Abstracts Earth and Environmental Sciences

Earth : an efficient architecture for running threads

by Kevin Bryan. Theobald

The two current approaches to increasing computer speed are giving individual processors the ability to exploit instruction-level parallelism (ILP), and harnessing multiple processors to work together on a single problem. The first approach has worked well so far, but more and more investments in instruction scheduling hardware seem to be producing diminishing returns. Parallel processing, on the other hand, has tremendous potential, but has been disappointing in practice. One reason is that parallel computer designers cannot afford to put the same resources into building processors as the makers of commodity chips, who command a much larger market. Therefore, most parallel architects have turned to off-the-shelf microprocessors, which do not support parallel processing well. Multithreaded systems based on dataflow principles promise a solution to the problems inherent in many of today's parallel machines. A multithreaded processor supporting a program execution model designed according to such principles would be a better building block for parallel machines than today's ILP processors, because multithreading and dataflow address the problems facing contemporary parallel machines, such as latency and synchronization. Acceptance of such systems in a world ruled by commodity systems requires taking an evolutionary approach, in which the multithreaded machine is initially emulated on an existing multiprocessor based on off-the-shelf microprocessors, while still achieving good performance, and then transformed into more powerful versions by gradually replacing the stock components with custom hardware. This dissertation is about EARTH (Efficient Architecture for Running Threads), a multithreading model suitable for such an evolutionary approach. The thesis begins with a study of program parallelism which identifies important properties which affect the design of a multithreaded system. The EARTH model is defined at several layers of abstraction, and several implementations along the evolutionary path are discussed. The Threaded-C language, used for expressing algorithms in the EARTH model, is presented, and several applications are written in this language to illustrate its use. Experimental results show that EARTH lives up to its name, that it can run parallel programs efficiently, and that this efficiency improves as the machine moves along the evolutionary path and custom multithreading hardware replaces off-the-shelf hardware piece by piece.