[Devel] Re: [RFC] [PATCH 0/3] Add group fairness to CFS

[William Lee Irwin III]

* William Lee Irwin III <wli at holomorphy.com> wrote:
>> [...] As an interface it may be poor and worse yet poorly specified, 
>> [...]

On Wed, May 23, 2007 at 09:26:54PM +0200, Ingo Molnar wrote:
> that's the only thing that matters to fundamental design questions like 
> this.

I'm not sure where it comes in as a design question. Tuning whatever's
done for sched_yield() is just one of the drudgery tasks of rounding
out the implementation as I see it. It can be painful depending on
what's being changed, though it's not clear to me precisely how painful
tuning it will be for this change.

If it turns out to be a large enough issue, I can enforce reasonable
semantics for it with no impact on the core scheduling algorithm anyway.


* William Lee Irwin III <wli at holomorphy.com> wrote:
>> The content of my comment was that the patch does something to 
>> sched_yield() semantics, so it raises the question of what will happen 
>> in benchmarks and other performance affairs that are sensitive to 
>> sched_yield() semantics changes.

On Wed, May 23, 2007 at 09:26:54PM +0200, Ingo Molnar wrote:
> the correct aproach to the "sys_sched_yield() is an API that sucks" 
> problem is to simply _not use it_. User-space is figuring that out now, 
> fortunately.

We do need API's to displace it for that to really take off. I think
the yield_to() API's will help there. Sadly, I expect it to be a very
long time for the transition to really happen.


-- wli
_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers