| Dependability has become a critical challenge for commodity distributed storage systems. Toward addressing dependability issues in distributed storage systems, the first step is to comprehensively understand and evaluate the dependability of these systems. However, the lack of a comprehensive dependability benchmark limits people's ability to understand the dependability of distributed storage systems.;In this dissertation, a dependability benchmark, D-BENCH is presented to evaluate the dependability of distributed storage systems. D-BENCH provides sufficient flexibility and generality for users to emulate different faulty scenarios to evaluate these systems. In order to analyze the dependability of a distributed storage system in a faulty scenario, an HMM-based model is developed to identify the most likely sequence of system internal states in the faulty scenario and quantify the dependability of a distributed storage system based on the sequence. Meanwhile, a performance anomaly detector is developed to detect performance anomalies in a computer node of a distributed storage system, and it is also used to identify the faulty state of a computer node to assist the training process of the HMM-based model. Finally, a series of experiments were conducted to evaluate a distributed storage system with two different file systems, PVFS and Lustre, via D-BENCH, respectively. The experimental results demonstrated that D-BENCH was able to accurately evaluate system dependability, with an accuracy of appropriately 90% on average. |
