A new transpose split method for three-dimensional FFTs: performance on an Origin2000 and Alphaserver cluster

We discuss a new transpose split method for parallel computations of three-dimensional Fourier transforms. By splitting the data along two dimensions over the processors it allows for a higher degree of parallelization than the original transpose split method. The traditional transpose split method involves one alltoall communication. The new method involves two communication steps in which each processor performs an ”alltoall” communication in groups. The performance of the new method has been evaluated using MPI on an Origin2000 and an Alphaserver cluster and compared with the traditional transpose split method. We found that the extra communication step introduced in the new method only slightly increases communication time. However, an efficient parallelization depends critically on how fast the communications can be performed.