IMPROVING THE PERFORMANCE OF A PARFLOW PROXY APPLICATIONTHROUGH CUDA PARALLELISM

As hydrological data becomes more in-depth and is measured at higher resolutions, the need for a fast, efficient software framework to perform analysis and simulation grows. HydroFrame is one such framework that allows users to subset inputs and outputs for watersheds in the United States, simulate the movement of water, and analyze the results of their simulations. An important problem to solve within a framework like HydroFrame is producing simulations in a timely manner. Through the effective use of parallel architectures, we can begin to solve this problem. However, with a large codebase, it can be difficult to map out the extensive set of changes necessary. To minimize the scope of the problem and make a bigger difference in a shorter amount of time, the most computationally intensive code is first extracted. After extracting the code, we can begin to test the effects of applying different computing paradigms to it. Multiple variants of the same original code are then compared against each other to determine not only the performance differences, but also the difference in code complexity and practicality. Some variants utilize multiple CPU cores to perform tasks in parallel through OpenMP, while others utilize GPU cores through OpenCL or CUDA. The results show that while parallelism does increase performance on a CPU, it is more effective on a GPU. In terms of the speedup over a baseline serial variant, the OpenMP variant with 8 threads had a 3.57x speedup, the OpenCL variant had a 6.87x speedup, and the fastest CUDA variant had a 11.6x speedup. The metrics gathered show that CUDA offers a far greater increase in performance over OpenCL and should be utilized if a user’s machine contains a NvidiaGPU