DISTRIBUTED-PROCESSING OF A REGIONAL PREDICTION MODEL

This paper describes the parallelization of a mesoscale-cloud-scale nu merical weather prediction model and experiments conducted to assess i ts performance. The model used is the Advanced Regional Prediction Sys tem (ARPS), a limited-area, nonhydrostatic model suitable for cloud-sc ale and mesoscale studies. Because models such as ARPS are usually mem ory and CPU bound, the motivation here is to decrease the computer tim e required for running the model and/or increase the size of the probl em that can be run. Adomain decomposition strategy using a network of workstations produced a significant decrease in problem size relative to a single-workstation run. The performance of the resulting program is described by derived formulas (collectively known as a performance model), which predict the execution time and speedup for different num bers of processors and problem sizes. The interprocessor communication speeds are shown to be the major obstacle to achieving full processor use. The effect of faster communication networks on parallel performa nce is predicted based on this performance model. Parallelization expe riments using the ARPS code were run on a cluster of IBM RS6000 workst ations connected via Ethernet. The message-passing paradigm implemente d here made use of the library of routines from the Parallel Virtual M achine software package.