Towards Optimized Hydrological Forecast Prediction of WRF-Hydro on GPU

Hydrological models are designed to simulate and predict the dynamic cycling of surface and groundwater. These models represent simplifications of complex hydrological pro-cesses and play an integral role in hydrological research and real-world applications. The WRF Hydro model is widely used due to its multi-scale capabilities for land-atmosphere modeling. Recent studies have focused on enhancing the performance of different modules in WRF through GPU parallel optimization techniques. However, WRF - Hydro has not been able to utilize the GPU for computations which leads to lower computational efficiency when simulating hydrological calculations over large regions for extended periods. In this paper, we first analyze the computational and communication bottlenecks in WRF Hydro. We then propose several methods for performance optimization accordingly. To address the computational bottleneck, we develop efficient GPU-based overland routing and channel routing, eliminating redundant calculations. Additionally, we adopt an improved block partitioning strategy to enhance memory access efficiency on the GPU. To tackle the communicational bottleneck, we adopt a packetized communication strategy to reduce the number of communication operation calls in channel routing and simplify the preprocessing operations for communication. Meanwhile, for overland routing, we employ multi-stream to minimize redundant synchronization operations and overlap communication with computation to reduce communication overhead. The evaluation results demonstrate that WRF -Hydro can achieve significant speedups in several simulation cases after applying our proposed optimizations, resulting in a maximum speedup of 6.67 ×.