Small-College Supercomputing: Building A Beowulf Cluster At A Comprehensive College† Joel Adams David Vos Department of Computer Science Calvin College Grand Rapids, MI 49546 {adams, dvos12} @calvin.edu Abstract A Beowulf cluster is a MIMD multiprocessor built from commodity off-the-shelf personal computers connected via a dedicated network, running free open-source software. Such a cluster can provide a supercomputer’s performance at a small fraction of one’s cost. For small colleges and universities, the relatively low cost of a Beowulf cluster makes it an attractive alternative to a commercial supercomputer. This paper details our experience building a Beowulf cluster at a four-year comprehensive college. 1 Introduction In the early 1990s, NASA Goddard Space Flight Center researchers Donald Becker and Thomas Sterling needed more computing power, but funding for their projects was decreasing. They saw themselves as needing to be “liberated” from supercomputer vendors, for three reasons: • Commercial supercomputers were very expensive. • The proprietary (binary) software that came with commercial supercomputers could not be customized. • The small market for supercomputers combined with the high R&D costs to develop them was driving most supercomputer vendors bankrupt — voiding their maintenance contracts and making upgrades impossible. To gain their “liberty,” Becker and Sterling built their own supercomputer by turning “a pile of PCs” into a MIMD multiprocessor, using ethernet and free open-source software (e.g., Linux, MPI [6], PVM [5]). To reflect their “liberation” theme, they chose names for their machines from the medieval epic in which Beowulf liberated the Danes from the monster Grendel. Such multiprocessors have since come to be known as Beowulf clusters [3][8]. †This work was supported by NSF grant MRI-0079739. Becker and Sterling’s goal was to achieve at least 1 Gflop performance for roughly $50,000. Their 1994 cluster — Wiglaf — consisted of 16 100-MHz 486 DX-4 PCs, connected with 10-Mbit/sec ethernet in a triple-bus topology, and customized Linux ethernet drivers to spread the communications traffic evenly across the three networks. Wiglaf achieved a top speed of 42 Mflops. Their 1995 attempt — Hrothgar — substituted 100-MHz Pentium PCs for the 486s, and 100-Mbit/sec ethernet for the 10 Mbit/sec ethernet. These modifications improved Hrothgar’s performance to 280 Mflops. In 1996, Mike Warren from Los Alamos National Labs built Loki [10] out of 16 200-MHz Pentium Pro PCs connected with 100 Mbit/sec ethernet using a hybrid star and hypercube topology. Loki cost $63,000 and achieved 1.2 Gflops. Within a year, the price for its components had dropped to $28,000, making it one of the first clusters to achieve 1 Gflop for less than $60,000. Since that time, the decreasing prices of components have made it feasible for institutions, departments, and even individuals to build their own Beowulf clusters that achieve a supercomputer’s performance for a fraction of one’s cost. Figure 1 shows the growth in cluster-numbers over time: Figure 1. Clusters at www.beowulf.org by Year. 0 20 40 60 80 100 120 1995 1997 1999 2001