Autonomic behavioural framework for structural parallelism over heterogeneous multi-core systems.

With the continuous advancement in hardware technologies, significant research has been devoted to design and develop high-level parallel programming models that allow programmers to exploit the latest developments in heterogeneous multi-core/many-core architectures. Structural programming paradigms propose a viable solution for e ciently programming modern heterogeneous multi-core architectures equipped with one or more programmable Graphics Processing Units (GPUs). Applying structured programming paradigms, it is possible to subdivide a system into building blocks (modules, skids or components) that can be independently created and then used in di erent systems to derive multiple functionalities. Exploiting such systematic divisions, it is possible to address extra-functional features such as application performance, portability and resource utilisations from the component level in heterogeneous multi-core architecture. While the computing function of a building block can vary for di erent applications, the behaviour (semantic) of the block remains intact. Therefore, by understanding the behaviour of building blocks and their structural compositions in parallel patterns, the process of constructing and coordinating a structured application can be automated. In this thesis we have proposed Structural Composition and Interaction Protocol (SKIP) as a systematic methodology to exploit the structural programming paradigm (Building block approach in this case) for constructing a structured application and extracting/injecting information from/to the structured application. Using SKIP methodology, we have designed and developed Performance Enhancement Infrastructure (PEI) as a SKIP compliant autonomic behavioural framework to automatically coordinate structured parallel applications based on the extracted extra-functional properties related to the parallel computation patterns. We have used 15 di erent PEI-based applications (from large scale applications with heavy input workload that take hours to execute to small-scale applications which take seconds to execute) to evaluate PEI in terms of overhead and performance improvements. The experiments have been carried out on 3 di erent Heterogeneous (CPU/GPU) multi-core architectures (including one cluster machine with 4 symmetric nodes with one GPU per node and 2 single machines with one GPU per machine). Our results demonstrate that with less than 3% overhead, we can achieve up to one order of magnitude speed-up when using PEI for enhancing application performance