High Performance Parallelism Pearls Volume Two

Chapter 16

MPI-3 Shared Memory Programming Introduction

Mikhail Brinskiy*; Mark Lubin^†; James Dinan^†    * Intel Corporation, Russia^† Intel Corporation, USA

Abstract

This chapter discusses an interprocess shared memory extension, which was added in MPI 3.0 standard, and how it can be used to enhance communication efficiency and to enable memory footprint optimization. A simple 1-D ring “hello world” example is shown and then extended for several node runs. Using a modified MPPTEST benchmark, up to 4.9× improvement over a standard point-to-point communication approach on one Intel Xeon Phi coprocessor is shown. This approach is shown to have potential benefits with halo exchanges even for multinode cases. Improvements up to 1.8× with two Intel ...