Modeling Of Micro-architecture Independent Software Characteristics For Android Applications

The performance evaluation of embedded processors for mobile terminals requires a representative Benchmark,and the study of the software behavior of mainstream applications has a very important role in generating Benchmark.At present the android application is one of the mainstream applications.Therefore,this thesis analyzes the software behavior of microarchitecture-indepedent workload characterization for android application stage,and selects the program fragment which can represent the whole application.According to the android application execution behavior characteristic,based on the existing microarchitecture-independent characteristic parameters,serial instructions and its distribution,the distribution of jump addresses in the branch behavior,the average number of register reads and writes and the data spatial locality four types of microarchitecture-independent workload characterization were added in this thesis,a total of 227 microarchitecture-independent workload characterization.At the same time,a dimensionality reduction method based on correlation elimination method and genetic algorithm were used to reduce the 227 microarchitecture-independent workload characterization,and 77 microarchitecture-independent workload characterization were obtained.The average correlation coefficient of microarchitecture-independent workload characterization before and after dimension reduction was 0.835.Finally the principal component analysis and K-means clustering algorithm were used to cluster all the feature fragments,and representative feature fragments were extracted to represent the whole android application execution workload characterization.In this thesis,the Gem5 simulator is used as a research platform to cut the nine android applications in Moby Benchmark based on the fixed number of cycles(10 million)to extract the microarchitecture-independent characteristics and the microarchitecture-dependent characteristics of all feature fragments.The experimental results show that the average error of the instruction per cycle(IPC)is 1.29%for representive feature fragments with multiple android applications,and the average error of L1 level Instruction Cache miss rate(L1 I-Cache miss)is 3.84%,the average error of L1 level Data Cache miss rate(LI D-Cache miss)is 3.73%,and the average error of Branch prediction error rate(Branch miss)is 7.85%.Therefore,the final selection of the typical feature fragments can better represent the whole android application execution workload characterization.