Design And Verification Of Performance Analysis Model For Mobile Application Memory Access

In recent years,with the development of the integrated circuit technology,the computing performance of CPU is getting better and the gap between CPU and memory is becoming more obvious.Memory wall becomes the key problem.Cache is introduced to alleviatethe problem to some extent.Applications show different Cache behavior under different Cache configurations.Analyzing and understanding the Cache behavior of applications is a prerequisite to improve the performance of program memory access.In order to help hardware and software developers quickly and accurately analyze the application Cache behavior,this thesis presents an analysis model of Cache behavior based on traditional stack distance theory for mobile applications.This thesis presents a calculating algorithm of limited-stack distance and designstheanalysismodelof program memory access performance.The main work of this thesis includes two parts.Firstly,this thesis analyzes the shortcomings of the traditional approaches to analyzing the performance of program memory access.Secondly,this thesis puts the limited-stack distance algorithm into the analysis model of Cache behavior for mobile applications.It consists of three core modules:online access sequence analysis,Cache miss type analysis and application cache capacity sensitivity analysis.The online analysis module of memory access sequence is used to decode the memory access instruction and obtain the data address.The Cache miss type analysis module is used to classify the Cache miss based on the limited-stack distance theory and the data address.The Cache capacity sensitivity analysis module is used to analysis the possible causes of the Cache miss and guides the hardware and software developers to design and select Cache.After that,the memory access performance of the applications gets better.To assess the accuracy of the model,this thesis chooses Mibench and Mediabench benchmark suite.The experimental results show that the model can classify Cache miss types under different Cache configurations.Compared the sum of Cache miss statistics withthe statistics from Gem5,the mean error is 1.33%and the maximum absolute error is 2.9%.The number of Cache miss of mpeg2.dec and Typeset applications becomes smaller under the guidance of the model.