Auditing the Structural Reliability of the Clouds

Large scale systems, common in cloud computing, rely on redundancy for reliability and availability. Modern clouds have become ever-increasingly complex and diverse creating large messes that experience long outages when failures occur. While there exist significant effort in resolving faults after they occur, we propose a novel approach to untangling this mess before it occurs by auditing the underlying structure of a cloud, which we call the cloud Structural Reliability Auditor (SRA). SRA achieves our goal by auditing a cloud with the following steps: 1) collecting comprehensive component and its dependency information, 2) using this data to construct a system-wide fault tree, 3) and leveraging fault tree analysis algorithms to determine and rank sets of components based on the likelihood of causing a cloud service outage. SRA enables a cloud administrator to be able to evaluate risks within the cloud beforehand and improve the reliability of her service deployments before the occurrences of critical failure events. We have built a prototype implementation that performs all three tasks. Using this prototype, our experimental evaluation shows that SRA is practical: auditing a cloud containing 13,824 servers and 3,000 switches spends about 6 hours.