Research On Cloud Server Power Consumption Forecasting And Energy-saving Technology Of Power Capping

The power management of cloud servers is one of the most importantant research topics in the direction of cloud computing energy-saving technology.Accurate and efficient server power consumption forecasting can provide forward-looking information to realize intelligent and adaptive server power management,and achieve the goal of energy-efficient server operation.Traditional machine learning methods(ML)require sufficient and high-quality data in order to deliver excellent predictive performance.However,the cloud server in datacenter has the characteristics of heterogeneous,frequent expansion and iteration,and complex and changeable workload.Training an accurate forecasting model for each server requires huge time and space costs in data collection and storage.The power over-provision of the cloud server ensures the stability and reliability of the service,but some analysis shows that the server is not fully loaded most of the time,leading to a waste of energy utilization.Power capping can restrict power budget of server and mitigate power over-provisioning problem,but it will also cause the degradation of performance.There exists a tradeoff between ensuring service quality and achieving energy saving.To address the problems and challenges mentioned above,this dissertation proposes corresponding solutions and verifies their effectiveness through experiments.The main research contents and contributions are shown as follows:(1)Analyze the current research status of mainstream prediction tasks in cloud computing scenarios.Summarize the basic concepts,related definitions,technical methods,and application scenarios of transfer learning(TL)method,as the basis for the prorpoesed power consumption forecasting model based on TL method.Investigate and analyze common server power management mechanisms and energy-saving technologies,which mainly focus on the DVFS and power capping technology.(2)The TL-based model can achieve a similar performance campare to ML-based model by reusing the similar domain data and combining a small amount of target domain data for training.In order to reduce the cost of collecting and storaging data,a novel power consumption forecasting model(named TLSPP)is proposed,which is based on instance-based transfer learning method and can reuse the power series data of similar servers.TLSPP,a multi-value forecasting model,is designed and implemented in the aspect of the feature construction,selection algorithm of source domain data,the instance transfer algorithm and the update menthod of model.The experimental results show that compared with ML methods,TLSPP can save 50% of the target domain data while achieving about 7% prediction error,which can effectively save the time and space cost of collecting and storing a large amount of target domain data.(3)Power capping can can reduce the power consumption of servers and improve energy efficiency of servers,but it will also bring about performance degradation.With the help of the server power consumption forecasting model,a power prediction driven power capping technology(PDPC)is proposed.The essence of PDPC is a “four-band” method that the target capping value is dynamically adjusted based on the power forecasting values,and the capping upper threshold,lower threshold and uncapping threshold are used to ensure that the degradation of service quality is within the scope of acceptable.The experimental results show that PDPC achieves better energy-saving effects and maintains better service quality when performing in different types of workload and compare to similar methods,respectively.