Chapter 6. ParaStation MPI variables & application performance

Table of Contents

Scope and audience
ParaStation MPI run-time tuning parameters
General scaling tuning variables
InfiniBand specific tuning variables
MPI tuning for connected service types - addressing scalability problems with InfiniBand

Scope and audience

This chapter is intended for use by scientists and engineers who wish to explore the performance of their application codes using parameters that modify the behavior of the Massage-Passing Interface (MPI) and/or the placement of processes within a distributed memory cluster. This document details the parameters available to the user's environment that will impact application performance for the ParaStation MPI release of the Message-Passing Interface (MPI).

Any MPI and pscom variables found to be universally beneficial with respect to application performance have already been selected as defaults in the ParaStation MPI environment. The relative benefit of employing any of the parameters detailed herein is dependent upon specific application characteristics. In particular, the degree of inter-node message passing, message size and extent, frequency and type of collective operations used will all vary from one application to the next.

Prior to modifying any of the variables detailed in this chapter, the user is encouraged to make efforts to understand the communication characteristics of their code and how they relate to the physical hardware and interconnect topology.