Implementation of Parallel 1-D FFT on GPU Clusters

In this paper, we propose an implementation of a parallel one-dimensional fast Fourier transform (FFT) on GPU clusters. This implementation is based on the six-step FFT algorithm. Because the parallel one-dimensional FFT requires three all-to-all communications, one goal for parallel FFTs on GPU clusters is to minimize the PCI Express transfer time and the MPI communication time. We demonstrate that the advanced features of MVAPICH2-GPU make it easy to overlap PCI Express transfers and MPI communication. Performance results of one-dimensional FFTs on a GPU cluster are reported. We successfully achieved a performance of over 763 GFlops on 128 nodes of the HA-PACS (268 nodes, 2.99 TFlops/node, 802 TFlops peak performance) for 2³⁴-point FFT.