Research On Server Energy Efficiency Benchmark And Power Consumption Prediction

In recent years,more and more services have been deployed in the cloud,and the scale of the data centers has expanded rapidly.The rapid development of data centers also brings the problem of high energy consumption.The improvement of energy costs and the requirements of green environmental protection force the industry to shift its focus from performance to the energy efficiency of data centers.Although the PUE value of the data center has been steadily improved in recent years,the energy efficiency of servers is still a problem.To improve the energy efficiency of servers,we need to be able to measure and evaluate servers’ energy efficiency,and benchmarking is a scientific method to evaluate servers’ energy efficiency.There are two challenges in current benchmarking: 1)redundancy between worklet in the benchmark suite will increase the time for computer system performance evaluation,but a randomly selected subset will lead to misleading conclusions;2)the localization of domestic chips gave birth to the domestic server energy efficiency benchmark-Bench SEE,but compared with SERT,the effectiveness of Bench SEE needs to be verified.In addition,the research shows that the power consumption of different applications on the same server is different due to the different proportion of hardware resources used.Therefore,the results of the server energy efficiency benchmark can not intuitively represent the power consumption characteristics of specific applications.For data center service providers,how to buy the most energy-saving servers for specific applications has great research significance.Given the problems and challenges in the existing research work,this thesis puts forward the corresponding solutions and verifies their effectiveness through experiments.The main research contents and contributions of this thesis are as follows:(1)Aiming at the redundancy problem of benchmark suites,this thesis proposes a benchmark subset selection framework-Bench Subset.The framework includes three modules:group principal component analysis,consensus clustering,and evaluation method.Compared to other methods,the framework takes the validation problem of benchmark subset selection for unlabeled benchmark suites into account.According to the benchmark subset selection experiment on SPEC CPU2017,compared with similar methods,the benchmark subset selected by Bench Subset is more representative.(2)To verify the effectiveness of Bench SEE,this thesis proposes a worklet similarity analysis method based on the design principle of Bench Subset to compare and analyze the worklet of Bench SEE and SERT.In terms of energy efficiency evaluation,this thesis uses SERT and Bench SEE to test and compare servers belonging to two different architectures,including X86 and Arm.The experimental results show that Bench SEE and SERT have high similarities in worklet and energy efficiency evaluation.Therefore,Bench SEE can evaluate the server’s energy efficiency fairly.(3)Existing server power consumption prediction methods based on energy efficiency benchmarks have not considered the generalization ability of the prediction model in the target server,and there are shortcomings in prediction granularity and accuracy.To solve this problem,this thesis implements a new framework for Energy Efficiency Benchmark Server Power Prediction(2EBSPP).The 2EBSPP framework includes three modules: power interpolation,source server selection,and prediction algorithm based on server energy efficiency benchmark.Compared with other methods,the 2EBSPP framework combined two domestic and foreign server energy efficiency benchmarks,including SERT and Bench SEE,to achieve a more fine-grained prediction of server power consumption.The method not only increases the datasets but also avoids the one-sidedness caused by using only SERT power data.In addition,considering the generalization ability of the model,2EBSPP adds the source server selection mechanism.The experimental results show that the 2EBSPP framework can not only obtain more fine-grained prediction results but also achieve better prediction performance.