Improving the O-GEHL branch prediction accuracy using analytical results

Abstract

The O-GEHL branch predictor has outperformed other prediction schemes using the same set of benchmarks in an international branch prediction contest, CBP-1. In this paper, we present the analysis results on each of the OGEHL branch predictor tables and also on the optimal number of predictor tables. Two methods are subsequently proposed to help increase the O-GEHL prediction accuracy. The first one aims to increase the space utilization of the first predictor table by dynamically adjusting the lengths of branch history regarding to the type of a benchmark currently in execution. The second one adds an extra table into the O-GEHL predictor using the space saved from the sharing of hysteresis bits. Experimental results have confirmed that both schemes improve the accuracy of two different predictor configurations, leading to two promising research directions for future explorations.

Keywords

• branch predictor
• perceptron
• predictor analysis
• o-gehl
• neural networks