lkml.org 
[lkml]   [2022]   [Sep]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [RFC PATCH 08/23] sched/fair: Compute task-class performance scores for load balancing
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;
> +}

That's just plain broken; also lots of cleanups done...

---
--- 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;

- /*
- * 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;
+ }

- 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;
}

/**
@@ -8832,18 +8816,19 @@ static bool sched_asym_class_pick(struct
}

#else /* CONFIG_SCHED_TASK_CLASSES */
-static void update_rq_task_classes_stats(struct sg_lb_task_class_stats *class_sgs,
- struct rq *rq)
+
+static void init_sg_lb_task_class_stats(struct sg_lb_task_class_stats *sgcs)
{
}

-static void init_rq_task_classes_stats(struct sg_lb_task_class_stats *class_sgs)
+static void update_sg_lb_task_class_stats(struct sg_lb_task_class_stats *sgcs,
+ struct rq *rq)
{
}

-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)
{
}

@@ -8854,7 +8839,6 @@ static bool sched_asym_class_pick(struct
{
return false;
}
-
#endif /* CONFIG_SCHED_TASK_CLASSES */

/**
@@ -8979,11 +8963,11 @@ static inline void update_sg_lb_stats(st
struct sg_lb_stats *sgs,
int *sg_status)
{
- struct sg_lb_task_class_stats class_stats;
+ struct sg_lb_task_class_stats sgcs;
int i, nr_running, local_group;

memset(sgs, 0, sizeof(*sgs));
- init_rq_task_classes_stats(&class_stats);
+ init_sg_lb_task_class_stats(&sgcs);

local_group = group == sds->local;

@@ -9034,7 +9018,7 @@ static inline void update_sg_lb_stats(st
sgs->group_misfit_task_load = load;
}

- update_rq_task_classes_stats(&class_stats, rq);
+ update_sg_lb_task_class_stats(&sgcs, rq);
}

sgs->group_capacity = group->sgc->capacity;
@@ -9045,7 +9029,7 @@ static inline void update_sg_lb_stats(st
if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
sched_asym(env, sds, sgs, group)) {
- compute_ilb_sg_task_class_scores(&class_stats, sgs, env->dst_cpu);
+ update_sg_lb_stats_scores(&sgcs, sgs, env->dst_cpu);
sgs->group_asym_packing = 1;
}

\
 
 \ /
  Last update: 2022-09-27 11:18    [W:0.462 / U:0.200 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site