[GE users] Limits in queue configuration prevent fair job distribution

Reuti reuti at staff.uni-marburg.de
Fri Jun 6 16:34:45 BST 2008


Am 06.06.2008 um 13:28 schrieb Arne Brutschy:

> On Fr, 2008-06-06 at 12:49 +0200, Reuti wrote:
>> you mean you set "h_vmem 2GB" in the queue definitions? Did the jobs
>> request any h_rt?
> Yes and no.
>> But I don't see a difference - if you have three queues on each
>> machine, it doesn't matter in which queue it's running (for a short
>> job).
> What we currently have:
>  - three queues on all machines.
>  - when users submits job without specifying a specific queues, jobs
>    are added to all three queues (thus utilizing all available slots)
>> You set a proper slot count on each exechost to avoid
>> oversubscription by these three queues?
> What do you mean by oversubscription? We're using currently three  
> queues
> per core. Each queue can run one job. So there are up to 3 jobs  
> running
> a processor at a given time.

so you are oversubscribing the machines by intention. Usually the  
approach is to have just one job per core running, otherwise three  
porcesses per core will steal computing time from each other and will  
run only at a one third speed. If you want this, it's okay of course,  
but rarely see such a configuration.

-- Reuti

To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
For additional commands, e-mail: users-help at gridengine.sunsource.net

More information about the gridengine-users mailing list