Joshua Dunfield

The authors describe how to replace expensive bottleneck
DFS servers with ``Network-Attached Secure Disks''.
These allow (1) direct transfers from the device to
client, (2) secure interfaces, (3) ``asynchronous oversight'',
that is, clients can send commands to the device (which
enforces the security policy) directly, 
(4) a device interface providing a data model of
variable-length ``objects'' with associated ``attributes''
(metadata).  The latter includes attributes manipulated
by the device directly (such as an object's length) and
arbitrary metadata needed by the client FS.

The approach is enabled by the possibility of putting
sufficient processing power inside disks.

Standard DFS's---namely NFS and AFS---have been ``mapped onto''
NASDs without much difficulty; NFS performs similarly to a conventional
setting.  AFS numbers are not reported.

Striping becomes very effective: a sample application (parallel data
mining) scales linearly with the number of disks.

The use of an approximate NASD (an older workstation with an ordinary
peripheral-network disk) seems troubling at first, but it is
sufficient to demonstrate basic viability; disks have onboard
computers and computers have onboard disks, so is there much
difference? (Actual NASDs would themselves vary in performance
anyway.) In fact, narrowing the difference leads to ``active disks''.