Re: [PATCH 1/7] CPU controller V1 - split runqueue

[Srivatsa Vaddagiri <vatsa@in.ibm.com>]

On Mon, Aug 28, 2006 at 10:31:18PM +1000, Nick Piggin wrote:
> I still haven't had much time to look at the implementation, but this
> design seems cleanest I've considered, IMO.
> 
> Of course I would really hope we don't need any special casing in the
> SMP balancing (which may be the tricky part). However hopefully if
> things don't work well in that department, they can be made to by
> improving the core code to be more general rather than special casing.
> 
> Do you have a better (/another) idea for the design?

I dont' know if it is a better idea - but I have been trying to
experiment with some token-based system where task-groups run until
exhausted out of their tokens. Of course, this will be work-conserving
in the sense that expired task-groups continue running if there arent
others who want to use their share. Token are renewed at periodic
intervals. I believe that is how vserver scheduler works (though havent
looked at their code).

And I was thinking of using something similar to smpnice for
load-balance purposes.

The main point here is that scheduling next-task-group decision is local
to each CPU (very similar to how next-task is picked up currently), with
some load-balance code expected to balance tasks/task-groups across all
CPUs.

In what Kirill is proposing, this "scheduling next-task-group decision"
on each CPU perhaps takes a global view and because of the
physical/virtual CPU separation, any CPU can be running any other CPU's
tasks (smp_processor_id/get_cpu etc now returning virtual CPU number rather than
the actual CPU on which they are running). Kirill is that description correct?


-- 
Regards,
vatsa