NPAC Technical Report SCCS-533

The Parallelization of a Weather Predition Model

Gregor von Laszewski

Submitted October 01 1993


Abstract

Recent devastating events caused by storms show how important the development of a reliable storm prediction system is. One such effort was started at the Center for Analysis and Prediction of Storms (CAPS) in Oklahoma. This institute tries to develop techniques for the practical prediction of weather phenomena on scales ranging from a few kilometers to hundreds of kilometers. The use of a prediction model is possible in near future because a network of 175 Doppler radars will be installed around the U.S in order to gather the necessary initial data for the model. The initialization is an important part of this application since the precise initial data leads to more accurate predictions. It is easy to imagine that this data accumulation leads to huge storage requirements. One way to deal with this problem is to distribute the prediction of a storm at the area of interest. Since, many sites have different computers it is useful to develop a highly portable source code. Another important requirement is that a modification of the program should be possible with moderate effort in order to incooperate new computational schemes determining the modeling equations. Naturally a storm has to be predicted as fast and long as possible with very high accuracy in order to avoid damages by invoking early prevention methods. This can be achieved by, e.g. using supercomputers for solving the modeling equations. This report describes the parallelization of a weather prediction code using the dataparallel programming scheme. In addition, it is shown how to obtain a version for message passing computers while using High Performance Fortran. Benchmarking for the dataparallel program is done on a CM5 with 32 nodes and vector units.


PostScript version of the paper