A comparative study of shared memory parallelism on regular and irregular data structures using OpenMP and Galois

Task and/or data parallelism can be exploited in most applications on today's multi-core processors. However, inefficient data organization, data dependencies, and hardware constraints limit scalable parallelization of these applications. In this thesis, performance and the impact of some optimizations is compared and evaluated for simple tasks using two parallel frameworks, OpenMP and Galois. Additionally, their performance on three real life applications, High Accuracy Relativistic Magnetohydrodynamics (HARM) which operates on a grid data structure; Delaunay Triangulation, which refines a triangulated mesh; and Dynamic Fracture Propagation, which operates on a triangulated mesh with adaptive refinement; is evaluated. It is found that OpenMP is a simple yet powerful tool for parallelization of most regular applications and workloads. Galois, which is specially designed for irregular data patterns, performs well for graph like structures. However, neither of them are well suited for all tasks and other frameworks must be explored to find one that is simple to use and yet powerful for all possible applications