Project Overview

MicroStAT (A Microarchitect’s Statistically-based Analysis Toolkit), which is a suite of statistically-based simulation and performance analysis tools. The components of this toolkit are shown in Figure 1. In particular, this project will initially focus on developing three key tools. The first tool (listed under “Fractional Multifactorial Designs (FMD)” in the figure) will be a statistically-based benchmark that is architecture independent that can identify a processor’s performance bottlenecks (and their relative ordering) and the effect of a new processor feature. This benchmark will be based on a fractional multifactorial design-of-experiments. A distinguishing novelty of this tool is that it does not require a simulator, which makes it possible to determine the performance bottlenecks in silicon processors. The requisite bridge between this tool and prior work will be a sampling-based version of the fractional multifactorial design that requires only a single simulation run – as opposed to N simulations, where N is the number of performance bottlenecks being measured. The second tool uses signal-processing and statistical concepts to determine, statically and dynamically, respectively, the ideal sampling frequency for a benchmark (“Determining Optimal Sampling Rate”). Sampling a benchmark at the ideal sampling frequency minimizes the simulation time while maintaining a high level of accuracy. Finally, the third tool is a framework that will automatically create reduced input sets for feedback-directed optimization and to stress different parts of the processor (“Automaticallycreated reduced input sets”).

MicroStAT architecture overview
Figure 1. The components of MicroStAT

A unique aspect of this project that will significantly enhance its relevance to real-world computing environments is the collaboration of the Freescale Semiconductor Corporation’s performance evaluation group in Austin, Texas. This group has long been responsible for the performance evaluation of processors for both commercial and technical workloads. Thus, their involvement adds an important dimension to this project.

This project is supported by NSF grant #0541162.