The Development and Integration of a Distributed 3D FFT for a Cluster of Workstations

In this paper, the authors discuss the steps taken in the formulation of a parallel 3D FFT with good scalability on a cluster of fast workstations connected via commodity 100 Mb/s ethernet. The motivation for this work is to improve the performance and scalability of the Distributed Particle Mesh Ewald (DPME) N-body solver. Scalability issues in the FFT and DPME as an application are presented separately. Also discussed are scalability issues related to the networking hardware used in the cluster. Results indicate that the existence of a parallel FFT significantly improves performance in DPME from a maximum of 5 processors to at least 24 processors on a cluster of workstations. This has an associated increase in speedup from 4 to 12 times faster than the serial version.