Please ignore my last email.
Roberto
On 26.05.2017 15:21, Roberto Nunnari wrote:
On 24.05.2017 21:28, Reuti wrote:
Am 24.05.2017 um 21:17 schrieb Roberto Nunnari:
Il 24.05.2017 20:45, Reuti ha scritto:
Am 24.05.2017 um 19:09 schrieb Roberto Nunnari
<[email protected]>:
Hi Reuti.
wow.. that enhancement request was opened 10 years ago! :-D
Anyways.. that means that at present it's not possible to grant
access to a node (or queue) only if the user ask for a gpu with -l
gpu=<n> ?
It is possible, but someone could specify -l gpu=0 and it counts as
requested.
One could use a JSV of course, to be sure that the limit is higher
than zero (or remove the request completely in case it’s zero, so
that the FORCED criteria is no longer satisfied and such a job would
be routed to the conventional nodes again).
In case you use "-w e": this will be checked before the JSV is
called and after (AFAICS independent from the action taken in the
JSV). To avoid this: one can submit always with "-w n", but set set
"-w e" inside the JSV.
— Reuti
I see.
So the problem is only that a user could cheat and get anyways access
to the nodes with GPUs.
Yep.
For now I'd be glad to trust my users and set this limit. Could you
please tell me how to set it?
True, there was a mix-up: one has to change "requestable" from YES to
FORCED for the gpu complex with `qconf -mc`, not the "consumable" one
I mentioned initally.
-- Reuti
Humm.. I just realized that will enforce all users of all nodes and
queues to
Thank you.
Roberto
At present, based on my configuration, the scheduler can (and do)
give access to those nodes (equipped with GPUs) even if the user
don't request any GPU.
Thank you and best regards.
Roberto
On 24.05.2017 16:47, Reuti wrote:
Am 24.05.2017 um 14:50 schrieb Roberto Nunnari
<[email protected]>:
Hello.
I am using Oracle Grid Engine 6.2u7.
Is there a way to configure the OGE so that it will give access
to some nodes only if a gpu is requested? I have a few nodes that
are equipped with gpus and I would like that only jobs that
require at least one gpu can access them. gpu is a JOB consumable
in complex configuration that users can request upon job
submission with -l gpu=<n>
I knew I entered an issue about it:
https://arc.liv.ac.uk/trac/SGE/ticket/391
-- Reuti
_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users
