Review: An Analysis of Internet Content Delivery Systems

Review: An Analysis of Internet Content Delivery Systems

Reviewer: Di Niu

This paper presents measurement results characterizing the shape of
Internet content delivery traffic from a trace of all incoming and
outgoing network traffic at the University of Washington. Four
different classes of content delivery systems were considered,
including HTTP web traffic, the Akamai content delivery network, and
the Kazaa and Gnutella peer-to-peer file sharing systems. Based on
this trace, the paper shows a rapid increasing of P2P traffic,
characterizes the behavior of content delivery systems from the
perspective of clients, objects and servers, and also drew
implications for caching in these systems.

First, as a high-level picture of Internet content delivery traffic,
it is shown that content delivery systems account for a dominant part
of Internet traffic, and P2P traffic has overtaken WWW traffic in the
last few years. HTTP, Akamai, Kazaa and Gnutella account for 57% of
total TCP traffic, with Kazaa being the largest contributer,
consuming 36.9% of TCP bytes. It was also observed that both WWW and
Kazaa have diurnal cycles, with WWW peaking in the middle of the day
and Kazaa peaking late at night. Besides, the object types for
transfer has changed in the past few years. Compared to the study in
1999, HTML traffic has decreased 43% and GIF/JPG has decreased 59%,
whereas AVI/MPG traffic has increased by nearly 400% and MP# traffic
has increased by nearly 300%. In particular, AVI and MPG account for
79% of all the traffic in 2002, while 13.6% of the bytes are MP3.

Detailed characterization about content delivery systems were also
provided, including the characteristics of objects, clients, servers
and the scalability of P2P systems. As far as objects are concerned,
it was observed that objects of P2P systems have a much larger size
than the web content. However, it seems that this conclusion is a
straightforward and trivial truth which people do not need
measurement to know. In the characterization of clients, it was
pointed out a small number of clients account for a large portion of
the traffic, which is a quite useful observation known as a power-law
of P2P systems later. For example, the 200 top Kazaa clients are
responsible for 50% of Kazaa bytes downloaded. Perhaps the paper was
among the first few which have observed this important phenomenon. It
was also observed that only a small amount of peers carry a large
part of burden of all P2P traffics. Thus, the paper concludes that
P2P systems may still suffer from scalability problems. However, this
point of argument lacks justification in some sense. The fact that
600 of 281,026 UW-external Kazaa peers provided 26% of the bytes only
means the clients are sharing different amounts of burdens, and does
not necessarily leads to a decrease of download efficiency as network
size grows.

Moreover, caching in CDNs and P2P systems was also discussed based on
the trace study. The paper is so well-written that the digest of the
content is fairly easy. It was among the first few papers that offer
important measurement results regarding the workload characterization
of P2P systems.
Received on Tue Nov 21 2006 - 00:45:15 EST