File System Design for an NFS File Server Appliance
Dave Hitz, James Lau, Michael Malcolm

The authors describe the special purpose filesystem WAFL
(Write Anywhere File Layout), as well as its integration
into Netapp's FAServer(tm) NFS file server appliance,
which implements Snapshots(tm).  In addition to working well
with the copy-on-write semantics necessary for Snapshots,
WAFL's characteristics make it well-suited for high-availability
and good NFS performance.  Like LFS, not only data but meta-data
blocks can be relocated to anywhere on the disk in a WAFL, and both
filesystems keep track of their inodes using an inode map.  WAFL
differs significantly from LFS, though, in that it does not have a
cleaner.  The paper does not describe the allocation algorithm in
detail, but WAFL does have a "block-map" file (unlike LFS,
and different from FFS's block bitmap since it carries extra
information about snapshots); having a cleaner would likely
be incompatible with the consistently high performance required
of an appliance.  WAFL maintains a NVRAM log of the actual NFS
requests, instead of processing them first and logging disk writes.
This allows it to store many more operations, prevents NVRAM failure
from corrupting the disk (vital since RAIDs are more dependable than
NVRAM), and also improves the FAServer's response time.

Since WAFL is organized as a full tree, taking a Snapshot means
simply cloning the root inode.  WAFL uses this mechanism to keep
checkpoints every few seconds so that the filesystem can be restored
immediately after a crash.  In addition, it appears to outperform
other general purpose servers in the normal case -- in terms of NFS
operations/second -- though it would be unwise to make any hard
statistical claims based on the one graph provided in the paper.
Now, it is unclear how much of this performance edge comes from
features of the WAFL filesystem, from the particular hardware
configuration used (especially NVRAM), or even from the heavy
software optimization Netapp no doubt performed.  Still, the basic
strategy behind WAFL -- copy-on-write and logging for easy snapshots,
recovery, and asynchronicity -- certainly seems to play a key role
in its success.  However, since the allocator details are not
given, it is unclear how WAFL would perform after a long period
of use -- whether it would become fragmented more quickly than
other filesystems.
Title
File System Design for an NFS File Server Appliance

Author
Dave Hitz, James Lau, Michael Malcolm

Date
1994

Location
Usenix

This conference paper details the WAFL file system and conveys the
benefit of specialized applicance systems through its eight-fold
improvement over general-purpose systems which attempt to supply the
same service.  It is the followup to Hitz's 1993 paper "introducting
the NFS Server Appliance."  After 

While this paper has a major research impact in its focus on
specialized systems as opposed to general-purpose systems, it does
include several interesting technical ideas.