From: Ricardo Neri <ricardo.neri-calderon@linux.intel.com>
To: Peter Zijlstra <peterz@infradead.org>
Cc: Juri Lelli <juri.lelli@redhat.com>,
Vincent Guittot <vincent.guittot@linaro.org>,
Ricardo Neri <ricardo.neri@intel.com>,
"Ravi V. Shankar" <ravi.v.shankar@intel.com>,
Ben Segall <bsegall@google.com>,
Daniel Bristot de Oliveira <bristot@redhat.com>,
Dietmar Eggemann <dietmar.eggemann@arm.com>,
Len Brown <len.brown@intel.com>, Mel Gorman <mgorman@suse.de>,
"Rafael J. Wysocki" <rafael.j.wysocki@intel.com>,
Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>,
Steven Rostedt <rostedt@goodmis.org>,
Tim Chen <tim.c.chen@linux.intel.com>,
Valentin Schneider <vschneid@redhat.com>,
x86@kernel.org, linux-kernel@vger.kernel.org,
"Tim C . Chen" <tim.c.chen@intel.com>
Subject: Re: [RFC PATCH 08/23] sched/fair: Compute task-class performance scores for load balancing
Date: Tue, 25 Oct 2022 20:57:24 -0700 [thread overview]
Message-ID: <20221026035724.GA21523@ranerica-svr.sc.intel.com> (raw)
In-Reply-To: <YzK/QisKmix6hrKG@hirez.programming.kicks-ass.net>
On Tue, Sep 27, 2022 at 11:15:46AM +0200, Peter Zijlstra wrote:
> On Fri, Sep 09, 2022 at 04:11:50PM -0700, Ricardo Neri wrote:
>
> > +static void compute_ilb_sg_task_class_scores(struct sg_lb_task_class_stats *class_sgs,
> > + struct sg_lb_stats *sgs,
> > + int dst_cpu)
> > +{
> > + int group_score, group_score_without, score_on_dst_cpu;
> > + int busy_cpus = sgs->group_weight - sgs->idle_cpus;
> > +
> > + if (!sched_task_classes_enabled())
> > + return;
> > +
> > + /* No busy CPUs in the group. No tasks to move. */
> > + if (!busy_cpus)
> > + return;
> > +
> > + score_on_dst_cpu = arch_get_task_class_score(class_sgs->p_min_score->class,
> > + dst_cpu);
> > +
> > + /*
> > + * The simpest case. The single busy CPU in the current group will
> > + * become idle after pulling its current task. The destination CPU is
> > + * idle.
> > + */
> > + if (busy_cpus == 1) {
> > + sgs->task_class_score_before = class_sgs->sum_score;
> > + sgs->task_class_score_after = score_on_dst_cpu;
> > + return;
> > + }
> > +
> > + /*
> > + * Now compute the group score with and without the task with the
> > + * lowest score. We assume that the tasks that remain in the group share
> > + * the CPU resources equally.
> > + */
> > + group_score = class_sgs->sum_score / busy_cpus;
> > +
> > + group_score_without = (class_sgs->sum_score - class_sgs->min_score) /
> > + (busy_cpus - 1);
> > +
> > + sgs->task_class_score_after = group_score_without + score_on_dst_cpu;
> > + sgs->task_class_score_before = group_score;
> > +}
(I am sorry Peter, I just found that several emails were sitting on my drafts
directory).
>
> That's just plain broken; also lots of cleanups done...
Thank you very much for your suggestions. They make sense to me. I only
have a comment...
Do you want me to add your Signed-off-by and Co-developed-by tags?
>
> ---
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -8405,12 +8405,14 @@ struct sg_lb_stats {
> enum group_type group_type;
> unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
> unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
> - long task_class_score_after; /* Prospective task-class score after load balancing */
> - long task_class_score_before; /* Task-class score before load balancing */
> #ifdef CONFIG_NUMA_BALANCING
> unsigned int nr_numa_running;
> unsigned int nr_preferred_running;
> #endif
> +#ifdef CONFIG_SCHED_TASK_CLASSES
> + long task_class_score_after; /* Prospective task-class score after load balancing */
> + long task_class_score_before; /* Task-class score before load balancing */
> +#endif
> };
>
> /*
> @@ -8689,58 +8691,54 @@ group_type group_classify(unsigned int i
> }
>
> struct sg_lb_task_class_stats {
> - /*
> - * Score of the task with lowest score among the current tasks (i.e.,
> - * runqueue::curr) of all runqueues in the scheduling group.
> - */
> - int min_score;
> - /*
> - * Sum of the scores of the current tasks of all runqueues in the
> - * scheduling group.
> - */
> - long sum_score;
> - /* The task with score equal to @min_score */
> - struct task_struct *p_min_score;
> + int min_score; /* Min(rq->curr->score) */
> + int min_class;
> + long sum_score; /* Sum(rq->curr->score) */
> };
>
> #ifdef CONFIG_SCHED_TASK_CLASSES
> -static void init_rq_task_classes_stats(struct sg_lb_task_class_stats *class_sgs)
> +static void init_sg_lb_task_class_stats(struct sg_lb_task_class_stats *sgcs)
> {
> - class_sgs->min_score = INT_MAX;
> - class_sgs->sum_score = 0;
> - class_sgs->p_min_score = NULL;
> + *sgcs = (struct sg_lb_task_class_stats){
> + .min_score = INT_MAX,
> + };
> }
>
> /** Called only if cpu_of(@rq) is not idle and has tasks running. */
> -static void update_rq_task_classes_stats(struct sg_lb_task_class_stats *class_sgs,
> - struct rq *rq)
> +static void update_sg_lb_task_class_stats(struct sg_lb_task_class_stats *sgcs,
> + struct rq *rq)
> {
> - int score;
> + struct task_struct *curr;
> + int class, score;
>
> if (!sched_task_classes_enabled())
> return;
>
> + curr = rcu_dereference(rq->curr);
> + if (!curr || (curr->flags & PF_EXITING) || is_idle_task(curr))
> + return;
> +
> /*
> * TODO: if nr_running > 1 we may want go through all the tasks behind
> * rq->curr.
> */
> - score = arch_get_task_class_score(rq->curr->class, cpu_of(rq));
> -
> - class_sgs->sum_score += score;
> + class = curr->class;
> + score = arch_get_task_class_score(class, cpu_of(rq));
>
> - if (score >= class_sgs->min_score)
> - return;
> + sgcs->sum_score += score;
>
> - class_sgs->min_score = score;
> - class_sgs->p_min_score = rq->curr;
> + if (score < sgcs->min_score) {
> + sgcs->min_score = score;
> + sgcs->min_class = class;
> + }
> }
>
> -static void compute_ilb_sg_task_class_scores(struct sg_lb_task_class_stats *class_sgs,
> - struct sg_lb_stats *sgs,
> - int dst_cpu)
> +static void update_sg_lb_stats_scores(struct sg_lb_task_class_stats *sgcs,
> + struct sg_lb_stats *sgs,
> + int dst_cpu)
> {
> - int group_score, group_score_without, score_on_dst_cpu;
> int busy_cpus = sgs->group_weight - sgs->idle_cpus;
> + long before, after;
>
> if (!sched_task_classes_enabled())
> return;
> @@ -8749,32 +8747,18 @@ static void compute_ilb_sg_task_class_sc
> if (!busy_cpus)
> return;
>
> - score_on_dst_cpu = arch_get_task_class_score(class_sgs->p_min_score->class,
> - dst_cpu);
> + score_on_dst_cpu = arch_get_task_class_score(sgcs->min_class, dst_cpu);
>
> - /*
> - * The simpest case. The single busy CPU in the current group will
> - * become idle after pulling its current task. The destination CPU is
> - * idle.
> - */
> - if (busy_cpus == 1) {
> - sgs->task_class_score_before = class_sgs->sum_score;
> - sgs->task_class_score_after = score_on_dst_cpu;
> - return;
> - }
> + before = sgcs->sum_score
> + after = before - sgcs->min_score + score_on_dst_cpu;
This works when the sched group being evaluated has only one busy CPU
because it will become idle if the destination CPU (which was idle) pulls
the current task.
>
> - /*
> - * Now compute the group score with and without the task with the
> - * lowest score. We assume that the tasks that remain in the group share
> - * the CPU resources equally.
> - */
> - group_score = class_sgs->sum_score / busy_cpus;
> -
> - group_score_without = (class_sgs->sum_score - class_sgs->min_score) /
> - (busy_cpus - 1);
> + if (busy_cpus > 1) {
> + before /= busy_cpus;
> + after /= busy_cpus;
However, I don't think this works when the sched group has more than one
busy CPU. 'before' and 'after' reflect the total throughput score of both
the sched group *and* the destination CPU.
One of the CPUs in the sched group will become idle after the balance.
Also, at this point we have already added score_on_dst_cpu. We are incorrectly
scaling it by the number of busy CPUs in the sched group.
We instead must scale 'after' by busy_cpus - 1 and then add score_on_dst_cpu.
> + }
>
> - sgs->task_class_score_after = group_score_without + score_on_dst_cpu;
> - sgs->task_class_score_before = group_score;
> + sgs->task_class_score_before = before;
> + sgs->task_class_score_after = after;
>
Thanks and BR,
Ricardo
>
next prev parent reply other threads:[~2022-10-26 3:50 UTC|newest]
Thread overview: 78+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-09-09 23:11 [RFC PATCH 00/23] sched: Introduce classes of tasks for load balance Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 01/23] sched/task_struct: Introduce classes of tasks Ricardo Neri
2022-09-14 13:46 ` Peter Zijlstra
2022-09-16 14:41 ` Ricardo Neri
2022-09-27 13:01 ` Peter Zijlstra
2022-10-02 22:32 ` Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 02/23] sched: Add interfaces for " Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 03/23] sched/core: Initialize the class of a new task Ricardo Neri
2022-09-26 14:57 ` Joel Fernandes
2022-09-26 21:53 ` Ricardo Neri
2022-09-27 13:04 ` Peter Zijlstra
2022-09-27 15:48 ` Joel Fernandes
2022-10-01 20:32 ` Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 04/23] sched/core: Add user_tick as argument to scheduler_tick() Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 05/23] sched/core: Move is_core_idle() out of fair.c Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 06/23] sched/core: Update the classification of the current task Ricardo Neri
2022-09-14 13:44 ` Peter Zijlstra
2022-09-16 14:42 ` Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 07/23] sched/fair: Collect load-balancing stats for task classes Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 08/23] sched/fair: Compute task-class performance scores for load balancing Ricardo Neri
2022-09-27 9:15 ` Peter Zijlstra
2022-10-26 3:57 ` Ricardo Neri [this message]
2022-10-26 8:55 ` Peter Zijlstra
2022-10-27 3:30 ` Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 09/23] sched/fair: Use task-class performance score to pick the busiest group Ricardo Neri
2022-09-27 11:01 ` Peter Zijlstra
2022-10-05 23:38 ` Ricardo Neri
2022-10-06 8:37 ` Peter Zijlstra
2022-10-06 19:07 ` Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 10/23] sched/fair: Use classes of tasks when selecting a busiest runqueue Ricardo Neri
2022-09-27 11:25 ` Peter Zijlstra
2022-10-07 23:36 ` Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 11/23] thermal: intel: hfi: Introduce Hardware Feedback Interface classes Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 12/23] thermal: intel: hfi: Convert table_lock to use flags-handling variants Ricardo Neri
2022-09-27 11:34 ` Peter Zijlstra
2022-09-27 11:36 ` Peter Zijlstra
2022-10-26 3:59 ` Ricardo Neri
2022-10-26 3:58 ` Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 13/23] x86/cpufeatures: Add the Intel Thread Director feature definitions Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 14/23] thermal: intel: hfi: Update the class of the current task Ricardo Neri
2022-09-27 11:46 ` Peter Zijlstra
2022-10-07 20:34 ` Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 15/23] thermal: intel: hfi: Report per-cpu class-specific performance scores Ricardo Neri
2022-09-27 11:59 ` Peter Zijlstra
2022-10-05 23:59 ` Ricardo Neri
2022-10-06 8:52 ` Peter Zijlstra
2022-10-06 9:14 ` Peter Zijlstra
2022-10-06 15:05 ` Brown, Len
2022-10-06 16:14 ` Peter Zijlstra
2022-10-07 11:20 ` Len Brown
2022-09-09 23:11 ` [RFC PATCH 16/23] thermal: intel: hfi: Define a default classification for unclassified tasks Ricardo Neri
2022-09-09 23:11 ` [RFC PATCH 17/23] thermal: intel: hfi: Enable the Intel Thread Director Ricardo Neri
2022-09-27 12:00 ` Peter Zijlstra
2022-10-06 1:50 ` Ricardo Neri
2022-09-09 23:12 ` [RFC PATCH 18/23] sched/task_struct: Add helpers for task classification Ricardo Neri
2022-09-27 11:52 ` Peter Zijlstra
2022-10-08 0:38 ` Ricardo Neri
2022-09-09 23:12 ` [RFC PATCH 19/23] sched/core: Initialize helpers of " Ricardo Neri
2022-09-09 23:12 ` [RFC PATCH 20/23] thermal: intel: hfi: Implement model-specific checks for " Ricardo Neri
2022-09-09 23:12 ` [RFC PATCH 21/23] x86/cpufeatures: Add feature bit for HRESET Ricardo Neri
2022-09-09 23:12 ` [RFC PATCH 22/23] x86/hreset: Configure history reset Ricardo Neri
2022-09-27 12:03 ` Peter Zijlstra
2022-10-02 22:34 ` Ricardo Neri
2022-09-09 23:12 ` [RFC PATCH 23/23] x86/process: Reset hardware history in context switch Ricardo Neri
2022-09-27 12:52 ` Peter Zijlstra
2022-10-03 23:07 ` Ricardo Neri
2022-10-06 8:35 ` Peter Zijlstra
2022-10-06 22:55 ` Ricardo Neri
2022-09-27 12:53 ` Peter Zijlstra
2022-10-02 22:02 ` Ricardo Neri
2022-09-27 13:15 ` Borislav Petkov
2022-10-02 22:12 ` Ricardo Neri
2022-10-02 22:15 ` Borislav Petkov
2022-10-03 19:49 ` Ricardo Neri
2022-10-03 19:55 ` Borislav Petkov
[not found] ` <20220910072120.2651-1-hdanton@sina.com>
2022-09-16 14:51 ` [RFC PATCH 06/23] sched/core: Update the classification of the current task Ricardo Neri
2022-10-11 19:12 ` Trying to apply patch set Carlos Bilbao
2022-10-18 2:31 ` Ricardo Neri
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20221026035724.GA21523@ranerica-svr.sc.intel.com \
--to=ricardo.neri-calderon@linux.intel.com \
--cc=bristot@redhat.com \
--cc=bsegall@google.com \
--cc=dietmar.eggemann@arm.com \
--cc=juri.lelli@redhat.com \
--cc=len.brown@intel.com \
--cc=linux-kernel@vger.kernel.org \
--cc=mgorman@suse.de \
--cc=peterz@infradead.org \
--cc=rafael.j.wysocki@intel.com \
--cc=ravi.v.shankar@intel.com \
--cc=ricardo.neri@intel.com \
--cc=rostedt@goodmis.org \
--cc=srinivas.pandruvada@linux.intel.com \
--cc=tim.c.chen@intel.com \
--cc=tim.c.chen@linux.intel.com \
--cc=vincent.guittot@linaro.org \
--cc=vschneid@redhat.com \
--cc=x86@kernel.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).