Hello everyone, We have implemented a set of modules to sandbox OpenMPI code within a SimGrid simulation. As we believe that emulation can be convenient to test some OpenMPI parts at a reduced cost, we would like to share the modules with you. They could for example be used to test collective algorithms on various platforms, using a single machine instead of requiring access to real platforms.
The modules are freely available (https://framagit.org/simgrid/openmpi-rsg-plugins) and are in a quite experimental state for the moment. We have modified the ODLS and PML components, respectively from the default and ob1 modules in OpenMPI's code base. In short, the approach taken here is to manage the "simulation world" in a dedicated orchestration process, to connect all usual processes to it and to make the processes controllable by the orchestration one. Our PML directly handles the calls it receives and does not delegate tasks to other components (bml, btl...). We plan to increase the robustness and the usability of our modules (and the underlying tools they lean on), but their design should not face dramatic changes now. I'll contact you again after a cleanup phase but feel free to give us any feedback in the meantime :). Best regards, Millian Poquet
_______________________________________________ devel mailing list devel@lists.open-mpi.org https://lists.open-mpi.org/mailman/listinfo/devel