Exploring the Potential of Large Scale Distributed Modeling of Snow Accumulation and Melt on GPUs

Water from snow melt is a critical resource in watersheds of the western US, Canada, and other similar regions of the world. The distribution of snow and melt-water controls the temporal and spatial distributions of soil moisture, evapo-transpiration (ET), recharge, stream-aquifer interaction and other hydrologic processes within the watershed. It also influences the quantity and timing of water availability in downstream areas. In spite of the serious impacts on the water resources at multiple scales, the knowledge base for prediction of snow accumulation and melt in mountainous watersheds is notably weak. Physics-based, distributed snow models such as UEB, SNTHERM, SHAW and ISNOBAL, have positioned themselves as an appropriate tool for understanding of snow-process interactions and prediction of melt, and have been applied in numerous watersheds to varying degrees of success. In spite of the significant advances in hardware speed and programming efficiency, the application of the above-mentioned snow models has mostly been limited to small watersheds. Application of these models at finer spatio-temporal resolution, in large domains, and for longer time periods, to address problems such as quantifying the response of snow-dominated watersheds to climate change scenarios, is restrictive due to the large computational cost involved. Additionally, the computational requirement of current generation snow models is expected to rise as improved snow-depth characterization and a tighter coupling with hydrologic processes are incorporated. This poses considerable challenge to their application in feasible time. We suggest alleviating this problem by taking advantage of high performance computing (HPC) systems based on Graphics Processing Unit (GPU) processors. High performance GPUs work like SIMD processors, but can take advantage of larger number of cores thus providing higher throughput. As of June 2010, the second fastest supercomputer in the world uses NVidia Tesla GPU;s and it is expected that the future supercomputers will be based on hybrid architecture involving GPUs. In this work we develop a distributed, two-layered, physics-based snow model (of kernel structure very similar to ISNOBAL), that works on GPUs. We demonstrate the effectiveness of the model and compare its computational efficiency in relation to CPU based systems.