We run EMI-3 using 2.5.13-1cri.9nik.el6 and SL6. About half our workload are grid jobs (mainly ATLAS) and half are jobs run by local users directly using from the head node (qsub). One of our worker nodes has a GPU -- a student needs to have some libraries installed that are next to impossible to install on Linux kernel 2 machines but trivial on SL7.
One solution is just to remove this node from the cluster and to install SL7. It's only for a relatively short period and we can tolerate the loss ofthe machine wrt workload.
Better would be to set it up using torque (our ancient version anyway) so that grid jobs cannot run this node but head node-submittd jobs can. These jobs can clearly be identified by the queues that they run on.
Can anyone suggest a quick fix to this or any suggestions?
In the medium-term we'll upgrade to UMD-4 but I need something now that will not be onerous on my time