Parallel FFT Package Performance Testing
TimeTuesday, July 246:30pm - 8:30pm
DescriptionThe fast Fourier transform (FFT) algorithm for computing discrete Fourier transform (DFT) of a sequence is one of the most important algorithms in digital signal processing and data analysis. Its importance is rooted in the fact that it reduces the number of arithmetical operations from O(N2) to O(N log N), so that DFT computations are made practical in real applications. Despite its algorithmic success on single processor, a parallel FFT is usually needed for multi-dimensional data, and it requires careful implementation of parallel algorithms since data are distributed on multiprocessors. For high performance computing, parallel FFT has found wide applications in three-dimensional (3D) data processing, from computational fluid dynamics to first principles electronic structure calculations for solids. Given the importance of parallel FFT for 3D data, we will test the performance of several open source parallel FFT packages, including P3DFFT, PFFT, and openFFT, on XSEDE platforms. Since these packages employ different data distribution and communication strategies, their performance results on XSEDE platforms for different problem sizes and different number of processors will provide useful guideline for application developers and XSEDE users in making optimal choice of data distribution strategies.