Multi-attribute Anomaly Analysis Of Virtual Machines In The Cloud Computing Environment

The cloud computing technology is facing tremendous challenges for its virtualization, hierarchical, dynamic, large-scale, and other characteristics. The monitoring of virtualized nodes resource is an essential one of them. The anomaly trend forecasting of virtual resource usage plays a more important role in practical applications. Therefore, we focus on the multiple attribute anomaly trends analysis of virtual resource usage based on the Hadoop cloud compting platform.First of all, a modelling method of virtual machines running states based on K-means is proposed. Three different levels of states,"Normal","Anomaly" and "Failure", can be achieved by training. The main procedure is as follow:(1) setting the default initialization of the cluster centres, and calculating Hopkins statistics of training data;(2) obtaining clustering results by K-means theory;(3) further amendments to the centre point of the clustering result is required if the Hopkins statistics stayed in [0.4,0.6] or the existence of an initial cluster centres did not changed. The modelling method of virtual machines running states can be applied to different types of training data.Secondly, an anomaly trend analysis mechanism is presented based on cluster centers and the non-parametric CUSUM algorithm. The kernel is to monitor trends of follow-up data whose classified result is "Anomaly" in order to detect continuing abnormalities which would lead to a "Failure" state using non-parametric CUSUM algorithm.The main procedure is as follow:(1) calculating the distance between short interval sampling data and the "Failure"category center;(2) accumulating anomaly status when the inspection sequence changes from negative to positive;(3) a forecast alert will be issued when the anomaly accumulation value reaches the threshold limit. And the threshold and prediction time-delay are analyzed.Thirdly, in the modeling stage, a set of experiments are performed to assess the reliability of clustering result under different training data by calculating the coefficient contours. In the preliminary detecting stage, classification decision algorithm KNN is selected to compare with the proposed method, and the advantages and disadvantages of the algorithm are analyzed from two aspects which are complexty and applicability. In trend forecasting stage, a series of experiments are designed for single property and multi-attribute conditions separately. The result shows that the proposed model has a better performance in forecast accuracy and prediction delay. Eventually, a location method of anomalous properties based on (Squared Prediction Error) SPE is introduced. By diagnosing data points whose classified result is "Failure", we can infer the most-likely property that lead to the exception, and providing evidence for post-processing. The result of the experiment shows that the location method of anomalous properties based on SPE can reach0.8889.