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

[Hubertus Franke <frankeh@watson.ibm.com>]

* Davide Libenzi <davidel@xmailserver.org> [20011031 18;53]:"
> On Wed, 31 Oct 2001, Mike Kravetz wrote:
> 
> > I'm going to try and merge your 'cache warmth' replacement for
> > PROC_CHANGE_PENALTY into the LSE MQ scheduler, as well as enable
> > the code to prevent task stealing during IPI delivery.  This
> > should still be significantly different than your design because
> > MQ will still attempt to make global decisions.  Results should
> > be interesting.
> 
> I'm currently evaluating different weights for that.
> Right now I'm using :
> 
>     if (p->cpu_jtime > jiffies)
>         weight += p->cpu_jtime - jiffies;
> 
> that might be too much.
> Solutions :
> 
> 1)
>     if (p->cpu_jtime > jiffies)
>         weight += (p->cpu_jtime - jiffies) >> 1;
> 
> 2)
>     int wtable[];
> 
>     if (p->cpu_jtime > jiffies)
>         weight += wtable[p->cpu_jtime - jiffies];
> 
> Speed will like 1).
> Other optimization is jiffies that is volatile and forces gcc to always
> reload it.
> 
> static inline int goodness(struct task_struct * p, struct mm_struct
> *this_mm, unsigned long jiff)
> 
> might be better, with jiffies taken out of the goodness loop.
> Mike I suggest you to use the LatSched patch to 1) know how really is
> performing the scheduler 2) understand if certain test gives certain
> results due wierd distributions.
> 

One more. Throughout our MQ evaluation, it was also true that 
the overall performance particularly for large thread counts was
very sensitive to the goodness function, that why a na_goodness_local 
was introduced.

-- Hubertus