CSAR-2: a Case Study of Parallel File System Dependability

Modern cluster file systems such as PVFS that stripe files across multiple nodes have shown to provide high aggregate I/O bandwidth but are prone to data loss since the failure of a single disk or server affects the whole file system. To address this problem a number of distributed data redundancy schemes have been proposed that represent different trade-offs between performance, storage efficiency and level of fault tolerance. However the actual level of dependability of an enhanced striped file system is determined by more than just the redundancy scheme adopted, depending in general on other factors such as the type of fault detection mechanism, the nature and the speed of the recovery. In this paper we address the question of how to assess the dependability of CSAR, a version of PVFS augmented with a RAID5 distributed redundancy scheme we described in a previous work. First, we address the issues encountered in adding fault detection and recovery mechanisms to CSAR in order to produce CSAR-2. Second, we build a reliability model of the new system with parameters obtained from a CSAR-2 prototype and from the literature. Finally, we assess the system and discuss some interesting observations that can be made with the help of the model. According to our analysis, a representative configuration shows a four nine reliability; the sensitivity analysis shows that a reduction of 16% of the system outage time can be obtained by increasing the speed of the reconstruction by a faster network.