Review of a Distributed File System (Andrew)

From: Muench, Joanna (jmuench_at_fhcrc.org)
Date: Mon Feb 23 2004 - 16:08:28 PST

  • Next message: David Coleman: "Review of Scale and Performance"

    The paper by Howard et al. (1988) on a distributed file system presented an
    interesting discussion of the design and implementation of a scalable
    system. The paper suffered somewhat from a lack of clear motivation and odd
    organization (two different prototypes). It did, however, re-iterate several
    concepts that we have come across in readings on RPC and virtual memory
    management over a distributed network and provided a useful comparison with
    the Unix NFS.

    The basic concept of Andrew is a set of servers (called Vice) that stores
    the shared files. Each server has a list of all the shared files with either
    the address of the server owning that file or the file itself. After a
    client has requested a file, the server will notify that client before
    allowing a modification by another workstation. One of the key concepts
    discussed in the paper included a whole-file transfer approach. This
    significantly reduces the number of calls required between the client and
    server, but does present a problem for files larger than the local disk
    cache.

    I didn't quite understand one of the consistency semantics, that "Multiple
    workstations can perform the same operation on a file concurrently." This
    sounds wonderful except for the final sentence, "Application programs have
    to cooperate in performing the necessary synchronization if they care about
    the serialization of these operations". But doesn't that require special
    application development and really break the concept of transparency?

    Because the system was built using an RPC package had many of the issues we
    have already discussed such as name resolution, transparency, the
    client-server relationship and authentication. Unfortunately the
    authentication issue was mentioned only in passing.

    The paper demonstrates that Andrew outperforms the Unix NFS due to a number
    of factors including lack of robustness of the NFS RPC protocol under load.
    The comparison didn't seem entirely reasonable, given the different goals of
    NFS vs. Andrew.


  • Next message: David Coleman: "Review of Scale and Performance"

    This archive was generated by hypermail 2.1.6 : Mon Feb 23 2004 - 16:10:19 PST