Comprehensive air pollution modeling on a multiprocessor system

W.-C Shin; G.R Carmichael

doi:10.1016/0098-1354(92)80062-E

Comprehensive air pollution models belong to the class of computationally-intensive coupled transport/chemistry problems, whose analysis remains restricted by today's computer technology. Results from the parallel implementation of a regional-scale acid deposition model on the Alliant FX/8 are presented and discussed. The Alliant FX/8 offers both multi-processing and vectorization on eight processors. Multi-tasking efforts are focused on the chemistry module. Exploitation of the spatial parallelism in the chemistry computations results in a speed-up ratio of 2.5. Most of the gain in speed-up is achieved through concurrent calculation rather than vectorization. Parallelization of the STEM-II comprehensive air pollution model was performed on the ALLIANT FX/8. Multi-tasking efforts focused on the chemistry module of the STEM-II model by restructuring the code and utilizing compiler directives. Exploitation of the spatial parallelism in the chemistry computations considerably reduced the total computing time. The speed-up ratios ranged from 1.8 to 2.5, depending on the meteorological conditions simulated. Most of the speed-up was obtained through concurrent calculation rather than vectorization. Even though vector processing is becoming commonplace for most large-scale scientific and engineering applications, it was difficult to vectorize the model because of data dependencies in the most critical section (radical and ionic species module), which consumes a large fraction of the total computing time. For more efficient vectorization, alternative algorithms, which do not impose any data dependency, must be sought, and short vectors must be rearranged to give long vectors. By combining vector processing and multi-tasking of independent processes, it should be possible to achieve higher speed-up ratios. It is clear that the computational requirements of comprehensive air pollution models will continue to increase as our knowledge of the atmospheric processes increases. The increased exploitation of parallel processing seems inevitable. An important question that remains is will this parallelism be of the finegrain mode (thousands of processors) (Carmichael et al., 1989) or the coarse grain mode. The maximum theoretical speed-up using the present approach of parallelizing only the chemistry portions of the model is 7.1 (assuming 86% parallelized and a fixed size of the computational problem). This reflects the present limitations of the algorithm used. Further improvements in execution speed will require exploitation of the spatial parallelism in the horizontal dimensions as well as the transport portions of the code, and the implementation of new algorithms. The successful exploitation of multi-tasking in air pollution applications will depend on the degree to which new designs, software tools, and algorithms are developed.

Comprehensive air pollution modeling on a multiprocessor system

View Online

Abstract

Details

Metrics