Review: Analysis of Internet Content Delivery Systems

This paper presents the analysis of four different content delivery
networks by using the traces of network traffic at University of
Washington. The four content delivery networks being considered are HTTP
web traffic, Akamai content delivery network, Kazaa and Gnutella. The
traces consist of 9 days of passive network traffic monitoring that
constitute 500 million transactions with over 20 terabyte of HTTP data.
Since Kazaa and Gnutella also uses HTTP to exchange files so a single
kernel level filter is being used for identifying the traffic of different
CDNs based on the unique port numbers.

The analysis shows that in current internet scenario P2P traffic dominates
the WWW traffic in terms of total number of HTTP bytes transferred. This
is due to the fact that an object size of P2P system is approximately 4MB
which is much larger than the 2KB object used in WWW. This makes
relatively small number of objects in P2P systems accounting for larger
portion of bytes transferred. Also a small number of Kazaa clients
transfer most of the bytes because of the fact that 80% of Kazaa traffic
is dominated by videos. The results show that during the traces an average
web client consumed 41.9 MB of bandwidth as compared to 3.6GB bandwidth
consumed by Kazaa client which clearly suggests that most of HTTP traffic
is now P2P traffic.

This paper has presented a comprehensive analysis of partition of HTTP
traffic between four different CDNs being used in University of Washington
but I feel that it fails to point out the implications. The author has
suggested that deploying cache could increase the traffic download from
Akamai and also could impact the performance of P2P sharing system. Though
for Akamai cache system it seems easy to implement caching but for P2P
systems, to me it seems it is not a straight forward job. The author has
not mentioned any details of how this cache could be implemented for P2P
systems and what could be the performance issues. Secondly I am curious
whether the study is taking into consideration the P2P traffic being
generated within the campus because it could significantly impact the
ratio of P2P. Also at one point the author is comparing the number of
simultaneous connections opened by P2P and WWW which I believe not a very
good comparison as the number of clients for each system is different. It
could have been better if author has taken the ratios of connections to
total clients and then compared them. Similarly at one point author has
mentioned that the 30% of WWW and Akamai requests return an error but he
has not shown that any data that quantify the requests or particular
servers that generate these errors. But even though there are some missing
points but I still believe it was a very good effort to quantify the
network traffic division between different CDNs that could help ISPs to
implement better policies to cater for changing network traffic pattern
because of these P2P systems.
Received on Tue Nov 21 2006 - 03:22:11 EST