high performance computing on graphics processing units: hgpu.org

This thesis describes how the performance of job management systems on heterogeneous computing grids can be increased with Graphics Processing Units (GPU). The focus lies on describing what is required to extend the grid to support the Open Computing Language (OpenCL) and how an OpenCL application can be implemented for the heterogeneous grid. Additionally, already existing applications and libraries utilizing GPU computation are discussed. The thesis begins by presenting the key differences between regular CPU computation and GPU computation from which it progresses to the OpenCL architecture. After presenting the underlying theory of GPU computation, the hardware and software requirements of OpenCL are discussed and how these can be met by the grid environment. Additionally a few recommendations are made how the grid can be configured for OpenCL. The thesis will then discuss at length how an OpenCL application is implemented and how it is run on a specific grid environment. Attention is paid to details that are impacted by the heterogeneous hardware in the grid. The theory presented by the thesis is put into practice by a case study in computational biology. The case study shows that significant performance improvements are achieved with OpenCL and dedicated graphics cards.