Weather Projections and Dynamical Downscaling for the Republic of Panama: Evaluation of Implementation Methods via GPGPU Acceleration

Abstract

Climate change could have a critical impact on the Republic of Panama where a major segment of the economy is dependent on the operation of the Panama Canal. New capabilities to do targeted research around climate change impacts on Panama is therefore being established. This includes anew GPU-cluster infrastructure called Iberogun, based around 2 DGX1 servers (each running 16 NVIDIA Tesla P100 GPUs). This infrastructure will be used to evaluate potential climate models and models of extreme weather events. In this review we therefore present an evaluation of the GPGPU (General Purpose Graphic Processing Unit, here abbreviated GPU) implementation methods for the study of weather projections and dynamical downscaling in the Republic of Panama. Different methods are discussed, including: domain-specific languages (DSLs), directive-based porting methods, granularity optimization methods, and memory layout transforming methods. One of these approaches that has yielded interesting previous results is further discussed, a directive-based code transformation method called ‘Hybrid Fortran’ that permits a high-performance GPU port for arranged lattice Fortran codes. Finally, we suggest a method akin to previous investigations related to climate change done for the Republic of Panama, but with acceleration via GPU capabilities.