Reliable file storage in a distributed computing system

Abstract

This thesis considers the problem of how to construct a file-store which is reliable in terms of high accessibility of data and low likelihood of data loss. In particular the problem is discussed in the context of a local area network computer architecture. A novel approach is taken to the implementation of the naming network for files, which controls multiple copy redundancy. The naming network is a single hierarchy which incorporates redundancy in the access paths to files as well as in the files themselves, thus improving accessibility as well as reducing likelihood of file loss. A prototype file-store was developed and implemented; to facilitate this the author had to develop a simple distributed operating system which evolved as an interesting research project in its own right. A distributed name-server algorithm was developed, and interesting insight gained into the design of local area network computer systems. The property of local area networks important to the file-store is scope for dynamic redundancy; the file-store is constructed out of a number of independent file-servers. Inconsistencies between multiple copies of a file are resolved automatically. Levels of redundancy and data location are controlled through the naming network, allowing replication of files to any degree thought necessary, bounded by the number of independent storage volumes in the system. Deadlock avoidance and automatic reconfiguration on hardware component failure are included. Some simple combinatorial mathematics is included to highlight the reliability of multiple independent copies of a file; the need for a more quantitative approach is indicated.