Abstract
Fast Fourier Transforms (FFTs) are an essential part of many scientic codes: from Molecular Dynamics to Climate Modelling. It is, therefore, evident that HPCx requires effcient methods forperforming FFTs and related calculations. This study compares the performance of the two main FFT libraries on HPCx: IBM's ESSL/PESSL and FFTW. Both serial and parallel (distributed-memory only) 3D complex-to-complex FFT routines are investigated, and the performance of the two different libraries is investigated.
In general, the ESSL and FFTW serial 3D FFT routines are comparable. For parallel FFTs, the PESSL library is, in general, slightly faster, however, FFTW has better parallel effciency. FFTW measured plans are extremely expensive to compute and only give a modest improvement in performance over estimated plans.
Some further comments are made about the overall performance of HPCx, and its impact of the use of FFT library routines.
In general, the ESSL and FFTW serial 3D FFT routines are comparable. For parallel FFTs, the PESSL library is, in general, slightly faster, however, FFTW has better parallel effciency. FFTW measured plans are extremely expensive to compute and only give a modest improvement in performance over estimated plans.
Some further comments are made about the overall performance of HPCx, and its impact of the use of FFT library routines.
Original language | English |
---|---|
Publisher | UoE HPCx Ltd. |
Publication status | Published - Jun 2003 |