Abstract :
The topology functionality of the Message Passing Interface (MPI) provides a portable, architecture-independent means for adapting application programs to the communication architecture of the target hardware. However, current MPI implementations rarely go beyond the most trivial implementation, and simply performs no process remapping. We discuss the potential of the topology mechanism for systems with a hierarchical communication architecture like clusters of SMP nodes. The MPI topology functionality is a weak mechanism, and we argue about some of its shortcomings. We formulate the topology optimization problem as a graph embedding problem , and show that for hierarchical systems it can be solved by graph partitioning . We state the properties of a new heuristic for solving both the embedding problem and the "easier" graph partitioning problem. The graph partitioning based framework has been fully implemented in MPI/SX for the NEC SX-series of parallel vector computers. MPI/SX is thus one of very few MPI implementations with a non-trivial topology functionality. On a 4 node NEC SX-6 significant communication performance improvements are achieved with synthetic MPI benchmarks.
