Performance Modeling Of Branch Predictors In Out-of-order Processors

With the development of micro-processor technology,a variety of mechanisms have been introduced to improve the overall performance.Branch predictors are introduced to mitigate the control hazards and reduce the pipeline bubble.How to evaluate the performance of a branch predictor,including the number of branch misses and the branch resolution time,has become a hot research topic.Researchers dig the relationship between software micro-architecture independent characters and branch prediction mechanism,and construct analytical models to evaluate branch misses.However,compared to the branch prediction simulation tools,the analytical model's accuracy is relatively poor.On the other hand,the existing branch resolution time models can not cover all the influence factors,and the accuracy of the prediction is aslo unsatisifying.In this thesis,a branch prediction simulator and an ANN branch resolution time model,are constructed to improve the accuracies.The main work of this thesis is divided into three parts.In the first part,three branch predictor simulators(Gshare,Bimode and Tournament)are constructed in python language.Given the configuration options,these simulators can accurately simulate the prediction mechanism of branch predictors.In the second part,we analyze the factors that affect the branch resolution time and the coupling effects between them,and introduce the ANN model to overcome the disadvantage of previous modles.In the third part,the work of the former two parts is integrated into a complete tool,and the graphical user interface(GUI)is provided.Compared to the Gem5 simulations in the AtomicCPU mode,the average accuracy of branch miss prediction is about 98%with the SPEC CPU2000 benchmark and nine branch predictor configurations.Compared to the results from previous analytical module,the accuracy is improved by 28%.In the time domain,the Trace stream has the characteristics of one simulation and many uses,so the more the micro architecture is examined,the lower the cost per unit time,which means the speedup of the branch prediction simulator is proportional to the number of the branch predictor configurations.With the MobyBench benchmark and nine branch predictor configurations,the average accuracy of branch resolution time module,which is based on ANN,is 82%,and compared to the previous none-linear module and simple-assumed module,the average precision is increased by 19%and 27%,respectively.