Design Of Template-based CPU Benchmark Synthesis Method

With the rapid development of mobile intelligent terminal devices such as smart phones and tablet computers,the evaluation of the hardware performance of mobile smart devices has become focus of the industry.Benchmarks are widely used in processor designs to reduce simulation time and provide guidance for exploring architectural design space.This thesis analyzes the typical benchmark into the methodological mechanism,useing a mixed genetic simulated annealing algorithm to optimize the template selection method,and generating benchmarks for real-world application scenarios on ARM and MIPS instruction set platforms for experimental verification.The current benchmark synthesis methods are mainly divided into two categories: direct synthesis methods and template-based synthesis methods.The direct synthesis method is difficult to construct program statistics flow graphs for large-scale and complex real applications,resulting in poor accuracy of synthesis benchmark.Template-based synthesis methods,according to the workload characteristics of the target application,some features matching template programs are selected from the template library for splicing and a benchmark is synthesized.Compared to the direct synthesis method,the template-based synthesis method is more suitable for large-scale and complex-featured target applications.Therefore,this thesis studies the template-based benchmark synthesis method.For the case of more workload feature dimensions and non-orthogonal features,the template selection method in the template-based synthesis method has a poor effect.In this thesis,a mixed genetic simulated annealing algorithm is used to improve the template selection method,and the effectiveness of the improved template selection method is verified through comparative experiments.Finally,the benchmark synthesis method in this thesis,including three stages of data extraction,template selection and template splicing,is used to construct the benchmarks on the ARM and MIPS instruction set platform for specific application scenarios to verify the accuracy of this synthesis method.Experimental results show,on the ARM instruction set platform,for the Android system startup scenario,BBench and BaiduMap respectively construct benchmark.The average similarity between the selected template of the template selection stage and the original load in the microarchitecture-independent feature is 0.802,which is better than the result of the previous template selection method.The average similarity of the microarchitecture-independent features between the generated benchmark and the target application is 0.801,and the average CPI error is 10.05%.On the MIPS instruction set platform,for SPEC CPU 2017,the average similarity between the synthesized benchmarks and the original workload in the locality characteristics of access memory is 0.784.