What MPI Could (and Cannot) Do for Mesh-Partitioning on Non-homogeneous Networks

摘要

We discuss the mesh-partitioning load-balancing problem for non-homogeneous communication systems, and investigate whether the MPI process topology functionality can aid in solving the problem. An example kernel shows that specific communication patterns can benefit substantially from a non-trivial MPI topology implementation, achieving improvements beyond a factor of five for certain system configurations. Still, the topology functionality lacks expressivity to deal effectively with the mesh-partitioning problem. A mild extension to MPI is suggested, which, however, still cannot exclude possibly sub-optimal partitioning results. Solving instead the mesh-partitioning problem outside of MPI requires knowledge of the communication system. We discuss ways in which such could be provided by MPI in a portable way. Finally, we formulate and discuss a more general affinity scheduling problem.