Energy Anomaly Detection Based On Power Modeling In Cloud Computer

With the growth of cloud computing services and the development of the Internet,the scale of cloud data centers which supporting cloud computing services is increasing,and its energy consumption has become a serious problem.How to reduce energy consumption becomes an important task in the process of operating cloud data centers.Cloud data centers use virtualization technology,and the granularity of their energy management changes from physical machines to virtual machines.Although energy consumption of physical machines is associated with virtual machines,traditional server energy metering and energy-saving methods are not suitable for virtual machines in cloud computing environments,which brings new challenges to data center energy management.In order to better improve the energy efficiency of the cloud data center,it is necessary to know the actual power consumed by the virtual machine,and at the same time,it is necessary to accurately discover the high energy consumption tasks and abnormal power behaviors in the cloud computing environment.In this context,this paper proposes an effective virtual machine power consumption anomaly detection framework in the cloud computing environment,which realizes the monitoring and measurement of virtual machine performance indicators in the cloud computing environment,and estimates energy consumption of virtual machine by effective modeling techniques.At the same time,the power consumption anomaly detection algorithm for the cloud computing environment is proposed,and alarm when virtual machine with abnormal power consumption.The two complements use together and jointly realize energy conservation in the cloud data center.Specifically,the contributions are as follows:(1)A virtual machine power modeling method in cloud computing environment is proposed.The virtual machine's use of various resource metrics is directly related to its power consumption,thereby eliminating the dependence on the specific sub-component power model and adapting naturally to workload and hardware changes;The metrics reflecting resource usage are identified by feature selection algorithms.Hardware performance counters that significantly affects the power consumption of the virtual machine are used to simplify the model and facilitate cross-platform migration of the power model.The power model combines the CPU power state C-States and use the decision tree regression based on automatic segmentation to improve the accuracy of the power model.(2)A power consumption anomaly detection algorithm in cloud computing environment is constructed.In view of the variable power consumption mode,various types of abnormalities,and no data labeling during the operation of the virtual machine,the proposed algorithm dynamically identifies different types of virtual machine power consumption modes based on cluster analysis algorithm;Unsupervised machine learning method LSTM without any prior knowledge learns the energy usage law in normal state of power consumption data to reconstruct the "normal" power consumption time series;and to make outliers analysis by comparing the difference between predicted power consumption and actual power consumption to implements anomaly detection.Based on the OpenStack cloud computing environment,this paper designs a power-aware anomaly detection system in cloud computing environment.Through a lot of experiments,virtual machines in different load modes and injected with specific exceptions are tested.For the virtual machine power modeling method in cloud computing environment,the experimental results show that the proposed model can predict the power of single or multiple virtual machines under various workloads in real time,and their average relative errors are 2.81% and 4.47% respectively.Both are superior to the traditional sub-component power model.For the power consumption anomaly detection algorithm in the cloud computing environment,the experiment proves that compared with the traditional anomaly detection algorithm,the algorithm can detect more abnormalities and have a lower false positive rate than the traditional anomaly detection algorithm.