Improving Branch Prediction Accuracy on Multi-Core Architectures for Big Data

Branch prediction is an effective way to improve instruction supply and throughput of processor pipeline. How-ever, branch predictors usually rely on certain amount of branch history to achieve high prediction accuracy through warming up, which is time consuming. Besides, studies on improving performance of big data workloads show that significant commonality exists in instruction streams across different cores running the same workload. Based on the above observation, in this paper, we propose a technique for migrating branch history through thread migration to reduce the training process in the newly starting core. We conduct a variety of simulations to verify the effectiveness of our method. Results show that compared to not migrating, migrating branch history can improve the prediction accuracy in the new core by 24.26% and 34.08% on average for migrating history in T0 and migrating history in T0-T4 respectively. Besides, our method performs well across different cores.