Research On Branch Prediction Migration For Multi-core Architectures

Branch predictor is an important component of modern superscalar multi-core processors.Increasing the prediction accuracy of the branch predictor can reduce the delay caused by misprediction penalty to increase the throughput of the pipeline.Even increasing a small number of prediction accuracy,there is a huge performance gain for microprocessors under a huge command base.The modern advanced branch predictor is based on a large number of branch prediction history.In the multi-core architecture,the thread from one core to another,the new core's branch prediction history is missing.It makes the branch predictor's hit rate drop significantly,which affects the performance of the core.Based on the background of thread migration in multi-core environment,this paper migrates branch history information to improve the hit rate of branch predictor.Based on the branch simulation competition simulated environment,we migrate branch history information and analyze the migration hit rate statistically.Based on different circumstances,we propose different migration strategies.We evaluate different migration strategies to find the optimal migration strategy.Specific work and innovations are shown as follows:1.In the multi-core environment,we set the state-of-the-art TAGE branch predictor as the research target,and migrate the branch prediction historical information between cores.After branch information migration,we increased branch prediction hit rate.In addition,we investigate the possible impact parameters of branch prediction migration.We test and analysis these parameters by simulation experiment.2.According to the characteristics and problems of branch prediction migration,some migration strategies are proposed in this paper.These migration strategies are simulated and analyzed with extensive simulation experiment.In this paper,with a large number of theoretical analysis and simulation experiments,we conclude that when a thread migrates to an idle core,the accuracy of the branch prediction can be improved by migrating the branch prediction history information.In addition,several migration strategies are proposed in this paper to make the best of performance improvement and try to solve the problem of data inconsistency caused by the migration process.Experimental results show that,compared with the experimental benchmark without branch prediction information,branch prediction migration can reduce the average MPKI value by 43.46% on the new core.