Re: [PATCH v2] sched/fair: Skip sched_balance_running cmpxchg when balance is not due

[Tim Chen <tim.c.chen@linux.intel.com>]

On Wed, 2025-10-29 at 09:47 +0100, Peter Zijlstra wrote:
> On Tue, Oct 28, 2025 at 01:23:30PM -0700, Tim Chen wrote:
> > The NUMA sched domain sets the SD_SERIALIZE flag by default, allowing
> > only one NUMA load balancing operation to run system-wide at a time.
> > 
> > Currently, each MC group leader in a NUMA domain attempts to acquire
> > the global sched_balance_running flag via cmpxchg() before checking
> > whether load balancing is due or whether it is the designated leader for
> > that NUMA domain. On systems with a large number of cores, this causes
> > significant cache contention on the shared sched_balance_running flag.
> > 
> > This patch reduces unnecessary cmpxchg() operations by first checking
> > whether the balance interval has expired. If load balancing is not due,
> > the attempt to acquire sched_balance_running is skipped entirely.
> > 
> > On a 2-socket Granite Rapids system with sub-NUMA clustering enabled,
> > running an OLTP workload, 7.8% of total CPU cycles were previously spent
> > in sched_balance_domain() contending on sched_balance_running before
> > this change.
> > 
> >          : 104              static __always_inline int arch_atomic_cmpxchg(atomic_t *v, int old, int new)
> >          : 105              {
> >          : 106              return arch_cmpxchg(&v->counter, old, new);
> >     0.00 :   ffffffff81326e6c:       xor    %eax,%eax
> >     0.00 :   ffffffff81326e6e:       mov    $0x1,%ecx
> >     0.00 :   ffffffff81326e73:       lock cmpxchg %ecx,0x2394195(%rip)        # ffffffff836bb010 <sched_balance_running>
> >          : 110              sched_balance_domains():
> >          : 12234            if (atomic_cmpxchg_acquire(&sched_balance_running, 0, 1))
> >    99.39 :   ffffffff81326e7b:       test   %eax,%eax
> >     0.00 :   ffffffff81326e7d:       jne    ffffffff81326e99 <sched_balance_domains+0x209>
> >          : 12238            if (time_after_eq(jiffies, sd->last_balance + interval)) {
> >     0.00 :   ffffffff81326e7f:       mov    0x14e2b3a(%rip),%rax        # ffffffff828099c0 <jiffies_64>
> >     0.00 :   ffffffff81326e86:       sub    0x48(%r14),%rax
> >     0.00 :   ffffffff81326e8a:       cmp    %rdx,%rax
> > 
> > After applying this fix, sched_balance_domain() is gone from
> > the profile and there is a 8% throughput improvement.
> > 
> 
> this..
> 
> > v2:
> > 1. Rearrange the patch to get rid of an indent level per Peter's
> >    suggestion.
> > 2. Updated the data from new run by OLTP team.
> > 
> > link to v1: https://lore.kernel.org/lkml/e27d5dcb724fe46acc24ff44670bc4bb5be21d98.1759445926.git.tim.c.chen@linux.intel.com/
> 
> ... stuff goes under the '---' sign.
> 
> Also, what happened to my other suggestion:
> 
>   https://lkml.kernel.org/r/20251014092436.GK4067720@noisy.programming.kicks-ass.net
> 
> ? That seemed like a better place to put things.

Yes, I agree it is better to lock sched_balance_running after should_we_balance().
That will keep all the CPUs that shouldn't do balancing from locking sched_balance_running, as Shrikanth
also pointed out. Let me put that version of the patch through some OLTP testing
to validate it.

Tim