Exploiting Hierarchy in Parallel Computer Networks to Optimize Collective Operation Performance

• Venue: IPDPS '00 Proceedings of the 14th International Symposium on Parallel and Distributed Processing
• Year: 2000

Initial efforts to efficiently implement collective communication operations model network communication and produced 驴optimal驴 trees based on those models. Those models assumed equal point-to-point latencies between any two processes. This assumption is violated in heterogeneous systems such as clusters of SMPs, and therefore, collective operations that utilize those trees perform suboptimally. Work that is more recent has focused on creating two-layer network topology-aware trees for collective operations that minimize communication across slower channels.We present a strategy based upon a multilayer view of the network, which takes advantage of communication cost differences at every level in the network. We used this strategy to implement topology-aware versions of MPI collective operations in MPICH-G, the Globus-enabled version of the MPICH implementation of MPI. Using topology information discovered by Globus, we construct topology-aware trees dynamically. We present results demonstrating the advantages of our multilevel approach.