Algorithm Selection of MPI Collectives Considering System Utilization

MPI collective communications play an important role in coordinating and exchanging data among parallel processes in high performance computing. Various algorithms exist for implementing MPI collectives, each of which exhibits different characteristics, such as message overhead, latency, and scalability, which can significantly impact overall system performance. Therefore, choosing a suitable algorithm for each collective operation is crucial to achieve optimal performance. In this paper, we present our experience with MPI collectives algorithm selection on a large-scale supercomputer and highlight the impact of network traffic and system workload as well as other previously-investigated parameters such as message size, communicator size, and network topology. Our analysis shows that network traffic and system workload can make the performance of MPI collectives highly variable and, accordingly, impact the algorithm selection strategy.