OpenMPI Rankfile Support Requested

[Larry_A_]

With the number of cores increasing on the various OS platforms, in order to maximize efficiency when there are multiple jobs running on the same physical machine, I believe it would be in Intels best interest to support the rankfile syntax used by OpenMPI.  This allows upstream schedulers to minimize numa-node fragmentation as jobs are onboarded and offboarded with differing timings.  It would also allow upstream schedulers to allow serial workload to intermingle with parallel workload when the parallel workload can coexist with the serial workload.

The current binding options appear to be good, but there appears to be no capability for the upstream scheduler to define specific cores for execution on a per host basis.

Thanks,

TheWitness

[TimP]

In what way does Intel not support OpenMPI? Do you mean you want Intel MPI rebased on OpenMPI rather than mpich and mvapich? The Intel MPI defaults assume that an entire node is dedicated to a single job. Marketers consider that a competitive advantage, as in the usual case, options required e.g. by OpenMPI aren't needed. There are options to change that. If you wish to discuss the scheduling option of Intel MPI, the forum section on HPC and clusters is the place.

[Larry_A_]

Tim, I'll post a followup at the location you specified, in the mean time, please review this link for information on how OpenMPI handles the rankfile. It's quite flexible and give the user 100% control, which means the scheduler can dictate the binding. And yes, just an MPI option... http://www.open-mpi.org/faq/?category=tuning#using-paffinity-v1.3