Efficient filtering techniques for finite-difference atmospheric general circulation models on parallel processors

摘要

Finite-difference-based atmospheric general circulation models using latitude/longitude meshes typically invoke filtering to suppress numerical instabilities occurring near the poles. We present a more efficient methodology for performing this filtering on parallel processors. This involves redistribution of the computational workload among all available processing elements, along with either the invocation of fast Fourier transforms to replace the slower numerical convolutions, or the smoothing of the transfer functions of enable use of fewer terns in the convolution sum. Results are presented on contemporary parallel architectures such as the Cray-T3D.