An Efficent Methodology for Realization of Parallel FFT for Large Data Set

Abstract

The paper presents an efficient methodology for the parallel realization of FFT on multiprocessor architecture for larger data sets. FFT algorithm demands batch processing on data; large data sizes thus gives rise to performance related issues namely latency, execution time degradation triggered by cache misses, buffer management etc. The realization of FFT on multiprocessor platform presented here is based on two dimensional formulation, wherein an N point radix 2 FFT is re-ordered as a two dimensional matrix organization.The implementation employs data partitioning and parallelism so as to minimize the number of cache misses thereby ensuring cache friendliness.The implementation also employs parallel pipelined architecture. Parallel pipelined architecture and two dimensional formulation of FFT coupled with vectorization improves the performance figure of FFT realization of large data sets. A detailed analysis is presented in this paper.