Pushing the Limits of Transparency in Distributed File Systems
Rick Floyd, S C Ellis · 1991
\indent The last few years have seen an explosion in the research and development of distributed file systems, and an accompanying increase in user dependence on remotely stored files. Network transparency has become a major theme in this area. Full network transparency implies not only that the file system interface and naming present the appearance of a single file system, but also that the behavior of the system does not betray its distribution. Existing systems provide only a limited degree of network transparency. While full transparency promises advantages such as automatic reconfiguration, increased availability, effective use of resources, and a simplified file system model, there are arguments about its inherent costs and questions about scalability. Attempts to address these concerns have been limited by a lack of understanding of the range of possible solutions to transparency issues. We have designed and implemented a prototype of a highly transparent distributed file system to explore the feasibility of full network transparency. Our distributed file system, Casper, supports a substantially higher degree of network transparency than earlier systems, and is able to do this in a heterogeneous environment. Casper appears to users as a single, globally accessible file system providing highly available, consistent files. Casper uses a name space that is decoupled from the file systems of the hosts supporting it. This allows Casper to provide a coherent framework for uniting techniques in the areas of naming, replication, consistency control, file and directory placement, and file and directory migration in a way that provides full network transparency. In this paper, we present the design of the Casper file system and describe our experiences (both positive and negative) with the implementation.