Home // INFOCOMP 2017, The Seventh International Conference on Advanced Communications and Computation // View article

What Do Scientific Applications Need? An Empirical Study of Multirail Network Bandwidth

Authors:
Edgar A. Leon
Chris Chambreau
Matthew L. Leininger

Keywords: HPC; performance measurement; multirail networking; network characterization; scientific applications.

Abstract:
High performance computing applications are commonly executed on large parallel machines composed of commodity components. These commodity clusters utilize high-speed interconnects that provide low latency and high bandwidth such as InfiniBand. Understanding the characteristics of scientific applications is important to properly configure and tune these machines and their software stacks. Are applications limited by network performance? Can they leverage increased network bandwidth? What type of network operations and message sizes do they use? This work provides a better understanding of the communication requirements of scientific applications by investigating the impact of multirail networking on their performance. We measure the performance of a suite of high performance computing mini-applications under different multirail configurations to determine their sensitivity to network bandwidth. The selected mini-applications provide simplified source code containing data access patterns and computational characteristics of larger production codes. The type of analysis presented in this paper can be applied to inform the procurement of future systems maximizing application productivity within a given capital budget.

Pages: 35 to 39

Copyright: Copyright (c) The Government of USA, 2017. Used by permission to IARIA.

Publication date: June 25, 2017

Published in: conference

ISSN: 2308-3484

ISBN: 978-1-61208-567-8

Location: Venice, Italy

Dates: from June 25, 2017 to June 29, 2017