Modeling superscalar processors via statistical simulation

Abstract

Statistical simulation is a technique for fast performance evaluation of superscalar processors. First, intrinsic statistical information is collected from a single detailed simulation of a program. This information is then used to generate a synthetic instruction trace that is fed to a simple processor model, along with cache and branch prediction statistics. Because of the probabilistic nature of the simulation, it quickly converges to a performance rate. The simplicity and simulation speed make it useful for fast design space exploration; as such, it is a good complement to conventional detailed simulation. The accuracy of this technique is evaluated for different levels of modeling complexity. Both errors and convergence properties are studied in detail. A simple instruction model yields an average error of 8% compared with detailed simulation. A more detailed instruction model reduces the error to 5% but requires about three times as long to converge.