Re: Fwd: Re: RFC: cgroups aware proc

Re: Fwd: Re: RFC: cgroups aware proc

[Peter Zijlstra]

On Mon, Jan 13, 2014 at 06:23:50PM +0200, Marian Marinov wrote:
> Hello Peter,
> 
> I need help with the scheduler.
> 
> I'm currently trying to patch the /proc/loadavg to show the load that is
> only related to the processes from the current cgroup.
> 
> I looked trough the code and I was hoping that tsk->sched_task_group->cfs_rq
> struct will give me the needed information, but unfortunately for me, it did
> not.
> 
> Can you advise me, how to approach this problem?

Yeah, don't :-) Really, loadavg is a stupid metric.

> I'm totally new to the scheduler code.

Luckily you won't actually have to touch much of it. Most of the actual
loadavg code lives in the first ~400 lines of kernel/sched/proc.c, read
and weep. Its one of the best documented bits around.

Your proposition however is extremely expensive, you turn something
that's already expensive O(nr_cpus) into something O(nr_cpus *
nr_cgroups).

I'm fairly sure people will not like that, esp. for something of such
questionable use as the loadavg -- its really only a pretty number that
doesn't mean all that much.

> -------- Original Message --------
> From: Li Zefan <lizefan@huawei.com>
>
> Then you should add Peter, Ingo and LKML to your Cc list. :)

You failed that, let me fix that.

Re: RFC: cgroups aware proc

[Marian Marinov]

On 01/13/2014 07:12 PM, Peter Zijlstra wrote:
> On Mon, Jan 13, 2014 at 06:23:50PM +0200, Marian Marinov wrote:
>> Hello Peter,
>>
>> I need help with the scheduler.
>>
>> I'm currently trying to patch the /proc/loadavg to show the load that is
>> only related to the processes from the current cgroup.
>>
>> I looked trough the code and I was hoping that tsk->sched_task_group->cfs_rq
>> struct will give me the needed information, but unfortunately for me, it did
>> not.
>>
>> Can you advise me, how to approach this problem?
>
> Yeah, don't :-) Really, loadavg is a stupid metric.

Yes... stupid, but unfortunately everyone is looking at it :(

>
>> I'm totally new to the scheduler code.
>
> Luckily you won't actually have to touch much of it. Most of the actual
> loadavg code lives in the first ~400 lines of kernel/sched/proc.c, read
> and weep. Its one of the best documented bits around.

I looked trough it but I don't understand how to introduce the per cgroup calculation.

I looked trough the headers and found the following, which is already implemented.

   task->sched_task_group->load_avg
   task->sched_task_group->cfs_rq->load_avg
   task->sched_task_group->cfs_rq->load.weight
   task->sched_task_group->cfs_rq->runnable_load_avg

Unfortunately there is almost no documentation for these elements of the cfs_rq and task_group structs.

It seams to me that part of the per task group loadavg code is already present.

>
> Your proposition however is extremely expensive, you turn something
> that's already expensive O(nr_cpus) into something O(nr_cpus *
> nr_cgroups).
>
> I'm fairly sure people will not like that, esp. for something of such
> questionable use as the loadavg -- its really only a pretty number that
> doesn't mean all that much.

I know that its use is questionable but in my case I need to have it, or I will not be able to offer correct loadavg 
values in the containers.


>
>> -------- Original Message --------
>> From: Li Zefan <lizefan@huawei.com>
>>
>> Then you should add Peter, Ingo and LKML to your Cc list. :)
>
> You failed that, let me fix that.
>
>

Re: RFC: cgroups aware proc

[Peter Zijlstra]

On Tue, Jan 14, 2014 at 02:58:14AM +0200, Marian Marinov wrote:
> On 01/13/2014 07:12 PM, Peter Zijlstra wrote:
> >On Mon, Jan 13, 2014 at 06:23:50PM +0200, Marian Marinov wrote:
> >>Hello Peter,
> >>
> >>I need help with the scheduler.
> >>
> >>I'm currently trying to patch the /proc/loadavg to show the load that is
> >>only related to the processes from the current cgroup.
> >>
> >>I looked trough the code and I was hoping that tsk->sched_task_group->cfs_rq
> >>struct will give me the needed information, but unfortunately for me, it did
> >>not.
> >>
> >>Can you advise me, how to approach this problem?
> >
> >Yeah, don't :-) Really, loadavg is a stupid metric.
> 
> Yes... stupid, but unfortunately everyone is looking at it :(
> 
> >
> >>I'm totally new to the scheduler code.
> >
> >Luckily you won't actually have to touch much of it. Most of the actual
> >loadavg code lives in the first ~400 lines of kernel/sched/proc.c, read
> >and weep. Its one of the best documented bits around.
> 
> I looked trough it but I don't understand how to introduce the per cgroup calculation.
> 
> I looked trough the headers and found the following, which is already implemented.
> 
>   task->sched_task_group->load_avg
>   task->sched_task_group->cfs_rq->load_avg
>   task->sched_task_group->cfs_rq->load.weight
>   task->sched_task_group->cfs_rq->runnable_load_avg
> 
> Unfortunately there is almost no documentation for these elements of the cfs_rq and task_group structs.
> 
> It seams to me that part of the per task group loadavg code is already present.

No, those are actual load metrics and completely unrelated to loadavg.

Loadavg requires per-cgroup-per-cpu variants of nr_running and
nr_uninterruptible.

Those are the only metrics used in kernel/sched/proc.c for loadavg.