Avalon: An Alpha/Linux Cluster Achieves 10 Gflops for $150k Finalist for the 1998 Gordon Bell Price/Performance Prize Michael S. Warren Timothy C. Germann Peter S. Lomdahl David M. Beazley John K. Salmon Abstract As an entry for the 1998 Gordon Bell price/performance prize, we present two calculations from the disciplines of condensed matter physics and astrophysics. The simulations were performed on a 70 pro- cessor DEC Alpha cluster (Avalon) constructed entirely from commodity personal computer technology and freely available software, for a cost of 152 thousand dollars. Avalon performed a 60 million particle molecular dynamics (MD) simulation of shock-induced plas- ticity using the SPaSM MD code. The beginning of this simulation sustained approximately 10 Gflops over a 44 hour period, and saved 68 Gbytes of raw data. The resulting price/performance is $15/Mflop, or equivalently, 67 Gflops per million dollars. This is more than a factor of three better than last year’s Gordon Bell price/performance winners. This simulation is similar to those which won part of the 1993 Gordon Bell performance prize using a 1024-node CM-5. This simulation continued to run for a total of 332 hours on Avalon, computing a total of floating point operations. This puts it among the few scientific simulations to have ever involved more than 10 Petaflops of computation. Avalon also performed a gravitational treecode N-body simulation of galaxy formation using 9.75 million particles, which sustained an average of 6.78 Gflops over a 26 hour period. This simulation is exactly the same as that which won a Gordon Bell price/performance prize last year on the Loki cluster, at a total performance 7.7 times that of Loki, and a price/performance 2.6 times better than Loki. Further, Avalon ranked at 315th on the June 1998 TOP500 list, by obtaining a result of 19.3 Gflops on the parallel Linpack benchmark. Keywords: Molecular Dynamics, SPaSM, N-body problem, treecode, Beowulf, price/performance 1 Introduction Building upon the foundation of the BEOWULF project [4] and our own success with Loki [9], it has become possible to construct high-performance computers entirely out of commodity components and open source software, thus obtaining a significant price/performance advantage over typical parallel machines. Last year Loki and Hyglac, clusters of 16 Pentium Pro processors, were the first such machines to win a Gordon Bell price/performance prize [14]. This year, changing to the the DEC Alpha microprocessor (which is also found in the Cray T3E series) and using a more advanced fast ethernet switch, we have improved total performance by almost a factor of ten, and improved price/performance by over a factor of three. In 1992, Warren and Salmon were awarded a Gordon Bell Performance Prize [10] for “Astrophysical N- body Simulations Using Hierarchical Tree Data Structures.” In 1993, Lomdahl and Beazley were awarded a Gordon Bell Performance Prize [8] for “50 GFlops Molecular Dynamics on the Connection Machine 5.” It Theoretical Astrophysics, Los Alamos National Laboratory, Los Alamos, NM 87545, Email: msw@lanl.gov Condensed Matter Theory, Los Alamos National Laboratory, Los Alamos, NM 87545, Email: tcg@lanl.gov Condensed Matter Theory, Los Alamos National Laboratory, Los Alamos, NM 87545, Email: pxl@lanl.gov Department of Computer Science, University of Chicago, Chicago, IL 60637, Email: beazley@cs.uchicago.edu CACR, California Institute of Technology, Mail Code 206-49, Pasadena, CA 91125, Email: johns@cacr.caltech.edu 1