Linux kernel support for Micro-Heterogeneous Computing

Heterogeneous Computing (HC) is a technique speeds the computation of large tasks by utilizing multiple computers or supercomputers, each of which is best suited to a particular type of computation. Micro-Heterogeneous Computing (MHC) has been proposed to bring this practice to individual computers. With the aid of the higher speed, short distance interconnects found within the next generation of personal computers and commodity servers, it should be possible to apply HC to relatively small grain sizes. MHC draws on the observations that in order for a new computing technology to become widely accepted, and cost effective, there must be a suitable abstraction layer that frees the application writers from the need of precise technical knowledge about the system. MHC provides such an abstraction layer, referred to as the MHC framework, which provides automated solutions to many of the problems that must be overcome when utilizing HC. The framework was designed with the goals of user transparency, flexibility, and performance. The problems addressed include matching tasks to devices and scheduling them (collectively known as mapping), dependency analysis, and parallelization of serial code. All of these problems are solved dynamically at run time by the framework whose implementation is discussed herein. In support of this framework, this thesis specifies a format for libraries that provide common functions and free their users from the tasks of code profiling and analytical benchmarking. This thesis provides the first implementation for such an abstraction layer by utilizing the Linux operating system. This thesis provides not only the kernel level support necessary to schedule tasks to hardware, but also implements the entire core framework, with functioning solutions to the problems mentioned above. This thesis provides well-defined interfaces and methods to expand the MHC system with new scheduling heuristics, function libraries, and device drivers