THE PARALLEL COMPUTATION OF GUANGZHOU AREA NUMERICAL PREDICTION MODEL

The parallel computation of the recently developed Guangzhou area numerical prediction model was researched by using Message Passing Interface (MPI) method. According to the characters of model structure and model computing process, the model is suited to parallel computation by using horizontal section-method. The parallel computation of the model has been realized by using one dimensional and two dimensional section-method, respectively, and parallel efficiency, parallel acceleration-rate and percent of parallel communication-time of the model have been tested on DAWN-3000 parallel computer. Results show that two parallel methods can finish 72-hour forecast in one hour by using 8 CPUs, the parallel efficiency of one dimensional section-method keeps about 90%, which can meet the operational need. When more than 8 CPUs are used in the model, the communication-time of the model increases rapidly and beyond 50% of the computation time, the parallel efficiency of the model decreases dramatically. One dimensional section-parallel-method of the model is simpler and more efficient by using the same CPU than two dimensional section-parallel-method of the model.