From Cluster to Grid: A Case Study in Scaling-Up a Molecular Electronics Simulation Code

This paper describes an ongoing project whose goal is to significantly improve the performance and applicability of a molecular electronics simulation code. The specific goals are to (1) increase computational performance on the simulation problems currently being solved by our physics collaborators; (2) allow much larger problems to be solved in reasonable time; and (3) expand the set of resources available to the code, from a single homogeneous cluster to a campus-wide computational grid, while maintaining acceptable performance across this larger set of resources. We describe the sequential performance of the code, the performance of two parallel versions, and the benefits of problem-specific load balancing strategies. The grid context motivates the need for runtime algorithm selection; we present a component-based software framework that makes this possible.