[RFC PATCH 0/2] sched: Load Balancing using Per-entity-Load-tracking

[RFC PATCH 0/2] sched: Load Balancing using Per-entity-Load-tracking

[preeti]

Hi Guys,

Can you please have a look at the below patchset? Your review comments
are very necessary and valuable.Thanks in advance.

> This patchset uses the per-entity-load-tracking patchset which will soon be
> available in the kernel.It is based on the tip/master tree and the first 8
> latest patches of sched:per-entity-load-tracking alone have been imported to
> the tree to avoid the complexities of task groups and to hold back the
> optimizations of this patch for now.
> 
> This patchset is an attempt to begin the integration of Per-entity-load-
> metric for the cfs_rq,henceforth referred to as PJT's metric,with the load
> balancer in a step wise fashion,and progress based on the consequences.
> 
> The following issues have been considered towards this:
> [NOTE:an x% task referred to in the logs and below is calculated over a
> duty cycle of 10ms.]
> 
> 1.Consider a scenario,where there are two 10% tasks running on a cpu.The
>   present code will consider the load on this queue to be 2048,while
>   using PJT's metric the load is calculated to be <1000,rarely exceeding this
>   limit.Although the tasks are not contributing much to the cpu load,they are
>   decided to be moved by the scheduler.
> 
>   But one could argue that 'not moving one of these tasks could throttle
>   them.If there was an idle cpu,perhaps we could have moved them'.While the
>   power save mode would have been fine with not moving the task,the
>   performance mode would prefer not to throttle the tasks.We could strive
>   to strike a balance by making this decision tunable with certain parameters.
>   This patchset includes such tunables.This issue is addressed in Patch[1/2].
> 
> 2.We need to be able to do this cautiously,as the scheduler code is too
>   complex.This patchset is an attempt to begin the integration of PJT's
>   metric with the load balancer in a step wise fashion,and progress based on
>   the consequences.
>   I dont intend to vary the parameters used by the load balancer.Some
>   parameters are however included anew to make decisions about including a
>   sched group as a candidate for load balancing.
> 
>   This patchset therefore has two primary aims.
>          Patch[1/2]: This patch aims at detecting short running tasks and
> 	 prevent their movement.In update_sg_lb_stats,dismiss a sched group
> 	 as a candidate for load balancing,if load calculated by PJT's metric
> 	 says that the average load on the sched_group <= 1024+(.15*1024).
> 	 This is a tunable,which can be varied after sufficient experiments.
> 
>          Patch[2/2]:In the current scheduler greater load would be analogous
>          to more number of tasks.Therefore when the busiest group is picked
>          from the sched domain in update_sd_lb_stats,only the loads of the
>          groups are compared between them.If we were to use PJT's metric,a
>          higher load does not necessarily mean more number of tasks.This
> 	 patch addresses this issue.
> 
> 3.The next step towards integration should be in using the PJT's metric for
>   comparison between the loads of the busy sched group and the sched
>   group which has to pull the tasks,which happens in find_busiest_group.
> ---
> 
> Preeti U Murthy (2):
>       sched:Prevent movement of short running tasks during load balancing
>       sched:Pick the apt busy sched group during load balancing
> 
> 
>  kernel/sched/fair.c |   38 +++++++++++++++++++++++++++++++++++---
>  1 file changed, 35 insertions(+), 3 deletions(-)
> 
> --
The links to PATCH[1/2] https://lkml.org/lkml/2012/10/12/13
             PATCH[2/2] https://lkml.org/lkml/2012/10/12/11
Regards
Preeti U Murthy

[RFC PATCH 0/2] sched: Load Balancing using Per-entity-Load-tracking

[preeti]

Hi Morten,
Thank you very much for your review.

>> 1.Consider a scenario,where there are two 10% tasks running on a cpu.The
>>   present code will consider the load on this queue to be 2048,while
>>   using PJT's metric the load is calculated to be <1000,rarely exceeding this
>>   limit.Although the tasks are not contributing much to the cpu load,they are
>>   decided to be moved by the scheduler.
> 
> I guess that you assume, for now, that all tasks have default (nice 0)
> priority? Both the old load and the PJT metric (tracked load) depends on
> priority.
Thats right.I have assumed default priority of the tasks.
> 
>>
>>   But one could argue that 'not moving one of these tasks could throttle
>>   them.If there was an idle cpu,perhaps we could have moved them'.While the
>>   power save mode would have been fine with not moving the task,the
>>   performance mode would prefer not to throttle the tasks.We could strive
>>   to strike a balance by making this decision tunable with certain parameters.
>>   This patchset includes such tunables.This issue is addressed in Patch[1/2].
>>
> 
> One could also argue that long as there are spare cpu cycles in each
> schedule period then all tasks have received the cpu time they needed.
> So from that point of view performance isn't affected by not balancing
> the tasks as long as the cpu is not fully utilized. If we look at the
> problem from a latency point of view then packing tasks on a single cpu
> will increase latency but the increase will be bounded by the schedule
> period.
> 
Assume that at the end of one scheduling period,there are a few spare
cycles on the cpu.this is fine from both the performance and latency
point of view at *this* point.nobody is waiting for the cpu.

The issue arises if it is detected that these spare cycles are due to
*sleeping tasks* and not due to no tasks.

At this point a decision needs to be made as to: if a scenario arises
where all these tasks wake up at the same time in the future,and wait on
the cpu,then are we ok with them waiting.Both performance and latency
views could be against this,as this also means less throughput.But
performance view could go slightly easy on this to argue,that its ok if
2-3 tasks wait,if more,then there is a need to move them.
>>   This patchset therefore has two primary aims.
>>          Patch[1/2]: This patch aims at detecting short running tasks and
>> 	 prevent their movement.In update_sg_lb_stats,dismiss a sched group
>> 	 as a candidate for load balancing,if load calculated by PJT's metric
>> 	 says that the average load on the sched_group <= 1024+(.15*1024).
>> 	 This is a tunable,which can be varied after sufficient experiments.
> 
> Your current threshold implies that there must be at least two (nice 0)
> tasks running breach the threshold and they need to be quite busy. This
> makes sense to me. When you have more tasks they are more likely to be
> waiting on the runqueue even if it is only 10% tasks. Let's say you have
> five 10% tasks and they all become runnable at the same instant. In that
> case some of the tasks would have a tracked load which is much higher
> than if we only had two 10% tasks running. So if I'm not mistaken, it
> would be possible to breach the threshold even though the overall cpu
> utilization is only 50% and it would have been safe not to load-balance
> that cpu.
> 
> Do you think it would make sense to let the threshold depend on the
> number of task on the cpu somehow? 

You are right,Morten.In fact I have included this viewpoint in both my
first and second patch enclosed by this. So lets take up the above
scenario.if there are 5 10% tasks running,they will surely cross the
threshold,but the cpu might have spare cycles at the end of a scheduling
period.Now that is your concern.

Again we have two different viewpoints.This threshold is like a tuning
knob.we could increase it if we feel that this threshold gets reached
very quickly with as few tasks as 5, although the cpu utilization is
poor.we prefer not to wake up another cpu unless the present cpu is
aptly loaded.we could call this the power saving view.

Else we could say that,we are not intending to affect the throughput of
tasks,so we prefer the knob be at this value,so that we qualify such a
load as a candidate for load balancing.we could call this the
performance view.
> 
> Alternative, the decision could be based on the cpu idle time over the
> last schedule period. A cpu with no or very few spare cycles in the last
> schedule period would be a good candidate for load-balancing. Latency
> would be affected as mentioned earlier.
> 
Exactly.idle_time == spare_cpu_cycles == less cpu_utilization.I hope i
am not wrong in drawing this equivalence.if thats the case then the same
explanation as above holds good here too.
> 
> Morten

Thank you

Regards
Preeti

[RFC PATCH 0/2] sched: Load Balancing using Per-entity-Load-tracking

[preeti]

Sorry guys this mail had problems getting sent.hence the repost.

Hi Morten,
Thank you very much for your review.

>> 1.Consider a scenario,where there are two 10% tasks running on a cpu.The
>>   present code will consider the load on this queue to be 2048,while
>>   using PJT's metric the load is calculated to be <1000,rarely exceeding this
>>   limit.Although the tasks are not contributing much to the cpu load,they are
>>   decided to be moved by the scheduler.
>
> I guess that you assume, for now, that all tasks have default (nice 0)
> priority? Both the old load and the PJT metric (tracked load) depends on
> priority.

Thats right.I have assumed default priority of the tasks.
>>
>>   But one could argue that 'not moving one of these tasks could throttle
>>   them.If there was an idle cpu,perhaps we could have moved them'.While the
>>   power save mode would have been fine with not moving the task,the
>>   performance mode would prefer not to throttle the tasks.We could strive
>>   to strike a balance by making this decision tunable with certain parameters.
>>   This patchset includes such tunables.This issue is addressed in Patch[1/2].
>>
>
> One could also argue that long as there are spare cpu cycles in each
> schedule period then all tasks have received the cpu time they needed.
> So from that point of view performance isn't affected by not balancing
> the tasks as long as the cpu is not fully utilized. If we look at the
> problem from a latency point of view then packing tasks on a single cpu
> will increase latency but the increase will be bounded by the schedule
> period.
>
Assume that at the end of one scheduling period,there are a few spare
cycles on the cpu.this is fine from both the performance and latency
point of view at *this* point.nobody is waiting for the cpu.

The issue arises if it is detected that these spare cycles are due to
*sleeping tasks* and not due to no tasks.

At this point a decision needs to be made as to: if a scenario arises
where all these tasks wake up at the same time in the future,and wait on
the cpu,then are we ok with them waiting.Both performance and latency
views could be against this,as this also means less throughput.But
performance view could go slightly easy on this to argue,that its ok if
2-3 tasks wait,if more,then there is a need to move them.

>>   This patchset therefore has two primary aims.
>>          Patch[1/2]: This patch aims at detecting short running tasks and
>> 	 prevent their movement.In update_sg_lb_stats,dismiss a sched group
>> 	 as a candidate for load balancing,if load calculated by PJT's metric
>> 	 says that the average load on the sched_group <= 1024+(.15*1024).
>> 	 This is a tunable,which can be varied after sufficient experiments.
>
> Your current threshold implies that there must be at least two (nice 0)
> tasks running breach the threshold and they need to be quite busy. This
> makes sense to me. When you have more tasks they are more likely to be
> waiting on the runqueue even if it is only 10% tasks. Let's say you have
> five 10% tasks and they all become runnable at the same instant. In that
> case some of the tasks would have a tracked load which is much higher
> than if we only had two 10% tasks running. So if I'm not mistaken, it
> would be possible to breach the threshold even though the overall cpu
> utilization is only 50% and it would have been safe not to load-balance
> that cpu.
>
> Do you think it would make sense to let the threshold depend on the
> number of task on the cpu somehow? 
> 
You are right,Morten.In fact I have included this viewpoint in both my
first and second patch enclosed by this. So lets take up the above
scenario.if there are 5 10% tasks running,they will surely cross the
threshold,but the cpu might have spare cycles at the end of a scheduling
period.Now that is your concern.

Again we have two different viewpoints.This threshold is like a tuning
knob.we could increase it if we feel that this threshold gets reached
very quickly with as few tasks as 5, although the cpu utilization is
poor.we prefer not to wake up another cpu unless the present cpu is
aptly loaded.we could call this the power saving view.

Else we could say that,we are not intending to affect the throughput of
tasks,so we prefer the knob be at this value,so that we qualify such a
load as a candidate for load balancing.we could call this the
performance view.

> Alternative, the decision could be based on the cpu idle time over the
> last schedule period. A cpu with no or very few spare cycles in the last
> schedule period would be a good candidate for load-balancing. Latency
> would be affected as mentioned earlier.
>
Exactly.idle_time == spare_cpu_cycles == less cpu_utilization.I hope i
am not wrong in drawing this equivalence.if thats the case then the same
explanation as above holds good here too.

> Morten

Thank you

Regards
Preeti