Hi, Am 16.08.2016 um 22:59 schrieb Chester Langin:
> Hello, > > We implemented a new 40-unit HPCC October, 2015, and have been using the SGE > scheduler with the default configuration. The faculty have been submitting > jobs to the scheduler without a specifying a queue, and the queue defaults to > all.q. I would like to add more queues and force the faculty to specify one > of the new queues. The new queues are gpu.q with just the GPU nodes, hiram.q > with just the high-RAM nodes, and standard.q for the remaining queues. By > using @gpu_hosts, @hiram_hosts, and @standard_hosts, I have this running on a > development cluster. I am ok with the faculty using the honor system in > determining which queue to use, so long as they do not use all.q. Instead of specifying a queue it would be better to request resources and SGE will choose an appropriate queue for your job. E.g. for a gpu you could define a boolean complex and if users need a gpu they have to request something like "-l gpu". If you don't want to run non-gpu jobs on such a machine this complex could also be defined as "forced" (see `man complex`). BTW: all.q is nothing special, it's only created for convenience. In case you don't need all.q any longer, you can delete it without harm. > The problem is that if a faculty does not specify a queue, then all.q is used > by the scheduler, and a node becomes over-subscribed. Why? Do you have more than one queue per node now? Did you define more slots than installed cores? -- Reuti > What is the best/easiest way to prevent this from happening? > > -- > Chet Langin > SIU > > _______________________________________________ > users mailing list > users@gridengine.org > https://gridengine.org/mailman/listinfo/users _______________________________________________ users mailing list users@gridengine.org https://gridengine.org/mailman/listinfo/users
