Taming Aggressive Replication in the Pangaea Wide-Area File System 
Yasushi Saito, Christos Karamanolis, Magnus Karlsson, and Mallik
Mahalingam, HP Labs 

The goal is take advantage of locality of corporate campuses.  The
authors assume that servers trust each other and that "eventual
consistency" is ok.  This consistency is achieved through "Pervasive
replication": there are many replicas for a file, and replicas are
allowed to change (without notification to other replicas); these
changes are gradually propagated to other copies of the file.  The
local server process is Implemented with SFS loopback.

A random graph is created independently for each file/directory.  Edges in graph are used for update
propagation (How is a random graph created without centralized
knowledge?).  A few replicas of the file are denoted "golden" and are
used as the focal points of backup.  "Bronze" replicas are created on
each open.

Adding and deleting replicas is a challenge.

"Harbingers" are used to handle propagation delay.  They are
transferred before actual updates, letting the upper (server) layer
know that update is coming.

Evaluation:  with 4 replicas it is comparable to NFS when copying.
Another experiment looks at Usenet-style collaboration (like
newsgroups).  The graphs showed time to access decreasing as number of
replicas increases (this is good).  They also looked at bandwidth usage during
experiments.