Re: Definition of fairness (was Re: [patch] CFS scheduler, -v11)

[Srivatsa Vaddagiri <vatsa@in.ibm.com>]

On Wed, May 09, 2007 at 01:24:18PM -0700, William Lee Irwin III wrote:
> It's not even feasible much of the time. Suppose your task ran for
> 100ms then slept for 900ms. It can't get more than 10% of the CPU in
> any scheduler, work-conserving or not.

sure. The question of fairnes arises when such a task has to "fight" for
space/cputime in those 100ms, resulting in lesser than 10% bandwidth.
Having some statistics shown on cpu demand vs obtained bandwidth may be
good?

> Fairness applies only to running tasks. The fair share of CPU must be
> granted while the task is running. As for sleep, it has to use its
> fair share of CPU or lose it. The numerous of pathologies that arise
> from trying to credit tasks for sleep in this fashion this are why
> fairness in the orthodox sense I describe is now such a high priority.
> 
> However, it is possible to credit tasks for sleep in other ways. For
> instance, EEVDF (which is very close to CFS) has a deadline parameter
> expressing latency in addition to one expressing bandwidth that could,
> in principle, be used for the purpose of crediting sleeping tasks. It's
> not clear to me whether trying to use it for such purposes would be
> sound, or, for that matter, whether tasks should receive latency
> credits for sleep as opposed to other sorts of behaviors even if
> utilizing the latency parameter for credits and bonuses for various
> sorts of behavior is sound.

I guess fairness interval also matters, as Mike pointed out. Over
shorter intervals, it may appear more fair to such sleeping tasks than over
longer intervals.  Btw CFS seems to leave this fairness interval undefined (its 
dependent on N, number of tasks and sysctl_sched_granularity?).

I suspect container/database/webserver workloads would like to receive
such latency/bandwidth credits for sleep. Will check with our
database/workload management folks on this.

-- 
Regards,
vatsa