Commodity Cluster Computing for Computational Chemistry

Access to high-performance computing power remains crucial for many computational chemistry problems. Unfortunately, traditional supercomputers or cluster computing solutions from commercial vendors remain very expensive, even for entry level configurations, and are therefore often beyond the reach of many small to medium-sized research groups and universities. Clusters of networked commodity computers provide an alternative computing platform that can offer substantially better price/performance than commercial supercomputers. We have constructed a networked PC cluster, or Beowulf, dedicated to computational chemistry problems using standard ab initio molecular orbital software packages such as Gaussian and GAMESS-US. This paper introduces the concept of Beowulf computing clusters and outlines the requirements for running the ab initio software packages used by computational chemists at the University of Adelaide. We describe the economic and performance trade-offs and design choices made in constructing the Beowulf system, including the choice of processors, networking, storage systems, operating system and job queuing software. Other issues such as throughput, scalability, software support, maintenance, and future trends are also discussed. We present some benchmark results for the Gaussian 98 and GAMESS-US programs, in order to compare the processor performance (and price/performance) with other computing platforms. We also analyse the efficiency and scalability of the parallel versions of these programs on a commodity Beowulf cluster. We believe that the Beowulf cluster we have constructed offers the best price/performance ratio for our computational chemistry applications, and that commodity clusters can now provide dedicated supercomputer performance within the budget of most university departments.