Efficient Utilization of Fine-Grained Parallelism using a microHeterogeneous Environment

The goal of this thesis is to propose a new computing paradigm, called micro- Heterogeneous computing or mHC, which incorporates PCI (or other high speed local system bus) based processing elements (vector processors, digital signal processors, etc) into a general purpose machine. In this manner the benefits of heterogeneous computing on scientific applications can be achieved while avoiding some of the lim itations. Overall performance is increased by exploiting fine-grained parallelism on the most efficient architecture available, while reducing the high communication over head and costs of traditional heterogeneous environments. Furthermore, mHC based machines can be combined into a cluster, allowing both the coarse-grained and fine grained parallelism to be fully exploited in order to achieve even greater levels of performance. An existing high performance computing API (GSL) was chosen as the interface to the system to allow for easy integration with applications that were previously developed using this API. The ensuing chapters will provide the motivation for this work, an overview of heterogenous computing, and the details pertaining to microHeterogeneous comput ing. The framework implemented to demonstrate a microHeterogeneous computing environment will be examined as well as the results. Finally, the future of micro Heterogeneous computing will be discussed