sched: move enough load to balance average load per task
PATCH 1/1 sched: move enough load to balance average load per task
Problem:
The current implementation of find_busiest_group() recognizes that approximately equal average loads per task for each group/queue are desirable (e.g. this condition will increase the probability that the top N highest priority tasks on an N CPU system will be on different CPUs) by being slightly more aggressive when *imbalance is small but the average load per task in "busiest" group is more than that in "this" group. Unfortunately, the amount moved from "busiest" to "this" is too small to reduce the average load per task on "busiest" (at best there will be no change and at worst it will get bigger).
Solution:
Increase the amount of load moved from "busiest" to "this" in these circumstances while making sure that the amount of load moved won't increase the (absolute) difference in the two groups' total weighted loads. A task with a weighted load greater than the average needs to be moved to cause the average to be reduced.
NB This makes no difference to load balancing for the case where all tasks have nice==0.
Signature · 4 lines
Peter Williams [email protected] "Learning, n. The kind of ignorance distinguishing the studious." -- Ambrose Bierce
kernel/sched.c +10 -6
@@ -2258,16 +2258,20 @@ find_busiest_group(struct sched_domain *if (*imbalance < busiest_load_per_task) {unsigned long pwr_now =0 , pwr_move =0 ;unsigned long tmp; -unsigned int imbn =2 ; -if (this_nr_running) { +if (this_nr_running) this_load_per_task /= this_nr_running; -if (busiest_load_per_task > this_load_per_task) - imbn =1 ; - }else +else this_load_per_task =SCHED_LOAD_SCALE ; -if (max_load - this_load >= busiest_load_per_task * imbn) { +if (busiest_load_per_task > this_load_per_task) { +unsigned long dld = max_load - this_load; + +if (dld > busiest_load_per_task) { + *imbalance = (dld + busiest_load_per_task) /2 ; +return busiest; + } + }else if (max_load - this_load >= busiest_load_per_task *2 ) { *imbalance = busiest_load_per_task;return busiest; }
Quoted 11 lines · Peter Williams
Problem:
The current implementation of
find_busiest_group()recognizes that approximately equal average loads per task for each group/queue are desirable (e.g. this condition will increase the probability that the top N highest priority tasks on an N CPU system will be on different CPUs) by being slightly more aggressive when *imbalance is small but the average load per task in "busiest" group is more than that in "this" group. Unfortunately, the amount moved from "busiest" to "this" is too small to reduce the average load per task on "busiest" (at best there will be no change and at worst it will get bigger).
Peter, We don't need to reduce the average load per task on "busiest" always. By moving a "busiest_load_per_task", we will increase the average load per task of lesser busy cpu (there by trying to achieve the equality with busiest...)
Can you give an example scenario where this patch helps? And doesn't the normal imabalance calculations capture those issues?
thanks,
suresh
Quoted 51 lines
Solution:
Increase the amount of load moved from "busiest" to "this" in these circumstances while making sure that the amount of load moved won't increase the (absolute) difference in the two groups' total weighted loads. A task with a weighted load greater than the average needs to be moved to cause the average to be reduced.
NB This makes no difference to load balancing for the case where all tasks have nice==0.
Signed-off-by Peter Williams <[email protected]>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 +10
-6@@ -2258,16 +2258,20 @@ find_busiest_group(struct sched_domain *if (*imbalance < busiest_load_per_task) {unsigned long pwr_now =0 , pwr_move =0 ;unsigned long tmp; -unsigned int imbn =2 ; -if (this_nr_running) { +if (this_nr_running) this_load_per_task /= this_nr_running; -if (busiest_load_per_task > this_load_per_task) - imbn =1 ; - }else +else this_load_per_task =SCHED_LOAD_SCALE ; -if (max_load - this_load >= busiest_load_per_task * imbn) { +if (busiest_load_per_task > this_load_per_task) { +unsigned long dld = max_load - this_load; + +if (dld > busiest_load_per_task) { + *imbalance = (dld + busiest_load_per_task) /2 ; +return busiest; + } + }else if (max_load - this_load >= busiest_load_per_task *2 ) { *imbalance = busiest_load_per_task;return busiest; }
Quoted 11 lines · Suresh B
Quoted 3 lines · Peter Williams
Problem:
The current implementation of
find_busiest_group()recognizes that approximately equal average loads per task for each group/queue are desirable (e.g. this condition will increase the probability that the top N highest priority tasks on an N CPU system will be on different CPUs) by being slightly more aggressive when *imbalance is small but the average load per task in "busiest" group is more than that in "this" group. Unfortunately, the amount moved from "busiest" to "this" is too small to reduce the average load per task on "busiest" (at best there will be no change and at worst it will get bigger).Peter, We don't need to reduce the average load per task on "busiest"
always. By moving a "busiest_load_per_task", we will increase the average load per task of lesser busy cpu (there by trying to achieve the equality with busiest...)Can you give an example scenario where this patch helps? And doesn't the normal imabalance calculations capture those issues?
Yes, I think that the normal imbalance calculations (in find_busiest_queue()) will generally capture the aim of having approximately equal average loads per task on run queues. But this bit of code is a special case in that the extra aggression being taken by the load balancer (in response to a scenario raised by you) is being justified by the imbalance in the average loads per task so it behooves us to do the best we can to ensure that that imbalance is addressed.
I don't think this is true for try_to_wake_up() and some changes may be desirable there. However, any such changes would interact with the RT load balancing that Ingo is working on and would need to be considered in conjunction with that.
Why I think "approximately equal average loads per task" is worthwhile secondary aim for the load balancer is because it helps restore an implicit aim (approximately equal numbers of tasks per run queue) that was present in the original version. This in turn means that the distribution of priorities within the queues will be similar and this increases the chances that (on an N CPU system) the N highest priority tasks will be on different CPUs. This is a desirable state of affairs.
Peter
Signature · 4 lines
Peter Williams [email protected] "Learning, n. The kind of ignorance distinguishing the studious." -- Ambrose Bierce
Quoted 9 lines · Peter Williams
Quoted 2 lines · Suresh B
Can you give an example scenario where this patch helps? And doesn't the normal imabalance calculations capture those issues?
Yes, I think that the normal imbalance calculations (in
find_busiest_queue()) will generally capture the aim of having approximately equal average loads per task on run queues. But this bit of code is a special case in that the extra aggression being taken by the load balancer (in response to a scenario raised by you) is being
Can you give a specific example which shows the problem and which you are trying to fix with this particular patch..
thanks,
suresh
Quoted 8 lines · Suresh B
Quoted 3 lines · Peter Williams
Problem:
The current implementation of
find_busiest_group()recognizes that approximately equal average loads per task for each group/queue are desirable (e.g. this condition will increase the probability that the top N highest priority tasks on an N CPU system will be on different CPUs) by being slightly more aggressive when *imbalance is small but the average load per task in "busiest" group is more than that in "this" group. Unfortunately, the amount moved from "busiest" to "this" is too small to reduce the average load per task on "busiest" (at best there will be no change and at worst it will get bigger).Peter, We don't need to reduce the average load per task on "busiest"
always. By moving a "busiest_load_per_task", we will increase the average load per task of lesser busy cpu (there by trying to achieve the equality with busiest...)
Well, first off, we don't always move busiest_load_per_task we move UP TO busiest_load_per_task so there is no way you can make definitive statements about what will happen to the value "this_load_per_task" as a result of setting *imbalance to busiest_load_per_task. Load balancing is a probabilistic endeavour and we need to take steps that increase the probability that we get the desired result.
Without this patch there is no chance that busiest_load_per_task will get smaller and whether this_load_per_task will get bigger is indeterminate. With this patch there IS a chance that busiest_load_per_task will decrease and an INCREASED chance that this_load_per_task will get bigger. Ergo we have increased the probability that the (absolute) difference between this_load_per_task and busiest_load_per_task will decrease. This is a desirable outcome.
Peter
Signature · 4 lines
Peter Williams [email protected] "Learning, n. The kind of ignorance distinguishing the studious." -- Ambrose Bierce
Quoted 19 lines · Peter Williams
Quoted 13 lines · Suresh B
Quoted 11 lines · Peter Williams
Problem:
The current implementation of
find_busiest_group()recognizes that approximately equal average loads per task for each group/queue are desirable (e.g. this condition will increase the probability that the top N highest priority tasks on an N CPU system will be on different CPUs) by being slightly more aggressive when *imbalance is small but the average load per task in "busiest" group is more than that in "this" group. Unfortunately, the amount moved from "busiest" to "this" is too small to reduce the average load per task on "busiest" (at best there will be no change and at worst it will get bigger).Quoted 4 lines
Peter, We don't need to reduce the average load per task on "busiest" always. By moving a "
busiest_load_per_task", we will increase the average load per task of lesser busy cpu (there by trying to achieve the equality with busiest...)
Quoted 6 lines
Well, first off, we don't always move
busiest_load_per_taskwe move UP TObusiest_load_per_taskso there is no way you can make definitive statements about what will happen to the value "this_load_per_task" as a result of setting *imbalance tobusiest_load_per_task. Load balancing is a probabilistic endeavour and we need to take steps that increase the probability that we get the desired result.
I agree with you. But the previous code was more conservative and may slowly (just from theory pt of view... I don't have an example to show this..) balance towards the desired state. With this code, I feel we are aggressive. for example, on a DP system: if I run one high priority and two low priority processes, they keep hopping from one processor to another... you may argue it is because of the "top" or some other process... I agree that it is the case.. But same thing doesn't happen with the previous version.. I like the conservative approach...
Quoted 2 lines
Without this patch there is no chance that
busiest_load_per_taskwill get smaller
Is there an example for this?
Quoted 6 lines
and whether
this_load_per_taskwill get bigger is indeterminate. With this patch there IS a chance thatbusiest_load_per_taskwill decrease and an INCREASED chance thatthis_load_per_taskwill get bigger. Ergo we have increased the probability that the (absolute) difference betweenthis_load_per_taskandbusiest_load_per_taskwill decrease. This is a desirable outcome.
All I am saying is we are more aggressive.. I don't have any issue with the desired outcome..
thanks,
suresh
Quoted 20 lines · Suresh B
Quoted 9 lines · Peter Williams
Quoted 7 lines · Suresh B
Quoted 3 lines · Peter Williams
Problem:
The current implementation of
find_busiest_group()recognizes that approximately equal average loads per task for each group/queue are desirable (e.g. this condition will increase the probability that the top N highest priority tasks on an N CPU system will be on different CPUs) by being slightly more aggressive when *imbalance is small but the average load per task in "busiest" group is more than that in "this" group. Unfortunately, the amount moved from "busiest" to "this" is too small to reduce the average load per task on "busiest" (at best there will be no change and at worst it will get bigger).Peter, We don't need to reduce the average load per task on "busiest"
always. By moving a "busiest_load_per_task", we will increase the average load per task of lesser busy cpu (there by trying to achieve the equality with busiest...)Well, first off, we don't always move
busiest_load_per_taskwe move UP TObusiest_load_per_taskso there is no way you can make definitive statements about what will happen to the value "this_load_per_task" as a result of setting *imbalance tobusiest_load_per_task. Load balancing is a probabilistic endeavour and we need to take steps that increase the probability that we get the desired result.I agree with you. But the previous code was more conservative and may slowly (just from theory pt of view... I don't have an example to show this..) balance towards the desired state. With this code, I feel we are aggressive. for example, on a DP system: if I run one high priority and two low priority processes, they keep hopping from one processor to another... you may argue it is because of the "top" or some other process... I agree that it is the case.. But same thing doesn't happen with the previous version.. I like the conservative approach...
Quoted 3 lines
Quoted 1 line
Without this patch there is no chance that
busiest_load_per_taskwill get smallerIs there an example for this?
Yes, we just take a slight variation of your scenario that prompted the first patch (to which this patch is a minor modification) by adding one normal priority task to each of the CPUs. This gives us a 2 CPU system with CPU-0 having 2 high priority tasks plus 1 normal priority task and CPU-1 having two normal priority tasks. Clearly, the desirable load balancing outcome would be for the two high priority tasks to be on different CPUs otherwise we have a high priority task stuck on a run queue while a normal priority is running on another (less heavily loaded) CPU.
In order to analyze what happens during load balancing, let's use W as the load weight for a normal task and suppose that the load weights of the two high priority tasks are (W + k) and that "this" == CPU-1 in find_busiest_queue(). This will result in "busiest" == CPU-0 and:
this_load = 2Wthis_load_per_task = Wmax_load = 3W + 2kbusiest_load_per_task = W + 2k / 3avg_load = 5W / 2 + kmax_pull = W / 2 + k
*imbalance = W / 2 + k
Whenever k < (3W / 2) this will result in *imbalance < busiest_load_per_task and we end up in the small imbalance code.
(max_load - this_load) = W + 2k which is greater than busiest_load_per_task so we decide that we want to move some load from "busiest" to "this".
Without this patch we would set *imbalance to busiest_load_per_task and the only task on "busiest" that has a weighted load less than or equal to this value is the normal task so this is the one that will be moved resulting:
this_load = 3Wthis_load_per_task = Wmax_load = 2W + 2kbusiest_load_per_task = W + k
Even if you reverse the roles of "busiest" and "this", this will be considered balanced and the system will stabilize in this undesirable state. NB, as predicted, the average load per task on "this" hasn't changed and the average load per task on "busiest" has increased. We still have the situation where a high priority task is stuck on a run queue while a low priority task is running on another CPU -- we've failed :-(.
With this patch, *imbalance will be set to (W + 4k / 3) which is bigger than the weighted load of the high priority tasks so one of them will be moved resulting in:
this_load = 3W + kthis_load_per_task = W + k / 3max_load = 2W + kbusiest_load_per_task = W + k / 2
Quoted 5 lines
Quoted 2 lines
and whether
this_load_per_taskwill get bigger is indeterminate. With this patch there IS a chance thatbusiest_load_per_taskwill decrease and an INCREASED chance thatthis_load_per_taskwill get bigger. Ergo we have increased the probability that the (absolute) difference betweenthis_load_per_taskandbusiest_load_per_taskwill decrease. This is a desirable outcome.All I am saying is we are more aggressive.. I don't have any issue with the desired outcome..
We need to be more aggressive but not too aggressive and I think this patch achieves the required balance.
NB busiest_load_per_task < *imbalance < (max_load - this_load) is true for this path through the code. To be precise, *imbalance will be half way between busiest_load_per_task and (max_load - this_load).
Peter
Signature · 4 lines
Peter Williams [email protected] "Learning, n. The kind of ignorance distinguishing the studious." -- Ambrose Bierce
Quoted 43 lines · Peter Williams
Quoted 1 line · Suresh B
Is there an example for this?
Yes, we just take a slight variation of your scenario that prompted the first patch (to which this patch is a minor modification) by adding one normal priority task to each of the CPUs. This gives us a 2 CPU system with CPU-0 having 2 high priority tasks plus 1 normal priority task and CPU-1 having two normal priority tasks. Clearly, the desirable load balancing outcome would be for the two high priority tasks to be on different CPUs otherwise we have a high priority task stuck on a run queue while a normal priority is running on another (less heavily loaded) CPU.
In order to analyze what happens during load balancing, let's use W as the load weight for a normal task and suppose that the load weights of the two high priority tasks are (W + k) and that "this" == CPU-1 in
find_busiest_queue(). This will result in "busiest" == CPU-0 and:
this_load= 2Wthis_load_per_task= Wmax_load= 3W + 2kbusiest_load_per_task= W + 2k / 3avg_load= 5W / 2 + kmax_pull= W / 2 + k
*imbalance = W / 2 + kWhenever k < (3W / 2) this will result in *imbalance <
busiest_load_per_taskand we end up in the small imbalance code.(
max_load-this_load) = W + 2k which is greater thanbusiest_load_per_taskso we decide that we want to move some load from "busiest" to "this".Without this patch we would set *imbalance to
busiest_load_per_taskand the only task on "busiest" that has a weighted load less than or equal to this value is the normal task so this is the one that will be moved resulting:
this_load= 3Wthis_load_per_task= Wmax_load= 2W + 2kbusiest_load_per_task= W + k
Quoted 7 lines
Even if you reverse the roles of "busiest" and "this", this will be considered balanced and the system will stabilize in this undesirable state. NB, as predicted, the average load per task on "this" hasn't changed and the average load per task on "busiest" has increased. We still have the situation where a high priority task is stuck on a run queue while a low priority task is running on another CPU -- we've failed :-(.
for such a 'k' value, we fail anyhow. For example, how does the normal load balance detect an imbalance in this below situation?
this_load = 3Wthis_load_per_task = Wmax_load = 2W + 2kbusiest_load_per_task = W + k
if we really want to distribute 'N' higher priority tasks(however small or big is the priority difference between low and high priority tasks) on to 'N' different cpus, we will need really different approach for load balancing..
thanks,
suresh
Quoted 34 lines
With this patch, *imbalance will be set to (W + 4k / 3) which is bigger than the weighted load of the high priority tasks so one of them will be moved resulting in:
this_load= 3W + kthis_load_per_task= W + k / 3max_load= 2W + kbusiest_load_per_task= W + k / 2Quoted 10 lines
Quoted 7 lines
and whether
this_load_per_taskwill get bigger is indeterminate. With this patch there IS a chance thatbusiest_load_per_taskwill decrease and an INCREASED chance thatthis_load_per_taskwill get bigger. Ergo we have increased the probability that the (absolute) difference betweenthis_load_per_taskandbusiest_load_per_taskwill decrease. This is a desirable outcome.All I am saying is we are more aggressive.. I don't have any issue with the desired outcome..
We need to be more aggressive but not too aggressive and I think this patch achieves the required balance.
NB
busiest_load_per_task< *imbalance < (max_load-this_load) is true for this path through the code. To be precise, *imbalance will be half way betweenbusiest_load_per_taskand (max_load-this_load).Peter
Signature · 4 lines
Peter Williams [email protected] "Learning, n. The kind of ignorance distinguishing the studious." -- Ambrose Bierce
Quoted 28 lines · Suresh B
Quoted 18 lines · Peter Williams
Quoted 1 line · Suresh B
Is there an example for this?
Yes, we just take a slight variation of your scenario that prompted the first patch (to which this patch is a minor modification) by adding one normal priority task to each of the CPUs. This gives us a 2 CPU system with CPU-0 having 2 high priority tasks plus 1 normal priority task and CPU-1 having two normal priority tasks. Clearly, the desirable load balancing outcome would be for the two high priority tasks to be on different CPUs otherwise we have a high priority task stuck on a run queue while a normal priority is running on another (less heavily loaded) CPU.
In order to analyze what happens during load balancing, let's use W as the load weight for a normal task and suppose that the load weights of the two high priority tasks are (W + k) and that "this" == CPU-1 in
find_busiest_queue(). This will result in "busiest" == CPU-0 and:
this_load= 2Wthis_load_per_task= Wmax_load= 3W + 2kbusiest_load_per_task= W + 2k / 3avg_load= 5W / 2 + kmax_pull= W / 2 + k
*imbalance = W / 2 + kWhenever k < (3W / 2) this will result in *imbalance <
busiest_load_per_taskand we end up in the small imbalance code.(
max_load-this_load) = W + 2k which is greater thanbusiest_load_per_taskso we decide that we want to move some load from "busiest" to "this".Quoted 8 lines
Without this patch we would set *imbalance to
busiest_load_per_taskand the only task on "busiest" that has a weighted load less than or equal to this value is the normal task so this is the one that will be moved resulting:
this_load= 3Wthis_load_per_task= Wmax_load= 2W + 2kbusiest_load_per_task= W + kEven if you reverse the roles of "busiest" and "this", this will be considered balanced and the system will stabilize in this undesirable state. NB, as predicted, the average load per task on "this" hasn't changed and the average load per task on "busiest" has increased. We still have the situation where a high priority task is stuck on a run queue while a low priority task is running on another CPU -- we've failed :-(.
Quoted 7 lines
for such a 'k' value, we fail anyhow. For example, how does the normal load balance detect an imbalance in this below situation?
this_load= 3Wthis_load_per_task= Wmax_load= 2W + 2kbusiest_load_per_task= W + k
Yes, it's hard to get out of such a situation if you get into one so that's why changes to try_to_wake_up() may be needed. We certainly have to stop the load balancing code from creating these situations as well.
Quoted 3 lines
if we really want to distribute 'N' higher priority tasks(however small or
big is the priority difference between low and high priority tasks) on to 'N' different cpus, we will need really different approach for load balancing..
Yes, I've said similar in another thread but I agreed with Ingo when he said that this wasn't really a problem for the load balancer to solve. I expressed the same opinion as above, namely that this problem needs to be addressed in try_to_wake_up() (which isn't really load balancing) and suggested that (for high priority tasks) try_to_wake_up() should be modified to find either an idle CPU or (if it can't find an idle one) the CPU with the lowest priority current task.
However, it should be noted that Ingo is working on something for ensuring the distribution of RT tasks across CPUs and this is likely to overlap with this idea so consultation is necessary.
Peter
Signature · 4 lines
Peter Williams [email protected] "Learning, n. The kind of ignorance distinguishing the studious." -- Ambrose Bierce
A message that is not in the archive (sent off-list or not yet received): [email protected]
smpnice: issues with finding busiest queue
Quoted 3 lines · Peter Williams
There be dragons here :-(.
At more places in this part of the world (smpnice) :)
We need to relook at find_busiest_queue()... With the current weighted calculations, it doesn't always make sense to look at the highest weighted runqueue in the busy group..
for example on a DP with HT system, how does the load balance behave with Package-0 containing one high priority and one low priority, Package-1 containing one low priority(with other thread being idle)..
Package-1 thinks that it need to take the low priority thread from Package-0. And find_busiest_queue() returns the cpu thread with highest priority task.. And ultimately(with help of active load balance) we move high priority task to Package-1. And same continues with Package-0 now, moving high priority task from package-1 to package-0..
Even without the presence of active load balance, load balance will fail to balance(having two low priority tasks on one package, and high priority task on another package) the above scenario....
We probably need to use imbalance(and more factors) to determine the busiest queue in the group.....
thanks,
suresh
Re: smpnice: issues with finding busiest queue
Quoted 20 lines · Suresh B
Quoted 2 lines · Peter Williams
There be dragons here :-(.
At more places in this part of the world (smpnice) :)
We need to relook at
find_busiest_queue()... With the current weighted calculations, it doesn't always make sense to look at the highest weighted runqueue in the busy group..for example on a DP with HT system, how does the load balance behave with Package-0 containing one high priority and one low priority, Package-1 containing one low priority(with other thread being idle)..
Package-1 thinks that it need to take the low priority thread from Package-0. And
find_busiest_queue()returns the cpu thread with highest priority task.. And ultimately(with help of active load balance) we move high priority task to Package-1. And same continues with Package-0 now, moving high priority task from package-1 to package-0..
Quoted 5 lines
Even without the presence of active load balance, load balance will fail to
balance(having two low priority tasks on one package, and high priority task on another package)the above scenario....We probably need to use imbalance(and more factors) to determine the busiest queue in the group.....
A patch would be nice.
Signature · 4 lines
Peter Williams [email protected] "Learning, n. The kind of ignorance distinguishing the studious." -- Ambrose Bierce