Performance Anomaly Prediction Monitoring In Distributed Computing Environment

In the past several decads, the distributed system environment has beenwidely explored in the research field. Distributed computing is a computerscience subject, it studies how to divide a task consuming lots of resourceinto many small components, assign these components to many computers todeal with, and gather the components results to a final results. In order toimprove the dependability of the distributed system, the traditional method isusing the high availability (HA) technology of the redundant components toguarantee the continous service. But if we can apply prediction mechanismsto detect anomalies before the node fails, and concern the level of anomaliesto maintain the system, the unexpected system suspend can be avoided, so thedependability of the distributed system can be largely improved. In reality,performance anomaly prediction can be used in many fields to reduce losses，and achieve specific monitoring purposes. The performance of system can bereflected as CPU usage, memory usage, disk usage and so on, so in this paperwe focus on the anomaly prediction technology on the time series formed by these data.Currently, in the anomaly research field there are many performanceanomaly detection technologies, but the performance anomaly prediction isnot widely studied yet. Using anomaly dectection we can only check andanalyze the anomalies after they already happened, and anomaly predictiontechnology is the key to realy avoid the anomalies happening.In order to solve the problem, we propsed two kinds of anomalyprediction models, single variable anomaly prediction algorithm and multiplevariable anomaly prediction algorithm. The single variable anomalyprediction algorithm takes a single performance metric as an object, and byfinding the most similar pattern and machine learning approach to judge theanomaly symptom. The multiple variable anomaly prediction algorithms takemultiple metrics of the system into account, and predict anomalies by revisedKNN method, and we proposed two different multiple variable anomalyprediction algorithms based on the KNN method. We also designed manyexperiments to evaluate the proposed models, and compared the experimentresults between them and with other anomaly prediction algorithms. Theresults show these models can effectively predict performance anomalieswith certain fault tolerance.At last, an a semantic event based distributed monitoring system is presented. In this system monitoring information is released in event forms,and the events are predefined by users. so this monitoring system is moreflexible and is easy to be extended. Besides, we integrated anomalyprediction models into this distributed monitoring system. Exprements showthis monitoring system can work well for monitoring and anomaly prediction.