Date of Award
Department of Computer Science
Thomas H. Cormen
In applications using large amounts of data, hiding the latency inherent in accessing data far from the processor is often necessary in order to achieve high performance. Several researchers have observed that one way to address the challenge of latency is by using a common structure: in a series of passes, the program reads in the data, performs various operations on it, and writes out the data. Passes often consist of a pipeline structure composed of different stages. In order to achieve high performance, the stages are frequently overlapped, for example, by using asynchronous threads. Out-of-core parallel programs provide one such example of this pattern. The development and debugging time resulting from coordinating overlapping stages, however, can be substantial. Moreover, modifying the structure of the overlap in an attempt to achieve higher performance can require significant additional time on the part of the programmer. This thesis presents FG, a Framework Generator designed to coordinate the stages of a pipeline and allow the programmer to easily experiment with the pipeline's structure, thus significantly reducing time to solution. We also discuss preliminary results of using FG in an out-of-core sorting program.
Hamon, Elizabeth Anne, "Enhancing Asynchronous Parallel Computing" (2003). Dartmouth College Undergraduate Theses. 206.