Predictive Anomaly Management for Resilient Virtualized Computing Infrastructures.

摘要

Large-scale distributed virtualized computing infrastructures have become important platforms for many critical real-world systems such as cloud computing, big data processing, and intelligence analysis. However, due to its inherent complexity and sharing nature, virtualized computing infrastructures are inevitably prone to various system anomalies caused by software bugs, hardware failures, and resource contentions. The situation exacerbates if the system is also exposed to malicious attacks. Moreover, although some anomaly symptoms such as machine crash are easy to detect, many other anomalies (e.g., performance degradation, processing bottlenecks, memory leak bugs) are hard to detect and diagnosis, which often have latent impact to the system. In this project, we have explored various online system anomaly prediction and cause inference schemes using unsupervised machine learning methods. We tested our algorithms using extensive real system experiments and our results show that we can achieve high fidelity anomaly prediction (i.e., greater than 95% true positive rate with less than 1% false positive rate) with low overhead (less than 1% CPU load).