high performance computing on graphics processing units: hgpu.org

A project was undertaken to improve the performance of a traditional CPU-based sequential program by modifying it for parallel execution in a GPU environment. A speedup of at least 1.5x and the preservation of the program’s accuracy and integrity were outlined as the two key goals of the project. Deal.II, a differential applications analysis library, was selected as the program to be modified. Modifications to the vector multiplication functionality of the program were made which materially improved the execution time of the solve class within deal.II. After installing a baseline and modified version of the deal.II program on the Electrical and Computer engineering department cluster, multiple runs of a selected benchmark program were conducted in order to gather performance data. The data obtained provided clear evidence of both a speedup and preservation of program accuracy. An average speedup of 1.59x was obtained across all measured workloads, with a peak speedup of 2.44x under optimal conditions. The program’s accuracy was verified by comparing the output of the benchmark program with a baseline output. The project was successful in accomplishing all of its initially stated goals while incurring no cost to the department; all resources were provided free of charge to the group.