[GE users] Oversubscribe hosts on demand.

reuti reuti at staff.uni-marburg.de
Fri Sep 11 11:04:07 BST 2009


Am 11.09.2009 um 11:41 schrieb jesperkrogh:

> I have some "high priority" jobs.. I would like to just send some  
> flag on
> the qsub line telling Gridengine to just schedule this one (and
> oversubscribe the host to get it done).
> Is that possible?

in principle: yes. But not by a simple flag.

You need:

- one forced boolean complex, which you will request for these  
special jobs and attach it to:
- one queue e.g. urgent.q for this special jobs (priority 0 = nice 0)
- queues for other jobs on the same host with (priority 19 = nice 19)  
which you have already
- limiting the number of slots per host not in the exechost  
definition, but in an RQS:
   limit queues !urgent.q to slots=8

(hence the urgent.q won't be honored for the total slots)

Jobs submitted with the urgent flag can run only in the urgent.q and  
are allowed to oversubscribe the node. You can even limit the access  
to this special queue by users_lists in the queue configuration.

(In fact: we use such a setup to allow interactive access to the  
nodes to peek around. Users must use SGE, as a simple ssh isn't  
allowed. The interactive queue for this purpose has a h_cpu limit of  
60 seconds to avoid abuse of this feature.)

-- Reuti

> I would prefer not to use checkpointing for this and just suspend  
> jobs is
> "not good" for the other jobs currently running on the nodes. But  
> pushing
> a node to 9 jobs/8 cores is acceptable for getting it executes "now".
> Thanks.
> Jesper
> ------------------------------------------------------
> http://gridengine.sunsource.net/ds/viewMessage.do? 
> dsForumId=38&dsMessageId=216851
> To unsubscribe from this discussion, e-mail: [users- 
> unsubscribe at gridengine.sunsource.net].


To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].

More information about the gridengine-users mailing list