\reviewtitle{An Analysis of Internet Content Delivery Systems}
\reviewlabel{saroiu02analysis}
\reviewauthor{Stefan Saroiu, Krishna P. Gummadi, Richard J. Dunn, Steven D. Gribble, and Henry M. Levy}

Got Best Student Paper at OSDI 02.

How has content delivery changed?  More data, new data (audio, video).
Four types of systems: (1) web, (2) akamia (cdn), (3) Kazaa, (4) gnutella.
Used sniffers at borders of University of Washington, which is classified
into these four categories.  90 days worth of traffic.

What is bandwidth impact of new systems?  p2p makes up most of
traffic.  non-HTTP TCP traffic makes up just about as much
amount. (grayed out parts of a table in a nice presentation
technique).

He divides traffic into inbound and outbound.  Web outbound greater
than inbound (strange); inbound greater than inbound for p2p.  Byte
categorization in bar chart.  In p2p, video dominates.  Do they have
any ideas on the distribution of video? (yes, answered later).
Are they mostly the same files?  p2p objects are 3 orders of mag
greater than web.  Huge number of server errors (503 error - server
busy) in p2p.  Many more unique objects in web.  Many of the largest p2p
objects were 700MB in size (isn't that the size of a CD?).  Caching
might work very well in a system like Kazaa.  Tiny number of Kazaa
users use most of the bandwidth.

He looks at how an ideal cache might perform (in both directions in
both directions).  Long warm up for cache.  Is Kazaa scalable?

Questions: it would be hard to legally cache Kazaa traffic because
then you acknowledge what it is; but if you shut it down, Kazaa will
try to work around port closers.  Saroiu says that Internet backbone
study produced similar results.