Skip to contentSkip to conversation

sched/core: Prioritize migrating eligible tasks in sched_balance_rq()

PATCH Applied Hao Jia 5 participants 16 messages Dec 23, 2024 – Feb 17, 2025

Subsystems (2)
lore
Share & tools
lore

Hao Jia

author
#0 lore raw

PATCH v2 1/1 sched/core: Prioritize migrating eligible tasks in sched_balance_rq()

From: Hao Jia <[email protected]>

When the PLACE_LAG scheduling feature is enabled and dst_cfs_rq->nr_queued is greater than 1, if a task is ineligible (lag < 0) on the source cpu runqueue, it will also be ineligible when it is migrated to the destination cpu runqueue. Because we will keep the original equivalent lag of the task in place_entity(). So if the task was ineligible before, it will still be ineligible after migration.

So in sched_balance_rq(), we prioritize migrating eligible tasks, and we soft-limit ineligible tasks, allowing them to migrate only when nr_balance_failed is non-zero to avoid load-balancing trying very hard to balance the load.

Below are some benchmark test results. From my test results, this patch shows a slight improvement on hackbench.

Benchmark
=========

All of the benchmarks are done inside a normal cpu cgroup in a clean environment with cpu turbo disabled, and test machine is:

Single NUMA machine model is 13th Gen Intel(R) Core(TM) i7-13700, 12 Core/24 HT.

Based on master b86545e02e8c.

Results
=======

hackbench-process-pipes
vanilla patched

Amean     1       0.5837 (   0.00%)      0.5733 (   1.77%)
Amean     4       1.4423 (   0.00%)      1.4503 (  -0.55%)
Amean     7       2.5147 (   0.00%)      2.4773 (   1.48%)
Amean     12      3.9347 (   0.00%)      3.8880 (   1.19%)
Amean     21      5.3943 (   0.00%)      5.3873 (   0.13%)
Amean     30      6.7840 (   0.00%)      6.6660 (   1.74%)
Amean     48      9.8313 (   0.00%)      9.6100 (   2.25%)
Amean     79     15.4403 (   0.00%)     14.9580 (   3.12%)
Amean     96     18.4970 (   0.00%)     17.9533 (   2.94%)

hackbench-process-sockets
vanilla patched

Amean     1       0.6297 (   0.00%)      0.6223 (   1.16%)
Amean     4       2.1517 (   0.00%)      2.0887 (   2.93%)
Amean     7       3.6377 (   0.00%)      3.5670 (   1.94%)
Amean     12      6.1277 (   0.00%)      5.9290 (   3.24%)
Amean     21     10.0380 (   0.00%)      9.7623 (   2.75%)
Amean     30     14.1517 (   0.00%)     13.7513 (   2.83%)
Amean     48     24.7253 (   0.00%)     24.2287 (   2.01%)
Amean     79     43.9523 (   0.00%)     43.2330 (   1.64%)
Amean     96     54.5310 (   0.00%)     53.7650 (   1.40%)

tbench4 Throughput
vanilla patched

Hmean     1       255.97 (   0.00%)      275.01 (   7.44%)
Hmean     2       511.60 (   0.00%)      544.27 (   6.39%)
Hmean     4       996.70 (   0.00%)     1006.57 (   0.99%)
Hmean     8      1646.46 (   0.00%)     1649.15 (   0.16%)
Hmean     16     2259.42 (   0.00%)     2274.35 (   0.66%)
Hmean     32     4725.48 (   0.00%)     4735.57 (   0.21%)
Hmean     64     4411.47 (   0.00%)     4400.05 (  -0.26%)
Hmean     96     4284.31 (   0.00%)     4267.39 (  -0.39%)

---
Previous discussion link: https://lore.kernel.org/all/[email protected]
Link to v1: https://lore.kernel.org/all/[email protected]

v1 to v2:
- Modify dst_cfs_rq->nr_running to dst_cfs_rq->nr_queued to resolve conflicts with commit 736c55a02c47 ("sched/fair:
Rename cfs_rq.nr_running into nr_queued").

 kernel/sched/fair.c | 34 ++++++++++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)
kernel/sched/fair.c +34 -0
index 5599b0c1ba9b..c884bf631e66 100644
@@ -9396,6 +9396,30 @@ static inline int migrate_degrades_locality(struct task_struct *p,
 }
 #endif
 
+/*
+ * Check whether the task is ineligible on the destination cpu
+ *
+ * When the PLACE_LAG scheduling feature is enabled and
+ * dst_cfs_rq->nr_queued is greater than 1, if the task
+ * is ineligible, it will also be ineligible when
+ * it is migrated to the destination cpu.
+ */
+static inline int task_is_ineligible_on_dst_cpu(struct task_struct *p, int dest_cpu)
+{
+	struct cfs_rq *dst_cfs_rq;
+
+#ifdef CONFIG_FAIR_GROUP_SCHED
+	dst_cfs_rq = task_group(p)->cfs_rq[dest_cpu];
+#else
+	dst_cfs_rq = &cpu_rq(dest_cpu)->cfs;
+#endif
+	if (sched_feat(PLACE_LAG) && dst_cfs_rq->nr_queued &&
+	    !entity_eligible(task_cfs_rq(p), &p->se))
+		return 1;
+
+	return 0;
+}
+
 /*
  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
  */
@@ -9420,6 +9444,16 @@ int can_migrate_task(struct task_struct *p, struct lb_env *env)
 	if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
 		return 0;
 
+	/*
+	 * We want to prioritize the migration of eligible tasks.
+	 * For ineligible tasks we soft-limit them and only allow
+	 * them to migrate when nr_balance_failed is non-zero to
+	 * avoid load-balancing trying very hard to balance the load.
+	 */
+	if (!env->sd->nr_balance_failed &&
+	    task_is_ineligible_on_dst_cpu(p, env->dst_cpu))
+		return 0;
+
 	/* Disregard percpu kthreads; they are where they need to be. */
 	if (kthread_is_per_cpu(p))
 		return 0;
Signature · 1 line
2.34.1
#1 lore raw

…

Quoted 2 lines

All of the benchmarks are done inside a normal cpu cgroup in a clean environment with cpu turbo disabled, and test machine is:

…
CPU?

You may occasionally put more than 63 characters into text lines of such a change description.

Regards,
Markus

author
#2 lore raw
Quoted 5 lines · Markus Elfring

…

Quoted 2 lines

All of the benchmarks are done inside a normal cpu cgroup in a clean environment with cpu turbo disabled, and test machine is:

…
CPU?

Thanks for your review, will fix it.

Quoted 3 lines

You may occasionally put more than 63 characters into text lines of such a change description.

I checked the patch using ./scripts/checkpatch.pl before sending it, and found no warnings or errors. The commit log should preferably have less than 75 characters per line. If I'm wrong, please correct me, thank you.

Thanks,
Hao

#3 lore raw
Quoted 4 lines
Quoted 2 lines

You may occasionally put more than 63 characters into text lines of such a change description.

I checked the patch using ./scripts/checkpatch.pl before sending it, and found no warnings or errors. The commit log should preferably have less than 75 characters per line.

Can any texts look nicer if word wrapping would be accordingly adjusted a bit more?

Regards,
Markus

author
#4 lore raw
In reply to Hao Jia

Friendly ping...

Quoted 134 lines · Hao Jia

From: Hao Jia <[email protected]>

When the PLACE_LAG scheduling feature is enabled and dst_cfs_rq->nr_queued is greater than 1, if a task is ineligible (lag < 0) on the source cpu runqueue, it will also be ineligible when it is migrated to the destination cpu runqueue. Because we will keep the original equivalent lag of the task in place_entity(). So if the task was ineligible before, it will still be ineligible after migration.

So in sched_balance_rq(), we prioritize migrating eligible tasks, and we soft-limit ineligible tasks, allowing them to migrate only when nr_balance_failed is non-zero to avoid load-balancing trying very hard to balance the load.

Below are some benchmark test results. From my test results, this patch shows a slight improvement on hackbench.

Benchmark
=========

All of the benchmarks are done inside a normal cpu cgroup in a clean environment with cpu turbo disabled, and test machine is:

Single NUMA machine model is 13th Gen Intel(R) Core(TM) i7-13700, 12 Core/24 HT.

Based on master b86545e02e8c.

Results
=======

hackbench-process-pipes
vanilla patched

Amean     1       0.5837 (   0.00%)      0.5733 (   1.77%)
Amean     4       1.4423 (   0.00%)      1.4503 (  -0.55%)
Amean     7       2.5147 (   0.00%)      2.4773 (   1.48%)
Amean     12      3.9347 (   0.00%)      3.8880 (   1.19%)
Amean     21      5.3943 (   0.00%)      5.3873 (   0.13%)
Amean     30      6.7840 (   0.00%)      6.6660 (   1.74%)
Amean     48      9.8313 (   0.00%)      9.6100 (   2.25%)
Amean     79     15.4403 (   0.00%)     14.9580 (   3.12%)
Amean     96     18.4970 (   0.00%)     17.9533 (   2.94%)

hackbench-process-sockets
vanilla patched

Amean     1       0.6297 (   0.00%)      0.6223 (   1.16%)
Amean     4       2.1517 (   0.00%)      2.0887 (   2.93%)
Amean     7       3.6377 (   0.00%)      3.5670 (   1.94%)
Amean     12      6.1277 (   0.00%)      5.9290 (   3.24%)
Amean     21     10.0380 (   0.00%)      9.7623 (   2.75%)
Amean     30     14.1517 (   0.00%)     13.7513 (   2.83%)
Amean     48     24.7253 (   0.00%)     24.2287 (   2.01%)
Amean     79     43.9523 (   0.00%)     43.2330 (   1.64%)
Amean     96     54.5310 (   0.00%)     53.7650 (   1.40%)

tbench4 Throughput
vanilla patched

Hmean     1       255.97 (   0.00%)      275.01 (   7.44%)
Hmean     2       511.60 (   0.00%)      544.27 (   6.39%)
Hmean     4       996.70 (   0.00%)     1006.57 (   0.99%)
Hmean     8      1646.46 (   0.00%)     1649.15 (   0.16%)
Hmean     16     2259.42 (   0.00%)     2274.35 (   0.66%)
Hmean     32     4725.48 (   0.00%)     4735.57 (   0.21%)
Hmean     64     4411.47 (   0.00%)     4400.05 (  -0.26%)
Hmean     96     4284.31 (   0.00%)     4267.39 (  -0.39%)

---
Previous discussion link: https://lore.kernel.org/all/[email protected]
Link to v1: https://lore.kernel.org/all/[email protected]

v1 to v2:
- Modify dst_cfs_rq->nr_running to dst_cfs_rq->nr_queued to resolve conflicts with commit 736c55a02c47 ("sched/fair:
Rename cfs_rq.nr_running into nr_queued").

  kernel/sched/fair.c | 34 ++++++++++++++++++++++++++++++++++
  1 file changed, 34 insertions(+)
kernel/sched/fair.c +34 -0
index 5599b0c1ba9b..c884bf631e66 100644
@@ -9396,6 +9396,30 @@ static inline int migrate_degrades_locality(struct task_struct *p,
  }
  #endif
 
+/*
+ * Check whether the task is ineligible on the destination cpu
+ *
+ * When the PLACE_LAG scheduling feature is enabled and
+ * dst_cfs_rq->nr_queued is greater than 1, if the task
+ * is ineligible, it will also be ineligible when
+ * it is migrated to the destination cpu.
+ */
+static inline int task_is_ineligible_on_dst_cpu(struct task_struct *p, int dest_cpu)
+{
+	struct cfs_rq *dst_cfs_rq;
+
+#ifdef CONFIG_FAIR_GROUP_SCHED
+	dst_cfs_rq = task_group(p)->cfs_rq[dest_cpu];
+#else
+	dst_cfs_rq = &cpu_rq(dest_cpu)->cfs;
+#endif
+	if (sched_feat(PLACE_LAG) && dst_cfs_rq->nr_queued &&
+	    !entity_eligible(task_cfs_rq(p), &p->se))
+		return 1;
+
+	return 0;
+}
+
  /*
   * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
   */
@@ -9420,6 +9444,16 @@ int can_migrate_task(struct task_struct *p, struct lb_env *env)
  	if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
  		return 0;
 
+	/*
+	 * We want to prioritize the migration of eligible tasks.
+	 * For ineligible tasks we soft-limit them and only allow
+	 * them to migrate when nr_balance_failed is non-zero to
+	 * avoid load-balancing trying very hard to balance the load.
+	 */
+	if (!env->sd->nr_balance_failed &&
+	    task_is_ineligible_on_dst_cpu(p, env->dst_cpu))
+		return 0;
+
  	/* Disregard percpu kthreads; they are where they need to be. */
  	if (kthread_is_per_cpu(p))
  		return 0;
maintainer
#5 lore raw
Quoted 20 lines · Hao Jia

Friendly ping...

Quoted 15 lines · Hao Jia

From: Hao Jia <[email protected]>

When the PLACE_LAG scheduling feature is enabled and dst_cfs_rq->nr_queued is greater than 1, if a task is ineligible (lag < 0) on the source cpu runqueue, it will also be ineligible when it is migrated to the destination cpu runqueue. Because we will keep the original equivalent lag of the task in place_entity(). So if the task was ineligible before, it will still be ineligible after migration.

So in sched_balance_rq(), we prioritize migrating eligible tasks, and we soft-limit ineligible tasks, allowing them to migrate only when nr_balance_failed is non-zero to avoid load-balancing trying very hard to balance the load.

Could you explain why you think it's better to balance eligible tasks in priority and potentially skip a load balance ?

I can see an interest for idle and newly_idle load balance in order to favor fairness as tasks will become eligible but I don't see why it would be helpful if dst already has some runnable tasks. Furthermore, when a cpu is idle or newly idle, we really want to migrate a task even an non eligible one instead of possibly skipping this load balance round. With your patch, we might end up not pulling any task, increasing the nr_balance_failed and waiting next load balance

Quoted 119 lines
Quoted 119 lines

Below are some benchmark test results. From my test results, this patch shows a slight improvement on hackbench.

Benchmark
=========

All of the benchmarks are done inside a normal cpu cgroup in a clean environment with cpu turbo disabled, and test machine is:

Single NUMA machine model is 13th Gen Intel(R) Core(TM) i7-13700, 12 Core/24 HT.

Based on master b86545e02e8c.

Results
=======

hackbench-process-pipes
vanilla patched

Amean     1       0.5837 (   0.00%)      0.5733 (   1.77%)
Amean     4       1.4423 (   0.00%)      1.4503 (  -0.55%)
Amean     7       2.5147 (   0.00%)      2.4773 (   1.48%)
Amean     12      3.9347 (   0.00%)      3.8880 (   1.19%)
Amean     21      5.3943 (   0.00%)      5.3873 (   0.13%)
Amean     30      6.7840 (   0.00%)      6.6660 (   1.74%)
Amean     48      9.8313 (   0.00%)      9.6100 (   2.25%)
Amean     79     15.4403 (   0.00%)     14.9580 (   3.12%)
Amean     96     18.4970 (   0.00%)     17.9533 (   2.94%)

hackbench-process-sockets
vanilla patched

Amean     1       0.6297 (   0.00%)      0.6223 (   1.16%)
Amean     4       2.1517 (   0.00%)      2.0887 (   2.93%)
Amean     7       3.6377 (   0.00%)      3.5670 (   1.94%)
Amean     12      6.1277 (   0.00%)      5.9290 (   3.24%)
Amean     21     10.0380 (   0.00%)      9.7623 (   2.75%)
Amean     30     14.1517 (   0.00%)     13.7513 (   2.83%)
Amean     48     24.7253 (   0.00%)     24.2287 (   2.01%)
Amean     79     43.9523 (   0.00%)     43.2330 (   1.64%)
Amean     96     54.5310 (   0.00%)     53.7650 (   1.40%)

tbench4 Throughput
vanilla patched

Hmean     1       255.97 (   0.00%)      275.01 (   7.44%)
Hmean     2       511.60 (   0.00%)      544.27 (   6.39%)
Hmean     4       996.70 (   0.00%)     1006.57 (   0.99%)
Hmean     8      1646.46 (   0.00%)     1649.15 (   0.16%)
Hmean     16     2259.42 (   0.00%)     2274.35 (   0.66%)
Hmean     32     4725.48 (   0.00%)     4735.57 (   0.21%)
Hmean     64     4411.47 (   0.00%)     4400.05 (  -0.26%)
Hmean     96     4284.31 (   0.00%)     4267.39 (  -0.39%)

---
Previous discussion link: https://lore.kernel.org/all/[email protected]
Link to v1: https://lore.kernel.org/all/[email protected]

v1 to v2:
- Modify dst_cfs_rq->nr_running to dst_cfs_rq->nr_queued to resolve conflicts with commit 736c55a02c47 ("sched/fair:
Rename cfs_rq.nr_running into nr_queued").

  kernel/sched/fair.c | 34 ++++++++++++++++++++++++++++++++++
  1 file changed, 34 insertions(+)
kernel/sched/fair.c +34 -0
index 5599b0c1ba9b..c884bf631e66 100644
@@ -9396,6 +9396,30 @@ static inline int migrate_degrades_locality(struct task_struct *p,
  }
  #endif
 
+/*
+ * Check whether the task is ineligible on the destination cpu
+ *
+ * When the PLACE_LAG scheduling feature is enabled and
+ * dst_cfs_rq->nr_queued is greater than 1, if the task
+ * is ineligible, it will also be ineligible when
+ * it is migrated to the destination cpu.
+ */
+static inline int task_is_ineligible_on_dst_cpu(struct task_struct *p, int dest_cpu)
+{
+     struct cfs_rq *dst_cfs_rq;
+
+#ifdef CONFIG_FAIR_GROUP_SCHED
+     dst_cfs_rq = task_group(p)->cfs_rq[dest_cpu];
+#else
+     dst_cfs_rq = &cpu_rq(dest_cpu)->cfs;
+#endif
+     if (sched_feat(PLACE_LAG) && dst_cfs_rq->nr_queued &&
+         !entity_eligible(task_cfs_rq(p), &p->se))
+             return 1;
+
+     return 0;
+}
+
  /*
   * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
   */
@@ -9420,6 +9444,16 @@ int can_migrate_task(struct task_struct *p, struct lb_env *env)
      if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
              return 0;
 
+     /*
+      * We want to prioritize the migration of eligible tasks.
+      * For ineligible tasks we soft-limit them and only allow
+      * them to migrate when nr_balance_failed is non-zero to
+      * avoid load-balancing trying very hard to balance the load.
+      */
+     if (!env->sd->nr_balance_failed &&
+         task_is_ineligible_on_dst_cpu(p, env->dst_cpu))
+             return 0;
+
      /* Disregard percpu kthreads; they are where they need to be. */
      if (kthread_is_per_cpu(p))
              return 0;
author
#6 lore raw
Quoted 17 lines · Vincent Guittot
Quoted 16 lines · Hao Jia

Friendly ping...

Quoted 11 lines · Hao Jia

From: Hao Jia <[email protected]>

When the PLACE_LAG scheduling feature is enabled and dst_cfs_rq->nr_queued is greater than 1, if a task is ineligible (lag < 0) on the source cpu runqueue, it will also be ineligible when it is migrated to the destination cpu runqueue. Because we will keep the original equivalent lag of the task in place_entity(). So if the task was ineligible before, it will still be ineligible after migration.

Quoted 7 lines
Quoted 4 lines
Quoted 4 lines

So in sched_balance_rq(), we prioritize migrating eligible tasks, and we soft-limit ineligible tasks, allowing them to migrate only when nr_balance_failed is non-zero to avoid load-balancing trying very hard to balance the load.

Could you explain why you think it's better to balance eligible tasks in priority and potentially skip a load balance ?

In place_entity(), we maintain the task's original equivalent lag, even if we migrate the task to dst_rq, this does not change its eligibility attribute.

When there are multiple tasks on src_rq, and the dst_cpu has some runnable tasks, migrating ineligible tasks to dst_rq will not allow them to run. Therefore, such task migration is inefficient. We should prioritize migrating tasks that can run on dst_rq.

In other words, migrating ineligible tasks is merely moving them to another runqueue to wait until they become eligible.

Quoted 9 lines

I can see an interest for idle and newly_idle load balance in order to favor fairness as tasks will become eligible but I don't see why it would be helpful if dst already has some runnable tasks. Furthermore, when a cpu is idle or newly idle, we really want to migrate a task even an non eligible one instead of possibly skipping this load balance round. With your patch, we might end up not pulling any task, increasing the nr_balance_failed and waiting next load balance

If I understand correctly, when the destination CPU is idle, my patch does not change the original behavior. it only prevents the migration of ineligible tasks when dst_cfs_rq->nr_queued is greater than 1.

If I missed something, please correct me.

Thanks,
Hao

Quoted 119 lines
Quoted 119 lines
Quoted 119 lines

Below are some benchmark test results. From my test results, this patch shows a slight improvement on hackbench.

Benchmark
=========

All of the benchmarks are done inside a normal cpu cgroup in a clean environment with cpu turbo disabled, and test machine is:

Single NUMA machine model is 13th Gen Intel(R) Core(TM) i7-13700, 12 Core/24 HT.

Based on master b86545e02e8c.

Results
=======

hackbench-process-pipes
vanilla patched

Amean     1       0.5837 (   0.00%)      0.5733 (   1.77%)
Amean     4       1.4423 (   0.00%)      1.4503 (  -0.55%)
Amean     7       2.5147 (   0.00%)      2.4773 (   1.48%)
Amean     12      3.9347 (   0.00%)      3.8880 (   1.19%)
Amean     21      5.3943 (   0.00%)      5.3873 (   0.13%)
Amean     30      6.7840 (   0.00%)      6.6660 (   1.74%)
Amean     48      9.8313 (   0.00%)      9.6100 (   2.25%)
Amean     79     15.4403 (   0.00%)     14.9580 (   3.12%)
Amean     96     18.4970 (   0.00%)     17.9533 (   2.94%)

hackbench-process-sockets
vanilla patched

Amean     1       0.6297 (   0.00%)      0.6223 (   1.16%)
Amean     4       2.1517 (   0.00%)      2.0887 (   2.93%)
Amean     7       3.6377 (   0.00%)      3.5670 (   1.94%)
Amean     12      6.1277 (   0.00%)      5.9290 (   3.24%)
Amean     21     10.0380 (   0.00%)      9.7623 (   2.75%)
Amean     30     14.1517 (   0.00%)     13.7513 (   2.83%)
Amean     48     24.7253 (   0.00%)     24.2287 (   2.01%)
Amean     79     43.9523 (   0.00%)     43.2330 (   1.64%)
Amean     96     54.5310 (   0.00%)     53.7650 (   1.40%)

tbench4 Throughput
vanilla patched

Hmean     1       255.97 (   0.00%)      275.01 (   7.44%)
Hmean     2       511.60 (   0.00%)      544.27 (   6.39%)
Hmean     4       996.70 (   0.00%)     1006.57 (   0.99%)
Hmean     8      1646.46 (   0.00%)     1649.15 (   0.16%)
Hmean     16     2259.42 (   0.00%)     2274.35 (   0.66%)
Hmean     32     4725.48 (   0.00%)     4735.57 (   0.21%)
Hmean     64     4411.47 (   0.00%)     4400.05 (  -0.26%)
Hmean     96     4284.31 (   0.00%)     4267.39 (  -0.39%)

---
Previous discussion link: https://lore.kernel.org/all/[email protected]
Link to v1: https://lore.kernel.org/all/[email protected]

v1 to v2:
- Modify dst_cfs_rq->nr_running to dst_cfs_rq->nr_queued to resolve conflicts with commit 736c55a02c47 ("sched/fair:
Rename cfs_rq.nr_running into nr_queued").

   kernel/sched/fair.c | 34 ++++++++++++++++++++++++++++++++++
   1 file changed, 34 insertions(+)
kernel/sched/fair.c +34 -0
index 5599b0c1ba9b..c884bf631e66 100644
@@ -9396,6 +9396,30 @@ static inline int migrate_degrades_locality(struct task_struct *p,
   }
   #endif
 
+/*
+ * Check whether the task is ineligible on the destination cpu
+ *
+ * When the PLACE_LAG scheduling feature is enabled and
+ * dst_cfs_rq->nr_queued is greater than 1, if the task
+ * is ineligible, it will also be ineligible when
+ * it is migrated to the destination cpu.
+ */
+static inline int task_is_ineligible_on_dst_cpu(struct task_struct *p, int dest_cpu)
+{
+     struct cfs_rq *dst_cfs_rq;
+
+#ifdef CONFIG_FAIR_GROUP_SCHED
+     dst_cfs_rq = task_group(p)->cfs_rq[dest_cpu];
+#else
+     dst_cfs_rq = &cpu_rq(dest_cpu)->cfs;
+#endif
+     if (sched_feat(PLACE_LAG) && dst_cfs_rq->nr_queued &&
+         !entity_eligible(task_cfs_rq(p), &p->se))
+             return 1;
+
+     return 0;
+}
+
   /*
    * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
    */
@@ -9420,6 +9444,16 @@ int can_migrate_task(struct task_struct *p, struct lb_env *env)
       if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
               return 0;
 
+     /*
+      * We want to prioritize the migration of eligible tasks.
+      * For ineligible tasks we soft-limit them and only allow
+      * them to migrate when nr_balance_failed is non-zero to
+      * avoid load-balancing trying very hard to balance the load.
+      */
+     if (!env->sd->nr_balance_failed &&
+         task_is_ineligible_on_dst_cpu(p, env->dst_cpu))
+             return 0;
+
       /* Disregard percpu kthreads; they are where they need to be. */
       if (kthread_is_per_cpu(p))
               return 0;
maintainer
#7 lore raw
Quoted 26 lines · Hao Jia
Quoted 22 lines · Vincent Guittot
Quoted 20 lines · Hao Jia

Friendly ping...

Quoted 15 lines · Hao Jia

From: Hao Jia <[email protected]>

When the PLACE_LAG scheduling feature is enabled and dst_cfs_rq->nr_queued is greater than 1, if a task is ineligible (lag < 0) on the source cpu runqueue, it will also be ineligible when it is migrated to the destination cpu runqueue. Because we will keep the original equivalent lag of the task in place_entity(). So if the task was ineligible before, it will still be ineligible after migration.

So in sched_balance_rq(), we prioritize migrating eligible tasks, and we soft-limit ineligible tasks, allowing them to migrate only when nr_balance_failed is non-zero to avoid load-balancing trying very hard to balance the load.

Quoted 6 lines
Quoted 2 lines

Could you explain why you think it's better to balance eligible tasks in priority and potentially skip a load balance ?

In place_entity(), we maintain the task's original equivalent lag, even if we migrate the task to dst_rq, this does not change its eligibility attribute.

Yes, but you don't answer the question why it's better to select an eligible task vs a non eligible task.

Quoted 4 lines

When there are multiple tasks on src_rq, and the dst_cpu has some runnable tasks, migrating ineligible tasks to dst_rq will not allow them to run. Therefore, such task migration is inefficient. We should

Why is it inefficient ? load balance is about evenly balancing the number of tasks or the load between CPUs, it never says that the newly migrated task should run immediately

Quoted 4 lines

prioritize migrating tasks that can run on dst_rq.

In other words, migrating ineligible tasks is merely moving them to another runqueue to wait until they become eligible.

But I don't get why it's a problem. Migrating an eligible task might delay its scheduling because of its deadline vs other tasks already eligible on the dst_rq. Eligible and non eligible tasks are all runnable, it's just how much they have already run. In addition, migrating an eligible task will clear its positive vlag with DELAY_ZERO which is unfair IMO

Quoted 12 lines
Quoted 11 lines

I can see an interest for idle and newly_idle load balance in order to favor fairness as tasks will become eligible but I don't see why it would be helpful if dst already has some runnable tasks. Furthermore, when a cpu is idle or newly idle, we really want to migrate a task even an non eligible one instead of possibly skipping this load balance round. With your patch, we might end up not pulling any task, increasing the nr_balance_failed and waiting next load balance

Quoted 3 lines

If I understand correctly, when the destination CPU is idle, my patch does not change the original behavior. it only prevents the migration of ineligible tasks when dst_cfs_rq->nr_queued is greater than 1.

It changes the behavior. My concern is that migrating an eligible task when dst rq already has runnable tasks, doesn't assure you that it will give any advantage to this eligible task.

On an idle or newly idle cpu, any runnable tasks that will be pulled will immediately start to run whatever it was eligible or not on src cpu. In such a situation, we could consider that selecting an eligible task which has less running time (positive lag) than others could be more fair because the eligible task will immediately run. But this is true as long as we migrate only 1 task. If you migrate several tasks, an eligible task on src rq could even become ineligible quicker on dst cpu than on src cpu has it lost its lag

Quoted 127 lines

If I missed something, please correct me.

Thanks,
Hao

Quoted 119 lines
Quoted 119 lines
Quoted 119 lines

Below are some benchmark test results. From my test results, this patch shows a slight improvement on hackbench.

Benchmark
=========

All of the benchmarks are done inside a normal cpu cgroup in a clean environment with cpu turbo disabled, and test machine is:

Single NUMA machine model is 13th Gen Intel(R) Core(TM) i7-13700, 12 Core/24 HT.

Based on master b86545e02e8c.

Results
=======

hackbench-process-pipes
vanilla patched

Amean     1       0.5837 (   0.00%)      0.5733 (   1.77%)
Amean     4       1.4423 (   0.00%)      1.4503 (  -0.55%)
Amean     7       2.5147 (   0.00%)      2.4773 (   1.48%)
Amean     12      3.9347 (   0.00%)      3.8880 (   1.19%)
Amean     21      5.3943 (   0.00%)      5.3873 (   0.13%)
Amean     30      6.7840 (   0.00%)      6.6660 (   1.74%)
Amean     48      9.8313 (   0.00%)      9.6100 (   2.25%)
Amean     79     15.4403 (   0.00%)     14.9580 (   3.12%)
Amean     96     18.4970 (   0.00%)     17.9533 (   2.94%)

hackbench-process-sockets
vanilla patched

Amean     1       0.6297 (   0.00%)      0.6223 (   1.16%)
Amean     4       2.1517 (   0.00%)      2.0887 (   2.93%)
Amean     7       3.6377 (   0.00%)      3.5670 (   1.94%)
Amean     12      6.1277 (   0.00%)      5.9290 (   3.24%)
Amean     21     10.0380 (   0.00%)      9.7623 (   2.75%)
Amean     30     14.1517 (   0.00%)     13.7513 (   2.83%)
Amean     48     24.7253 (   0.00%)     24.2287 (   2.01%)
Amean     79     43.9523 (   0.00%)     43.2330 (   1.64%)
Amean     96     54.5310 (   0.00%)     53.7650 (   1.40%)

tbench4 Throughput
vanilla patched

Hmean     1       255.97 (   0.00%)      275.01 (   7.44%)
Hmean     2       511.60 (   0.00%)      544.27 (   6.39%)
Hmean     4       996.70 (   0.00%)     1006.57 (   0.99%)
Hmean     8      1646.46 (   0.00%)     1649.15 (   0.16%)
Hmean     16     2259.42 (   0.00%)     2274.35 (   0.66%)
Hmean     32     4725.48 (   0.00%)     4735.57 (   0.21%)
Hmean     64     4411.47 (   0.00%)     4400.05 (  -0.26%)
Hmean     96     4284.31 (   0.00%)     4267.39 (  -0.39%)

---
Previous discussion link: https://lore.kernel.org/all/[email protected]
Link to v1: https://lore.kernel.org/all/[email protected]

v1 to v2:
- Modify dst_cfs_rq->nr_running to dst_cfs_rq->nr_queued to resolve conflicts with commit 736c55a02c47 ("sched/fair:
Rename cfs_rq.nr_running into nr_queued").

   kernel/sched/fair.c | 34 ++++++++++++++++++++++++++++++++++
   1 file changed, 34 insertions(+)
kernel/sched/fair.c +34 -0
index 5599b0c1ba9b..c884bf631e66 100644
@@ -9396,6 +9396,30 @@ static inline int migrate_degrades_locality(struct task_struct *p,
   }
   #endif
 
+/*
+ * Check whether the task is ineligible on the destination cpu
+ *
+ * When the PLACE_LAG scheduling feature is enabled and
+ * dst_cfs_rq->nr_queued is greater than 1, if the task
+ * is ineligible, it will also be ineligible when
+ * it is migrated to the destination cpu.
+ */
+static inline int task_is_ineligible_on_dst_cpu(struct task_struct *p, int dest_cpu)
+{
+     struct cfs_rq *dst_cfs_rq;
+
+#ifdef CONFIG_FAIR_GROUP_SCHED
+     dst_cfs_rq = task_group(p)->cfs_rq[dest_cpu];
+#else
+     dst_cfs_rq = &cpu_rq(dest_cpu)->cfs;
+#endif
+     if (sched_feat(PLACE_LAG) && dst_cfs_rq->nr_queued &&
+         !entity_eligible(task_cfs_rq(p), &p->se))
+             return 1;
+
+     return 0;
+}
+
   /*
    * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
    */
@@ -9420,6 +9444,16 @@ int can_migrate_task(struct task_struct *p, struct lb_env *env)
       if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
               return 0;
 
+     /*
+      * We want to prioritize the migration of eligible tasks.
+      * For ineligible tasks we soft-limit them and only allow
+      * them to migrate when nr_balance_failed is non-zero to
+      * avoid load-balancing trying very hard to balance the load.
+      */
+     if (!env->sd->nr_balance_failed &&
+         task_is_ineligible_on_dst_cpu(p, env->dst_cpu))
+             return 0;
+
       /* Disregard percpu kthreads; they are where they need to be. */
       if (kthread_is_per_cpu(p))
               return 0;
author
#8 lore raw
Quoted 38 lines · Vincent Guittot
Quoted 26 lines · Hao Jia
Quoted 22 lines · Vincent Guittot
Quoted 20 lines · Hao Jia

Friendly ping...

Quoted 15 lines · Hao Jia

From: Hao Jia <[email protected]>

When the PLACE_LAG scheduling feature is enabled and dst_cfs_rq->nr_queued is greater than 1, if a task is ineligible (lag < 0) on the source cpu runqueue, it will also be ineligible when it is migrated to the destination cpu runqueue. Because we will keep the original equivalent lag of the task in place_entity(). So if the task was ineligible before, it will still be ineligible after migration.

So in sched_balance_rq(), we prioritize migrating eligible tasks, and we soft-limit ineligible tasks, allowing them to migrate only when nr_balance_failed is non-zero to avoid load-balancing trying very hard to balance the load.

Quoted 6 lines
Quoted 2 lines

Could you explain why you think it's better to balance eligible tasks in priority and potentially skip a load balance ?

In place_entity(), we maintain the task's original equivalent lag, even if we migrate the task to dst_rq, this does not change its eligibility attribute.

Yes, but you don't answer the question why it's better to select an eligible task vs a non eligible task.

Quoted 1 line
Quoted 7 lines
Quoted 3 lines

When there are multiple tasks on src_rq, and the dst_cpu has some runnable tasks, migrating ineligible tasks to dst_rq will not allow them to run. Therefore, such task migration is inefficient. We should

Why is it inefficient ? load balance is about evenly balancing the number of tasks or the load between CPUs, it never says that the newly migrated task should run immediately

My initial thought is that when we need to migrate some tasks during load balancing, at the current point in time, migrating ineligible tasks to dst_cpu means they definitely cannot run there. Therefore, I prefer to keep them on src_cpu to reduce the overhead of dequeueing and enqueueing ineligible tasks.

Migrating eligible tasks to dst_cpu does not guarantee that they will run earlier than on src_cpu. it depends on too many factors.

Quoted 6 lines
Quoted 5 lines

prioritize migrating tasks that can run on dst_rq.

In other words, migrating ineligible tasks is merely moving them to another runqueue to wait until they become eligible.

Quoted 6 lines

But I don't get why it's a problem. Migrating an eligible task might delay its scheduling because of its deadline vs other tasks already eligible on the dst_rq. Eligible and non eligible tasks are all runnable, it's just how much they have already run. In addition, migrating an eligible task will clear its positive vlag with DELAY_ZERO which is unfair IMO

Sorry, I'd like to ask you a question that confuses me: Why does migrating eligible task will clear the positive vlag?

In detach_task(), the ENQUEUE_DELAYED and DEQUEUE_SLEEP flags are not set, and in dequeue_entity(), eligible tasks will not set sched_delayed, so they will be dequeued normally with se->on_rq being 0.

Similarly, attach_task() does not set the ENQUEUE_DELAYED and DEQUEUE_SLEEP flags, and since se->on_rq is 0, it will not call requeue_delayed_entity().

Thanks,
Hao

Quoted 30 lines
Quoted 13 lines
Quoted 4 lines
Quoted 8 lines

I can see an interest for idle and newly_idle load balance in order to favor fairness as tasks will become eligible but I don't see why it would be helpful if dst already has some runnable tasks. Furthermore, when a cpu is idle or newly idle, we really want to migrate a task even an non eligible one instead of possibly skipping this load balance round. With your patch, we might end up not pulling any task, increasing the nr_balance_failed and waiting next load balance

Quoted 3 lines

If I understand correctly, when the destination CPU is idle, my patch does not change the original behavior. it only prevents the migration of ineligible tasks when dst_cfs_rq->nr_queued is greater than 1.

It changes the behavior. My concern is that migrating an eligible task when dst rq already has runnable tasks, doesn't assure you that it will give any advantage to this eligible task.

On an idle or newly idle cpu, any runnable tasks that will be pulled will immediately start to run whatever it was eligible or not on src cpu. In such a situation, we could consider that selecting an eligible task which has less running time (positive lag) than others could be more fair because the eligible task will immediately run. But this is true as long as we migrate only 1 task. If you migrate several tasks, an eligible task on src rq could even become ineligible quicker on dst cpu than on src cpu has it lost its lag

maintainer
#9 lore raw
Quoted 51 lines · Hao Jia
Quoted 38 lines · Vincent Guittot
Quoted 26 lines · Hao Jia
Quoted 22 lines · Vincent Guittot
Quoted 20 lines · Hao Jia

Friendly ping...

Quoted 15 lines · Hao Jia

From: Hao Jia <[email protected]>

When the PLACE_LAG scheduling feature is enabled and dst_cfs_rq->nr_queued is greater than 1, if a task is ineligible (lag < 0) on the source cpu runqueue, it will also be ineligible when it is migrated to the destination cpu runqueue. Because we will keep the original equivalent lag of the task in place_entity(). So if the task was ineligible before, it will still be ineligible after migration.

So in sched_balance_rq(), we prioritize migrating eligible tasks, and we soft-limit ineligible tasks, allowing them to migrate only when nr_balance_failed is non-zero to avoid load-balancing trying very hard to balance the load.

Quoted 6 lines
Quoted 2 lines

Could you explain why you think it's better to balance eligible tasks in priority and potentially skip a load balance ?

In place_entity(), we maintain the task's original equivalent lag, even if we migrate the task to dst_rq, this does not change its eligibility attribute.

Yes, but you don't answer the question why it's better to select an eligible task vs a non eligible task.

Quoted 1 line
Quoted 7 lines
Quoted 3 lines

When there are multiple tasks on src_rq, and the dst_cpu has some runnable tasks, migrating ineligible tasks to dst_rq will not allow them to run. Therefore, such task migration is inefficient. We should

Why is it inefficient ? load balance is about evenly balancing the number of tasks or the load between CPUs, it never says that the newly migrated task should run immediately

Quoted 5 lines

My initial thought is that when we need to migrate some tasks during load balancing, at the current point in time, migrating ineligible tasks to dst_cpu means they definitely cannot run there. Therefore, I prefer to keep them on src_cpu to reduce the overhead of dequeueing and enqueueing ineligible tasks.

Sorry but I still don't get why it's important and would make a difference. They are all runnable but ineligible tasks got more runtime than other at that point in time so there is no real difference

Quoted 20 lines

Migrating eligible tasks to dst_cpu does not guarantee that they will run earlier than on src_cpu. it depends on too many factors.

Quoted 6 lines
Quoted 5 lines

prioritize migrating tasks that can run on dst_rq.

In other words, migrating ineligible tasks is merely moving them to another runqueue to wait until they become eligible.

Quoted 6 lines

But I don't get why it's a problem. Migrating an eligible task might delay its scheduling because of its deadline vs other tasks already eligible on the dst_rq. Eligible and non eligible tasks are all runnable, it's just how much they have already run. In addition, migrating an eligible task will clear its positive vlag with DELAY_ZERO which is unfair IMO

Quoted 2 lines

Sorry, I'd like to ask you a question that confuses me: Why does migrating eligible task will clear the positive vlag?

sorry I mess up everything that only for delayed dequeue task

Quoted 43 lines

In detach_task(), the ENQUEUE_DELAYED and DEQUEUE_SLEEP flags are not set, and in dequeue_entity(), eligible tasks will not set sched_delayed, so they will be dequeued normally with se->on_rq being 0.

Similarly, attach_task() does not set the ENQUEUE_DELAYED and DEQUEUE_SLEEP flags, and since se->on_rq is 0, it will not call requeue_delayed_entity().

Thanks,
Hao

Quoted 30 lines
Quoted 13 lines
Quoted 4 lines
Quoted 8 lines

I can see an interest for idle and newly_idle load balance in order to favor fairness as tasks will become eligible but I don't see why it would be helpful if dst already has some runnable tasks. Furthermore, when a cpu is idle or newly idle, we really want to migrate a task even an non eligible one instead of possibly skipping this load balance round. With your patch, we might end up not pulling any task, increasing the nr_balance_failed and waiting next load balance

Quoted 3 lines

If I understand correctly, when the destination CPU is idle, my patch does not change the original behavior. it only prevents the migration of ineligible tasks when dst_cfs_rq->nr_queued is greater than 1.

It changes the behavior. My concern is that migrating an eligible task when dst rq already has runnable tasks, doesn't assure you that it will give any advantage to this eligible task.

On an idle or newly idle cpu, any runnable tasks that will be pulled will immediately start to run whatever it was eligible or not on src cpu. In such a situation, we could consider that selecting an eligible task which has less running time (positive lag) than others could be more fair because the eligible task will immediately run. But this is true as long as we migrate only 1 task. If you migrate several tasks, an eligible task on src rq could even become ineligible quicker on dst cpu than on src cpu has it lost its lag

author
#10 lore raw
Quoted 58 lines · Vincent Guittot
Quoted 51 lines · Hao Jia
Quoted 38 lines · Vincent Guittot
Quoted 26 lines · Hao Jia
Quoted 22 lines · Vincent Guittot
Quoted 20 lines · Hao Jia

Friendly ping...

Quoted 15 lines · Hao Jia

From: Hao Jia <[email protected]>

When the PLACE_LAG scheduling feature is enabled and dst_cfs_rq->nr_queued is greater than 1, if a task is ineligible (lag < 0) on the source cpu runqueue, it will also be ineligible when it is migrated to the destination cpu runqueue. Because we will keep the original equivalent lag of the task in place_entity(). So if the task was ineligible before, it will still be ineligible after migration.

So in sched_balance_rq(), we prioritize migrating eligible tasks, and we soft-limit ineligible tasks, allowing them to migrate only when nr_balance_failed is non-zero to avoid load-balancing trying very hard to balance the load.

Quoted 6 lines
Quoted 2 lines

Could you explain why you think it's better to balance eligible tasks in priority and potentially skip a load balance ?

In place_entity(), we maintain the task's original equivalent lag, even if we migrate the task to dst_rq, this does not change its eligibility attribute.

Yes, but you don't answer the question why it's better to select an eligible task vs a non eligible task.

Quoted 1 line
Quoted 7 lines
Quoted 3 lines

When there are multiple tasks on src_rq, and the dst_cpu has some runnable tasks, migrating ineligible tasks to dst_rq will not allow them to run. Therefore, such task migration is inefficient. We should

Why is it inefficient ? load balance is about evenly balancing the number of tasks or the load between CPUs, it never says that the newly migrated task should run immediately

Quoted 5 lines

My initial thought is that when we need to migrate some tasks during load balancing, at the current point in time, migrating ineligible tasks to dst_cpu means they definitely cannot run there. Therefore, I prefer to keep them on src_cpu to reduce the overhead of dequeueing and enqueueing ineligible tasks.

Quoted 4 lines

Sorry but I still don't get why it's important and would make a difference. They are all runnable but ineligible tasks got more runtime than other at that point in time so there is no real difference

I adopt a lazy strategy for ineligible tasks. At the current point in time, even if we migrate ineligible tasks to the dst CPU, they still have to wait on the dst CPU until they become eligible. We do not see clear benefits from migrating ineligible tasks, but their dequeueing and enqueueing would instead incur overhead.

Let them wait on the src CPU until they become eligible before migrating them. this can reduce the number of task migrations.

Thanks,
Hao

Quoted 38 lines
Quoted 21 lines

Migrating eligible tasks to dst_cpu does not guarantee that they will run earlier than on src_cpu. it depends on too many factors.

Quoted 6 lines
Quoted 5 lines

prioritize migrating tasks that can run on dst_rq.

In other words, migrating ineligible tasks is merely moving them to another runqueue to wait until they become eligible.

Quoted 6 lines

But I don't get why it's a problem. Migrating an eligible task might delay its scheduling because of its deadline vs other tasks already eligible on the dst_rq. Eligible and non eligible tasks are all runnable, it's just how much they have already run. In addition, migrating an eligible task will clear its positive vlag with DELAY_ZERO which is unfair IMO

Quoted 2 lines

Sorry, I'd like to ask you a question that confuses me: Why does migrating eligible task will clear the positive vlag?

sorry I mess up everything that only for delayed dequeue task

Quoted 12 lines

In detach_task(), the ENQUEUE_DELAYED and DEQUEUE_SLEEP flags are not set, and in dequeue_entity(), eligible tasks will not set sched_delayed, so they will be dequeued normally with se->on_rq being 0.

Similarly, attach_task() does not set the ENQUEUE_DELAYED and DEQUEUE_SLEEP flags, and since se->on_rq is 0, it will not call requeue_delayed_entity().

Thanks,
Hao

maintainer
#11 lore raw
Quoted 68 lines · Hao Jia
Quoted 58 lines · Vincent Guittot
Quoted 51 lines · Hao Jia
Quoted 38 lines · Vincent Guittot
Quoted 26 lines · Hao Jia
Quoted 22 lines · Vincent Guittot
Quoted 20 lines · Hao Jia

Friendly ping...

Quoted 15 lines · Hao Jia

From: Hao Jia <[email protected]>

When the PLACE_LAG scheduling feature is enabled and dst_cfs_rq->nr_queued is greater than 1, if a task is ineligible (lag < 0) on the source cpu runqueue, it will also be ineligible when it is migrated to the destination cpu runqueue. Because we will keep the original equivalent lag of the task in place_entity(). So if the task was ineligible before, it will still be ineligible after migration.

So in sched_balance_rq(), we prioritize migrating eligible tasks, and we soft-limit ineligible tasks, allowing them to migrate only when nr_balance_failed is non-zero to avoid load-balancing trying very hard to balance the load.

Quoted 6 lines
Quoted 2 lines

Could you explain why you think it's better to balance eligible tasks in priority and potentially skip a load balance ?

In place_entity(), we maintain the task's original equivalent lag, even if we migrate the task to dst_rq, this does not change its eligibility attribute.

Yes, but you don't answer the question why it's better to select an eligible task vs a non eligible task.

Quoted 1 line
Quoted 7 lines
Quoted 3 lines

When there are multiple tasks on src_rq, and the dst_cpu has some runnable tasks, migrating ineligible tasks to dst_rq will not allow them to run. Therefore, such task migration is inefficient. We should

Why is it inefficient ? load balance is about evenly balancing the number of tasks or the load between CPUs, it never says that the newly migrated task should run immediately

Quoted 5 lines

My initial thought is that when we need to migrate some tasks during load balancing, at the current point in time, migrating ineligible tasks to dst_cpu means they definitely cannot run there. Therefore, I prefer to keep them on src_cpu to reduce the overhead of dequeueing and enqueueing ineligible tasks.

Quoted 4 lines

Sorry but I still don't get why it's important and would make a difference. They are all runnable but ineligible tasks got more runtime than other at that point in time so there is no real difference

Quoted 5 lines

I adopt a lazy strategy for ineligible tasks. At the current point in time, even if we migrate ineligible tasks to the dst CPU, they still have to wait on the dst CPU until they become eligible. We do not see clear benefits from migrating ineligible tasks, but their dequeueing and enqueueing would instead incur overhead.

But your explanation doesn't make sense.
Not migrating an ineligible task only make sense for delayed_dequeue tasks because they don't really want to run but only exhaust their lag but this is already taken into account by
61b82dfb6b7e ("sched/fair: Do not try to migrate delayed dequeue task")

Did you run your benchmark on top of this change ?

Quoted 49 lines

Let them wait on the src CPU until they become eligible before migrating them. this can reduce the number of task migrations.

Thanks,
Hao

Quoted 38 lines
Quoted 21 lines

Migrating eligible tasks to dst_cpu does not guarantee that they will run earlier than on src_cpu. it depends on too many factors.

Quoted 6 lines
Quoted 5 lines

prioritize migrating tasks that can run on dst_rq.

In other words, migrating ineligible tasks is merely moving them to another runqueue to wait until they become eligible.

Quoted 6 lines

But I don't get why it's a problem. Migrating an eligible task might delay its scheduling because of its deadline vs other tasks already eligible on the dst_rq. Eligible and non eligible tasks are all runnable, it's just how much they have already run. In addition, migrating an eligible task will clear its positive vlag with DELAY_ZERO which is unfair IMO

Quoted 2 lines

Sorry, I'd like to ask you a question that confuses me: Why does migrating eligible task will clear the positive vlag?

sorry I mess up everything that only for delayed dequeue task

Quoted 12 lines

In detach_task(), the ENQUEUE_DELAYED and DEQUEUE_SLEEP flags are not set, and in dequeue_entity(), eligible tasks will not set sched_delayed, so they will be dequeued normally with se->on_rq being 0.

Similarly, attach_task() does not set the ENQUEUE_DELAYED and DEQUEUE_SLEEP flags, and since se->on_rq is 0, it will not call requeue_delayed_entity().

Thanks,
Hao

author
#12 lore raw
Quoted 75 lines · Vincent Guittot
Quoted 68 lines · Hao Jia
Quoted 58 lines · Vincent Guittot
Quoted 51 lines · Hao Jia
Quoted 38 lines · Vincent Guittot
Quoted 26 lines · Hao Jia
Quoted 22 lines · Vincent Guittot
Quoted 20 lines · Hao Jia

Friendly ping...

Quoted 15 lines · Hao Jia

From: Hao Jia <[email protected]>

When the PLACE_LAG scheduling feature is enabled and dst_cfs_rq->nr_queued is greater than 1, if a task is ineligible (lag < 0) on the source cpu runqueue, it will also be ineligible when it is migrated to the destination cpu runqueue. Because we will keep the original equivalent lag of the task in place_entity(). So if the task was ineligible before, it will still be ineligible after migration.

So in sched_balance_rq(), we prioritize migrating eligible tasks, and we soft-limit ineligible tasks, allowing them to migrate only when nr_balance_failed is non-zero to avoid load-balancing trying very hard to balance the load.

Quoted 6 lines
Quoted 2 lines

Could you explain why you think it's better to balance eligible tasks in priority and potentially skip a load balance ?

In place_entity(), we maintain the task's original equivalent lag, even if we migrate the task to dst_rq, this does not change its eligibility attribute.

Yes, but you don't answer the question why it's better to select an eligible task vs a non eligible task.

Quoted 1 line
Quoted 7 lines
Quoted 3 lines

When there are multiple tasks on src_rq, and the dst_cpu has some runnable tasks, migrating ineligible tasks to dst_rq will not allow them to run. Therefore, such task migration is inefficient. We should

Why is it inefficient ? load balance is about evenly balancing the number of tasks or the load between CPUs, it never says that the newly migrated task should run immediately

Quoted 5 lines

My initial thought is that when we need to migrate some tasks during load balancing, at the current point in time, migrating ineligible tasks to dst_cpu means they definitely cannot run there. Therefore, I prefer to keep them on src_cpu to reduce the overhead of dequeueing and enqueueing ineligible tasks.

Quoted 4 lines

Sorry but I still don't get why it's important and would make a difference. They are all runnable but ineligible tasks got more runtime than other at that point in time so there is no real difference

Quoted 5 lines

I adopt a lazy strategy for ineligible tasks. At the current point in time, even if we migrate ineligible tasks to the dst CPU, they still have to wait on the dst CPU until they become eligible. We do not see clear benefits from migrating ineligible tasks, but their dequeueing and enqueueing would instead incur overhead.

Quoted 6 lines

But your explanation doesn't make sense.
Not migrating an ineligible task only make sense for delayed_dequeue tasks because they don't really want to run but only exhaust their lag but this is already taken into account by
61b82dfb6b7e ("sched/fair: Do not try to migrate delayed dequeue task")

Thank you for your suggestion.

Yes, as you mentioned, this commit 61b82dfb6b7e ("sched/fair: Do not try to migrate delayed dequeue task") reduces the migration of delayed_dequeue tasks, but it doesn't work for ineligible RUNNING tasks and when the migration_type is migrate_load.

Quoted 1 line

Did you run your benchmark on top of this change ?

My previous benchmark tests were based on the torvalds/linux/master branch, which does not include commit 61b82dfb6b7e ("sched/fair: Do not try to migrate delayed dequeue task"). I will include this commit and retest on my machine after my leave ends.

Thanks,
Hao

Quoted 50 lines
Quoted 50 lines

Let them wait on the src CPU until they become eligible before migrating them. this can reduce the number of task migrations.

Thanks,
Hao

Quoted 38 lines
Quoted 21 lines

Migrating eligible tasks to dst_cpu does not guarantee that they will run earlier than on src_cpu. it depends on too many factors.

Quoted 6 lines
Quoted 5 lines

prioritize migrating tasks that can run on dst_rq.

In other words, migrating ineligible tasks is merely moving them to another runqueue to wait until they become eligible.

Quoted 6 lines

But I don't get why it's a problem. Migrating an eligible task might delay its scheduling because of its deadline vs other tasks already eligible on the dst_rq. Eligible and non eligible tasks are all runnable, it's just how much they have already run. In addition, migrating an eligible task will clear its positive vlag with DELAY_ZERO which is unfair IMO

Quoted 2 lines

Sorry, I'd like to ask you a question that confuses me: Why does migrating eligible task will clear the positive vlag?

sorry I mess up everything that only for delayed dequeue task

Quoted 12 lines

In detach_task(), the ENQUEUE_DELAYED and DEQUEUE_SLEEP flags are not set, and in dequeue_entity(), eligible tasks will not set sched_delayed, so they will be dequeued normally with se->on_rq being 0.

Similarly, attach_task() does not set the ENQUEUE_DELAYED and DEQUEUE_SLEEP flags, and since se->on_rq is 0, it will not call requeue_delayed_entity().

Thanks,
Hao

author
#13 lore raw
Quoted 85 lines · Hao Jia
Quoted 75 lines · Vincent Guittot
Quoted 68 lines · Hao Jia
Quoted 58 lines · Vincent Guittot
Quoted 51 lines · Hao Jia
Quoted 38 lines · Vincent Guittot
Quoted 26 lines · Hao Jia
Quoted 22 lines · Vincent Guittot
Quoted 20 lines · Hao Jia

Friendly ping...

Quoted 15 lines · Hao Jia

From: Hao Jia <[email protected]>

When the PLACE_LAG scheduling feature is enabled and dst_cfs_rq->nr_queued is greater than 1, if a task is ineligible (lag < 0) on the source cpu runqueue, it will also be ineligible when it is migrated to the destination cpu runqueue. Because we will keep the original equivalent lag of the task in place_entity(). So if the task was ineligible before, it will still be ineligible after migration.

So in sched_balance_rq(), we prioritize migrating eligible tasks, and we soft-limit ineligible tasks, allowing them to migrate only when nr_balance_failed is non-zero to avoid load-balancing trying very hard to balance the load.

Quoted 6 lines
Quoted 2 lines

Could you explain why you think it's better to balance eligible tasks in priority and potentially skip a load balance ?

In place_entity(), we maintain the task's original equivalent lag, even if we migrate the task to dst_rq, this does not change its eligibility attribute.

Yes, but you don't answer the question why it's better to select an eligible task vs a non eligible task.

Quoted 1 line
Quoted 7 lines
Quoted 3 lines

When there are multiple tasks on src_rq, and the dst_cpu has some runnable tasks, migrating ineligible tasks to dst_rq will not allow them to run. Therefore, such task migration is inefficient. We should

Why is it inefficient ? load balance is about evenly balancing the number of tasks or the load between CPUs, it never says that the newly migrated task should run immediately

Quoted 5 lines

My initial thought is that when we need to migrate some tasks during load balancing, at the current point in time, migrating ineligible tasks to dst_cpu means they definitely cannot run there. Therefore, I prefer to keep them on src_cpu to reduce the overhead of dequeueing and enqueueing ineligible tasks.

Quoted 4 lines

Sorry but I still don't get why it's important and would make a difference. They are all runnable but ineligible tasks got more runtime than other at that point in time so there is no real difference

Quoted 5 lines

I adopt a lazy strategy for ineligible tasks. At the current point in time, even if we migrate ineligible tasks to the dst CPU, they still have to wait on the dst CPU until they become eligible. We do not see clear benefits from migrating ineligible tasks, but their dequeueing and enqueueing would instead incur overhead.

Quoted 6 lines

But your explanation doesn't make sense.
Not migrating an ineligible task only make sense for delayed_dequeue tasks because they don't really want to run but only exhaust their lag but this is already taken into account by
61b82dfb6b7e ("sched/fair: Do not try to migrate delayed dequeue task")

Quoted 9 lines

Thank you for your suggestion.

Yes, as you mentioned, this commit 61b82dfb6b7e ("sched/fair: Do not try to migrate delayed dequeue task") reduces the migration of delayed_dequeue tasks, but it doesn't work for ineligible RUNNING tasks and when the migration_type is migrate_load.

Quoted 1 line

Did you run your benchmark on top of this change ?

My previous benchmark tests were based on the torvalds/linux/master branch, which does not include commit 61b82dfb6b7e ("sched/fair: Do not try to migrate delayed dequeue task"). I will include this commit and retest on my machine after my leave ends.

I'm really sorry for being away for so long. I have retested the hackbench on my machine, which includes the commit 61b82dfb6b7e ("sched/fair: Do not try to migrate delayed dequeue task"). Based on the hackbench test results, this patch still brings a slight performance improvement.

vanilla: Includes commit 61b82dfb6b7e ("sched/fair: Do not try to migrate delayed dequeue task"), but does not include my patch.

patched: Includes both the above commit and my patch.

hackbench-process-pipes
vanilla patched

Amean     1       0.4087 (   0.00%)      0.4003 (   2.04%)
Amean     4       1.7033 (   0.00%)      1.7100 (  -0.39%)
Amean     7       2.9020 (   0.00%)      2.8750 (   0.93%)
Amean     12      4.2543 (   0.00%)      4.2980 (  -1.03%)
Amean     21      5.8633 (   0.00%)      5.7507 (   1.92%)
Amean     30      7.3757 (   0.00%)      7.2887 (   1.18%)
Amean     48     10.5360 (   0.00%)     10.2647 (   2.58%)
Amean     79     16.5480 (   0.00%)     15.9820 (   3.42%)
Amean     96     19.7873 (   0.00%)     19.1347 (   3.30%)

hackbench-process-sockets
vanilla patched

Amean     1       0.7520 (   0.00%)      0.7377 (   1.91%)
Amean     4       2.5760 (   0.00%)      2.5103 (   2.55%)
Amean     7       4.3927 (   0.00%)      4.2653 (   2.90%)
Amean     12      7.3923 (   0.00%)      7.1427 (   3.38%)
Amean     21     12.3733 (   0.00%)     11.9760 (   3.21%)
Amean     30     17.2617 (   0.00%)     16.7987 (   2.68%)
Amean     48     28.8577 (   0.00%)     28.1980 (   2.29%)
Amean     79     50.0687 (   0.00%)     49.0887 (   1.96%)
Amean     96     62.1603 (   0.00%)     61.1177 (   1.68%)

FYI.

After the performance tests, I added some code in the can_migrate_task() to count the different reasons why tasks could not be migrated during the hackbench run.

hit_delayed_dequeue_cnt: Hit if (p->se.sched_delayed) && (env->migration_type != migrate_load)

hit_ineligible_cnt: Did not hit "hit_delayed_dequeue" && !env->sd->nr_balance_failed && task_is_ineligible_on_dst_cpu()

Count results:

hit_delayed_dequeue_cnt 378432
hit_ineligible_cnt 1862099

Thanks,
Hao

merge bot Applied
#14 lore raw
In reply to Hao Jia
The following commit has been merged into the sched/core branch of tip:

The following commit has been merged into the sched/core branch of tip:

Commit-ID:     873199d27bb25889ab7ccca03c8f30c67f18ae52
Gitweb:        https://git.kernel.org/tip/873199d27bb25889ab7ccca03c8f30c67f18ae52
Author:        Hao Jia <[email protected]>
AuthorDate:    Mon, 23 Dec 2024 17:14:46 +08:00
Committer:     Peter Zijlstra <[email protected]>
CommitterDate: Mon, 13 Jan 2025 14:10:23 +01:00

sched/core: Prioritize migrating eligible tasks in sched_balance_rq()

When the PLACE_LAG scheduling feature is enabled and dst_cfs_rq->nr_queued is greater than 1, if a task is ineligible (lag < 0) on the source cpu runqueue, it will also be ineligible when it is migrated to the destination cpu runqueue. Because we will keep the original equivalent lag of the task in place_entity(). So if the task was ineligible before, it will still be ineligible after migration.

So in sched_balance_rq(), we prioritize migrating eligible tasks, and we soft-limit ineligible tasks, allowing them to migrate only when nr_balance_failed is non-zero to avoid load-balancing trying very hard to balance the load.

Below are some benchmark test results. From my test results, this patch shows a slight improvement on hackbench.

Benchmark
=========

All of the benchmarks are done inside a normal cpu cgroup in a clean environment with cpu turbo disabled, and test machine is:

Single NUMA machine model is 13th Gen Intel(R) Core(TM) i7-13700, 12 Core/24 HT.

Based on master b86545e02e8c.

Results
=======

hackbench-process-pipes
vanilla patched

Amean     1       0.5837 (   0.00%)      0.5733 (   1.77%)
Amean     4       1.4423 (   0.00%)      1.4503 (  -0.55%)
Amean     7       2.5147 (   0.00%)      2.4773 (   1.48%)
Amean     12      3.9347 (   0.00%)      3.8880 (   1.19%)
Amean     21      5.3943 (   0.00%)      5.3873 (   0.13%)
Amean     30      6.7840 (   0.00%)      6.6660 (   1.74%)
Amean     48      9.8313 (   0.00%)      9.6100 (   2.25%)
Amean     79     15.4403 (   0.00%)     14.9580 (   3.12%)
Amean     96     18.4970 (   0.00%)     17.9533 (   2.94%)

hackbench-process-sockets
vanilla patched

Amean     1       0.6297 (   0.00%)      0.6223 (   1.16%)
Amean     4       2.1517 (   0.00%)      2.0887 (   2.93%)
Amean     7       3.6377 (   0.00%)      3.5670 (   1.94%)
Amean     12      6.1277 (   0.00%)      5.9290 (   3.24%)
Amean     21     10.0380 (   0.00%)      9.7623 (   2.75%)
Amean     30     14.1517 (   0.00%)     13.7513 (   2.83%)
Amean     48     24.7253 (   0.00%)     24.2287 (   2.01%)
Amean     79     43.9523 (   0.00%)     43.2330 (   1.64%)
Amean     96     54.5310 (   0.00%)     53.7650 (   1.40%)

tbench4 Throughput
vanilla patched

Hmean     1       255.97 (   0.00%)      275.01 (   7.44%)
Hmean     2       511.60 (   0.00%)      544.27 (   6.39%)
Hmean     4       996.70 (   0.00%)     1006.57 (   0.99%)
Hmean     8      1646.46 (   0.00%)     1649.15 (   0.16%)
Hmean     16     2259.42 (   0.00%)     2274.35 (   0.66%)
Hmean     32     4725.48 (   0.00%)     4735.57 (   0.21%)
Hmean     64     4411.47 (   0.00%)     4400.05 (  -0.26%)
Hmean     96     4284.31 (   0.00%)     4267.39 (  -0.39%)
---
 kernel/sched/fair.c | 34 ++++++++++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)
kernel/sched/fair.c +34 -0
index 7ec2587..52f7278 100644
@@ -9407,6 +9407,30 @@ static inline long migrate_degrades_locality(struct task_struct *p,
 #endif
 
 /*
+ * Check whether the task is ineligible on the destination cpu
+ *
+ * When the PLACE_LAG scheduling feature is enabled and
+ * dst_cfs_rq->nr_queued is greater than 1, if the task
+ * is ineligible, it will also be ineligible when
+ * it is migrated to the destination cpu.
+ */
+static inline int task_is_ineligible_on_dst_cpu(struct task_struct *p, int dest_cpu)
+{
+	struct cfs_rq *dst_cfs_rq;
+
+#ifdef CONFIG_FAIR_GROUP_SCHED
+	dst_cfs_rq = task_group(p)->cfs_rq[dest_cpu];
+#else
+	dst_cfs_rq = &cpu_rq(dest_cpu)->cfs;
+#endif
+	if (sched_feat(PLACE_LAG) && dst_cfs_rq->nr_queued &&
+	    !entity_eligible(task_cfs_rq(p), &p->se))
+		return 1;
+
+	return 0;
+}
+
+/*
  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
  */
 static
@@ -9432,6 +9456,16 @@ int can_migrate_task(struct task_struct *p, struct lb_env *env)
 	if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
 		return 0;
 
+	/*
+	 * We want to prioritize the migration of eligible tasks.
+	 * For ineligible tasks we soft-limit them and only allow
+	 * them to migrate when nr_balance_failed is non-zero to
+	 * avoid load-balancing trying very hard to balance the load.
+	 */
+	if (!env->sd->nr_balance_failed &&
+	    task_is_ineligible_on_dst_cpu(p, env->dst_cpu))
+		return 0;
+
 	/* Disregard percpu kthreads; they are where they need to be. */
 	if (kthread_is_per_cpu(p))
 		return 0;
#15 lore raw
In reply to Hao Jia
Quoted 124 lines · Hao Jia

From: Hao Jia <[email protected]>

When the PLACE_LAG scheduling feature is enabled and dst_cfs_rq->nr_queued is greater than 1, if a task is ineligible (lag < 0) on the source cpu runqueue, it will also be ineligible when it is migrated to the destination cpu runqueue. Because we will keep the original equivalent lag of the task in place_entity(). So if the task was ineligible before, it will still be ineligible after migration.

So in sched_balance_rq(), we prioritize migrating eligible tasks, and we soft-limit ineligible tasks, allowing them to migrate only when nr_balance_failed is non-zero to avoid load-balancing trying very hard to balance the load.

Below are some benchmark test results. From my test results, this patch shows a slight improvement on hackbench.

Benchmark
=========

All of the benchmarks are done inside a normal cpu cgroup in a clean environment with cpu turbo disabled, and test machine is:

Single NUMA machine model is 13th Gen Intel(R) Core(TM) i7-13700, 12 Core/24 HT.

Based on master b86545e02e8c.

Results
=======

hackbench-process-pipes
vanilla patched

Amean     1       0.5837 (   0.00%)      0.5733 (   1.77%)
Amean     4       1.4423 (   0.00%)      1.4503 (  -0.55%)
Amean     7       2.5147 (   0.00%)      2.4773 (   1.48%)
Amean     12      3.9347 (   0.00%)      3.8880 (   1.19%)
Amean     21      5.3943 (   0.00%)      5.3873 (   0.13%)
Amean     30      6.7840 (   0.00%)      6.6660 (   1.74%)
Amean     48      9.8313 (   0.00%)      9.6100 (   2.25%)
Amean     79     15.4403 (   0.00%)     14.9580 (   3.12%)
Amean     96     18.4970 (   0.00%)     17.9533 (   2.94%)

hackbench-process-sockets
vanilla patched

Amean     1       0.6297 (   0.00%)      0.6223 (   1.16%)
Amean     4       2.1517 (   0.00%)      2.0887 (   2.93%)
Amean     7       3.6377 (   0.00%)      3.5670 (   1.94%)
Amean     12      6.1277 (   0.00%)      5.9290 (   3.24%)
Amean     21     10.0380 (   0.00%)      9.7623 (   2.75%)
Amean     30     14.1517 (   0.00%)     13.7513 (   2.83%)
Amean     48     24.7253 (   0.00%)     24.2287 (   2.01%)
Amean     79     43.9523 (   0.00%)     43.2330 (   1.64%)
Amean     96     54.5310 (   0.00%)     53.7650 (   1.40%)

tbench4 Throughput
vanilla patched

Hmean     1       255.97 (   0.00%)      275.01 (   7.44%)
Hmean     2       511.60 (   0.00%)      544.27 (   6.39%)
Hmean     4       996.70 (   0.00%)     1006.57 (   0.99%)
Hmean     8      1646.46 (   0.00%)     1649.15 (   0.16%)
Hmean     16     2259.42 (   0.00%)     2274.35 (   0.66%)
Hmean     32     4725.48 (   0.00%)     4735.57 (   0.21%)
Hmean     64     4411.47 (   0.00%)     4400.05 (  -0.26%)
Hmean     96     4284.31 (   0.00%)     4267.39 (  -0.39%)

---
Previous discussion link: https://lore.kernel.org/all/[email protected]
Link to v1: https://lore.kernel.org/all/[email protected]

v1 to v2:
- Modify dst_cfs_rq->nr_running to dst_cfs_rq->nr_queued to resolve conflicts with commit 736c55a02c47 ("sched/fair:
Rename cfs_rq.nr_running into nr_queued").

 kernel/sched/fair.c | 34 ++++++++++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)
kernel/sched/fair.c +27 -0
index 5599b0c1ba9b..c884bf631e66 100644
@@ -9396,6 +9396,30 @@ static inline int migrate_degrades_locality(struct task_struct *p,
 }
 #endif
 
+/*
+ * Check whether the task is ineligible on the destination cpu
+ *
+ * When the PLACE_LAG scheduling feature is enabled and
+ * dst_cfs_rq->nr_queued is greater than 1, if the task
+ * is ineligible, it will also be ineligible when
+ * it is migrated to the destination cpu.
+ */
+static inline int task_is_ineligible_on_dst_cpu(struct task_struct *p, int dest_cpu)
+{
+	struct cfs_rq *dst_cfs_rq;
+
+#ifdef CONFIG_FAIR_GROUP_SCHED
+	dst_cfs_rq = task_group(p)->cfs_rq[dest_cpu];
+#else
+	dst_cfs_rq = &cpu_rq(dest_cpu)->cfs;
+#endif
+	if (sched_feat(PLACE_LAG) && dst_cfs_rq->nr_queued &&
+	    !entity_eligible(task_cfs_rq(p), &p->se))
+		return 1;
+
+	return 0;
+}
+
 /*
  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
  */
@@ -9420,6 +9444,16 @@ int can_migrate_task(struct task_struct *p, struct lb_env *env)
 	if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
 		return 0;
 
+	/*
+	 * We want to prioritize the migration of eligible tasks.
+	 * For ineligible tasks we soft-limit them and only allow

kernel/sched/fair.c int can_migrate_task(struct task_struct *p, struct lb_env *env) · line 9459

Quoted 10 lines
kernel/sched/fair.c +7 -0
+	 * them to migrate when nr_balance_failed is non-zero to
+	 * avoid load-balancing trying very hard to balance the load.
+	 */
+	if (!env->sd->nr_balance_failed &&
+	    task_is_ineligible_on_dst_cpu(p, env->dst_cpu))
+		return 0;
+
 	/* Disregard percpu kthreads; they are where they need to be. */
 	if (kthread_is_per_cpu(p))
 		return 0;

Just a general comment.

If we throw tasks with custom slices into the mix, I wonder what kinds of impacts we would see from migrating eligible tasks with very short slice lengths (they may run earlier) or very long ones.

We could've picked an inelegible task before, which wouldn't have an impact other than the overhead of the migration. But the patch might make it so we now pick an eligible one with a custom slice.