Latency-Balanced Optimization of MPI Collective Communication across Multi-clusters

Recently, the demands of resource sharing and cooperation across multiple HPC centers are growing rapidly. Such multi-cluster HPC systems provide much more computing resources than single cluster. It challenges the scalability and performance of current parallel computing tools, especially the collective communication of parallel applications. This paper focuses on the optimization of collective communications of parallel applications across multiple HPC centers. A scalable latency-balanced broadcast algorithm (LB-B) for MPI across multi-cluster is proposed in this paper. It can dynamically adapt to the topology of the system and improve the communication performance.