The Management of Consistency in Fault-Tolerant File Systems

Abstract

Critical data are often replicated in distributed systems to protect them against site failures and network partitions or to allow higher access rates. This trend benefits from recent advances in computer network technology and reductions in the cost of storage media, which have made the replication of important files on several sites a cost-effective proposition. The existence of several copies of the same file raises the issue of file consistency. Distributed file systems implementing file replication need to enforce an access policy isolating their users from that issue and maintaining a consistent view of all replicated data under all possible circumstances. Several consistency protocols have been discussed in the literature, including token-based schemes, active copy schemes and schemes based on quorum consensus. We present the tools used to evaluate their performance with respect to other consistency schemes, focusing on two key issues: copy locations and network topology play a primary role in the selection of a consistency protocol, and there is no need to compromise between file availability and network traffic overhead as all protocols can be efficiently implemented using possibly out-of-date information.