Toward massive distributed file systems

Massive scale in distributed file systems remains an elusive goal. Existing systems are generally limited to only a few hundred clients or make restrictive assumptions, such as that widely shared files are read-only. It is argued that a truly massive system must scale for all kinds of files; file access traces suggest that occasionally written files make up too large a proportion of shared reads to be ignored. The file access patterns that are likely to be important in the large-scale systems of the future are outlined, and a scheme for constructing dynamic client hierarchies that may work on a much greater scale than conventional approaches is described