From: Michael J Cafarella (mjc@cs.washington.edu)
Date: Sun Nov 07 2004 - 14:46:57 PST
An Analysis of Internet Content Delivery Systems
By Saroiu, Gummadi, Dunn, Gribble, and Levy
Review by Michael Cafarella
CSE561
November 8, 2004
Main result:
The authors instrumented the network link between UW and the rest of
the world. They present results of the trace in heavy detail, especially
on the topic of how various internet distribution schemes work. The
characterization of the workload is very different for the various
classes of traffic they analyze:
-- Standard HTTP
-- Akamai traffic
-- Kazaa
-- Gnutella
From just watching the bytestream, the authors can discern the
number of requests, the size of objects returned, and the connection
length, for both incoming and outgoing traffic. This data lets
them see a number of interesting facts.
The most interesting revelations are that peer-to-peer traffic now
accounts for most bytes transferred, and that the distribution of
P2P objects (mainly audio/video) is very different that seen with
objects handled by HTTP and Akamai.
The authors are very rigorous about presenting their data in every
form that might come in handy. It's surprising how just a few facts
about the traffic scheme (sender, receiver, object, size) can be
a very rich piece of data. It's especially interesting how this
can be used to suggest that Kazaa will have scalability problems.
As interesting as this is, I would have liked to see more on the
performance characteristics of the various systems. Instead of
just object and request distributions, it would be nice to see
info on response rates, jitter, etc. Especially since some of
the bytes were from streaming protocols, this data seems very
important.
I also would have liked some suggestions for improving the P2P
scalability problems, though perhaps that's a little outside the
scope of the paper.
This kind of paper is very important, as it gives guidance to network
designers and researchers about traffic demands. It's hard to
choose what sort of work to do without detailed workload
investigations like the one here.
This archive was generated by hypermail 2.1.6 : Sun Nov 07 2004 - 14:46:57 PST