GPU-Based Parallel Researches on RRTM Module of GRAPES Numerical Prediction System

Abstract

GRAPES (Global and Regional Assimilation and Prediction System) is a new generation of numerical weather prediction (NWP) system of China. As the system processes amount of data and requires high real-time,so it is always a hot research field of parallel computing.This is the first time that we use GPU (Graphics Processor Unit) general-purpose computing and CUDA technology on RRTM (Rapid Radiative transfer model) long-wave radiation module of GRAPES_Meso model for parallel processing, we rewrited the RRTM module with CUDA Fortran according to the characteristics of the GPU architecture.Enhancing the computational efficiency with optimization strategys such as the code tuning, asynchronous memory transfer,compiler option and etc. The optimization results indicate that a 14.3×speedup is obtained. Experiments are carried out on the multi-GPU platform,and can be easily extended to GPU clusters, the results show that the parallel computing algorithm is correct , stable and efficient. Index Terms—GPU, CUDA, GRAPES system, RRTM, Parallel computing