We have very little experience with OmniPath so far. To use the OmniPath interfaces via Intel MPI add "-iface ib0 -PSM" to your mpirun. Unsure about the rank/thread placement questions. Ben

On Aug 17, 2016, at 12:35 PM, Luo, Ye <[email protected]> wrote:

Hi all,

I just noticed a very strange behavior of the mpirun on KNL nodes. If I run 1 node with 1 MPI and 128 threads, everything goes well. If I run 2 nodes with 1 MPI and 128 threads on each nodes, all the 128 threads are actually running on a single core. This behavior is confirmed with KMP_AFFINITY=verbose.

An extra question, is the currect mpirun aware of the omni path? I'm using the intel mpi comes with the compiler.

Ye

=================== Ye Luo, Ph.D. Leadership Computing Facility Argonne National Laboratory _______________________________________________ Intel-nda mailing list [email protected] https://lists.jlse.anl.gov/mailman/listinfo/intel-nda