Wide-area Cooperative Storage with CFS
Dabek, Kaashoek, Karger, Morris, Stoica
SOSP, 2001

Design and prototype implementation of Cooperative File System (CFS):
a block-by-block, decentralized, read-only, peer-to-peer file system.
The file system and its semantics are built on top of "Dhash," which
stores the blocks themselves, and "Chord" which maintains the routing
tables to find blocks.  Blocks are stored in "virtual servers" of
which a given actual machine may have many.  These virtual servers are
the units of load balancing and distribution: blocks and naming are
distributed evenly among virtual servers, so a system which wants to
reduce its load should reduce its number of virtual servers.  In a
nice division of purpose, the CFS servers, made up of Dhash and Chord
layers, are oblivious to fs semantics.  This and any anonymity and
most system security are left to higher levels.  As they state in the
paper: "CFS relies on the assumption that large amounts of space disk
space are available."  No discussion is given to the possibility that
caches become full, and it appears that whenever a block passes
through a virtual server, the virtual server adds it to its cache.
"CFS uses the Chord protocol to locate blocks.  Chord supports just
one operation: given a key, it will determine the node responsible for
that key."  There is a nice description of the Chord lookup and
redundancy algorithm in 4.3.  The authors make a big deal about how
choosing the right server to get blocks from is important, but then
this "server selection" algorithm is only applied to twelve virtual
servers in their experiments.  They say they get about the same
performance as FTP, but their actual throughput numbers seem pretty poor.

File system structure comes from SFSRO.

Compares itself with PAST and OceanStore.

One item of particular future development is a scalable distributed
search engine.