Parallel triangular mesh refinement by longest edge bisection on GPUs

Abstract

In numerical analysis, mesh refinement techniques are used in many different areas such as in Finite Element Methods for the solution of partial differential equations. Formerly, a variety of different mesh refinement techniques have been proposed including both sequential and parallel implementations on clusters and multi-core CPUs. Since, today, both computational capacity and memory bandwidth of GPUs are better and still developing faster than CPUs, general purpose computing on GPUs (GPGPU) has become important in many application areas that require high computation and data throughput. In this thesis, we focus on refining non-uniform triangular meshes and present a new parallel adaptive mesh refinement technique that can be easily implemented on GPU architectures. We also present an implementation of our algorithm on CUDA architecture that achieve significant speed-ups. This thesis includes the algorithm and implementation details, as well as running time analysis and performance comparison of sequential implemetation on CPU and parallel implementation on GPU.