"Lane, William" <william.l...@cshs.org> writes: > Our issues with OpenMPI 1.8.7 and Son-of-Gridengine turned out to be > down to using the wrong Parallel Environment. Having a PE with > control_slaves set to TRUE and start_proc_args and stop_proc_args set > to NONE cleared up all our issues, at least for my MPI test code. > > Qsort_args was left set to NONE, which directly contradicts the FAQ > for running OpenMPI through Son-of-Gridengine so maybe the OpenMPI FAQ > WRT SOGE should be revised.
I don't see how. The example in the FAQ is correct, and exactly what is used here in some PEs (apart from the name). > Qsort_args is an API that involves writing > your own function in a dynamically shared object that determines what > nodes should be used to schedule a job. Yes, as in sge_pe(1), and it's NONE in the example. (The obvious application for it is to optimize the topology of the allocation.) This isn't openmpi-specific, though. It should be the same with any other MPI that provides tight integration directly, like those using hydra.
