面向云计算环境的动态故障检测与诊断方法

Dynamic Fault Detection and Diagnosis Method for Cloud Computing Environment

摘要: 在现代云计算系统中，通常会有成百甚至上千个云服务器通过多层网络相互连接，在如此大规模的复杂系统中，发生故障是一件很常见的事情。主动性故障管理是表征系统行为并预测云环境下故障动态的一项关键技术。为了预测云环境下的故障情况，我们需要监视系统的执行情况并收集与运行状况相关的运行时性能数据。但是，在新部署或者托管的云系统中，这些数据通常是没有标签的，在这种情况下，基于监督学习的方法是不合适的。在本文中，我们提出了一种使用贝叶斯模型集成的无监督故障检测方法，它可以表征系统的正常运行状态并检测异常行为，由系统管理员验证异常之后，对这些数据打上标签。在我们搭建的Hadoop集群下运用此种方法的实验结果表明，我们的方法可以实现较高的真阳率和较低的错误率。

Abstract: In a modern cloud computing system, there are usually hundreds or even thousands of cloud servers connected to each other through a multi-layer network. In such a large-scale and complex system, failure is a very common thing. Proactive fault management is a key technology that charac-terizes system behavior and predicts fault dynamics in cloud environments. In order to predict the failure of the cloud environment, we need to monitor the execution of the system and collect runtime performance data related to the operating status. However, in newly deployed or hosted cloud systems, these data are usually unlabeled. Methods based on supervised learning are inappropriate in this case. In this paper, we propose an unsupervised fault detection method using an ensemble of Bayesian models, which can characterize the normal operating state of the system and detect abnormal behavior. After the system administrator verifies the abnormality, labeled data are available. The experimental results of using this method under our Hadoop cluster show that our method can be implemented to achieve a higher true positive rate and a lower error rate.