Analytical Performance Modeling of Hierarchical Interconnect Fabrics

Abstract

The continuous scaling of nanoelectronics is increasing the complexity of chip multiprocessors (CMPs) and exacerbating the memory wall problem. As CMPs become more complex, the memory subsystem is organized into more hierarchical structures to better exploit locality. During the exploration and design of CMP architectures, it is essential to efficiently analyze their performance. However, performance is highly determined by the latency of the memory subsystem, which in turn has a cyclic dependency with the memory traffic generated by the cores. This paper proposes a scalable analytical method to estimate the performance of highly parallel CMPs (hundreds of cores) with hierarchical interconnect fabrics. The method can use customizable probabilistic models and solves the cyclic dependencies by using a fixed-point strategy. The technique is shown to be a very accurate and efficient strategy when compared to the results obtained by simulation.