An Analysis of Internet Content Delivery Systems

Motivation: An analysis paper trying to throw light on the relative
importance of HTTP web traffic, Akamai CDN and P2P traffic by analyzing a
trace collected at Univ of Washington. Interesting implications for caching
revealed.

Key Points:

1) The rising popularity of P2P sharing has enabled the usage of upload
bandwidth by normal clients and end users. This was prominent to the point
that the University was a net provider exporting 16.65TB and importing only
3.44TB. The bigger size of objects in P2P further aggravates the problem.

2) An interesting observation was that WWW traffic peaks during the day
while P2P traffic peaks during nights. This is because P2P downloads are
left active through long periods of time and users often prefer to have the
download take place at night when its not interfering with their own work.

3) It would have been interesting to see number of pages serving dynamic
content and pages serving static content along side the content types
ordered by size. Though the paper still makes a strong case for caching, but
the shift towards dynamic pages might be a concern for caching. Although,
the content still gets dominated by video, audio and images to enable
caching of these objects.

4) Few popularity objects and few users account for bulk of the traffic in
P2P sharing networks which itself forms a substantial portion of the TCP
traffic on the internet. ( 1000 kazaa objects account for 50% bytes
transferred). This provides a strong motivation for P2P sharing. Few users
cause the most severe hit which might justify ISPs for limiting the traffic
of these users within reasonable bounds in order to provide affordable
service to all the users.

5) An interesting find is that a reverse cache would benefit the University
network more than a forward cache. Since the outbound traffic is highly
dominated by P2P uploading and since upload has to be paid for more, it
might be worthwhile limiting the outbound P2P traffic or implement caching
at P2P level.

6) Since, P2P traffic provides content to a small number of users but
accounts for a large percentage of bandwidth consumption as compared to web
traffic, it seems increasingly challenging for both ISPs as well as
organizations to be able to provide such bandwidth in the long run in an
economic way.

7) Local cache seems to provide comparable hit rate as Akamai and thus,
dilutes the need for a separate CDN atleast for local benefits.

8) Scalability of P2P networks needs to be thought again. A small subset of
external peers provide most of the content in P2P networks. Does it then
make sense to think about highly scalable P2P networks or could it make
sense to have smaller clustered P2P networks for efficient lookup.

Its a well written straight forward paper that has analyzed the internet
traffic. Showing that few users and popular content account for majority
bandwidth consumption, a P2P cache seems likely to be the obvious choice.
Received on Tue Nov 21 2006 - 03:46:20 EST