Memory requirements to balance thus asymptotically full-speedup FFT computation on processor arrays

Abstract

The paper proves that for a linearly-connected array of alpha processors or a mesh-connected array of alpha /sup 2/ processors, where each processor has computation bandwidth C, I/O bandwidth I and C/I=logm, Omega (m/sup alpha /) memory size is required in each processor to minimize the I/O requirement in balancing the FFT computation. Then it presents balanced FFT algorithms on these arrays to meet their memory size lower bounds. These algorithms are time optimal exhibiting full speedups.<>