Re: [PATCH] kernel <linux-2.6.11.10> kernel/sched.c

[chen Shang <shangcs@gmail.com>]

Here is the statistics data to support requeue_task() inline  with the
counter on sched_cnt at the same period. In brief, every 10 schedule()
will trigger recalc_task_prio() a little more than 4 times.

CPU0: priority_changed (669 times), priority_unchanged(335,138 times),
schedule_cnt(787,085 times)
CPU1: priority_changed (784 times), priority_unchanged(342,419 times),
schedule_cnt(784,873 times)
CPU2: priority_changed (782 times), priority_unchanged(283,494 times),
schedule_cnt(681,917 times)
CPU3: priority_changed (872 times), priority_unchanged(365,865 times),
schedule_cnt(809,624 times)

On 5/20/05, Con Kolivas <kernel@kolivas.org> wrote:
> On Fri, 20 May 2005 17:21, Nick Piggin wrote:
> > chen Shang wrote:
> > >I minimized my patch and against to 2.6.12-rc4 this time, see below.
> > >
> > >The new schedstat fields are for the test propose only, so I removed
> > >them completedly from patch. Theoritically, requeue_task() is always
> > >cheaper than dequeue_task() followed by enqueue_task(). So, if 99% of
> > >priority recalculation trigger requeue_task(), it will save.
> > >
> > >In addition, my load is to build the kernel, which took around 30
> > >minutes with around 30% CPU usage on 2x2 processors (duel processors
> > >with HT enable).
> > >Here is the statistics:
> > >
> > >CPU0: priority_changed (669 times), priority_unchanged(335,138 times)
> > >CPU1: priority_changed (784 times), priority_unchanged(342,419 times)
> > >CPU2: priority_changed (782 times), priority_unchanged(283,494 times)
> > >CPU3: priority_changed (872 times), priority_unchanged(365,865 times)
> >
> > OK that gives you a good grounds to look at the patch, but _performance_
> > improvement is what is needed to get it included.
> 
> If you end up using requeue_task() in the fast path and it is hit frequently
> with your code you'll need to modify requeue_task to be inline as well.
> Currently it is hit only via sched_yield and once every 10 scheduler ticks
> which is why it is not inline. The performance hit will be demonstrable if it
> is hit in every schedule()
> 
> Cheers,
> Con
> 
> 
>