[GE users] parallel environments and multple queues
antony.cleave.ic at gmail.com
Thu May 14 19:17:38 BST 2009
[ The following text is in the "iso-8859-1" character set. ]
[ Your display is set for the "ISO-8859-10" character set. ]
[ Some special characters may be displayed incorrectly. ]
I have a small cluster of 6 nodes with some old myrinet 2G cards that I scavanged from a skip (with permission). They work really well but as I don't have a switch, I can only connect them point to point in seperate clusters of 2 nodes. I thought I could get around this by making 3 queues of paired nodes and then get SGE to allocate the nodes at run time but this doesn't work as it allocates nodes from queue 1 and 3 for the same job. I know I can fix this by having 3 different parallel environments but then I'm limiting each job to run on a specific pair of nodes which is not efficient at all.
I have configured 3 queues, queue1,queue2 and queue3 with the respective hostgroups as below and that all reference the OMPI (OpenMPI) PE which is setup to have 12 slots.
I'm pretty sure I must be missing a trick here as I can't be the only person trying to schedule jobs on groups of nodes that aren't on the same networks.
thanks in advance
More information about the gridengine-users