hulst at argoss.nl
Mon Sep 8 05:22:46 MDT 2008
I have a test cluster of 4 nodes that each have two quad core CPU's. Previous
benchmarks with an atmospheric model showed that there's an issue with the
quad cores in the sense that the speed of the front side bus is a limiting
factor. I would like to have control over the allocation of nodes.
Using the MPI TM interface I request:
#PBS -l nodes=4:ppn=4
What happens is that the job is run on two nodes, each node running two chunks
of 4 "cpus" (which are actually cores). That is not really desirable. I would
like to divide the job over 4 nodes.
My initial guess was that I could control node allocation with the
nodeallocationpolicy parameter, but I fail to get it working.
Does anyone have any experience with a similar issue? Is it possible?
More information about the mauiusers