Re: CFS: some bad numbers with Java/database threading [FIXED]

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Wed, 19 Sep 2007 23:41:05 +0200 Ingo Molnar <mingo@elte.hu> wrote:

> 
> * Linus Torvalds <torvalds@linux-foundation.org> wrote:
> 
> > Btw, the "enqueue at the end" could easily be a statistical thing, not 
> > an absolute thing. So it's possible that we could perhaps implement 
> > the CFS "yield()" using the same logic as we have now, except *not* 
> > calling the "update_stats()" stuff:
> > 
> > 	__dequeue_entity(..);
> > 	__enqueue_entity(..);
> > 
> > and then just force the "fair_key" of the to something that 
> > *statistically* means that it should be at the end of its nice queue.
> > 
> > I dunno.
> 
> i thought a bit about the statistical approach, and it's good in 
> principle, but it has an implementational problem/complication: if there 
> are only yielding tasks in the system, then the "queue rightwards in the 
> tree, statistically" approach cycles through the key-space artificially 
> fast. That can cause various problems. (this means that the 
> workload-flag patch that uses yield_granularity is buggy as well. The 
> queue-rightmost patch did not have this problem.)
> 
> So right now there are only two viable options i think: either do the 
> current weak thing, or do the rightmost thing. The statistical method 
> might work too, but it needs more thought and more testing - i'm not 
> sure we can get that ready for 2.6.23.
> 
> So what we have as working code right now is the two extremes, and apps 
> will really mostly prefer either the first (if they dont truly want to 
> use yield but somehow it got into their code) or the second (if they 
> want some massive delay). So while it does not have a good QoI, how 
> about doing a compat_yield sysctl that allows the turning on of the 
> "queue rightmost" logic? Find tested patch below.
> 
> Peter, what do you think?

I have to agree that for .23 we can't do much more than this. And tasks
moving to the right without actually doing work and advancing
fair_clock do scare me a little.

Also, while I agree with Linus' definition of sched_yield, I'm afraid
it will cause 'regressions' for all the interactivity people out there.
Somehow this yield thing has made it into all sorts of unfortunate
places like video drivers, so a heavy penalizing yield will hurt them.