Review "An Analysis of Internet Content Delivery Systems"

This paper presents a study of specialized content delivery systems:
HTTP web traffic, the Akamai CDN, and the Kazaa and Gnutella
peer-to-peer systems. The study's data was derived from a 9-day passive
trace of inbound and outbound traffic on the edge of the University of
Washington computer network. The collected data was analyzed to
quantify the rising importance of these services, evaluate the
feasibility of adding caching to reduce bandwidth requirements, and to
characterize the network behavior from the perspective of clients,
objects, and servers.

The authors observed that the university is a net exporter of HTTP data
despite the large number of clients within its network. This is
primarily because of p2p systems. Graphs of the logged TCP traffic
exhibit a typical diurnal cycle, with Kazaa contributing 36.9% of bytes,
more than any other system and almost 3 times more than WWW accesses.
In fact, the typical Kazaa client requires 90 times more bandwidth than
a typical WWW client.

GIF and JPEG account for the highest proportion of requests (42%), but
AVI and MPG videos account for 0.41% of all requests and 29.3% of the
bytes transferred. Despite the WWW clients' higher request rate (almost
2 orders of magnitude higher), Kazaa has about twice as many open HTTP
flows than WWW + Akamai, primarily because of the size of the objects
transferred.

The distribution of bytes in the identified Kazaa traffic is split
between 79% AVI and MPG videos and 13.6% MP3s. Clearly, p2p traffic,
especially video, commands a significant portion of the traffic going in
and out of the campus network. As a percentage of all HTTP bytes
transferred, HTML traffic has decreased by 43% in the 3 years since 1999
and GIF/JPEG has decreased by 59% while AVI/MPG and MP3 have enjoyed an
increase of 400% and 300% respectively.

Analysis of the objects transferred reveals that the median p2p object
size is 1000 times larger than the average Web object size with 5% of
the p2p objects larger than 100 MB. The top 200 Kazaa clients account
for 50% of Kazaa traffic while the top 200 Web clients account for only
13% of Web traffic. One would expect Kazaa clients to distribute their
request among a greater number of servers, but in reality, the
distribution for Kazaa requests is not dramatically different from the
distribution for Web clients. The top 300 Kazaa objects create almost
half of the total outbound Kazaa bandwidth. The authors then use
simulations to show that a local Web proxy could provide a similar hit
rate as an Akamai CDN and that a reverse cache proxy for outbound p2p
traffic could save the university network as much as 120 MBps.

This study has offered new insight into the behaviour of p2p systems and
also helped confirm and quantify existing beliefs about p2p systems.
Although the data was derived from a single university network over a
brief period of time, the trends presented in the study are likely to
hold in other environments. The methods and analysis could easily be
carried over to other environments to verify the patterns described in
this paper. The study does have one weakness: it ignores the p2p
traffic within the network. This may be significant because the authors
state that "Kazaa appears to direct nodes to nearby objects", therefore
there may be a significant amount of p2p traffic within the network that
has gone undetected.
Received on Mon Nov 20 2006 - 18:53:58 EST