Joshua Dunfield

This paper represents yet another arrangement of a distributed
file system: one with no centralized servers.  Rather than try
to improve server performance (as in the NFS ``appliance'') or
change the nature of the server (e.g., NASDs), everything is
left to the clients.  Salient aspects of the approach include:
- responsibility for files is allocated among the clients;
- data can ``live'' on any machine, leading to strategies
    to increase locality (``First Writer'': the first host
    to write a file stores that file);
-  LFS is used, requiring distributed cleaning.

Clearly, the network should be very fast to make this work well.
A high degree of trust is also important, since a compromised
host can (at least) read and write anyone's data that happens to
be stored on that host's local disks.  One can have server-like
physical security and administrator oversight over a ``core''
of machines, but then only the core's resources can be exploited
for the filesystem.

One deficiency (acknowledged in the paper) is that the
same questions about the practicality of cleaning arise
as for Rosenblum/Ousterhout's LFS.