Thanks Ralph, I'm trying to find out what can be accomplished in binding using the command-line and when I need to generate a mapping file. Using the command line I find is typically more robust. It is just too easy to forget to adapt a mapping script when moving between systems.
For the sake of argument one use case would be to map a hybrid MPI+threads application onto a set of dedicated nodes. That is, to have MPI processes evenly distributed over hosts, sockets and NuMA domains, but ranked compactly (MPI tasks that are close in rank in MPI_COMM_WORLD are more likely to be in the same NuMA domain). Processes should be bound to multiple cores in the NuMA domain, but not overlap cores with other MPI processes unless over-subscribed. The threading library would then rebind threads to cores within this subset. I just happened to choose LAMA because I found some not too old presentations on-line about it and it looked very adaptable. If there are other schemes that are in better shape I'd be happy to use whatever is recommended. With the general command line options -map-by -bind-to I have not been able to have the MPI processes bound to a subset of the cores in the NuMA region, only to the whole NuMA node or a single core. The more detailed --mca rmaps* options I have not yet found a good presentation about so I am not sure what I can do with it and how. Pointers to what can be done with parts of OpenMPI that is actively maintained are most welcome. Cheers
