Effect of CGMT Processors on Branch Prediction

CGMT can exert significant influence on the design and implementation of branch prediction. In this paper, the design of several typical branch predictors such as bimodel, GShare and McFarling is examined and branch prediction performance is improved by making some decisions with the help of TPC-W benchmark concerning whether to share some data structures among multiple threads or not. For a typical two-level adaptive branch predictor such as GShare, RAS & BHR should be kept thread private, BHT & BTB should be shared among threads. Also we make special effort to improve BTB hit rate besides keeping it shared. Algorithm called victim BTB with similar thought of victim cache is presented and a couple of variant algorithm, inter-thread victim BTB and intra-thread victim BTB are described in detail. Experimental results show that BTB hit rate is improved considerably when BTB size is not very large.