It's the correct behavior, because with jobs running, then there is load. And by default Grid Engine dispatches jobs to the lightest load machines first.
You can take a look at this blog entry, "N1GE 6 - Scheduler Hacks: "least used" / "fill up" configuration": http://wiki.gridengine.info/wiki/index.php/StephansBlog (Stephan was a Grid Engine developer at Sun at that time.) Rayson On Thu, Aug 9, 2012 at 2:54 PM, DeMattia, Edmond G. <[email protected]> wrote: > When users submit jobs to the grid, SGE is not using all the available cores > on each node, instead submitting a random number of jobs to each node. For > instance, when a user submits 1200 jobs, some nodes use 6 cores out of 12 > while others use 9 cores and still others 2 or 3 cores out of 12. All 1200 > jobs run but across many more physical nodes then would normally be needed > if the jobs would occupy all the available cores per node. > > Is this a configuration change I need to make at the qmaster or a job > submission change the users need to perform? The jobs are typical > embarrassingly parallel jobs. > > _______________________________________________ > users mailing list > [email protected] > https://gridengine.org/mailman/listinfo/users > _______________________________________________ users mailing list [email protected] https://gridengine.org/mailman/listinfo/users
