On Jun 6, 2013, at 9:29 AM, "Nima Aghajari" <email@example.com
first of all thanks for your reply. I tried specifying the --slot-list option like you proposed. This will unfortunately lead to the same mpirun with 5 cores. Adding another slot-list command for the second program, e.g.
mpirun -np 4 --slot-list 0-3 prog_1 : -np 1 --slot-list 0 prog_2
will actually run on only 4 cores, but now it takes more than triple the time as needed before on 5 cores. I suppose there should be some overhead because of the oversubscription but that definitely seems too much. Any suggestions?
Not really - oversubscription is pretty painful, depending on what mechanism you use to "idle" prog_2. It could be polling if you aren't careful. You also have to see how you are "idling" prog_1 once it completes its task - again, it could be polling hard if it is in a barrier in MPI_Finalize.
users mailing list