| dc.description.abstract | The very concept of offloading computationally complex routines to a graphics processing unit for general-purpose computing is a problem left wide open to the academic community, both in terms of application as well as implementation, with several different and popular interfaces exploding into popularity within the last twenty years. The OpenMP standard is among the elites in this category, standing as a parallelization interface that has stood the test of time. The goals that the inquiry presented herein seeks to answer are twofold: Firstly, we aim to assess the performance of common sorting algorithms parallelized and offloaded using OpenMP, offloaded to NVIDIA GPU hardware, and secondly, to critically analyze the programmer experience in using an implementation of the OpenMP standard (again, with offloading to NVIDIA GPU hardware) to implement these algorithms. For completeness, the empirical analysis contains a comparison to the unparallelized algorithms. From this data and the impression of the programming experience, strengths and weaknesses of usage of OpenMP for parallelizing and offloading sorting algorithms are derived. After discussing each benchmark in depth, as well as the data derived from the parallelized implementations of each, we found that OpenMP’s position as one of the forefront parallel programming standards is well-justified, with few, but notable, pitfalls for the average programmer. In terms of its performance in parallelizing common sorting algorithms with offloading to NVIDIA GPU hardware, it was found that OpenMP fails to deliver viable implementations of the algorithms that are advantageous over their single-threaded counterparts, though, this was found not to be the fault of OpenMP, but rather, of the inherent nature of offloading to NVIDIA GPU hardware. | en |
