Paper showing Linpack scalability of mainstream clusters

RICHARD,BRUNO (HP-France,ex1) bruno_richard at hp.com
Mon Sep 10 00:45:01 PDT 2001


Available from <http://lib.hpl.hp.com/techpubs/2001/HPL-2001-206.html>:

I-Cluster: Reaching TOP500 performance using mainstream hardware
 
By B. Richard (hp Laboratories Grenoble), P. Augerat, N. Maillard, S. Derr,
S. Martin, C. Robert (ID Laboratory)

Abstract: A common topic for PC clusters is the use of mainstream instead of
dedicated hardware i.e., using standard desktop PCs and standard network
connectivity, with technology to organize them so that they can be used as a
single computing entity.  Current work in this "off-the-shelf cluster"
domain usually focuses on how to reach a high availability infrastructure,
on how to efficiently balance the work between nodes of such clusters, or on
how to get the most computing power for loosely-coupled (large grained)
problems. hp Labs Grenoble, teaming with INRIA Rhône-Alpes, teamed up to
build a cluster out of 225 standard hp e-PC interconnected by standard
Ethernet, with the objective of getting the highest computational
performance and scaling from the simplest desktop PC to the most powerful
computers in the world. As an additional constraint, we decided to use a
cluster that models a modern enterprise network, using standard machines
interconnected through standard Ethernet connectivity. This paper describes
the issues and challenges we had to overcome in order to reach the 385th
rank in the TOP500 list of most powerful supercomputers in the world on June
21st, 2001, being the first mainstream cluster to enter TOP500 ever. Also we
provide hereafter some details about the software and middleware tuning we
have done, as well as the impact of different factors on performance such as
the network topology and infrastructure hardware.





More information about the Beowulf mailing list