On Wed, 2003-07-09 at 06:49, Guillaume Chazarain wrote:
> 08/07/03 23:13:22, Davide Libenzi <davidel@xmailserver.org> wrote:
> 
> >On Tue, 8 Jul 2003, Guillaume Chazarain wrote:
> >
> >> Hello,
> >>
> >> Currently the interactive points a process can have are in a [-5, 5] range,
> >> that is, 25% of the [0, 39] range. Two reasons are mentionned:
> >>
> >> 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
> >> 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
> >>
> >> But, using 50% of the range, instead of 25% the interactivity points are better
> >> spread and both rules are still respected.  Having a larger range for
> >> interactivity points it's easier to choose between two interactive tasks.
> >>
> >> So, why not changing PRIO_BONUS_RATIO to 50 instead of 25?
> >> Actually it should be in the [45, 49] range to maximize the bonus points
> >> range and satisfy both rules due to integer arithmetic.
> >
> >I believe these are the bits that broke the scheduler, that was working
> >fine during the very first shots in 2.5. IIRC Ingo was hit by ppl
> >complains about those 'nice' rules and he had to fix it. It'd be
> >interesting bring back a more generous interactive bonus and see how the
> >scheduler behave.
> 
> Thanks for the info.
> Before being 25% the interactivity range was 70%, thus breaking the rules. So
> I am now more convinced that a 50% range could be a good thing.
> 

Just a suggestion, why instead of changing the code you don't try the
attached patch? At least you don't have to recompile just to change a
few define's. Against 2.5.73, but applies in 2.5.74. Just the "long
sleep_time = jiffies - p->last_run;" isn't there.

I remember that I saw someone's patch nearly identical to this ( I think
it was Robert Love) but I don't remember the url.