Re: 回复：[Internet]Re: 回复：[PATCH 06/19] sched/fair: Assign preferred LLC ID to processes

[Tim Chen]

On Tue, 2025-10-14 at 15:07 +0800, vernhao(郝信) wrote:
> Hi Tim, 
> 
> > 
> > 
[snip]

> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 61c129bde8b6..d6167a029c47 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -1312,6 +1312,7 @@ void account_mm_sched(struct rq *rq, struct task_struct *p, s64 delta_exec)
> >   struct mm_struct *mm = p->mm;
> >   struct mm_sched *pcpu_sched;
> >   unsigned long epoch;
> > + int mm_sched_llc = -1;
> >  
> >   if (!sched_cache_enabled())
> >   return;
> > @@ -1342,6 +1343,12 @@ void account_mm_sched(struct rq *rq, struct task_struct *p, s64 delta_exec)
> >   if (mm->mm_sched_cpu != -1)
> >   mm->mm_sched_cpu = -1;
> >   }
> > +
> > + if (mm->mm_sched_cpu != -1)
> > + mm_sched_llc = per_cpu(sd_llc_id, mm->mm_sched_cpu);
> > 
> > In high-concurrency multi-threaded scenarios, not all threads handle same events, so their hot data in the LLC is not completely shared. 
> > Therefore, if every thread's preferred LLC is migrated to the LLC pointed to by mm->mm_sched_cpu, this would lead to the incorrect 
> > assumption that all threads prefer the same LLC, thereby intensifying competition between LLCs.
> 
> Yes, that's the reason why we stop aggregating to the preferred LLC once the the utilization of the
> LLC becomes too high relative to the other LLCs.
> 
> But this approach is only a compensatory measure after the fact. The threads have already undergone incorrect migration to they are not perferred LLC. 
> Is there a better way to handle this situation?

The threads would stay where they were instead of migrating to preferred LLC
that's overloaded.

> 
> If you know your threads characteristics before hand on which of them
> share data together, you probably can use cgroup/cpuset
> from user space to separate out the threads.  
> 
> Yes, this is a solution, and I am trying to implement it.
> 
> There's not enough info from occupancy data for OS to group
> the threads by data sharing. Perhaps an alternative if NUMA balancing
> is on is to group tasks by their task numa group instead of by mm.  
> 
> This may not be a good solution either, especially for virtual machine scenarios which has no NUMA.

If you are in a VM, the cache topology may not correspond to
real CPU cache topology and you probably should not enable cache
aware scheduling inside, unless you are doing some explicit
binding of VCPUs.

> 
> That would incur the page scanning overhead etc and make
> cache aware scheduling be dependent on NUMA balancing.
>  
> 
> > 
> > So I'm wondering, why not move ‘mm->mm_sched_cpu’ to ‘task_struct’, so that each thread can individually track its preferred LLC? What are the losses in doing so?
> 
> You would need a way to group related tasks together and put them
> on the same LLC.  Either group them by mm or some other means.
> 
> Yes, you are right, how about this, beside in 'mm',  add cgroup support too ？ 

Doing cgroup may not solve the original issue you brought
up, where a process may have a group of tasks wanting to go
into one cache and another group of tasks going to another cache.
I could be wrong but I don't think you can split up tasks in a process
in cgroup v2 to different cgroups.

Also the cgroup folks are quite resistant to adding new knobs.

Tim

> 
> > 
> > +
> > + if (p->preferred_llc != mm_sched_llc)
> > + p->preferred_llc = mm_sched_llc;
> >  }
> >  
> >  static void task_tick_cache(struct rq *rq, struct task_struct *p)