Memory Access Behavior Modeling For Out-of-order Processors

Due to the significant performance improvement,the Cache hit rate has become one of the most important metrics for performance evaluations.Cache researchers usually use simulations,hardware mesurements or analytical models to obtain Cache performance.However,the first two methods are either time-consuming or expensive.On the other hand,analytical model's accuracy is not good when they are used to forecast Cache hit number cross bench in out-of-oder proceccors,which reduces the value of analytical models.In this thesis,the method of modeling is based on Artificial Neural Networks(ANNs),which was propoased by our lab colleagues.The accuracy of the model is improved by optimizing the training set of the model.The main work of this thesis is divided into two parts.In the first part,a new method of selecting training set is established.In this thesis,a clustering algorithm is used to extract the code fragments,which will be used to build training set for the ANN model,with different memory access characteristics in the program.The accuracy of ANN model trained by the seleted training set is improvedin the aspect of predicting Cache hit rate across different programs.The second part is the implementation of a Cache behavior prediction tool.The previous Cache analytical models just stay in the level of research model,a analytical model that can provide effective help for Cache design should be short time consuming and easy to operate.In order to simplify the operation of the Cache model,this thesis uses Python to implement the model as a complete tool,and leave an extensible interface to add new functionality later.Compared to the results form cycle-accurate Gem5 simulations,the average accuracy of seven benches'Cache hits number predicted by models is about 93%under nine Cache configurations,and compared with the previous models,the precision is improved by 5%.In terms of time,the model propoased in this paer can save up to 90%runtime than that of cycle-accurate Gem5 simulations.In addition,the implemented Cache behavior tool is divorced from external software,which greatly reduces the difficulty of establishing and using Cache model.