancy is 'small' we don't care.
>> - */
>> - if (m_a_occ < (NICE_0_LOAD >> EPOCH_OLD))
>> - m_a_cpu = -1;
>> -
>> - mm->mm_sched_cpu = m_a_cpu;
>> + if (m_a_occ > (2 * last_m_a_occ)) {
>> + /* avoid the bouncing of mm_sched_cpu */
>> + mm->mm_sched_cpu = m_a_cpu;
>> + }
>
>
> The whole double thing doesn't seem right either. That means that
> anything over .5 will never be able to change, even when confronted with
> say a .8.
>
> Also, while this is a threshold, this is not in fact hysteresis.
>
I agree 2X might be too aggressive. The threshold here was intended
to avoid switching to a new mm_sched_cpu too quickly, but it is not
intended to address the back-and-forth bouncing issue that hysteresis
would resolve. Instead, the task migration bouncing issue is solved
in patch 7/28 via a hysteresis.
Maybe the comment above:
/* avoid the bouncing of mm_sched_cpu */
should be changed to:
/* avoid switching to mm_sched_cpu too fast */
I'll investigate an imbalance_pct value that is less aggressive than 2X.
thanks,
Chenyu