Re: [Lse-tech] Re: CPU affinity & IPI latency

[Andi Kleen]

On Sun, Jul 15, 2001 at 01:02:21PM -0700, Davide Libenzi wrote:
> The problem of the current scheduler is that it acts like an infinite feedback
> system.
> When we're going to decide if we've to move a task we analyze the status at the
> current time without taking in account the system status at previous time
> values.
> But the feedback we send ( IPI to move the task ) take a finite time to hit the
> target CPU and, meanwhile, the system status changes.
> So we're going to apply a feedback calculated in time T0 to a time Tn, and this
> will result in system auto-oscillation that we perceive as tasks bouncing
> between CPUs.
> This is kind of electronic example but it applies to all feedback systems.
> The solution to this problem, given that we can't have a zero feedback delivery
> time, is :
> 
> 1) lower the feedback amount, that means, try to minimize task movements
> 
> 2) a low pass filter, that means, when we're going to decide the sort ( move )
>         of a task, we've to weight the system status with the one that it had
>         at previous time values

Nice analysis. I think Mike's proposal of the 'saved_cpus_allowed' field
to temporarily bind the task to the target would just act as such an low 
pass filter.

-Andi