[Ksummit-2013-discuss] [ATTEND] Power-aware scheduling

[Preeti U Murthy]

Hi Tony,

On 07/16/2013 12:07 AM, Luck, Tony wrote:
>> While working on this issue, I noticed that, if we intend to observe the
>> behaviour of the scheduler with some clarity, we would have to restrict
>> the total number of cpus to very few. Else it is hard to understand how
>> the scheduler is behaving.
> 
> How many is "very few"?  And how do you then ensure that the changes you
> have validated on a small system work as expected on a large one?
> 
> I deal in 4 to 8 socket machines that currently have 10 cores time 2 hyperthreads
> per socket.  So up to 160 logical cpus today. The recent trend has been that the
> number of cores per socket increases with each generation.
> 
> It may well be very hard to develop tools to visualize what is going on inside
> a machine with hundreds of logical cpus - but I think we need to do  that
> rather than just look at small systems.

Exactly my point. We need to observe the behaviour of the scheduler on a
large scale. But it becomes a challenge to do so with increasing number
of sockets and cores on each.

If we are to observe the behaviour of the scheduler clearly today, we
probably need to restrict the number of sockets and the cores on each.
Atleast this is my experience. I intended to clearly observe load
balancing using perf sched map. I could get a clear picture of the
migrations as long as my experimental setup was a 2 socket, 4 core each
socket machine. Beyond that, it was difficult to observe the nature of
migrations.

One option was that I could parse the output of perf sched map into
intelligible data, but again with too many cpus, my observation was that
it is still a challenge to understand what is happening.

This is the issue that I wish to discuss. We cannot hope that what
happens at a smaller scale in scheduling will hold good on a large
scale, so what approaches can we take to observe scheduling on the
current systems?

> 
> -Tony
> 

Regards
Preeti U Murthy