Parallel DFT Implementation and Benchmarking in Cluster Architecture

Abstract

For the execution of very large DFT (Discrete Fourier Transform), where the implementation size is limited by the memory available in a single processor, it is still convenient to use the larger memory afforded by the use of cluster architectures. In this work we did different parallel MATLAB implementations of a onedimensional DFT through a two-dimensional DFT, which was coded using the row-column algorithm. One version of the code has client-based pre and post-processing stages. The cluster master node was used as the client computer. Since the pre and post-processing involves matrix transpositions, which could pose memory limitations for large data sets, we also did an implementation that distributes the data directly from disc to the cluster cores. This second approach allowed us to quadruple the largest length signal that we could tackle in the cluster architecture. Largest core memory in the nodes should allow even larger increases in signal size. We benchmarked both implementations and did scalability studies using up to 64 cores. Key Terms - FFT, Large Signals, Parallel Processing, Row-Column Algorithm