improved load_balance() tolerance for pinned tasks

PATCH In review John Hawkes 6 participants 8 messages Oct 20, 2004 – Oct 30, 2004 1 message not in the archive

Share & tools

John Hawkes

author

2004-10-20 19:37 UTC #0 lore raw

PATCH 1/1 improved load_balance() tolerance for pinned tasks

A large number of processes that are pinned to a single CPU results in every other CPU's load_balance() seeing this overloaded CPU as "busiest", yet move_tasks() never finds a task to pull-migrate. This condition occurs during module unload, but can also occur as a denial-of-service using sys_sched_setaffinity(). Several hundred CPUs performing this fruitless load_balance() will livelock on the busiest CPU's runqueue lock. A smaller number of CPUs will livelock if the pinned task count gets high. This simple patch remedies the more common first problem:
after a move_tasks() failure to migrate anything, the balance_interval increments. Using a simple increment, vs. the more dramatic doubling of the balance_interval, is conservative and yet also effective.

John Hawkes

Signed-off-by John Hawkes <[email protected]>

kernel/sched.c +12 -4

Index: linux/kernel/sched.c
===================================================================
@@ -2123,11 +2123,19 @@
 			 */
 			sd->nr_balance_failed = sd->cache_nice_tries;
 		}
-	} else
-		sd->nr_balance_failed = 0;
 
-	/* We were unbalanced, so reset the balancing interval */
-	sd->balance_interval = sd->min_interval;
+		/*
+		 * We were unbalanced, but unsuccessful in move_tasks(),
+		 * so bump the balance_interval to lessen the lock contention.
+		 */
+		if (sd->balance_interval < sd->max_interval)
+			sd->balance_interval++;
+	} else {
+                sd->nr_balance_failed = 0;
+
+		/* We were unbalanced, so reset the balancing interval */
+		sd->balance_interval = sd->min_interval;
+	}
 
 	return nr_moved;

Ingo Molnar

Acked-by on v1 1/1

2004-10-20 20:08 UTC #1 lore raw

Quoted 12 lines · John Hawkes

A large number of processes that are pinned to a single CPU results in every other CPU's load_balance() seeing this overloaded CPU as "busiest", yet move_tasks() never finds a task to pull-migrate. This condition occurs during module unload, but can also occur as a denial-of-service using sys_sched_setaffinity(). Several hundred CPUs performing this fruitless load_balance() will livelock on the busiest CPU's runqueue lock. A smaller number of CPUs will livelock if the pinned task count gets high. This simple patch remedies the more common first problem:
after a move_tasks() failure to migrate anything, the balance_interval increments. Using a simple increment, vs. the more dramatic doubling of the balance_interval, is conservative and yet also effective.

Quoted 4 lines

John Hawkes
Signed-off-by John Hawkes <[email protected]>

looks good to me.

Acked-by Ingo Molnar <[email protected]>

Ingo

Nick Piggin

2004-10-22 13:15 UTC #2 lore raw

In reply to John Hawkes

Quoted 12 lines · John Hawkes

A large number of processes that are pinned to a single CPU results in every other CPU's load_balance() seeing this overloaded CPU as "busiest", yet move_tasks() never finds a task to pull-migrate. This condition occurs during module unload, but can also occur as a denial-of-service using sys_sched_setaffinity(). Several hundred CPUs performing this fruitless load_balance() will livelock on the busiest CPU's runqueue lock. A smaller number of CPUs will livelock if the pinned task count gets high. This simple patch remedies the more common first problem:
after a move_tasks() failure to migrate anything, the balance_interval increments. Using a simple increment, vs. the more dramatic doubling of the balance_interval, is conservative and yet also effective.

Quoted 2 lines

John Hawkes

John and I also discussed a slightly more sophisticated algorithm for this that would try to be a bit smarter about the reason for balance failures (described in the changelog).

John, this works quite well here, would you have any objections using it instead? Does it cure your problems there?

John Hawkes explained the problem best:
A large number of processes that are pinned to a single CPU results in every other CPU's load_balance() seeing this overloaded CPU as "busiest", yet move_tasks() never finds a task to pull-migrate. This condition occurs during module unload, but can also occur as a denial-of-service using sys_sched_setaffinity(). Several hundred CPUs performing this fruitless load_balance() will livelock on the busiest CPU's runqueue lock. A smaller number of CPUs will livelock if the pinned task count gets high.

Expanding slightly on John's patch, this one attempts to work out whether the balancing failure has been due to too many tasks pinned on the runqueue. This allows it to be basically invisible to the regular blancing paths (ie. when there are no pinned tasks). We can use this extra knowledge to shut down the balancing faster, and ensure the migration threads don't start running which is another problem observed in the wild.

Signed-off-by Nick Piggin <[email protected]>

---

 linux-2.6-npiggin/kernel/sched.c |   34 +++++++++++++++++++++++-----------
 1 files changed, 23 insertions(+), 11 deletions(-)

kernel/sched.c renamed from kernel/sched.c~sched-lb-pinned +23 ~~-11~~

diff -puN kernel/sched.c~sched-lb-pinned kernel/sched.c
@@ -1626,7 +1626,7 @@ void pull_task(runqueue_t *src_rq, prio_
  */
 static inline
 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
-		     struct sched_domain *sd, enum idle_type idle)
+		     struct sched_domain *sd, enum idle_type idle, int *pinned)
 {
 	/*
 	 * We do not migrate tasks that are:

@@ -1636,8 +1636,10 @@ int can_migrate_task(task_t *p, runqueue
 	 */
 	if (task_running(rq, p))
 		return 0;
-	if (!cpu_isset(this_cpu, p->cpus_allowed))
+	if (!cpu_isset(this_cpu, p->cpus_allowed)) {
+		*pinned++;
 		return 0;
+	}
 
 	/* Aggressive migration if we've failed balancing */
 	if (idle == NEWLY_IDLE ||

@@ -1658,11 +1660,11 @@ int can_migrate_task(task_t *p, runqueue
  */
 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
 		      unsigned long max_nr_move, struct sched_domain *sd,
-		      enum idle_type idle)
+		      enum idle_type idle, int *all_pinned)
 {
 	prio_array_t *array, *dst_array;
 	struct list_head *head, *curr;
-	int idx, pulled = 0;
+	int idx, pulled = 0, pinned = 0;
 	task_t *tmp;
 
 	if (max_nr_move <= 0 || busiest->nr_running <= 1)

@@ -1706,7 +1708,7 @@ skip_queue:
 
 	curr = curr->prev;
 
-	if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle)) {
+	if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
 		if (curr != head)
 			goto skip_queue;
 		idx++;

@@ -1732,6 +1734,9 @@ skip_queue:
 		goto skip_bitmap;
 	}
 out:
+	*all_pinned = 0;
+	if (unlikely(pinned >= max_nr_move) && pulled == 0)
+		*all_pinned = 1;
 	return pulled;
 }

@@ -1906,7 +1911,7 @@ static int load_balance(int this_cpu, ru
 	struct sched_group *group;
 	runqueue_t *busiest;
 	unsigned long imbalance;
-	int nr_moved;
+	int nr_moved, all_pinned;
 
 	spin_lock(&this_rq->lock);
 	schedstat_inc(sd, lb_cnt[idle]);

@@ -1945,11 +1950,16 @@ static int load_balance(int this_cpu, ru
 		 */
 		double_lock_balance(this_rq, busiest);
 		nr_moved = move_tasks(this_rq, this_cpu, busiest,
-						imbalance, sd, idle);
+						imbalance, sd, idle,
+						&all_pinned);
 		spin_unlock(&busiest->lock);
 	}
-	spin_unlock(&this_rq->lock);
+	/* All tasks on this runqueue were pinned by CPU affinity */
+	if (unlikely(all_pinned))
+		goto out_balanced;
 
+	spin_unlock(&this_rq->lock);
+	
 	if (!nr_moved) {
 		schedstat_inc(sd, lb_failed[idle]);
 		sd->nr_balance_failed++;

@@ -2004,7 +2014,7 @@ static int load_balance_newidle(int this
 	struct sched_group *group;
 	runqueue_t *busiest = NULL;
 	unsigned long imbalance;
-	int nr_moved = 0;
+	int nr_moved = 0, all_pinned;
 
 	schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
 	group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE);

@@ -2024,7 +2034,7 @@ static int load_balance_newidle(int this
 
 	schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
 	nr_moved = move_tasks(this_rq, this_cpu, busiest,
-					imbalance, sd, NEWLY_IDLE);
+			imbalance, sd, NEWLY_IDLE, &all_pinned);
 	if (!nr_moved)
 		schedstat_inc(sd, lb_failed[NEWLY_IDLE]);

@@ -2085,6 +2095,7 @@ static void active_load_balance(runqueue
 	do {
 		runqueue_t *rq;
 		int push_cpu = 0;
+		int all_pinned;
 
 		if (group == busy_group)
 			goto next_group;

@@ -2106,7 +2117,8 @@ static void active_load_balance(runqueue
 		if (unlikely(busiest == rq))
 			goto next_group;
 		double_lock_balance(busiest, rq);
-		if (move_tasks(rq, push_cpu, busiest, 1, sd, SCHED_IDLE)) {
+		if (move_tasks(rq, push_cpu, busiest, 1,
+					sd, SCHED_IDLE, &all_pinned)) {
 			schedstat_inc(busiest, alb_lost);
 			schedstat_inc(rq, alb_gained);
 		} else {

John Hawkes

2004-10-22 19:44 UTC #3 lore raw

In reply to John Hawkes

Quoted 6 lines

John and I also discussed a slightly more sophisticated algorithm for this that would try to be a bit smarter about the reason for balance failures (described in the changelog).
John, this works quite well here, would you have any objections using it instead? Does it cure your problems there?

No, your variation doesn't solve the problem. This variation of your patch does, however, solve the problem. The difference is in move_tasks():

Signed-off-by John Hawkes <[email protected]>

 linux-2.6-npiggin/kernel/sched.c |   34 +++++++++++++++++++++++-----------
 1 files changed, 23 insertions(+), 11 deletions(-)

kernel/sched.c +23 ~~-11~~

Index: linux/kernel/sched.c
===================================================================
@@ -1770,7 +1770,7 @@
  */
 static inline
 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
-		     struct sched_domain *sd, enum idle_type idle)
+		     struct sched_domain *sd, enum idle_type idle, int *pinned)
 {
 	/*
 	 * We do not migrate tasks that are:

@@ -1780,8 +1780,10 @@
 	 */
 	if (task_running(rq, p))
 		return 0;
-	if (!cpu_isset(this_cpu, p->cpus_allowed))
+	if (!cpu_isset(this_cpu, p->cpus_allowed)) {
+		*pinned++;
 		return 0;
+	}
 
 	/* Aggressive migration if we've failed balancing */
 	if (idle == NEWLY_IDLE ||

@@ -1802,11 +1804,11 @@
  */
 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
 		      unsigned long max_nr_move, struct sched_domain *sd,
-		      enum idle_type idle)
+		      enum idle_type idle, int *all_pinned)
 {
 	prio_array_t *array, *dst_array;
 	struct list_head *head, *curr;
-	int idx, pulled = 0;
+	int idx, examined = 0, pulled = 0, pinned = 0;
 	task_t *tmp;
 
 	if (max_nr_move <= 0 || busiest->nr_running <= 1)

@@ -1850,7 +1852,8 @@
 
 	curr = curr->prev;
 
-	if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle)) {
+	examined++;
+	if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
 		if (curr != head)
 			goto skip_queue;
 		idx++;

@@ -1876,6 +1879,8 @@
 		goto skip_bitmap;
 	}
 out:
+	if (unlikely(examined && examined == pinned))
+		*all_pinned = 1;
 	return pulled;
 }

@@ -2056,7 +2061,7 @@
 	struct sched_group *group;
 	runqueue_t *busiest;
 	unsigned long imbalance;
-	int nr_moved;
+	int nr_moved, all_pinned;
 
 	spin_lock(&this_rq->lock);
 	schedstat_inc(sd, lb_cnt[idle]);

@@ -2095,11 +2100,16 @@
 		 */
 		double_lock_balance(this_rq, busiest);
 		nr_moved = move_tasks(this_rq, this_cpu, busiest,
-						imbalance, sd, idle);
+						imbalance, sd, idle,
+						&all_pinned);
 		spin_unlock(&busiest->lock);
 	}
-	spin_unlock(&this_rq->lock);
+	/* All tasks on this runqueue were pinned by CPU affinity */
+	if (unlikely(all_pinned))
+		goto out_balanced;
 
+	spin_unlock(&this_rq->lock);
+	
 	if (!nr_moved) {
 		schedstat_inc(sd, lb_failed[idle]);
 		sd->nr_balance_failed++;

@@ -2154,7 +2164,7 @@
 	struct sched_group *group;
 	runqueue_t *busiest = NULL;
 	unsigned long imbalance;
-	int nr_moved = 0;
+	int nr_moved = 0, all_pinned;
 
 	schedstat_inc(sd, lb_cnt[NEWLY_IDLE]);
 	group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE);

@@ -2174,7 +2184,7 @@
 
 	schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
 	nr_moved = move_tasks(this_rq, this_cpu, busiest,
-					imbalance, sd, NEWLY_IDLE);
+			imbalance, sd, NEWLY_IDLE, &all_pinned);
 	if (!nr_moved)
 		schedstat_inc(sd, lb_failed[NEWLY_IDLE]);

@@ -2236,6 +2246,7 @@
 		cpumask_t tmp;
 		runqueue_t *rq;
 		int push_cpu = 0;
+		int all_pinned;
 
 		if (group == busy_group)
 			goto next_group;

@@ -2261,7 +2272,8 @@
 		if (unlikely(busiest == rq))
 			goto next_group;
 		double_lock_balance(busiest, rq);
-		if (move_tasks(rq, push_cpu, busiest, 1, sd, IDLE)) {
+		if (move_tasks(rq, push_cpu, busiest, 1,
+				sd, IDLE, &all_pinned)) {
 			schedstat_inc(busiest, alb_lost);
 			schedstat_inc(rq, alb_gained);
 		} else {

A message that is not in the archive (sent off-list or not yet received): [email protected]

Nick Piggin

2004-10-23 04:45 UTC #5 lore raw

Quoted 11 lines · John Hawkes

From: "John Hawkes" <[email protected]>
Quoted 3 lines
No, your variation doesn't solve the problem. This variation of your patch does, however, solve the problem. The difference is in move_tasks():
Actually, there is another related problem that arises in active_load_balance() with a runqueue that holds hundreds of pinned processes. I'm seeing a migration_thread perpetually consuming 70% of its CPU.

That's what I was worried about, but in your most recent patch you just sent, the all_pinned path should skip over the active load balance completely... basically it shouldn't be running at all, and if it is then it is a bug I think?

John Hawkes

2004-10-25 16:36 UTC #6 lore raw

From: "Nick Piggin" <[email protected]>

Quoted 3 lines

Quoted 3 lines
From: "John Hawkes" <[email protected]> Actually, there is another related problem that arises in active_load_balance() with a runqueue that holds hundreds of pinned

processes.

Quoted 6 lines

Quoted 1 line
I'm seeing a migration_thread perpetually consuming 70% of its CPU.
That's what I was worried about, but in your most recent patch you just sent, the all_pinned path should skip over the active load balance completely... basically it shouldn't be running at all, and if it is then it is a bug I think?

To reiterate: this is probably reproducible on smaller SMP systems, too. Just do a 'runon' (using sys_sched_setaffinity) of ~200 (or more) small computebound processes on a single CPU.

My patch -- that has load_balance() skip over (busiest->active_balance = 1) trigger that starts up active_load_balance() -- does seem to reduce the frequency of bursts of long-running activity of the migration thread, but those burst of activity are still there, with migration_thread consuming 75-95% of its CPU for several seconds (as observed by 'top'). I have not yet determined what's happening. It might be an artifact of how long it takes to do those 'runon' startups of the computebound processes.

John Hawkes

Nick Piggin

2004-10-26 01:25 UTC #7 lore raw

Quoted 5 lines · John Hawkes

To reiterate: this is probably reproducible on smaller SMP systems, too. Just do a 'runon' (using sys_sched_setaffinity) of ~200 (or more) small computebound processes on a single CPU.

Yeah I tried that with a handful... I'll try again with 200.

Quoted 8 lines

My patch -- that has load_balance() skip over (busiest->active_balance = 1) trigger that starts up active_load_balance() -- does seem to reduce the frequency of bursts of long-running activity of the migration thread, but those burst of activity are still there, with migration_thread consuming 75-95% of its CPU for several seconds (as observed by 'top'). I have not yet determined what's happening. It might be an artifact of how long it takes to do those 'runon' startups of the computebound processes.

Hmm... it would be hard to believe that it is all to do with migrating tasks... but its possible I guess.

I'll also look into it if I can reproduce the problem here.

Matthew Dobson

2004-10-30 00:28 UTC #8 lore raw

In reply to John Hawkes

Quoted 16 lines · John Hawkes

From: "Nick Piggin" <[email protected]>
Quoted 3 lines
Quoted 3 lines
From: "John Hawkes" <[email protected]> Actually, there is another related problem that arises in active_load_balance() with a runqueue that holds hundreds of pinned
processes.
Quoted 6 lines
Quoted 1 line
I'm seeing a migration_thread perpetually consuming 70% of its CPU.
That's what I was worried about, but in your most recent patch you just sent, the all_pinned path should skip over the active load balance completely... basically it shouldn't be running at all, and if it is then it is a bug I think?
To reiterate: this is probably reproducible on smaller SMP systems, too. Just do a 'runon' (using sys_sched_setaffinity) of ~200 (or more) small computebound processes on a single CPU.

Quoted 7 lines

My patch -- that has load_balance() skip over (busiest->active_balance = 1) trigger that starts up active_load_balance() -- does seem to reduce the frequency of bursts of long-running activity of the migration thread, but those burst of activity are still there, with migration_thread consuming 75-95% of its CPU for several seconds (as observed by 'top'). I have not yet determined what's happening. It might be an artifact of how long it takes to do those 'runon' startups of the computebound processes.

You may want to try these tests again with linux-2.6.10-rc1-mm2. It's got 2 patches to fix some broken behavior of active_load_balance(). The version of active_load_balance() in 2.6.9 was not considering a great many CPUs as potential recipients of tasks due to some small logic problems in the code.

Cheers!

-Matt