high performance computing on graphics processing units: hgpu.org

In the autonomic computing context, we perceive the system as an ensemble of autonomous elements capable of self-managing, where endusers define high-level goals and the system shall adapt to achieve the desired behaviour. This runtime adaptation creates several optimisation opportunities, especially if we consider approximate computing applications, where it is possible to trade off the result accuracy and the performance. Given the power consumption limit on modern systems, autonomic computing is an appealing approach to increase the computation efficiency. I divided this PhD thesis into three main sections. The first section focuses on a dynamic autotuning framework, named mARGOt, which aims at enhancing the target application with an adaptation layer to provide selfoptimisation capabilities at the production phase. In this context, the enduser might specify complex high-level requirements, and the proposed approach automatically tunes the application accordingly. The second section evaluates the mARGOt framework, by leveraging its features in two different scenarios. On the one hand, we evaluated the orthogonality between resource managers and application autotuning. On the other hand, we proposed an approach to enhance the application with a kernel-level compiler autotuning and adaptation layer in a seamless way for application developers. The third section focuses on two application case studies, showing how it is possible to significantly improve computation efficiency, by applying approximate computing techniques and by using mARGOt to manage them.