Am 13.12.2011 um 21:05 schrieb Christoph Müller:

> Hi Reuti,
> 
>> -----Ursprüngliche Nachricht-----
>> Von: Reuti [mailto:[email protected]]
>> Gesendet: Dienstag, 13. Dezember 2011 20:25
>> An: Christoph Müller
>> Cc: [email protected]
>> Betreff: Re: AW: AW: AW: [gridengine users] Access complex resources from
>> prolog script
>> 
>> Am 13.12.2011 um 20:11 schrieb Christoph Müller:
> 
> 
>> How is the GPU selected right now? This is one of the cases where two queues
>> with one slot each might help. The environment variable $QUEUE will then
>> contain the name of the queue and by having gpu1/gpu2 the expression
>> gpu=${QUEUE##gpu} could do.
>> 
>> The submission command would have to include -q "gpu*" or be set in the JSV.
> 
> But this solution would not allow for requesting zero to two GPUs per slot, 
> right? Otherwise, I did not get the point.

Well, for parallel jobs a PE  with allocation rule $pe_slots can be used. In 
this case you will just get a slot from each queue. For non-GPU jobs I expected 
a "traditional" queue on the system being defined anyway.

Maybe even the GPU complex can be left out: submit also a serial GPU job into 
this GPU PE.

-- Reuti

PS: Looks like an RFE for a complex of type array. Internally maybe the core 
binding logic can be reused. Or OTOH allow the core binding to add custom 
sockets/cores.
_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users

Reply via email to