Decentralizing a global naming service for improved performance and fault tolerance

Reviewer: Jerzy J. A. Klaczak

This paper describes a successful attempt to create a unified and fully transparent naming system covering a whole organization. On top of frequently accessed “managerial” directories (stored in individual file servers), the authors superimpose an additional layer of “administrational” directories (implemented by cooperating managers). The multicasts used to communicate with these directories set a current limit of roughly 1000 hosts on the entire system. Further growth, even to worldwide scale, is enabled by connecting independent systems to an additional layer of directories implemented using well-known global naming techniques. The authors show how to make such a system secure. The system has no separate file and name servers, but each manager knows the full global names of its objects. Clients use small caches of directory names to bypass multicasts and to decide where to send the naming request. The hit ratio is well over 99 percent in typical systems. Caches are updated by deleting stale entries on use, allowing the system to achieve optimal fault tolerance, message traffic, and response time. The text is clear, concise, and well written. All the parts necessary to make a good research paper are present here. The authors not only describe the development, analysis, and field tests for theoreticians, they also give many practical guidelines for practitioners of distributed systems. Even beginners can start with the reference list and the discussion of related works. The only lack is a discussion of the direction for further research.