Re: [PATCH v4] sched/fair: Prefer fully idle cores for NOHZ balancing

[Andrea Righi]

Hi Shrikanth,

On Tue, Aug 04, 2026 at 08:49:01PM +0530, Shrikanth Hegde wrote:
> Hi Andrea,
> 
> On 8/4/26 8:43 PM, Andrea Righi wrote:
> > find_new_ilb() selects the first idle housekeeping CPU without
> > considering whether another thread is running on the same physical core.
> > On an SMT system, the idle load balancer can therefore activate both
> > siblings even when another housekeeping CPU has an entirely idle core.
> > 
> > On most SMT systems, this is not problematic because the idle load
> > balancer is a short-lived activity and the transient wakeup of a sibling
> > has negligible performance impact.
> > 
> > However, this can be particularly costly on NVIDIA Olympus cores used in
> > Vera. Briefly activating an otherwise idle sibling can reduce the
> > performance available to the other sibling and this effect does not
> > necessarily end once the activated sibling becomes idle: after the ILB
> > finishes and its CPU enters WFI, full single-thread performance is
> > restored only after the sibling has remained idle for a qualification
> > interval (10 Ki cycles on the tested Vera system). Repeated short
> > sibling wakeups can therefore sustain the interference even with little
> > actual overlap.
> 
> Is this hardware cost has been accounted in things like cpufreq or cpuidle?
> or have you set it already to max performance.

All the CPUs were using the cppc_cpufreq performance governor during the tests.
The scaling minimum frequency was between 95.5% and 97% of the maximum
frequency, depending on the CPU. The CPUs were not strictly pinned to their
maximum frequency, but the performance governor was in use.

The acpi_idle driver was enabled, with both LPI-0 and LPI-1 available. I didn't
disable the idle states.

> 
> > 
> > Prevent this by preferring an idle housekeeping CPU whose entire SMT
> > core is idle. Retain the first idle CPU as a fallback when no fully idle
> > core is available, so NOHZ balancing continues to make forward progress.
> > Once a partially busy core has been examined, skip its remaining SMT
> > siblings to avoid repeating the core-idle check on wide SMT systems.
> > 
> 
> Just Curious, making nohz_full=<except first core> yeilds similar numbers?

Yes, provided that the housekeeping core is also excluded from the workload.

I tested this booting with:

  nohz_full=1-175,177-351

This leaves CPUs 0 and 176 for housekeeping. I excluded that core from the
workload and ran 87 OpenMP tasks on the remaining node-0 cores (1 task per core,
excluding the housekeeping one):

 $ env OMP_NUM_THREADS=87 \
       OMP_DYNAMIC=false \
       OPENBLAS_LOOPS=10 \
       OPENBLAS_PARAM_M=16384 \
       OPENBLAS_PARAM_N=16384 \
       OPENBLAS_PARAM_K=16384 \
       numactl -C 1-87,177-263 --membind=0 \
       ./benchmark/sgemm.goto 1 1 1

Results:

    unpatched           : 5.246 TFLOP/s
    unpatched+nohz_full : 6.953 TFLOP/s
    patched             : 6.861 TFLOP/s
    patched+nohz_full   : 6.984 TFLOP/s

So, nohz_full seems to prevent the problematic ILB wakeups and can produce
similar results for this CPU-bound workload. However, it shouldn't be considered
a sobstiute for the ILB fix, since it requires explicit partitioning and
reserved housekeeping CPUs. Full dyntick also enables context tracking on the
isolated CPUs, adding kernel entry/exit overhead.

Thanks,
-Andrea

> 
> > Tests performed using an ad hoc GEMM benchmark running one CPU-intensive
> > task per SMT core within its CPU affinity mask improved from
> > approximately 6.2 TFLOP/s to 9.4 TFLOP/s.
> > 
> > Note that this preference may wake a fully idle physical core instead of
> > using an idle sibling of an active core, potentially increasing ILB
> > wakeup latency or energy consumption on some architectures. It may also
> > scan additional CPUs before selecting the one to run the ILB. The
> > selection falls back to the first idle CPU when no fully idle SMT core
> > is available. Non-SMT systems continue to select the first idle
> > housekeeping CPU.
> > 
> > Cc: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
> > Cc: K Prateek Nayak <kprateek.nayak@xxxxxxx>
> > Cc: Shrikanth Hegde <sshegde@xxxxxxxxxxxxx>
> > Signed-off-by: Andrea Righi <arighi@xxxxxxxxxx>
> > Reviewed-by: Mete Durlu <meted@xxxxxxxxxxxxx>
> > ---
> > Changes in v4:
> >   - Remove redundant this_cpu check (Prateek Nayak, Vincent Guittot)
> >   - Link to v3: https://lore.kernel.org/all/20260731191957.3199642-1-arighi@xxxxxxxxxx/
> > 
> > Changes in v3:
> >   - After finding an idle fallback, skip all siblings when a busy CPU is
> >     encountered, avoiding per-CPU traversal of known-busy cores (Mete Durlu)
> >   - Link to v2: https://lore.kernel.org/all/20260729163225.1987068-1-arighi@xxxxxxxxxx/
> > 
> > Changes in v2:
> >   - Avoid repeated is_core_idle() checks on wide SMT systems by pruning
> >     the remaining siblings of a partially busy core (Prateek Nayak)
> >   - Link to v1: https://lore.kernel.org/r/20260728214442.1648483-1-arighi@xxxxxxxxxx/
> > 
> >   kernel/sched/fair.c | 55 ++++++++++++++++++++++++++++++++++++---------
> >   1 file changed, 44 insertions(+), 11 deletions(-)
> > 
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 37001c63452e5..9a975a684b487 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -13964,29 +13964,62 @@ static inline int on_null_domain(struct rq *rq)
> >    */
> >   static inline int find_new_ilb(void)
> >   {
> > -	int this_cpu = smp_processor_id();
> > -	const struct cpumask *hk_mask;
> > -	int ilb_cpu;
> > +	struct cpumask *ilb_cpus;
> > +	int ilb_cpu, fallback = -1;
> > +
> > +	lockdep_assert_irqs_disabled();
> > +
> > +	/*
> > +	 * Reuse the per-CPU select_rq_mask, which is protected from concurrent
> > +	 * use on this CPU by having interrupts disabled.
> > +	 */
> > +	ilb_cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
> > +	cpumask_and(ilb_cpus, nohz.idle_cpus_mask,
> > +		    housekeeping_cpumask(HK_TYPE_KERNEL_NOISE));
> > +
> > +	for_each_cpu(ilb_cpu, ilb_cpus) {
> > +		if (!idle_cpu(ilb_cpu)) {
> > +			/*
> > +			 * Once an idle fallback exists, a busy CPU proves that
> > +			 * this core cannot be fully idle. Skip its siblings.
> > +			 */
> > +			if (sched_smt_active() && fallback >= 0)
> > +				cpumask_andnot(ilb_cpus, ilb_cpus, cpu_smt_mask(ilb_cpu));
> > +			continue;
> > +		}
> > -	hk_mask = housekeeping_cpumask(HK_TYPE_KERNEL_NOISE);
> > +		/*
> > +		 * Running the idle load balancer on an idle sibling of a busy
> > +		 * SMT core can reduce the capacity available to its sibling. Prefer
> > +		 * a CPU whose entire core is idle, but retain the first idle CPU as
> > +		 * a fallback so idle balancing can still make progress when no fully
> > +		 * idle core exists.
> > +		 */
> > +		if (sched_smt_active() && !is_core_idle(ilb_cpu)) {
> > +			if (fallback < 0)
> > +				fallback = ilb_cpu;
> > -	for_each_cpu_and(ilb_cpu, nohz.idle_cpus_mask, hk_mask) {
> > -		if (ilb_cpu == this_cpu)
> > +			/*
> > +			 * The core is not idle, so there is no need to check
> > +			 * any of its other SMT siblings.
> > +			 */
> > +			cpumask_andnot(ilb_cpus, ilb_cpus,
> > +				       cpu_smt_mask(ilb_cpu));
> >   			continue;
> > +		}
> > -		if (idle_cpu(ilb_cpu))
> > -			return ilb_cpu;
> > +		return ilb_cpu;
> >   	}
> > -	return -1;
> > +	return fallback;
> >   }
> >   /*
> >    * Kick a CPU to do the NOHZ balancing, if it is time for it, via a cross-CPU
> >    * SMP function call (IPI).
> >    *
> > - * We pick the first idle CPU in the HK_TYPE_KERNEL_NOISE housekeeping set
> > - * (if there is one).
> > + * Prefer a CPU on a fully idle core in the HK_TYPE_KERNEL_NOISE housekeeping
> > + * set. Fall back to the first idle CPU when no fully idle core exists.
> >    */
> >   static void kick_ilb(unsigned int flags)
> >   {
>