Performance analysis of memory hierachies in high performance systems.

摘要

This thesis studies memory bandwidth as a performance predictor of programs. The focus of this work is on computationally intensive programs. These programs are the most likely to access large amounts of data, stressing the memory system. Computationally intensive programs are also likely to use highly optimizing compilers to produce the fastest executables possible. Methods to reduce the amount of data traffic by increasing the average number of references to each item while it resides in the cache are explored. Increasing the average number of references to each cache item reduces the number of memory requests. Chapter 2 describes the DLX architecture. This is the architecture on which all the experiments were performed. Chapter 3 studies memory moves as a performance predictor for a group of application programs. Chapter 4 introduces a model to study the performance of programs in the presence of memory hierarchies. Chapter 5 explores some compiler optimizations that can help increase the references to each item while it resides in the cache.