Design and Performance Evaluation of Global History Dynamic Branch Predictors

The importance of accurate branch prediction to future processors has been widely recognized. The correct prediction of conditional branch outcomes can help avoid pipeline bubbles and attendant loss in performance. In order to achieve high prediction accuracy, numerous dynamic branch prediction schemes that exploit branch correlation have recently been proposed. Several of the best predictors are the gshare, agree, filter, skewed and bi-mode predictors. However, despite the intensive research work to propose these new schemes, there is no direct and comprehensive performance comparison among them. Such comparison is essential to guiding the design of future microprocessors. Therefore, in this paper, we conduct extensive experiments to calibrate the performance of each prediction scheme. Furthermore, we discuss the design philosophy and underlying mechanism for these schemes, and contrast their relative advantages and disadvantages. Among the schemes examined, we find that the skewed predictor performs the best for small budgets, while the bi-mode predictor outperforms others for large budgets.