A Statistical Exploration of the CMP Design Space

Abstract

Computer architects use simulation extensively to design computer systems, and they primarily rely on their intuition and detailed knowledge of a system to guide the simulation experiments that they perform. However, using a statistical design to construct an experiment and using statistical analysis to interpret the results will reduce the simulation runs necessary to optimize the system, enhance understanding of the main factors and interactions of the system, and increase confidence in the attained optimal design point. Moreover, the advent of Chip-Multiprocessors (CMPs) is challenging established ideas about multiprocessor design. The architect can use a statistical design to quickly explore the novel design space of CMPs and evaluate which factors are more important to improve performance or to reduce cost. This paper describes the method of constructing fractional factorial designs and analyzing the results to build a statistical model of the CMP design space.