GPU Computing for Longwave Radiation Physics: A RRTM_LW Scheme Case Study

GPU computing has been accepted as viable alternatives for compute-intensive applications. The radiative process is an important atmospheric physics process and the RRTM_LW radiative scheme has been chosen to be one of the five benchmark kernels by NCAR for atmospheric applications. We accelerate the RRTM_LW scheme on three different GPU platforms and obtain a speedup of 27.6×. Three decomposition strategies are utilized to exploit data parallelisms existing in various kernels. A systematic performance analysis is performed in the aspects of GPU hardware features, execution configurations, register file utilizations and the application characteristics. Several observations are achieved: (1) the performance of GPU applications is affected greatly by clock rates, (2) a 5.4% to 9.5% performance discrepancy exists between various execution configurations, and (3) hardly any benefit can be brought to the state-heavy atmospheric applications by bounding the register file usage.