Comparative study on behavior-based dynamic branch prediction using machine learning

Modern processors fetch and execute instructions speculatively based on the outcome of branch prediction for decreasing effect of control hazards. Many branch predictors are proposed in literature to increase accuracy of the branch prediction. Some ones use machine learning technique for improving accuracy of predicting conditional branches. In this paper, we investigate this issue by evaluating different branch predictors through using a well-designed set of correlation patterns. We built a framework for testing performance of different branch predictors. Our framework demonstrates efficiency of using machine learning in predicting conditional branches. This framework is designed for mimicking various behaviors of branch predictions and can be used easily by scholars to check performance of more branch predictors. Experimental results shown in this work illustrate performance of applying different approaches proposed for predicting conditional branches in comparison with employing machine learning technique. Our findings illustrate that using machine learning provides competitive results. However, employing machine learning does not help in predicting all behaviors of conditional branches.