[Gllug] Forcing CPU and Mem limits on a process

[John Hearns]

On Wed, 2005-02-16 at 00:57 +0000, Nix wrote:
> On Sun, 13 Feb 2005, John Hearns prattled cheerily:
> > The '25% of CPU' is an intriguing one.
> > I don't think it is possible looking at it with Gridengine eyes -
> > after all a scheduler is supposed to keep your cluster humming along
> > at maximum utilisation!
> 
> It still does: it's just that if more than one entity is contending for
> the CPU, and others can potentially take CPU time without exceeding
> their usage percentage, and you *are* exceeding your usage percentage,
> your usage drops until it's at no more than that percentage again.
> 
Yebbut :-)
The usage in Gridengine is time-averaged (the half-life is configurable)
The intention is that groups or individuals within a company or uni get
shares allocated. You can submit as many jobs as you want, and if the
CPU time is available they will get run. But your priority is adjusted
down till over time the usage averages out.
If the other groups aren't using their allocation you will still get to
run.


There is a feature of subordinate or low priority queues, which we've
set up in several places. With these, if a job is run in the higher
priority queue then the lower one is suspended.
I was thinking that mechanism could be mangled to achieve the 25% share.
But Daniel's link to CKRM makes much more sense.




-- 
Gllug mailing list  -  Gllug at gllug.org.uk
http://lists.gllug.org.uk/mailman/listinfo/gllug