high performance computing on graphics processing units: hgpu.org

High Performance Computing is essential to continued advancement in many scientific and engineering fields. In recent years, due to the scale of the platforms and the breakdown of laws which had long since supported rapid expansion, energy efficiency has emerged as a new design constraint on HPC platforms and applications. This constraint has increased the heterogeneity found in HPC nodes, seen in the form of higher CPU core counts and the adoption of specialised hardware accelerators. In an effort to address this new metric, models have been developed to capture and predict power and energy consumption, of application and systems. In this thesis, we survey the state of the art in modelling and perform an evaluation of the existing methods. We then address the problem of accurate application energy and power measurement at multiple levels of granularity, including node, components and cluster level. We contribute multiple measurement tools targeted at researchers, and include case studies demonstrating their application, to facilitate standard methods in our field. Finally, we introduce a methodology for selecting the ideal accelerator device with respect to performance and energy efficiency to execute a given algorithm.