From: Tyler Robison (trobison@cs.washington.edu)
Date: Mon Nov 08 2004 - 07:16:10 PST
The paper examines 4 types of HTTP traffic (WWW, Akamai, Kazaa,
and Gnutella) to and from the UW network, looking at the different usages
of each and considering how this information could be taken into account
in order to make the network more efficient. Their findings show that the
peer-to-peer networking uses by far the most bandwidth (especially Kazaa),
that a small number of the top bandwidth consumers use large percentages
of the bandwidth, and that peer-to-peer networks tend to deal in very
large objects; all of these contribute to the conclusion that they will
not scale well, and that this information could be used to help ease the
traffic. Specifically, they suggest a reverse cache system which they
then simulate with encouraging results.
To effectively deal with traffic issues we have to understand the
major sources, and this paper does a good job of identifying these. Even
if the results are somewhat intuitive, it is still necessary to derive
them empirically. It was especially nice to see a number of different
statistics, and not just bandwidth usage: data on the % usage of the top
x clients was very interesting, as was breakdown of file types. The
occasional comparisons with the results of '99 were nice, and I wish they
would have spent more time describing the change in trends since then.
The results are interesting, and their suggested reverse cache
sounds good, but more than a simulation using an ideal cache is needed
before it can be evaluated. They also don't really consider other
options; what about enforcing bandwidth restrictions on the computers in
the network, and dropping packets from a host if they've passed their
daily limit? Such a system would penalize the heavy users while leaving
the average user unaffected. In addition, the reverse cache wouldn't
necessarily work well in many other environments: networks with different
traffic patterns may not benefit much, and there may be issues with cache
consistency if there are multiple output routers; how well would this
apply outside the UW network?
This archive was generated by hypermail 2.1.6 : Mon Nov 08 2004 - 07:16:11 PST