From: T Scott Saponas (ssaponas@cs.washington.edu)
Date: Sun Nov 07 2004 - 14:36:34 PST
Review by T. Scott Saponas
“An Analysis of Internet Content Delivery Systems” looks at the current
state (as of 2002) of content delivery in the Internet from the
perspective of a large university using a 9 day trace of TCP traffic in
and out of the University of Washington (UW). They split all TCP
traffic into the categories of Akamai, WWW, Gnutella, Kazaa, P2P, and
non-HTTP TCP. They then compare how these different categories
contribute to the universities network load and comment on how much its
changes since the last measurement.
It was observed that while Akamai is meant to distribute the traffic
load for large web sites like CNN, most of what Akamai provides is the
caching of relatively static large data items. Thus, simple proxy
caching at the ISP or University level could provide the same benefit.
They also observed that in the last 5 years P2P traffic has grown so
much it has taken over as the largest consumer of network traffic at UW.
They also notice that most of this P2P traffic is outbound and thus
P2P nodes within the University serve a lot of content for clients
outside the University. They also looked at the distribution of WWW and
P2P bandwidth usage on a per client basis. They discovered that not
only are the top 200 consumers responsible for 50% of Kazaa bytes and
27% of all HTTP, but looking at the Kazaa overlay network as a whole
0.2% of outside clients provided 26% of UW client downloads. This
suggests that while these P2P systems are meant to be designed to scale
well – the way they scale might not be feasible in the future.
One of the limitations of this analysis is that it may not generalize
much beyond the context of the university setting. Certainly a large
university has many similar properties to a domestic ISP, a large
company, or even a small university. However, the makeup of the
universities user base may not be entirely representative. It seems
unlikely such ratios of web traffic to P2P would hold for a company
network since P2P probably has limited applicability to business needs.
Also, the width of their incoming and outgoing pipe may very well
contribute to the way it is used. For many of the university’s P2P
users this may be the first time they are not on cable/DSL or dial-up
and the first time parental restrictions are not in place about usage
time or content. For instance, the authors observed external Kazaa
clients consumed 7.6 times more bandwidth than UW clients; but in the
case of cable and dial-up this practice would not really be as fruitful
for the external clients because their upload is rate limited. One
might argue the reason why so much P2P comes out of a campus is because
the campus’ out-pipe is so much better than the alternatives (home
users). In fact, the dramatic rise in P2P traffic might very well be
due to the rich resource campus networks provide for P2P file sharing.
Despite my above comments about how well the observations may
generalize, certainly university networks experience this workload and
all those external Kazaa clients must come from somewhere. Thus, the
discussion on the possibility of using proxy caching to improve network
utilization is relevant and prudent for today and tomorrows Internet.
This archive was generated by hypermail 2.1.6 : Sun Nov 07 2004 - 14:36:44 PST