[Lse-tech] Re: [PATCH][RFC] Proposal For A More Scalable Scheduler ...

[Davide L. <da...@xm...>]

On Tue, 30 Oct 2001, Hubertus Franke wrote:

> > Real time processes, when wakeup up fall calling reschedule_idle() that
> > will either find the CPU idle or will be reschedule due a favorable
> > preemption_goodness().
> > One of balancing scheme I'm using tries to distribute RT tasks evenly on
> > CPUs.
> >
>
> I think that would be a problem. My understanding is that if two RT process
> are globally runnable, then one must run the one with higher priority.
> Am I missing something here ?

The only difference is when an RT task is currently running and another RT
task kicks in with a not favorable preemption_goodness().
In the current scheduler reschedule_idle() loops through the CPUs to find
one for the incoming RT tasks while the proposed scheduler actually
doesn't.
What I'm coding is to plug in get_best_cpu() a way to evenly spread RT
tasks between CPUs.
But even the RT task rapid move to another CPU is not too rapid due
IPI+schedule() latency.
Maybe it's faster the currently running RT tasks to reschedule instead of
the remote CPU, maybe :)
The current IPI method creates very funny/undesirable behavior due
IPI+schedule() latency.
When watching at the  schedcnt  dump of a  lat_ctx  of my 2 way SMP
system, I saw both tasks lying on the same CPU with the other one
tempested by reschedule IPI without being able to catch one task due latency.




- Davide