MicroStAT: A Microarchitect's Statistically-based Analytical Toolkit

点击翻译按钮获取中文摘要

Slow simulation speed and a vast number of user-configurable processor and compiler parameters preclude a detailed exploration of the design space. Although architects have proposed solutions such as reduced input sets and sampling to minimize the effect of slow simulation speed, what has not been proposed are systematic statistically-based approaches to pare the design space. Moreover, when evaluating the efficacy of a new processor feature or when comparing the performance of two processors, architects typically look at high-level single metrics or the behavior of a specific processor component in isolation. While these metrics and techniques can give the architect insight into the processor?s behavior, in addition to not being statistically-based, they are potentially limited by being too abstract or narrowly focused. The fundamental problem with both cases is that they do not systematically analyze the processor?s performance. To address the dual problems of poor simulation speed ? especially as it pertains to design space reduction ? and haphazard performance analysis, the goal of this project is to develop a suite of statistically-based simulation and performance analysis tools. Some of the key tools of this suite include a tool for identifying a processor?s performance bottlenecks (and their relative ordering); a tool for evaluating the effect of a new processor feature, and a tool for improving level of accuracy of simulation sampling techniques by finding an ideal sampling frequency using signal-processing and statistical concepts. Since MicroStAT will be publicly available at the end of this project, academic researchers and industry will be able to improve the statistical-rigor of their simulation methodology and, thus, the quality of their simulation results. Furthermore, this project will make substantial contributions to educational development. It will support the dissertation research of Ph.D. students in electrical and computer engineering and computer science, along with several master?s projects and research opportunities for several undergraduate students.