WIT Press


High Performance Multithreaded Message Passing On A Myrinet Cluster

Price

Free (open access)

Paper DOI

10.2495/HPC020241

Volume

27

Pages

Published

2002

Size

594 kb

Author(s)

A. Alves, A. Pina, J. Exposto & J. Rufino

Abstract

The main purpose of this paper is to present the impact of pCoR thread level message passing facilities on applications running in a Myrinet cluster. To exploit Myrinet technology we use the GM library, which provides a lim- ited number of ports as abstractions to name communication end-points. pCoR communication layer multiplexes GM ports by using a dispatcher thread to handle messages to/from a large number of communication entities (working threads). Our approach combines polling operations executed by multiple threads to avoid unnecessary context switching; a simple mechanism is used to engage working threads into the polling scheme. We aim to reduce the total number of polling operations required to hold message passing system performance. 1 Introduction Low latency communication technologies and SMP workstations make possible to build clusters that attain the performance of massively parallel machines. However, software development for such systems is still a major concern because it is hard to take full advantage of hardware capabilities. pCoR [1] communication layer focus on two main aspects: user level communication and our particular concept of distributed multithreaded applications. 1.1 User level communication systems Nowadays operating systems provide built in communication abstractions to inter- face a variety of network adapters. Operating systems communication primitives inflict unacceptable delays because of excessive context switching and memory

Keywords