Mixed mode programming on HPCx

Abstract

Clusters of shared memory nodes have become a system of choice for many research and enterprise projects. Mixed mode programming is a combination of shared and distributed programming models and naturally matches the SMP cluster architecture. It can potentially exploit features of the system by replacing the message exchanges within a node with faster direct reads and writes from memory, using message passing only to exchange information between the nodes. Several benchmark codes, based on a simple Jacobi relaxation algorithm, were developed; a pure MPI (Message Passing Interface) version and three mixed mode versions: Master-Only, Funneled, and Multiple. None of the mixed mode versions managed to outperform the pureMPI, mainly due to longer MPI point-to-point communication times. Results will be presented and reasons behind the performance losses discussed.