PC Clusters

Clusters of off-the-shelf PCs have been hailed by enthusiasts as inexpensive alternatives to supercomputers, and clusters have been demonstrated to work well for a number of scientific applications. PC clusters might develop a "mass market" for parallel computing, which today is done almost exclusively at large supercomputer centers. For example, clusters would be ideal for developing and testing parallel codes and for running small to medium-sized parallel applications, freeing up large capability resources at NERSC and other centers for what they do best.

However, because there is no standardized and comprehensive software package for clusters, building and maintaining a cluster is not easy. Every new cluster project has to reinvent the wheel, and they often wind up with the software equivalent of a system held together with duct tape.

NERSC is working to develop the infrastructure for plug-and-play, high performance PC clusters. The goals of our PC Cluster Project are to develop critical enabling software components, to ensure that there is a uniform HPC software environment on PC clusters and supercomputers, and to perform and collect objective analyses of hardware and software and disseminate that information so that the evaluation does not need to be repeated.

One of our tasks is to develop a Modular Virtual Interface Architecture (M-VIA). NERSC contributed to the development of VIA as part of our COMPS Project (Cluster of MultiProcessor Systems), and Intel, Microsoft, and Compaq have accepted VIA as the standard architecture for communication within clusters.

NERSC's M-VIA will be the first publicly available full implementation, the first portable high performance implementation, and the first implementation for Linux. We are working with Intel and Argonne National Laboratory to develop more applications, extensions, and implementations of VIA.

NERSC also participates in the Extreme Linux project, a collaboration of the DOE laboratories, NASA, NIST, and several universities and commercial vendors, which aims to improve Linux for high performance clusters.

Our own cluster consists of 20 single-processor 400 MHz Pentium II nodes, 2 four-processor Pentium Pro nodes, and 2 dual-processor 233 MHz Pentium II front-end servers.