Date Awarded


Document Type


Degree Name

Doctor of Philosophy (Ph.D.)


Computer Science


Xiaodong Zhang


The fast performance improvement of computer systems in the last decade comes with the consistent increase on power consumption. In recent years, power dissipation is becoming a design constraint even for high-performance systems. Higher power dissipation means higher packaging and cooling cost, and lower reliability. This Ph.D. dissertation will investigate several memory-related design and optimization issues of general-purpose computer microarchitectures, aiming at reducing the power consumption without sacrificing the performance. The memory system consumes a large percentage of the system's power. In addition, its behavior affects the processor power consumption significantly. In this dissertation, we propose two schemes to address the power-aware architecture issues related to memory: (1) We develop and evaluate low-power techniques for high-associativity caches. By dynamically applying different access modes for cache hits and misses, our proposed cache structure can achieve nearly lowest power consumption with minimal performance penalty. (2) We propose and evaluate look-ahead architectural adaptation techniques to reduce power consumption in processor pipelines based on the memory access information. The scheme can significantly reduce the power consumption of memory-intensive applications. Combined with other adaptation techniques, our schemes can effectively reduce the power consumption for both computer- and memory-intensive applications. The significance, potential impacts, and contributions of this dissertation are: (1) Academia and industry R & D has solely targeted the objective of high performance in both hardware and software designs since the beginning stage of building computer systems. However, the pursuit of high performance without considering energy consumption will inevitably lead to increased power dissipation and thus will eventually limit the development and progress of increasingly demanded mobile, portable, and high-performance computing systems. (2) Since our proposed method adaptively combines the merits of existing low-power cache designs, it approaches the optimum in terms of both retaining performance and saving energy. This low power solution for highly associative caches can be easily deployed with a low cost. (3) Using "a cache miss", a common program execution event, as a triggering signal to slow down the processor issue rate, our scheme can effectively reduce processor power consumption. This design can be easily and practically deployed in many processor architectures with a low cost.



© The Author