Abstract:In container-based virtualization platforms, detecting anomalies with fixed thresholds is not practical, because many kinds of applications completely shared physical resources. Furthermore, monitoring systems need to consider the tradeoff between the timeliness and overhead by adjusting the monitoring period. To address these issues, this paper proposes an online anomaly detection method based on Principle Component Analysis (PCA), and then adjusts the monitoring period according to the anomaly significance. First, it non-intrusively collects the monitoring data of containers to get a matrix for each container. Then, it uses PCA to present the main direction of each matrix, and calculates the abnomal significance by calculating the cosine similarity between the current direction and the last one. If the significance is out of the defined tolerability threshold, the monitoring system sends an alert message to administrators, and automatically adjusts the monitoring period according to the significance. The experimental results demonstrate that our method can detect typical faults injected in HDFS with the accuracy of 80%, the delay of alerts iswithin 5 seconds, and the monitoring overhead ismuch lower than that with a fixed monitoring period.