Evaluation of a Dynamic Load-Balancing Molecular Dynamics Application using Automated HW/SW Architecture Generation

The application of reconfigurable computing technology to solve high performance computing problems has received increasing attention in recent years. N-body problems such as Molecular Dynamics simulations are common candidates for such studies. The ability to perform dynamic load-balancing offers the potential of improving performance, but the implementation of hardware architectures implementing load balancing is often tedious and error prone. Further, the isolation and quantification of performance gains rendered by the integration of dynamic load balancing hardware presents a challenge for system developers. In prior work, the authors have proposed a dynamic load balancing architecture for supporting Molecular Dynamics computations. This paper presents an evaluation of the impact of dynamic load balancing on a Molecular Dynamics application by comparing several candidate architectures against a previously-developed dynamic load balancing architecture. A hardware/software generation tool has been employed to automatically create the candidate architectures and architecture simulators for comparison. The results indicate that while specialized architectures in certain circumstances can perform better than the dynamic load balancing architecture, the integration of dynamic load balancing offers a consistently efficient use of FPGA resources, independent of variations in proximity between molecules.