How to measure time spent in context switch

How to measure time spent in context switch

[Manavendra Nath Manav]

How to measure time spent in context switch in Linux i.e. the time the
application stays suspended while the kernel is doing something else. I
don't want to use any profiling tools. What approach should I take to write
my own C code, which can also run as kernel module if required, to measure
time spent accurately. Will it also require to modify the scheduler code?
Does /proc gives any statistics in this regard?

Ref:
http://stackoverflow.com/questions/2368384/write-a-c-program-to-measure-time-spent-in-context-switch-in-linux-os

--
Manavendra Nath Manav
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.kernelnewbies.org/pipermail/kernelnewbies/attachments/20120717/bee032ab/attachment.html 

How to measure time spent in context switch

[Victor Rodriguez]

Hi Manavendra

Have you think in a driver that you can call when the context switch
happen? If this can be detected (more investigation on this area is
required). If that is possible we can create the driver to measure the
time according to the interruption.

Thanks a lot and regards

Victor Rodriguez



On Mon, Jul 16, 2012 at 2:28 PM, Manavendra Nath Manav
<mnm.kernel@gmail.com> wrote:
> How to measure time spent in context switch in Linux i.e. the time the
> application stays suspended while the kernel is doing something else. I
> don't want to use any profiling tools. What approach should I take to write
> my own C code, which can also run as kernel module if required, to measure
> time spent accurately. Will it also require to modify the scheduler code?
> Does /proc gives any statistics in this regard?
>
> Ref:
> http://stackoverflow.com/questions/2368384/write-a-c-program-to-measure-time-spent-in-context-switch-in-linux-os
>
> --
> Manavendra Nath Manav
> _______________________________________________
> Kernelnewbies mailing list
> Kernelnewbies at kernelnewbies.org
> http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies
>

How to measure time spent in context switch

[Manavendra Nath Manav]

On Tue, Jul 17, 2012 at 1:20 AM, Victor Rodriguez <vm.rod25@gmail.com> wrote:
>
> Hi Manavendra
>
> Have you think in a driver that you can call when the context switch
> happen? If this can be detected (more investigation on this area is
> required). If that is possible we can create the driver to measure the
> time according to the interruption.
>
> Thanks a lot and regards
>
> Victor Rodriguez
>
>
>
> On Mon, Jul 16, 2012 at 2:28 PM, Manavendra Nath Manav
> <mnm.kernel@gmail.com> wrote:
> > How to measure time spent in context switch in Linux i.e. the time the
> > application stays suspended while the kernel is doing something else. I
> > don't want to use any profiling tools. What approach should I take to write
> > my own C code, which can also run as kernel module if required, to measure
> > time spent accurately. Will it also require to modify the scheduler code?
> > Does /proc gives any statistics in this regard?
> >
> > Ref:
> > http://stackoverflow.com/questions/2368384/write-a-c-program-to-measure-time-spent-in-context-switch-in-linux-os
> >
> > --
> > Manavendra Nath Manav
> > _______________________________________________

Hi Victor

I was thinking something in these terms:
1) Process is executing a non-critical section like a bottom-half.
2) At Nth line of code, I will measure the timestamp.
3) At (N+1)th line of code, I will voluntarily call schedule() to do
context switch.
4) At (N+2)th, I will measure the timestamp, and calculate the time difference.

Although the above procedure will measure context switch time
statically and not dynamically, it can give a fair idea but for
multi-core processors this measurement will heavily deviate from
accuracy. If we can set scheduler() process affinity to one single
core, will it affect the measurements and predictability of context
switches?

--
Manavendra Nath Manav

How to measure time spent in context switch

[Mulyadi Santosa]

Hi...

On Tue, Jul 17, 2012 at 2:28 AM, Manavendra Nath Manav
<mnm.kernel@gmail.com> wrote:
> How to measure time spent in context switch in Linux i.e. the time the
> application stays suspended while the kernel is doing something else. I
> don't want to use any profiling tools. What approach should I take to write
> my own C code, which can also run as kernel module if required, to measure
> time spent accurately. Will it also require to modify the scheduler code?
> Does /proc gives any statistics in this regard?

IMHO, you need to mimic how lmbench does it. And AFAIK too, what you
are going to do is very much what lmbench does. But please check their
code for better understanding.

-- 
regards,

Mulyadi Santosa
Freelance Linux trainer and consultant

blog: the-hydra.blogspot.com
training: mulyaditraining.blogspot.com

How to measure time spent in context switch

[Mulyadi Santosa]

Hi... :)

Sorry for jumping in...

On Tue, Jul 17, 2012 at 3:09 AM, Manavendra Nath Manav
<mnm.kernel@gmail.com> wrote:
......
>If we can set scheduler() process affinity to one single
> core, will it affect the measurements and predictability of context
> switches?

I think it will affect, up to some unknown degree. On inter processor
scheduling, there would happen cache remote invalidation and process
task_struct transfer (IIRC process descripter is per cpu).

-- 
regards,

Mulyadi Santosa
Freelance Linux trainer and consultant

blog: the-hydra.blogspot.com
training: mulyaditraining.blogspot.com