[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <Z9hZ9fgtGNd8DeEf@gpd3>
Date: Mon, 17 Mar 2025 18:20:53 +0100
From: Andrea Righi <arighi@...dia.com>
To: Joel Fernandes <joelagnelf@...dia.com>
Cc: linux-kernel@...r.kernel.org, Tejun Heo <tj@...nel.org>,
David Vernet <void@...ifault.com>,
Changwoo Min <changwoo@...lia.com>, Ingo Molnar <mingo@...hat.com>,
Peter Zijlstra <peterz@...radead.org>,
Juri Lelli <juri.lelli@...hat.com>,
Vincent Guittot <vincent.guittot@...aro.org>,
Dietmar Eggemann <dietmar.eggemann@....com>,
Steven Rostedt <rostedt@...dmis.org>,
Ben Segall <bsegall@...gle.com>, Mel Gorman <mgorman@...e.de>,
Valentin Schneider <vschneid@...hat.com>
Subject: Re: [PATCH RFC] sched_ext: Choose prev_cpu if idle and cache affine
without WF_SYNC
Hi Joel,
On Mon, Mar 17, 2025 at 04:28:02AM -0400, Joel Fernandes wrote:
> Consider that the previous CPU is cache affined to the waker's CPU and
> is idle. Currently, scx's default select function only selects the
> previous CPU in this case if WF_SYNC request is also made to wakeup on the
> waker's CPU.
>
> This means, without WF_SYNC, the previous CPU being cache affined to the
> waker and is idle is not considered. This seems extreme. WF_SYNC is not
> normally passed to the wakeup path outside of some IPC drivers but it is
> very possible that the task is cache hot on previous CPU and shares
> cache with the waker CPU. Lets avoid too many migrations and select the
> previous CPU in such cases.
>
> This change is consistent with the fair scheduler's behavior as well. In
> select_idle_sibling(), the previous CPU is selected if it is cache
> affined with the target. This is done regardless of WF_SYNC and before
> any scanning of fully idle cores is done.
>
> One difference still exists though between SCX and CFS in this regard, in CFS
> we first check if the target CPU is idle before checking if the previous CPU is
> idle. However that could be a matter of choice and in the future, that behavior
> could also be unified.
>
> Signed-off-by: Joel Fernandes <joelagnelf@...dia.com>
> ---
> kernel/sched/ext.c | 24 +++++++++++-------------
> 1 file changed, 11 insertions(+), 13 deletions(-)
>
> diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
> index 5a81d9a1e31f..3b1a489e2aaf 100644
> --- a/kernel/sched/ext.c
> +++ b/kernel/sched/ext.c
> @@ -3479,7 +3479,7 @@ static s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
> {
> const struct cpumask *llc_cpus = NULL;
> const struct cpumask *numa_cpus = NULL;
> - s32 cpu;
> + s32 cpu = smp_processor_id();
>
> *found = false;
>
> @@ -3507,22 +3507,20 @@ static s32 scx_select_cpu_dfl(struct task_struct *p, s32 prev_cpu,
> llc_cpus = llc_span(prev_cpu);
> }
>
> + /*
> + * If the waker's CPU is cache affine and prev_cpu is idle, then avoid
> + * a migration.
> + */
> + if (cpus_share_cache(cpu, prev_cpu) &&
> + test_and_clear_cpu_idle(prev_cpu)) {
> + cpu = prev_cpu;
> + goto cpu_found;
> + }
> +
One potential issue that I see is that when SMT is enabled, you may end up
using prev_cpu also when the other sibling is busy. Maybe we should check
if prev_cpu is set in the SMT idle cpumask.
Also, last time I tried a similar change I was regressing a lot of
benchmarks. Maybe we should repeat the tests and get some numbers.
> /*
> * If WAKE_SYNC, try to migrate the wakee to the waker's CPU.
> */
> if (wake_flags & SCX_WAKE_SYNC) {
> - cpu = smp_processor_id();
> -
> - /*
> - * If the waker's CPU is cache affine and prev_cpu is idle,
> - * then avoid a migration.
> - */
> - if (cpus_share_cache(cpu, prev_cpu) &&
> - test_and_clear_cpu_idle(prev_cpu)) {
> - cpu = prev_cpu;
> - goto cpu_found;
> - }
> -
> /*
> * If the waker's local DSQ is empty, and the system is under
> * utilized, try to wake up @p to the local DSQ of the waker.
> --
> 2.43.0
>
Thanks,
-Andrea
Powered by blists - more mailing lists