Re: Ok to change cpuset flags for sched domains? (was [PATCH 1/3] CPUMETER ...)

[Dinakar Guniguntala <dino@in.ibm.com>]

I have been wanting to follow the cpumeter discussion more closely,
but currently am tied up. I hope to have more time towards the end of
this week.

I had a few queries below, though

On Sun, Oct 02, 2005 at 12:01:59AM -0700, Paul Jackson wrote:
> Dinikar,
> 
> How much grief will it cause you if I make the following incompatible
> change to the special boolean files in each cpuset directory?
> 
> I think I goofed in encouraging you to overload "cpu_exclusive"
> with defining dynamic scheduler domains.  I should have asked for a
> separate flag to be added for that, say "sched_domain", which would
> require "cpu_exclusive=1" as a precondition.  Other attributes that
> require cpu_exclusive or mem_exclusive are showing up, and it makes
> more sense for each of them to get their own boolean, and leave the
> "*_exclusive" flags to specify just the exclusive (no overlap with
> sibling) attribute.


One of the reasons for overloading the cpu_exclusive flag was to
ensure that the rebalance code does not try to pull tasks unnecessarily

With the scheme that you are proposing that is a possibility if
you turn on the cpu_exclusive and meter_cpu for example and not
turn on sched_domain. Is there a reason why we would want to have
exclusive cpusets not attached to sched domains at all?

I am not entirely convinced that we can compare sched_domains and 
meter_cpus.

However I am still open if there is a convincing reason to have
exclusive cpusets that dont form sched domains.

	-Dinakar