Comparison And Analysis Of Branch Predictiors In Android Applications

Superscalardeep pipeline technology has been widely used in modern embedded processors. In these processors, efficient execution of the branch instructions is the premise for making full use of the hardware. Therefore, the branch prediction technique is widely used in processors to reduce the penalty caused by branch instructions. On the other hand,with the popularization of Android devices, Android applications become more and more plentiful. Android applications are different from traditional benchmarks, comparing the prediction accuracy of different branch predictors and exploring the influence of microarchitectural parameters of branch predictors in Android scenareos can provide design insights and suggestions to processor designers.This thesis analyzes and modifies the CPU model in Gem5 Simulator to extract branch trace for the trace-driven simulation. This thesis also analyzes the influence of microarchitectural parameters on prediction accuracy of GShare and Bimode predictors, and compare the prediction accuracy in variety of microarchitectural parameters. Also, this thesis analyzes the influence of microarchitectural characteristics on prediction accuracy of BTB and the causes of wrong predictions. Aiming at the problem that BTB can't effectively predict the indirect branch's targets, the thesis implements the VPC prediction mechanism to solve the problem. Through experiments, the thesis has proved the effectiveness of this approche.From results of experiments, this thesis finds that, in android applications, Bimode predictor exceeds the GShare predictor. For GShare predictor, enlarging PHT size can effectively increase the accuracy of prediction until the PHT size reaches to 4K. For Bimode predictor, enlarging PHT size can effectively increase the accuracy of prediction when PHT size is less than 6K. Assuming that direction predictions are all correct, increasing the number of entries can improve the BTB's prediction accuracy when BTB's entries is less than 4096. At this point, the problem that BTB can't effectively predict the indirect branch's target becomes the main problem. Compared to the Bimode and BTB prediction methods, VPC mechanism improves the average branch prediction accuracy from 90.14% to 92.16%, and the penalty caused by misprediction is decreased by 20.5%.