Adaptive Selection of Communication Methods to Optimize Collective MPI Operations

Many parallel applications from scientific computing use collective MPI communication operations to distribute or collect data. The execution time of collective MPI communication operations can be significantly reduced by a restructuring based on orthogonal processor structures or by using specific point-topoint algorithms based on virtual communication topologies. The performance improvement depends strongly on numerous factors, like the collective MPI communication operation, the specific group layout, the message size, the specific MPI library, and the architecture parameters of the parallel target platform. In this paper we describe an adaptive approach to determine and select a specific processor group layout or communication algorithm for the realization of collective communication operations with the objective of minimizing the communication overhead. In the case that a communication method is faster than the original implementation of the collective MPI communication operation, the specific communication method is applied to perform the communication operation.