Performance of Network Subsystems for a Technical Simulation on Linux Clusters

A. Boklund, C. Jiresjö, S. Mankefors-Christiernin, N. Namaki, L. Christiernin, and M. Ebbmar (Sweden)

Keywords

High Performance Computing and Networks, Cluster Computing, Network Performance

Abstract

The focus of this paper is to evaluate the impact of the network infrastructure on the total execution time for non-linear industrial finite element simulations and to investigate the applicability of micro-benchmarks on this area of applications. For this purpose we selected an industrial strength simulation application which utilizes the MPI message passing library. Three network technologies (Gigabit Ethernet, SCI and InfiniBand) were evaluated from both a micro-benchmark and an application level perspective. The overall result is that although there are applications that benefit from being able to send vast amounts of small messages fast as well as other applications that need to move several Gigabytes of data thus utilizing either the low latency or high bandwidth of specialized network interconnects, non linear FEM analysis is not one of them.

Important Links:



Go Back