Effective computation of two-dimensional FFT on a homogeneous or heterogeneous cluster

The paper considers performing two-dimensional FFT on a supercomputer. It investigates a dependance of FFT computation time from a matrix size for MVS-100K, MVS-10P and HybriLIT supercomputers. A method of CPU-GPU load balance for a heterogeneous cluster is proposed. For a TESLA K40 card it is shown, that two-dimensional FFT computation time is almost equal to data transferring time. The computation itself is 48 times faster when using GPU comparing to two-processors node. (In Russian).