Modeling Research On The Influence Of Speculative Load Mechanism On Out-of-order Processor Performance

Modern superscalar Out-of-Order processors use the Speculative load mechanism to execute Load instructions whose register-ready,but with an indeterminate memory address ahead of time to improve processor performance.The CPI stack model is widely used in the industry to evaluate the performance of the processor,but the previous research did not analyze the impact of the Speculative load mechanism on the modeling of CPI stack.By gem5 cycle accurate simulation,it is found that the Speculative load mechanism has a great impact on the performance of the superscalar Out-of-Order processor.Therefore,the research on the impact of the Speculative load mechanism on processor performance is important for building a more accurate CPI stack model.This thesis focuses on the impact of the Speculative load mechanism on the performance of superscalar Out-of-Order processors.There are two main speculation mechanisms commonly used to implement the Speculative load mechanism: Blind speculation mechanism and Store Sets speculation mechanism.For these two different speculation mechanisms,the main work of this thesis is divided into three parts:(1)study the effects of Speculative load mechanism(including Blind speculation mechanism,Store Sets speculation mechanism)on processor steady-state performance;(2)analysis and quantifies the impact of misspeculation overhead on processor performance under different speculation mechanisms(3)Combines the previous superscalar Out-of-Order processor performance model to build a model that considers the effects of the Speculative load mechanism under different speculation mechanisms.This thesis uses the SPEC CPU2006 benchmark to analyze the impact of the Speculative load mechanism on the performance of superscalar Out-of-Order processors.The artificial neural network is used to model the number of misspeculations under Blind speculation mechanism.The average relative error of the model in application self-prediction,cross-hardware-microarchitecture prediction,and cross-application-cross-hardware-microarchitecture prediction is less than 13%.It is found that when the Store Sets speculation mechanism is adopted,the replay overhead caused by the Speculative load failure can be neglected.The accuracy of our model is not much improved compared with the previous model.When the Blind speculation mechanism is adopted,the improved superscalar Out-of-Order processor performance model improves the accuracy of the previous model by an average of about 10%.