Re: [patch] scheduler cache affinity improvement for 2.4 kernels

[Davide Libenzi]

On Fri, 9 Nov 2001, Ingo Molnar wrote:

>
> On Thu, 8 Nov 2001, Mike Fedyk wrote:
>
> > It remains to be proven whether the coarser scheduling approach
> > (Ingo's) will actually help when looking at cache properties.... [...]
>
> have you seen the numbers/measurements i posted in my original email? 3%
> kernel compile speedup on an 'idle' 8-way system, 7% compilation speedup
> with HZ=1024 and background networking load on a 1-way system.

Ingo, i'm giving the timer_ticks patch a try in my proposed scheduler coz
i like the idea of skipping the if inside goodness(), and i can do this
safely because inside the proposed scheduler i don't have any cross CPU
goodnesses ( no "if (p->processor != this_cpu) weight -= p->timer_ticks;" ).
I made a change to it anyway, that is adding a water-mark in the decay
behavior ( timer.c ).
When counter is above this watermark ( currently 20 ) the counter decay as
usual while if counter <= watermark, ticks accumulates in timer_ticks.
This solution keeps the same good behavior for CPU bound tasks while it
gives a "human"/current decay to tasks that has got a lot of counter
accumulation inside the recalc loop ( I/O bound ).

- Davide