A High Performance Parallel Sparse Linear Equation Solver Using CUDA
Department of Computer Science, College of Arts and Sciences, Kent State University
Kent State University, 2011
The management of electric power systems requires continuously computing the powerflow of a power system in real-time. For large power systems, this task is often beyond the capabilities of modern CPUs. Concurrent computation is an attractive approach to accelerating it. However, the powerflow computation requires solving a large system of sparse linear equations. This problem is difficult to parallelize on classic parallel computer architectures. In this thesis I implement an algorithm that solves a system of sparse linear equations on the GPU using NVIDIA’s CUDA programming language. I develop software that randomly produces power system topologies that mimic real world systems. These power systems produce sparse matrices, which are then factorized using a method known as bi-factorization. I test my algorithm on the random power system topologies of progressively larger sizes. My experiments indicate that with a system size of 9,000 buses, my algorithm achieves a speedup of thirty-eight times over the CPU algorithm. My results demonstrate the practicality of GPU-based powerflow computation.
October 17, 2011 by hgpu