[PATCH sched_ext/for-7.1] sched_ext: idle: Prioritize idle SMT sibling

[PATCH sched_ext/for-7.1] sched_ext: idle: Prioritize idle SMT sibling

[Andrea Righi]

In the default built-in idle CPU selection policy, when @prev_cpu is
busy and no fully idle core is available, try to place the task on its
SMT sibling if that sibling is idle, before searching any other idle CPU
in the same LLC.

Migration to the sibling is cheap and keeps the task on the same core,
preserving L1 cache and reducing wakeup latency.

On large SMT systems this appears to consistently boost throughput by
roughly 2-3% on CPU-bound workloads (running a number of tasks equal to
the number of SMT cores).

Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 kernel/sched/ext_idle.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/kernel/sched/ext_idle.c b/kernel/sched/ext_idle.c
index c7e4052626979..e0c57355b33b8 100644
--- a/kernel/sched/ext_idle.c
+++ b/kernel/sched/ext_idle.c
@@ -616,6 +616,18 @@ s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu, u64 wake_flags,
 		goto out_unlock;
 	}
 
+	/*
+	 * Use @prev_cpu's sibling if it's idle.
+	 */
+	if (sched_smt_active()) {
+		for_each_cpu_and(cpu, cpu_smt_mask(prev_cpu), allowed) {
+			if (cpu == prev_cpu)
+				continue;
+			if (scx_idle_test_and_clear_cpu(cpu))
+				goto out_unlock;
+		}
+	}
+
 	/*
 	 * Search for any idle CPU in the same LLC domain.
 	 */
-- 
2.53.0

Re: [PATCH sched_ext/for-7.1] sched_ext: idle: Prioritize idle SMT sibling

[Cheng-Yang Chou]

Hi Andrea,

On Wed, Mar 18, 2026 at 01:38:42AM +0100, Andrea Righi wrote:
> In the default built-in idle CPU selection policy, when @prev_cpu is
> busy and no fully idle core is available, try to place the task on its
> SMT sibling if that sibling is idle, before searching any other idle CPU
> in the same LLC.
> 
> Migration to the sibling is cheap and keeps the task on the same core,
> preserving L1 cache and reducing wakeup latency.
> 
> On large SMT systems this appears to consistently boost throughput by
> roughly 2-3% on CPU-bound workloads (running a number of tasks equal to
> the number of SMT cores).
> 
> Signed-off-by: Andrea Righi <arighi@nvidia.com>
> ---
>  kernel/sched/ext_idle.c | 12 ++++++++++++
>  1 file changed, 12 insertions(+)
> 
> diff --git a/kernel/sched/ext_idle.c b/kernel/sched/ext_idle.c
> index c7e4052626979..e0c57355b33b8 100644
> --- a/kernel/sched/ext_idle.c
> +++ b/kernel/sched/ext_idle.c
> @@ -616,6 +616,18 @@ s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu, u64 wake_flags,
>  		goto out_unlock;
>  	}
>  
> +	/*
> +	 * Use @prev_cpu's sibling if it's idle.
> +	 */
> +	if (sched_smt_active()) {
> +		for_each_cpu_and(cpu, cpu_smt_mask(prev_cpu), allowed) {
> +			if (cpu == prev_cpu)
> +				continue;
> +			if (scx_idle_test_and_clear_cpu(cpu))
> +				goto out_unlock;
> +		}
> +	}
> +
>  	/*
>  	 * Search for any idle CPU in the same LLC domain.
>  	 */
> -- 
> 2.53.0
> 

Overall looks good, just a nit:

The block comment at the top of scx_select_cpu_dfl() still lists 5
steps. With this patch a new step should be added and the numbering
updated accordingly.

-- 
Thanks,
Cheng-Yang

Re: [PATCH sched_ext/for-7.1] sched_ext: idle: Prioritize idle SMT sibling

[Andrea Righi]

Hi Cheng-Yang,

On Wed, Mar 18, 2026 at 09:11:29AM +0800, Cheng-Yang Chou wrote:
> Hi Andrea,
> 
> On Wed, Mar 18, 2026 at 01:38:42AM +0100, Andrea Righi wrote:
> > In the default built-in idle CPU selection policy, when @prev_cpu is
> > busy and no fully idle core is available, try to place the task on its
> > SMT sibling if that sibling is idle, before searching any other idle CPU
> > in the same LLC.
> > 
> > Migration to the sibling is cheap and keeps the task on the same core,
> > preserving L1 cache and reducing wakeup latency.
> > 
> > On large SMT systems this appears to consistently boost throughput by
> > roughly 2-3% on CPU-bound workloads (running a number of tasks equal to
> > the number of SMT cores).
> > 
> > Signed-off-by: Andrea Righi <arighi@nvidia.com>
> > ---
> >  kernel/sched/ext_idle.c | 12 ++++++++++++
> >  1 file changed, 12 insertions(+)
> > 
> > diff --git a/kernel/sched/ext_idle.c b/kernel/sched/ext_idle.c
> > index c7e4052626979..e0c57355b33b8 100644
> > --- a/kernel/sched/ext_idle.c
> > +++ b/kernel/sched/ext_idle.c
> > @@ -616,6 +616,18 @@ s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu, u64 wake_flags,
> >  		goto out_unlock;
> >  	}
> >  
> > +	/*
> > +	 * Use @prev_cpu's sibling if it's idle.
> > +	 */
> > +	if (sched_smt_active()) {
> > +		for_each_cpu_and(cpu, cpu_smt_mask(prev_cpu), allowed) {
> > +			if (cpu == prev_cpu)
> > +				continue;
> > +			if (scx_idle_test_and_clear_cpu(cpu))
> > +				goto out_unlock;
> > +		}
> > +	}
> > +
> >  	/*
> >  	 * Search for any idle CPU in the same LLC domain.
> >  	 */
> > -- 
> > 2.53.0
> > 
> 
> Overall looks good, just a nit:
> 
> The block comment at the top of scx_select_cpu_dfl() still lists 5
> steps. With this patch a new step should be added and the numbering
> updated accordingly.

Ah yes, good catch, we should update the comment as well. Will send a v2.

Thanks,
-Andrea

Re: [PATCH sched_ext/for-7.1] sched_ext: idle: Prioritize idle SMT sibling

[Daniel Jordan]

Hi Andrea,

On Wed, Mar 18, 2026 at 01:38:42AM +0100, Andrea Righi wrote:
> In the default built-in idle CPU selection policy, when @prev_cpu is
> busy and no fully idle core is available, try to place the task on its
> SMT sibling if that sibling is idle, before searching any other idle CPU
> in the same LLC.
> 
> Migration to the sibling is cheap and keeps the task on the same core,
> preserving L1 cache and reducing wakeup latency.

Seems reasonable.

> On large SMT systems this appears to consistently boost throughput by
> roughly 2-3% on CPU-bound workloads (running a number of tasks equal to
> the number of SMT cores).

What workloads out of curiosity?

Re: [PATCH sched_ext/for-7.1] sched_ext: idle: Prioritize idle SMT sibling

[Andrea Righi]

Hi Daniel,

On Fri, Mar 20, 2026 at 11:57:02AM -0400, Daniel Jordan wrote:
> Hi Andrea,
> 
> On Wed, Mar 18, 2026 at 01:38:42AM +0100, Andrea Righi wrote:
> > In the default built-in idle CPU selection policy, when @prev_cpu is
> > busy and no fully idle core is available, try to place the task on its
> > SMT sibling if that sibling is idle, before searching any other idle CPU
> > in the same LLC.
> > 
> > Migration to the sibling is cheap and keeps the task on the same core,
> > preserving L1 cache and reducing wakeup latency.
> 
> Seems reasonable.
> 
> > On large SMT systems this appears to consistently boost throughput by
> > roughly 2-3% on CPU-bound workloads (running a number of tasks equal to
> > the number of SMT cores).
> 
> What workloads out of curiosity?

For the "server" side, I've used an internal benchmark suite, the workload
that is showing the best results (3% speedup) is based on NVBLAS, but doing
pure CPU activity.

I've also tested this locally (AMD Ryzen 9 laptop), usual gaming benchmarks
checking avg fps and tail latency. I noticed small improvements for this
use-case as well (still in the 1-2% range, nothing impressive, but it looks
consistent).

Thanks,
-Andrea