Skip to contentSkip to conversation

sched: modify move_tasks() to improve load balancing outcomes

PATCH In review Peter Williams 2 participants 9 messages Apr 13, 2006 – Apr 18, 2006

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sched: modify move_tasks() to improve load balancing outcomes

Problem:

The move_tasks() function is designed to move UP TO the amount of load it is asked to move and in doing this it skips over tasks looking for ones whose load weights are less than or equal to the remaining load to be moved. This is (in general) a good thing but it has the unfortunate result of breaking one of the original load balancer's good points: namely, that (within the limits imposed by the active/expired array model and the fact the expired is processed first) it moves high priority tasks before low priority ones and this means there's a good chance (see active/expired problem for why it's only a chance) that the highest priority task on the queue but not actually on the CPU will be moved to the other CPU where (as a high priority task) it may preempt the current task.

Solution:

Modify move_tasks() so that high priority tasks are not skipped when moving them will make them the highest priority task on their new run queue.

Signature · 4 lines
Peter Williams                                   [email protected]

"Learning, n. The kind of ignorance distinguishing the studious."
 -- Ambrose Bierce
kernel/sched.c +10 -2
Index: MM-2.6.17-rc1-mm2/kernel/sched.c
===================================================================
@@ -2043,7 +2043,7 @@ static int move_tasks(runqueue_t *this_r
 {
 	prio_array_t *array, *dst_array;
 	struct list_head *head, *curr;
-	int idx, pulled = 0, pinned = 0;
+	int idx, pulled = 0, pinned = 0, this_min_prio;
 	long rem_load_move;
 	task_t *tmp;
 
@@ -2052,6 +2052,7 @@ static int move_tasks(runqueue_t *this_r
 
 	rem_load_move = max_load_move;
 	pinned = 1;
+	this_min_prio = this_rq->curr->prio;
 
 	/*
 	 * We first consider expired tasks. Those will likely not be
@@ -2091,7 +2092,12 @@ skip_queue:
 
 	curr = curr->prev;
 
-	if (tmp->load_weight > rem_load_move ||
+	/*
+	 * To help distribute high priority tasks accross CPUs we don't
+	 * skip a task if it will be the highest priority task (i.e. smallest
+	 * prio value) on its new queue regardless of its load weight
+	 */
+	if ((idx >= this_min_prio && tmp->load_weight > rem_load_move) ||
 	    !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
 		if (curr != head)
 			goto skip_queue;
@@ -2113,6 +2119,8 @@ skip_queue:
 	 * and the prescribed amount of weighted load.
 	 */
 	if (pulled < max_nr_move && rem_load_move > 0) {
+		if (idx < this_min_prio)
+			this_min_prio = idx;
 		if (curr != head)
 			goto skip_queue;
 		idx++;
#1 lore raw
Quoted 7 lines · Peter Williams

Problem:

The move_tasks() function is designed to move UP TO the amount of load it is asked to move and in doing this it skips over tasks looking for ones whose load weights are less than or equal to the remaining load to be moved. This is (in general) a good thing but it has the unfortunate result of breaking one of the original load balancer's good points:

with previous load balancer code it was a good point.. because all tasks were weighted the same from load balancer perspective.. but now the imbalance represents what task to move (atleast in the working cases :)

Quoted 8 lines

namely, that (within the limits imposed by the active/expired array model and the fact the expired is processed first) it moves high priority tasks before low priority ones and this means there's a good chance (see active/expired problem for why it's only a chance) that the highest priority task on the queue but not actually on the CPU will be moved to the other CPU where (as a high priority task) it may preempt the current task.

Quoted 4 lines

Solution:

Modify move_tasks() so that high priority tasks are not skipped when moving them will make them the highest priority task on their new run queue.

you mean the highest priority task on the current active list of the new run queue, right?

There will be some unnecessary movements of high priority tasks because of this...

thanks,
suresh

Quoted 45 lines
Signature · 7 lines
Peter Williams                                   [email protected]

"Learning, n. The kind of ignorance distinguishing the studious."
  -- Ambrose Bierce

Index: MM-2.6.17-rc1-mm2/kernel/sched.c
===================================================================
kernel/sched.c +8 -2
@@ -2043,7 +2043,7 @@ static int move_tasks(runqueue_t *this_r
 {
 	prio_array_t *array, *dst_array;
 	struct list_head *head, *curr;
-	int idx, pulled = 0, pinned = 0;
+	int idx, pulled = 0, pinned = 0, this_min_prio;
 	long rem_load_move;
 	task_t *tmp;
 
@@ -2052,6 +2052,7 @@ static int move_tasks(runqueue_t *this_r
 
 	rem_load_move = max_load_move;
 	pinned = 1;
+	this_min_prio = this_rq->curr->prio;
 
 	/*
 	 * We first consider expired tasks. Those will likely not be
@@ -2091,7 +2092,12 @@ skip_queue:
 
 	curr = curr->prev;
 
-	if (tmp->load_weight > rem_load_move ||
+	/*
+	 * To help distribute high priority tasks accross CPUs we don't
+	 * skip a task if it will be the highest priority task (i.e. smallest
+	 * prio value) on its new queue regardless of its load weight
+	 */
+	if ((idx >= this_min_prio && tmp->load_weight > rem_load_move) ||
 	    !can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
 		if (curr != head)
 			goto skip_queue;

kernel/sched.c skip_queue: · line 2126

Quoted 8 lines
kernel/sched.c +2 -0
@@ -2113,6 +2119,8 @@ skip_queue:
 	 * and the prescribed amount of weighted load.
 	 */
 	if (pulled < max_nr_move && rem_load_move > 0) {
+		if (idx < this_min_prio)
+			this_min_prio = idx;
 		if (curr != head)
 			goto skip_queue;
 		idx++;

Peter, Are you sure that this is a converging solution? If we want to

#2 lore raw
Quoted 9 lines · Suresh B
Quoted 3 lines · Peter Williams

Problem:

The move_tasks() function is designed to move UP TO the amount of load it is asked to move and in doing this it skips over tasks looking for ones whose load weights are less than or equal to the remaining load to be moved. This is (in general) a good thing but it has the unfortunate result of breaking one of the original load balancer's good points:

with previous load balancer code it was a good point.. because all tasks were weighted the same from load balancer perspective.. but now the imbalance represents what task to move (atleast in the working cases :)

That's the option 4 case in my original mail. Are you suggesting that it would have been the better option to adopt? If so, why?

I don't like it because (leaving aside the active/expired array issues) it would move the X top priority tasks across where X is the number of tasks required to meet the requirement. If we could arrange for it to skip every second one without making the code too complicated (i.e. there'd be a possible need for multiple passes until the required load was moved).

Quoted 8 lines
Quoted 6 lines

namely, that (within the limits imposed by the active/expired array model and the fact the expired is processed first) it moves high priority tasks before low priority ones and this means there's a good chance (see active/expired problem for why it's only a chance) that the highest priority task on the queue but not actually on the CPU will be moved to the other CPU where (as a high priority task) it may preempt the current task.

Solution:

Modify move_tasks() so that high priority tasks are not skipped when moving them will make them the highest priority task on their new run queue.

you mean the highest priority task on the current active list of the new run queue, right?

Good point. this_min_prio should probably be initialized to the minimum of this_rq->curr->prio and this_rq->best_expired_prio rather just using this_rq->curr->prio.

Quoted 3 lines

There will be some unnecessary movements of high priority tasks because of this...

How so. At most one task per move_tasks() will be moved as a result of this code that wouldn't have been moved otherwise. That task will be a high priority task stuck behind a higher priority task on its current queue that will be the highest priority on its new queue causing a preempt and access to the CPU. I wouldn't call this unnecessary.

Quoted 1 line

Peter, Are you sure that this is a converging solution? If we want to

Yes, I think we're getting there.

I think we need changes to try_to_wake_up() to help high priority tasks find idle CPUs or CPUs where they would preempt when they wake up. Leaving this to the load balancer is bad for these tasks latencies. I think that this change needs to be done independently of smpnice as it would be useful even without smpnice. I'm hoping Ingo or Nick will comment on this proposal.

It would also help if you fixed the active load balance code so that it's not necessary to distort normal load balancing to accommodate it. I haven't had time to look at it myself (other than a quick glance) yet.

Peter

Signature · 4 lines
Peter Williams                                   [email protected]

"Learning, n. The kind of ignorance distinguishing the studious."
 -- Ambrose Bierce
#3 lore raw
Quoted 10 lines · Peter Williams
Quoted 9 lines · Suresh B
Quoted 7 lines · Peter Williams

Problem:

The move_tasks() function is designed to move UP TO the amount of load it is asked to move and in doing this it skips over tasks looking for ones whose load weights are less than or equal to the remaining load to be moved. This is (in general) a good thing but it has the unfortunate result of breaking one of the original load balancer's good points:

Quoted 7 lines
Quoted 4 lines

with previous load balancer code it was a good point.. because all tasks were weighted the same from load balancer perspective.. but now the imbalance represents what task to move (atleast in the working cases :)

That's the option 4 case in my original mail. Are you suggesting that it would have been the better option to adopt? If so, why?

No. I was not suggesting option-4. With this change in move_tasks, we will be overriding the decision what ever we made while calculating imbalance. Lets see for example, we have a simple DP system. With proc-0 running one high priority and one low priority task, Proc-1 running one low priority task. Ideally we would like to move low priority task from P0 to P1. But with this patch, we may end up moving high priority task from P0 to P1. But slowly after sometime(depending on high priority task is on active/expired list), we will converge to the expected solution..

Quoted 6 lines
Quoted 2 lines

you mean the highest priority task on the current active list of the new run queue, right?

Good point. this_min_prio should probably be initialized to the minimum of this_rq->curr->prio and this_rq->best_expired_prio rather just using this_rq->curr->prio.

yes.

Quoted 10 lines
Quoted 4 lines

There will be some unnecessary movements of high priority tasks because of this...

How so. At most one task per move_tasks() will be moved as a result of this code that wouldn't have been moved otherwise. That task will be a high priority task stuck behind a higher priority task on its current queue that will be the highest priority on its new queue causing a preempt and access to the CPU. I wouldn't call this unnecessary.

highest priority task can be in the expired list with normal priority task running.. as in my above example.

Quoted 12 lines
Quoted 2 lines

Peter, Are you sure that this is a converging solution? If we want to

Yes, I think we're getting there.

I think we need changes to try_to_wake_up() to help high priority tasks find idle CPUs or CPUs where they would preempt when they wake up. Leaving this to the load balancer is bad for these tasks latencies. I think that this change needs to be done independently of smpnice as it would be useful even without smpnice. I'm hoping Ingo or Nick will comment on this proposal.

Quoted 3 lines

It would also help if you fixed the active load balance code so that it's not necessary to distort normal load balancing to accommodate it. I haven't had time to look at it myself (other than a quick glance) yet.

The only special check in find_busiest_group() helping MT/MC balancing is pwr_now and pwr_move calculations.. These calculations will also help, in future when we are dealing with sched groups which are not symmetric. Asymmetries can be caused in scenarios like cpufreq, cpu logical hotplug..

I think we are unnecessarily behind active load balance...

thanks,
suresh

#4 lore raw
Quoted 12 lines · Suresh B
Quoted 10 lines · Peter Williams
Quoted 8 lines · Suresh B
Quoted 3 lines · Peter Williams

Problem:

The move_tasks() function is designed to move UP TO the amount of load it is asked to move and in doing this it skips over tasks looking for ones whose load weights are less than or equal to the remaining load to be moved. This is (in general) a good thing but it has the unfortunate result of breaking one of the original load balancer's good points:

with previous load balancer code it was a good point.. because all tasks were weighted the same from load balancer perspective.. but now the imbalance represents what task to move (atleast in the working cases :)

That's the option 4 case in my original mail. Are you suggesting that it would have been the better option to adopt? If so, why?

Quoted 8 lines

No. I was not suggesting option-4. With this change in move_tasks, we will be overriding the decision what ever we made while calculating imbalance. Lets see for example, we have a simple DP system. With proc-0 running
one high priority and one low priority task, Proc-1 running one
low priority task. Ideally we would like to move low priority task from
P0 to P1. But with this patch, we may end up moving high priority task
from P0 to P1. But slowly after sometime(depending on high priority task
is on active/expired list), we will converge to the expected
solution..

Yes, there are problems with the active/expired arrays but they're no worse with smpnice than they are without it.

Going to a single priority array would eliminate this problem not to mention other problems to do with when to excuse tasks from being put on the expired array at the end of their time slice (see Mike Galbraith's patches). But I haven't had much luck pushing that barrow. :-)

Quoted 4 lines
Quoted 3 lines
Quoted 2 lines

you mean the highest priority task on the current active list of the new run queue, right?

Good point. this_min_prio should probably be initialized to the minimum of this_rq->curr->prio and this_rq->best_expired_prio rather just using this_rq->curr->prio.

Quoted 8 lines

yes.

Quoted 3 lines
Quoted 2 lines

There will be some unnecessary movements of high priority tasks because of this...

How so. At most one task per move_tasks() will be moved as a result of this code that wouldn't have been moved otherwise. That task will be a high priority task stuck behind a higher priority task on its current queue that will be the highest priority on its new queue causing a preempt and access to the CPU. I wouldn't call this unnecessary.

highest priority task can be in the expired list with normal priority task running.. as in my above example.

Yes, but this will get it onto a CPU sooner so it's not all bad. Once again it's no worse than without smpnice.

I said in the patch description that there are active/expired array issues that make load balancing less than perfect with and without smpnice. How far we go trying to eliminate them is the question.

If you have a better suggestion for how move_tasks() does its job, how about providing a patch and with supporting arguments as to why its better. If it is better then we can use it.

Quoted 10 lines
Quoted 7 lines
Quoted 2 lines

Peter, Are you sure that this is a converging solution? If we want to

Yes, I think we're getting there.

I think we need changes to try_to_wake_up() to help high priority tasks find idle CPUs or CPUs where they would preempt when they wake up. Leaving this to the load balancer is bad for these tasks latencies. I think that this change needs to be done independently of smpnice as it would be useful even without smpnice. I'm hoping Ingo or Nick will comment on this proposal.

It would also help if you fixed the active load balance code so that it's not necessary to distort normal load balancing to accommodate it. I haven't had time to look at it myself (other than a quick glance) yet.

The only special check in find_busiest_group() helping MT/MC balancing is pwr_now and pwr_move calculations..

What about the very messy code I had to put in so that find_busiest_group() would return a group even if there were no queues in the group with more than one task. Similar for find_busiest_queue().

Quoted 3 lines

These calculations will also help,
in future when we are dealing with sched groups which are not symmetric. Asymmetries can be caused in scenarios like cpufreq, cpu logical hotplug..

I meant for you to move the (highly speculative) active load balancing trigger out of load_balance() so that we can stop returning busiest groups and queues that have no hope of participating in successful load balancing.

Quoted 2 lines

I think we are unnecessarily behind active load balance...

I'm not sure what you mean here. I'm not behind it at all. I can see the need for HT systems to be able to move the only running task of a CPU to another CPU in another package but I think the implementation is appalling. I think it needs a complete rewrite.

Peter

Signature · 4 lines
Peter Williams                                   [email protected]

"Learning, n. The kind of ignorance distinguishing the studious."
 -- Ambrose Bierce
#5 lore raw
Quoted 2 lines · Peter Williams

Yes, there are problems with the active/expired arrays but they're no worse with smpnice than they are without it.

With smpnice, if we make any wrong decision while moving the high and low priority tasks, we will enter an endless loop(as load balance will keep on showing the imbalance and move_tasks will always move the wrong tasks, instead of the correct ones pointed by imbalance..)

With the current code in the mainline, we don't have such a problem.

Quoted 4 lines

Going to a single priority array would eliminate this problem not to mention other problems to do with when to excuse tasks from being put on the expired array at the end of their time slice (see Mike Galbraith's patches). But I haven't had much luck pushing that barrow. :-)

I will take a loot at those patches.

Quoted 3 lines

If you have a better suggestion for how move_tasks() does its job, how about providing a patch and with supporting arguments as to why its better. If it is better then we can use it.

I think we can have a second pass(if the first pass fails to move any), in which we will not skip those tasks which will become highest priority on the target runqueue...

Quoted 6 lines
Quoted 2 lines

The only special check in find_busiest_group() helping MT/MC balancing is pwr_now and pwr_move calculations..

What about the very messy code I had to put in so that find_busiest_group() would return a group even if there were no queues in the group with more than one task. Similar for find_busiest_queue().

Thats messy for sure and that was introduced by you to fix an imabalance issue for a simple DP system(with out breaking HT systems). I will post a fix for that.

thanks,
suresh

#6 lore raw
Quoted 9 lines · Suresh B
Quoted 4 lines · Peter Williams

Yes, there are problems with the active/expired arrays but they're no

Quoted 2 lines

The only special check in find_busiest_group() helping MT/MC balancing is pwr_now and pwr_move calculations..

What about the very messy code I had to put in so that find_busiest_group() would return a group even if there were no queues in the group with more than one task. Similar for find_busiest_queue().

Thats messy for sure and that was introduced by you to fix an imabalance issue for a simple DP system(with out breaking HT systems). I will post a fix for that.

By moving the active load balance trigger out of load_balance() and providing an independent trigger that does not require sub optimal normal load balancing, I hope?

Peter

Signature · 4 lines
Peter Williams                                   [email protected]

"Learning, n. The kind of ignorance distinguishing the studious."
 -- Ambrose Bierce
#7 lore raw
In reply to Siddha, Suresh B
Quoted 6 lines · Suresh B
Quoted 1 line · Peter Williams

If you have a better suggestion for how move_tasks() does its job, how about providing a patch and with supporting arguments as to why its better. If it is better then we can use it.

I think we can have a second pass(if the first pass fails to move any), in which we will not skip those tasks which will become highest priority on the target runqueue...

That won't solve the problem that this patch is intended to address. As a reminder the problem is that a low priority task is being moved when we would prefer a high priority task to be moved. I.e. we want to move the high priority task because it's the best one to move not because we couldn't move any tasks.

NB (ignoring races and can_migrate_task() vetoing a task being moved) there should always be a task that can be moved as the minimum value of imbalance is the average load per task of the source queue and there must be tasks whose load weight is less than or equal to the average (it's axiomatic).

To put that another way, only a change in the state of the source queue since imbalance was determined (possible because that's all done without locks) or can_migrate_task() will stop at least one task being moved.

Peter

Signature · 4 lines
Peter Williams                                   [email protected]

"Learning, n. The kind of ignorance distinguishing the studious."
 -- Ambrose Bierce
#8 lore raw
In reply to Siddha, Suresh B
Quoted 7 lines · Suresh B
Quoted 1 line · Peter Williams

Yes, there are problems with the active/expired arrays but they're no worse with smpnice than they are without it.

With smpnice, if we make any wrong decision while moving the high and low priority tasks, we will enter an endless loop(as load balance will keep on showing the imbalance and move_tasks will always move the wrong tasks, instead of the correct ones pointed by imbalance..)

This will only happen if the load weight for the task moved is larger than the difference between the weighted loads for the two queues. So a check based on this observation can be used to prevent the loop.

Peter

Signature · 4 lines
Peter Williams                                   [email protected]

"Learning, n. The kind of ignorance distinguishing the studious."
 -- Ambrose Bierce