Author: Ross N. HoffmanPublisher:
ISBN:
Category :
Languages : en
Pages : 42
Book Description
A preliminary exploration is made of the uses of multiprocessing computers for large scale NWP using spectral models. In general if global communication between processors is relatively fast and easy, then implementing spectral models is feasible. The global spectral model is recast in terms of latitude and wavenumber tasks. This approach has a number of advantages: The entire algorithm is macrotasked. Only a handful of crucial pointers need to be locked. The spectral transform calculations are localized so that arithmetic always follows the same ordering and all results are exactly reproducible. The latitude wavenumber tasking scheme is implemented and tested on the Sequent Balance, a shared memory multiple instruction multiple data device. It is argued that this scheme could be easily extended and applied to larger machines of this class and provide a good starting point for distributed memory machines. The potential of single instruction multiple data machines is huge. A proposed algorithm for this class of machine uses a processor for each horizontal grid point. Keywords: Numerical weather prediction; Multiprocessors; Atmosphere models; Weather forecasting. (edc).
