Re: [PATCH 2/2] sched: Improve fairness of cpu allocation for task groups

[Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>]

On Mon, Nov 19, 2007 at 02:12:01PM +0100, Ingo Molnar wrote:
> >  include/linux/sched.h |    4 
> >  kernel/sched.c        |  292 +++++++++++++++++++++++++++++++++++++++++++++-----
> >  kernel/sched_fair.c   |   95 ++++++++++------
> >  kernel/sched_rt.c     |    2 
> >  kernel/sysctl.c       |   16 ++
> >  5 files changed, 348 insertions(+), 61 deletions(-)
> 
> i'm leaning towards making this v2.6.25 material, as it affects the 
> non-group-scheduling bits too and is rather large. When i tested it, 
> group scheduling worked pretty well - at least for CPU bound tasks - and 
> on SMP too. Could we live with what we have for now and defer this patch 
> to v2.6.25?

Hi Ingo,
	I would prefer this to go in 2.6.24 if possible. 2.6.24 would be the
first kernel to support a group scheduler in its entirety (user interface +
related support in scheduler) and also that works reasonably well :) It would 
also give me early test feedback.

> If not, could you split up this patch in a way to defer all 
> the FAIR_GROUP_SCHED relevant changes to a separate patch which will not 
> affect the !FAIR_GROUP_SCHED case at all? That will make the case much 
> clearer.

>From my inspection, here are the changes introduced by this patch
for !CONFIG_FAIR_GROUP_SCHED case:

- inc/dec_load() takes a load input instead of task pointer input as their
  2nd arg
- inc/dec_nr_running don't call inc/dec_load. Instead,
- enqueue/dequeue_task class callbacks call inc/dec_load
- [Unintended/will-fix change] min/max tunables added in
  /proc/sys/kernel

All of above changes (except last, which I will fix) should have zero
functional+runtime effect for !CONFIG_FAIR_GROUP_SCHED case. So I don't see how
I can split Patch 2/2 further.

Or do you prefer I introduce #ifdef's such that even these minor changes to 
inc/dec_load are avoided for !CONFIG_FAIR_GROUP_SCHED case? That would
make the code slightly ugly I suspect.

-- 
Regards,
vatsa