GPU-based high-performance computing.

摘要

In the past few years, multi-core and many-core CPU/GPU architectures have received unprecedented attention and were becoming the only hope to keep Moore's law from working in this age. Because of the change of the computer architecture and the corresponding programming model, software development for the new GPU architectures has not been fully researched and is far behind the development of hardware. Applications and algorithms need to be redesigned and optimized as GPU code so that reasonable comparison can be made between CPU and GPU versions in term of performance.;In this thesis work, the author collects four major research activities that have utilized NVidia Fermi GPU's computation power fully or partially and presents the acceleration procedures with rich details. Some of them, such as the genetic algorithm on traveling salesman problem and circular Hough transform, are more of algorithm test cases while the other works, exact matching approach based on Burrows-Wheelers transform and distributed file system based on secret sharing, are from the author's research, in which several contributions have also been made to the algorithm or system designs other than GPU accelerations.;In the end, based on the different types of applications and their performance results, summaries of the relationship between applications and speedup ratios are made and corresponding suggestions for GPU-based high-performance computing are given.