Branch Prediction X Performance: an analysis on Superscalar Processors

Abstract

Simulation is the most used and efficient method to design new processors. It can reproduce and consider the parameters and variables of a real processor execution. Branch prediction is one of these parameters, being very important in the research and design of newer and better processors. In this way, this paper presents a study on the impact of the branch prediction accuracy in the final performance of superscalar architectures. The results were obtained by simulation, using some of the integer and floating-point benchmarks (ammp, equake, gcc, gzip, mesa, vpr) provided by SPEC2000. Simprevar, a variable accuracy branch prediction simulator based on one of the simulators included in the SimpleScalar Tool Set, was used to simulate different prediction accuracies. Our simulations results leads us to conclude that, in some situations, it is better enlarging the hardware than trying to get better accuracy predictors, achieving very similar results.