Microsoft is interested in the commercial development of high
performance
distributed caching algorithms.  We view development of advanced network
caching as central to the scaling of the Internet to accommodate every
household.

The primary avenue that we have been exploring to solve the large scale
distributed caching problem is based on hash-based URL routing.  This
"queryless" technique uses hash-based partitioning of URL space amongst
a loosely coupled "array" of proxy caches.  

We take a deterministic hash function based on the URL to be resolved
and
couple it with a table of n upstream proxies.  The output of the URL
hash
function is used to determine with of n upstream proxies a request
should
be forwarded to.  We believe that this protocol creates many interesting
results including:

*  any cache object within an array can be addressed from 
	anywhere on the network given consistent application of the
	hash function + a consistent array membership table.
*  allows creation of a single, very large logical cache image within
	the proxy array;  choosing a good hash function will result in
	no object duplication by the machines in an array.
*  extremely lightweight and scaleable routing implementations for
	downstream clients/proxies;  correct hash function selection
	results in one additional hash computation per additional
	upstream route.
*  no additional network traffic / protocols;  this protocol avoids the
	"query" hop & eliminates the almost totally useless "cache miss"
	message.

More information on these techniques can be found in an internet draft
at:  http://www.nlanr.net/Cache/ICP/draft-vinod-icp-traffic-dist-00.txt.

This draft demonstrates a hypothetical implementation based on proxy
autoconfig (PAC) files which are now supported by Netscape & Microsoft
browsers.  An implementation in which a queryless hash-based routing 
function is downloaded by browsers & proxies in the form of a PAC file
is
compared & contrasted against current ICP technology.