high performance computing on graphics processing units: hgpu.org

The computational landscape is dominated by the use of a very high number of CPU resources; this has however provided diminishing returns in recent years, pushing for a paradigm shift in the choice for computational systems. The following work was aimed at determining the maturity of heterogeneous computer systems in terms of computational performance and their possible integration within High-Performance Computing resources through the use of the OpenCL parallel programming platform. An introduction is given in the existing hardware architectures targeted by the OpenCL platform, existing literature regarding the integration of heterogeneous systems for computational applications, and the OpenCL platform as a development environment. A number of applications are developed to benchmark the capabilities of the framework in multi-architecture environments, the results of which show up to 160 times performance gain when targeting GPU architectures, as opposed to CPU, for matrix multiplication algorithms. Based on this, an extensive test-bench is designed targeting the HTCondor resource pool for a Fast-Fourier Transform application. Results from these machines once again showed a significant performance increase against CPU systems, while also enabling the expansion of the HTCondor system and the uncovering of 30 Teraflops of dormant computing power. The FPGA architecture is also investigated for its potential in OpenCL computational acceleration, with a focus on the platforms ease of use. It is determined that the framework is mature enough for FPGA application development.