Design And Implementation Of CPU Benchmark Synthesizing Methods Targeted At Android Application

In order to confront new challenges raised by the rapid development of mobile smart devices,operating systems and application softwares,benchmarks are extensively applied in processor designs to reduce simulation duration and guide architecture design space exploration.Targeting at Android applications on ARM processors,the existing benchmark synthesizing methods merely cover several workload characteristics and introduce microarchitecture-dependent characteristics,therefore the generated benchmarks exhibit poor accuracy and cross-architecture performance.Based on microarchitecture independent characteristics,this thesis puts forward new synthesizing methods to cover more characteristics and therefore improves the accuracy and cross-architecture performance of benchmarks.Firstly,on the basis of existing characteristics including instruction mix,register dependence distance,data spatial locality and branch transition rate,this thesis adds approach to implement more microarchitecture independent characteristics targeting at android applications,including critical path length,instruction temporal and spatial locality,data temporal locality,branch and serial block distribution.Secondly,in order to eliminate the coupling between characteristics,this thesis puts forwards new approches of realizing all specified characteristic dimensions in one model program.With target application varying,this thesis designs model programs generation rules and generates amounding model programs as a library.Thirdly,this thesis adopts genetic algorithm to select proper model programs from model library;meanwhile,it uses function call to assemble all programs into benchmarks.In experiments,Mibench is used as Android application to synthesize benchmarks for Android 6.0 and ARMv8 instruction set.It is shown that the average MICA error between benchmarks and orginal workloads is 10.25%while the average CPI error is 9.63%.When altering the architecture configurations,the CPI error is 10.37%.The results demonstrate that synthesizing methods in this thesis performs well in accurancy and cross-architecture feature.