Studies On The Trace Sampling Method For Processor Analytical Modeling

There are two main methods for processor performance analysis,one is the micro-architecture simulation and the other is analytical modeling.The fastest cycle-accurate simulators today are still nearly five orders of magnitude slower than the actual hardware execution.The analytical models build equations based on mechanistic analyses or empirical fittings,and use the statistical information collected from the software trace as the models' input to quickly analyze the performance of different architectures.However,the analytical model needs to profile a large amount of software trace information.The trace profiling time is about three to four times slower than the trace generation time,and the previous research on the analytical model ignores this problem.Therefore,this thesis proposes a trace sampling method for the analytical model,which can further improve the efficiency of the analytical model.The analytical model of this thesis is based on the CPI stack model of interval analysis,including the steady state throughput model,the memory-level parallelism model and the cache miss rate model.Through research and experimental analysis of different sampling methods,this thesis proposes a two-level trace sampling method for the input of the model,Includes interval sampling based on program phase analysis and statistical sampling of trace information within the interval.The interval of the first-level sampling refers to the program fragment of the fixed instruction length,and the representative intervals are selected and sampled through the online phase analysis of the program;the second level sampling is to sample the trace analysis unit of a specific model and complete the corresponding trace information statistics.This thesis analyzes the sampling results under different parameter selection in detail,and determines the reasonable sampling parameters through the corresponding index analysis.In addition,this thesis quantifies the influence of two-level sampling parameters on performance estimation,and gives the calculation method of sampling error estimation.Twenty-one benchmarks from SPEC CPU2006 are used to verify the accuracy of the two-level trace sampling method,which is implemented in the AtomicSimpleCPU mode of Gem5.The corresponding performance estimation value is calculatedby the model usingsampled trace statisticalinformation.Finally,the sampling result is compared with the results of profiling all traces.Under the reasonable sampling parameter selection,the relative error of the performance estimation value corresponding to each performance analysis module is within 5%,which is nearly 10 times faster than the time of profiling all traces.