Optimization On The Trace Profling Method For Processor Analytical Modeling

In the processor performance analysis model of CPI stack theory,the steady state throughput model and the memory-level parallelism model need an analysis method based on the sliding instruction window.At this stage,whether it is a single instruction or a whole window sliding strategy,the size of the Reorder Buffer(ROB)is used as the size of the instruction window.However,according to the experimental observations in this thesis,ROB is at a low level full state for most of the time.Therefore,the ROB size as the instruction window size is essentially an approximation,this will cause some error in the accuracy of the subsequent steady state throughput model and the memory-level parallelism model.This thesis analyzes various factors affecting the ROB fullness,including hardware parameters such as ROB size,IQ,Cache,physical register,front-end width,number of back-end execution units,and software load characteristics,and analyzes the internal mechanism of the processor.Interpreted experimental data.Then an empirical model for evaluating ROB fullness is established based on both linear and nonlinear aspects.In terms of linearity,the first step is to reduce the dimension by Spearman analysis,so that the independent variables that are not related to the dependent variable are eliminated.The second step is to eliminate the independent and insignificant independent variables of the dependent variable by stepwise regression.The linear regression model reveals the positive and negative correlation and influence weight of different hardware and software parameters on the effective ROB.In the non-linear aspect,it is mainly based on the BP neural network training model.In this thesis,the neural network of various structural combinations is used to predict the error of self-application prediction.Finally,the 18 nodes of the hidden layer are selected,the excitation function tansig,and the training method traingd is the best combination.In this thesis,14 benchmarks from SPEC 2006 were used to verify the effective ROB model and its accuracy improvement on the theoretical model of CPI stack.In terms of the error of the effective ROB prediction model,the predicted average errors of the linear-based stepwise regression model in self-application,self-application cross-hardware configuration and cross-application cross-hardware configuration are 8.81%,8.94% and 9.33%,respectively.The predicted average errors of the nonlinear neural network model based on self-application,self-application cross-hardware configuration and cross-application cross-hardware configuration are 6.71%,7.41% and 8.21%,respectively.Compared with the stepwise regression model,the neural network model has no significant reduction in prediction error,but the training time is three orders of magnitude higher.Therefore,this thesis selects a stepwise regression model to predict the final model of effective ROB.The effective ROB predicted by stepwise regression model replaces the default ROB as window size for Trace analysis.The average accuracy of steady-state throughput model,memory-level parallelism model and total CPI stack theoretical model is improved by 9.31%,2.78% and 10.62% respectively.