Re: [patch] O(1) scheduler, -G1, 2.5.2-pre10, 2.4.17 (fwd)

["François Cami" <stilgar2k@wanadoo.fr>]

Ingo Molnar wrote:

> On Thu, 10 Jan 2002, Mike Kravetz wrote:
> 
> 
>>If I run 3 cpu-hog tasks on a 2 CPU system, then 1 task will get an
>>entire CPU while the other 2 tasks share the other CPU (easily
>>verified by a simple test program). On previous versions of the
>>scheduler 'balancing' this load was achieved by the global nature of
>>time slices. No task was given a new time slice until the time slices
>>of all runnable tasks had expired.  In the current scheduler, the
>>decision to replenish time slices is made at a local (pre-CPU) level.
>>I assume the load balancing code should take care of the above
>>workload?  OR is this the behavior we desire? [...]
>>
> 
> Arguably this is the most extreme situation - every other distribution
> (2:3, 3:4) is much less problematic. Will this cause problems? We could
> make the fairness-balancer more 'sharp' so that it will oscillate the
> length of the two runqueues at a slow pace, but it's still caching loss.
> 
> 
>>We certainly have optimal cache use.
>>
> 
> indeed. The question is, should we migrate processes around just to get
> 100% fairness in 'top' output? The (implicit) cost of a task migration
> (caused by the destruction & rebuilding of cache state) can be 10
> milliseconds easily on a system with big caches.
> 
> 	Ingo


I do vote for optimal cache use. Using squid (200MB process in my case)
can be much faster if squid stays on the same CPU for a while, instead
of hopping from one CPU to another (dual PII350 machine).

François Cami