Cloud computing can provide users with on-demand,scalable,high-performance,and reliable IT resources that have been widely used.However,frequent cloud system failures in recent years have had a serious impact on the reliability and availability of cloud platforms.The research report of the University of California at Berkeley also pointed out that the availability of cloud services will be the number one factor hindering the development of cloud computing.Because current mainstream cloud computing providers use server virtualization technology to pool hardware resources and share them with users,the virtual machine(VM)is the core component of the cloud platform.Once an exception occurs,the cloud platform may fail.Therefore,it has become more and more popular to study the anomaly detection of VMs to improve the dependability of cloud platforms.This paper takes the VMs in the cloud environment as the research object,firstly compares and analyzes the similarities and differences between VMs under the cloud computing environment and under the traditional architecture,and summarizes the key problem to anomaly detection of the VMs under the cloud environment.For can not know detail of user behavior in the cloud environment,this paper proposes the use of VMs performance indicators data set to black box testing;For the large correlation between VMs performance indicator data set and operating environment,this paper proposes an anomaly detection framework for detecting contextual anomaly based on VMs operating environment attributes and performance attributes.Then,this paper focus on the feature extraction algorithm and anomaly detection algorithm in the anomaly detection framework.A nonlinear feature extraction algorithm combined with a kernel method is proposed to reduce the dimensions of VMs performance indicator data set that is not linear in the cloud environment and does not satisfy a certain probability distribution.For the high dynamic characteristics of cloud environment,this paper proposes a real-time anomaly detection combined with online learning ideas,and research parameter optimization techniques and imbalanced data processing techniques.Finally,the algorithm proposed in this paper is tested using the data set obtained from the our cloud platform.The experimental results show that the proposed feature extraction algorithm can achieve better separability while reducing more data redundancy and correlation.Anomaly detection algorithm is proposed to ensure detection accuracy in a shorter time while single samples for training to update the model to achieve real-time detection,and occupies less memory.Therefore,the detection framework and related algorithms designed in this paper can guarantee the credibility of the cloud platform. |