Using GPU processing to solve large-scale scientific problems

As the processing power available in computers grows, so do the applications for using that power for modeling. The growing field of General Purpose Graphics Processing Unit (GPGPU) programming has provided a significant increase in the amount of processing power available in a desktop computer. This thesis will analyze the amount of power provided by GPGPU Programming and its applications in large-scale scientific modeling. The construction and use of computers capable of carrying out this processing will be discussed, followed by the development of some GPU-based modeling as well as compilation of some existing GPU-based models. First the design of the machines used will be discussed. There are several factors in the construction of the computer being used that can impact the performance of modeling. Linear systems of equations are fundamental in almost all areas of scientific modeling and can take a significant amount of time to solve. A GPU-based solver was developed to take advantage of multiple GPUs for extremely large-scale linear systems of equations. A particle based model was also developed to better analyze the data transfer process of CUDA. This model was used to model simple fluids on one or more GPUs. Lastly some existing models were looked at in the scope of GPGPU. CHILD is an existing model used to model erosion and material movement. CHILD was modified to be multithreaded and analyzed for its possible GPU applications. The Weather Research and Forecasting Model (WRF) has existing GPU modules for use; the compilation and use of one of these modules is described