From: Shobhit Raj Mathur (shobhit@cs.washington.edu)
Date: Sun Nov 07 2004 - 01:26:48 PST
An Analysis of Internet Content Delivery Systems
================================================
This paper examines content delivery in the Internet today by focusing on
four content delivery systems: HTTP Web traffic, Akamai CDN, Kazaa and
Gnutella P2P file sharing systems. A trace of all incoming and outgoing
internet traffic at the UW was collected over a period of 9 days to
perform this study. The results quantify the extent to which P2P traffic
dominates the Internet bandwidth consumption and the vast differences in
the object sizes transferred. The paper also concludes that P2P systems
are not scaling despite their explicit scaling design due to their
bandwidth demands.
The paper presents a detailed analysis of Internet content delivery
systems. I will mention some of the results I found interesting. The
median object size for a P2P request for found to be 4MB which is 1000
times larger than an average WWW object. P2P is not widely used but it
still accounts for over three quarters of HTTP traffic due to to the large
object sizes. As a result, a small number of P2P users consume a very high
fraction of the bandwidth. Moreover, the long transfer times of P2P
objects result in many simultaneous connections. From the graphs it looks
like, most P2P transfers are done overnight as the transfer times are
long. Overall, the bandwidth requirement of a single Kazaa peer is ninety
times that of a single web client. To prevent such users from hogging all
the bandwidth, the ISPs could reduce their bandwidth share.
The paper analyzes the Internet traffic at the University of Washington,
but does not mention whether the results are applicable to non-University
networks. I expect the P2P traffic in a University network to be much more
than in a corporate network. Hence these results may not be valid for
other types of networks. The authors observe that a small number of P2P
clients and servers are responsible for the majority of traffic and find
this contrary to what is expected from a P2P network. Since Kazaa's
architecture is proprietary we can only guess the possible reasons for
this. The authors do not attempt to reason this phenomenon.
The paper also proposes a reverse caching mechanism to absorb the outbound
P2P traffic from a university network. It concludes that due to the high
bandwidth requirements of a P2P system, it will never scale in a
university environment. Though I feel that most of the results regarding
P2P traffic are applicable only to University networks, the paper does a
good job overall. If the traces were collected from a wider variety of
networks, the results would have been more convincing.
This archive was generated by hypermail 2.1.6 : Sun Nov 07 2004 - 01:26:49 PST