twu...@goodyear.com writes: > In the past (v 1.6.4-) we used mpirun args of > > --mca mpi_paffinity_alone 1 --mca btl openib,tcp,sm,self > > with lsf 7.0.6, and this was enough to make cores not be oversubscribed when > submitting 2 or more jobs to the same node.
[I'm puzzled by that. It should have made sure they _were_ oversubscribed, since it was supposed to start binding at core 0, and did as far as I remember.] > Now I am using 1.8.8 and thus far don't have the right combination of args to > make sure cores don't oversubscribe. > > I am currently using: > > -report-bindings --map-by core --bind-to core --nooversubscribe --mca btl > openib,tcp,sm,self > > Do I have the incorrect options or am I missing an option? As I don't see an answer: That looks OK, assuming LSF binds the processes in the first place or allocates complete nodes. If you want to make core binding the default and can do the LSF bit similarly to (or better than) SGE, the OMPI part of the recipe at <http://arc.liv.ac.uk/SGE/howto/sge-configs.html#_core_binding> should apply. Can someone say why the default changed to mapping by socket?
