From: Charles Reis (creis@u.washington.edu)
Date: Sun Nov 07 2004 - 15:22:40 PST
An Analysis of Internet Content Delivery Systems
Saroiu, Gummadi, Dunn, Gribble, Levy, 2002.
The paper describes a measurement study of all web, Akamai, P2P, and other TCP traffic in and out of UW. Comparing against previous studies of web traffic, the authors found that P2P traffic has vastly taken over the web as the majority of bytes transferred. A large set of graphs and observations for each graph help convey the characteristics of P2P traffic (especially that of Kazaa), such as enormous objects, long lived transfers, and relatively few clients and requests as compared to the web.
A few observations are particularly interesting, including the fact that the P2P traffic isn't spread more evenly among peers, but instead a large percentage of requests are satisfied by a small percentage of peers. It's also interesting to note that UW is a net provider (not consumer) of data by almost an order of magnitude. Many of the authors' observations are then used to propose that an outbound P2P cache for the university may be extremely effective at reducing bandwidth demands, rather than relying solely on filtering and rate limiting techniques.
The legal implications of such a cache, however, are less clear, given that the majority of P2P traffic is likely to be illegal music and movie trading. The introduction of online music stores changes the nature of media file transfers back to a client/server model, while recent RIAA (and the potential for MPAA) court cases may somewhat stunt the growth of P2P networks.
I am concerned with the conclusion that P2P does not scale based on the fact that load is not spread evenly among clients, since that likely speaks more about existing implementations than the potential for P2P (eg. structured DHTs). Also, the study excludes traffic between peers within UW, which is likely significant for Kazaa (which supposedly emphasizes locality). This is less important for the caching proposal than the model of P2P clients, though. Overall, it is clear that mechanisms designed for improving the web alone will not be sufficient in the future.
This archive was generated by hypermail 2.1.6 : Sun Nov 07 2004 - 15:22:40 PST