A recent breakthrough in high-performance computing is the application of general-purpose graphics processing units (GPGPUs) to scientific computing. This paper presents a case study of a project to reformulate mathematical computations within a weather model, the Advanced Regional Prediction System (ARPS), to work with GPGPU hardware. This hardware typically consists of hundreds of simple processors, compared to the conventional central processing units (CPUs) with fewer than eight processors (as found in most computers). While GPGPUs are extremely powerful, their usage requires specialized programming to achieve their full potential. As a forecasting tool, ARPS is capable of producing very high-resolution weather simulations. Using ARPS to simulate detailed atmospheric disturbances necessitates the use of large-scale distributed-memory parallel computing clusters. The adaptation of a critical numerical kernel within ARPS for GPGPUs resulted in a six-fold speedup over the CPU version. These optimizations dramatically reduce simulation time, thereby leading to faster weather predictions that may benefit society as well as the research community.

Caption.  Initial data decomposition and thread mapping for ARPS Finite Difference Methods (FDM) kernel.

Acknowledgements. This study is supported by the Coastal Carolina University Academic Enhancement Grant

Referred Conference Proceeding

Whetstone, B., V. Limpasuvan, D. B. Larkins, 2014: GPU Acceleration of the Advanced Regional Prediction System (ARPS), Proceedings to the 52nd Association for Computing Machinery Southeast, Kennesaw, GA.