[SGE-discuss] Advice sort on SGE Queue Configurations
reuti at staff.uni-marburg.de
Wed May 21 19:34:07 BST 2014
Am 21.05.2014 um 11:27 schrieb Dr Andrew Smith:
> We currently building a small cluster that is mainly for users to develop their code. What we aim to provide is a system that supports:
> a fast turn around for small time limited jobs - for development and debugging
> the ability to run longer jobs overnight and at the weekends
In principle also short jobs can run in a queue without any time limit. There are two options:
Having two queues would allow to set up a resource quota set (RQS) to have a fixed number of maximum jobs of each type. During submission, it's best to request the necessary wall clock time for the job, and SGE will select an appropriate queue for this job. There is no need to specify a queue in the `qsub ...` command.
While there is a calendar to switch on/off/suspend queues, there no such thing in RQS. But you could adjust the number specified there with a cron job issuing `qconf -mattr resource_quota ...` to have a different number for each type at the weekends.
In case you prefer one queue, you could set up two custom consumable complexes and the jobs have to request them like "short" and "long" which were attached to the global exechost configuration with an arbitrary high value. The consumption of each consumable could then be limited by an RQS again.
I would go for a), as requesting time is more natural and users can't bypass it to run more long jobs although they shouldn't. You could of course attach an h_rt setting to each job in JSV. The users would learn in case b) to request the correct consumable.
NB: short jobs can also start in the long queue. This is normal as there is no minimum request setting for any limit (unless you implement this by an JSV and route jobs to dedicated queues on your own).
Sorry if this sounds confusing, but there are often several ways to achieve a goal, and it might be personal taste what one prefers.
> the ability to prioritise jobs for groups who "own" a share of the nodes
You can set up a share-tree policy, so that over a certain time frame the group gets the computing time equivalent to their share of the complete cluster.
> I am looking for best practice on how we can achieve this. Are separate queues to address the three aims the way to go, or it is possible to have only one queue configured for this?
> Any advise, or pointers to documentation on how to do this would be much appreciated.
> Thank you,
> SGE-discuss mailing list
> SGE-discuss at liv.ac.uk
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the SGE-discuss