Review - An Analysis of Internet Content Delivery Systems

Review of An Analysis of Internet Content Delivery Systems
by Ivan Hernández

The paper analyzes the use of four content delivery systems (CDS) in
an University environment. The examined CDS are: HTTP web traffic,
Akamai CDN, and Kazaa and Gnutella. The authors collected traffic
traces from outgoing and incoming traffic in the University of
Washington for nine days gathering more than 20TB of HTTP data, which
is a considerable amount of data to do a good analysis.

Next I will list some of the results that I found interesting. The UW
is more a provider of content rather than a consumer; on explanation
is that the observed outbound traffic is dominated by
application/octet-streams, possibly UW-supplied software, binary data,
and video streams from its TV stations or web-broadcast technical
talks; besides to this traffic, maybe the UW hosts some mirrors to
popular software. In addition, the UW it is also an important P2P
provider, but this content is hardly academic. Another interesting
result is that a small number of clients account for a large portion
of the traffic, in particular, only 4% of the clients account for 50%
of Kazaa traffic! That by the way, Kazaa is the largest bandwidth
consumer. One question is, why Kazaa traffic peaks late at night?
Maybe at night residential users browse for content and they are able
to find the desired content on Kazaa peers inside the UW. This would
support next the result, external Kazaa clients consume 7.6 times more
bandwidth than UW Kazaa clients. It is amazing to see that only one
10% of the Kazaa servers serve the 80% of all the Kazaa traffic. Thus,
even when Kazaa's design is of a distributed system, the content and
the utilization of the nodes is not that distributed. I found
interesting the discussion about the benefits of caching P2P content
and I think that this is a plausible solution to reduce the traffic
inside the UW, nevertheless, because this cache would be in the campus
(in the border) the UW would still get that traffic and there will be
no savings on WAN traffic.

Finally, it would be interesting to have a similar measurement
currently, given the increment in multimedia usage from the web
(Podcasts, youtube, flickr, ... ) and its correspondent traffic. In
addition we could see if BitTorrent is currently widely used.
Received on Mon Nov 20 2006 - 20:07:02 EST