Parallel computation of the bispectrum

This paper investigates the effectiveness of parallel computing in the calculation of the bispectrum. The bispectrum is estimated by using two different methods namely the direct and the indirect. The direct method employs 1-D FFT algorithms and the indirect method employs the 2-D FFT algorithm to estimate the bispectrum. Both methods have been implemented using 2 different parallel programming techniques: semi-automatic and fully automatic using the Power C Analyzer (PCA). The Silicon Graphics Power Challenge Multiprocessor System (with 12 CPUs) is used to run the parallel codes. Near-linear speedup was observed by employing both techniques. Overall, the maximum speedup of 10.84 at N=12 can be achieved for the direct method and of 8.7 at N=10 for the indirect method using the semi-automatic parallel technique. For the PCA fully parallel technique, the maximum measured speedup for the direct and indirect methods are 10.07 at N=11 and 7.67 at N=12 respectively.