Load-aware Power Oversubscription Optimization In Co-located Data Center

As the demand for cloud services continues to grow,data centers are expanding,leading to enormous energy consumption and low resource utilization.To improve resource utilization,data centers commonly use power oversubscription to increase capacity,which may result in power peaks exceeding rated power,leading to circuit breakers tripping or even data center power outages.Therefore,data centers are typically equipped with power capping systems to quickly reduce power consumption when it exceeds a certain threshold.In recent years,some cloud service providers have improved resource utilization by deploying online and offline workloads together in one cluster,known as colocation.Due to the existence of Service Level Agreements(SLAs)for online workloads,there can be significant penalties for breach of contract,while offline workloads only need to be completed before a deadline,resulting in a significant difference in their priority.Therefore,when implementing power capping in colocated data centers,it is necessary to distinguish between these two types of workloads.However,in cloud environments,cloud service providers often do not know the type of workload being run due to privacy concerns.Therefore,it is necessary to obtain more information related to the workload and accurately predict its power consumption.Unlike traditional data centers,colocated data centers cannot directly reduce power consumption by limiting or shutting down offline workloads on machines,due to the complexity of the workloads.Therefore,it is necessary to design a workload-aware power capping method for colocated data centers.This thesis focuses on the research of workload characteristics and power consumption prediction,as well as workload-aware power capping systems,with the following specific research content:(1)We propose a workload classification and workload intensity prediction model,as well as a workload power prediction model,to achieve load-aware power capping.Firstly,we processed the Google dataset to obtain the training set for the classification model and constructed a workload classification framework based on t SVC.The classifier’s recall rate for online workloads can reach 97.06%.Next,we proposed a workload prediction model based on a combination model to predict short-term workload intensity,with an~2 of 95.46%.Finally,we proposed a container power prediction model based on random forest,selecting the four features with the highest correlation to power consumption for power prediction.Experimental results show that the random forest model can capture nonlinear relationships between features and power consumption,making it suitable for most workload power predictions with a maximum accuracy of 97.13%.(2)We propose a dynamic oversubscription rate method and a load-aware three-stage power capping method to improve the power resource utilization of data centers and ensure their power safety.Firstly,we propose a dynamic oversubscription rate method that adjusts the allowable oversubscription power value based on historical power consumption records and the current operating status of the data center.Experimental results show that compared to conservative oversubscription methods,the dynamic oversubscription method can increase the oversubscription rate by an average of 4%and reduce the number of power capping events by 33%.Next,we propose a workload adjustment strategy based on server energy efficiency to select the appropriate machine to adjust power consumption when the power budget of the data center changes,achieving the highest efficiency.In experiments,this method can increase offline workload throughput by 7%compared to the average allocation method.Finally,we propose a load-aware three-stage power capping system called WTPC,which includes a high threshold,target value,and low threshold to determine the starting value,target value,and ending value of power capping,respectively.Experimental results show that this method can improve the resource utilization of data centers and ensure their power safety by allowing offline workloads to run at the highest efficiency possible without affecting online workloads during power capping.Achieving load-aware optimization of power consumption in colocated data centers is of great significance.In this thesis,we propose a workload classification model to achieve load-awareness,followed by workload intensity and power prediction models.Additionally,we propose a load-aware three-stage power capping system to further optimize data center energy consumption.