Re: [RFC] scheduler: improve SMP fairness in CFS

[Tong Li <tong.n.li@intel.com>]

I'd like to clarify that I'm not trying to push this particular code to 
the kernel. I'm a researcher. My intent was to point out that we have a 
problem in the scheduler and my dwrr algorithm can potentially help fix 
it. The patch itself was merely a proof-of-concept. I'd be thrilled if the 
algorithm can be proven useful in the real world. I appreciate the people 
who have given me comments. Since then, I've revised my algorithm/code. 
Now it doesn't require global locking but retains strong fairness 
properties (which I was able to prove mathematically).

Thank you,

  tong

On Thu, 26 Jul 2007, Li, Tong N wrote:

> On Thu, 2007-07-26 at 23:31 +0200, Ingo Molnar wrote:
> > * Tong Li <tong.n.li@intel.com> wrote:
> >
> > > > you need to measure it over longer periods of time. Its not worth
> > > > balancing for such a thing in any high-frequency manner. (we'd trash
> > > > the cache constantly migrating tasks back and forth.)
> > >
> > > I have some data below, but before that, I'd like to say, at the same
> > > load balancing rate, my proposed approach would allow us to have
> > > fairness on the order of seconds. I'm less concerned about trashing
> > > the cache. The important thing is to have a knob that allow users to
> > > trade off fairness and performance based on their needs. [...]
> >
> > such a knob already exists to a certain degree, but i havent tested its
> > full effects on SMP fairness yet. If you pull my scheduler tree:
> >
> >   git://git.kernel.org/pub/scm/linux/kernel/git/mingo/linux-2.6-sched.git
> >
> > and if you enable CONFIG_SCHED_DEBUG, then all the sched-domain
> > parameters become runtime tunable under /proc/sys/cpu*.
> >
> > Could you try to increase the cross-CPU rebalancing frequency and see
> > how it impacts the precision of your measurement? Tune 'min_interval'
> > and 'max_interval' down to increase the frequency of rebalancing.
> >
> > 	Ingo
>
> Yes, I'll do it when I find time. If anyone is willing to do the
> testing, please let me know and I can post my benchmark. On the other
> hand, I don't think tuning the existing knobs will help solve the
> problem. The root of the problem is that the current load balancing
> doesn't take into account how much time each task has been running and
> how much it's entitled. This is why I'm proposing a new approach for
> solving it. The new approach, as I said, will be much more fair/accurate
> than the current one even without tuning those balancing intervals
> (i.e., it's able to provide good fairness even if balancing is less
> frequent).
>
>  tong
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/