From: Nathan Dire (ndire_at_cs.washington.edu)
Date: Mon Mar 08 2004 - 14:14:52 PST
"Measurement, Modeling, and Analysis of a Peer-to-Peer File-Sharing Workload"
P. Gummandi, R. Dunn, S. Saroiu, S. Gribble, H. Levy, J. Zahorjan
The goals of this paper are: to understand the fundamental properties of
multimedia file-sharing systems; to explore the dynamics of the file-sharing
workload; and, to investigate the possibility of optimization based on
locality. The first goal is accomplished using a trace of Kazaa P2P traffic
collected at UW. This data is used to create a model which is used to meet
the second and third goals.
The Kazaa traffic data show a number of interesting characteristics about user
activity and downloaded objects. Kazaa users are very patient, with 20% will
to wait a week for downloads to complete. New clients generate most of the
load. Average session lengths are very short. With regards to the objects in
the system, a number of characteristics result from the immutability of
multimedia files: clients only them once, popularity is brief, popularity is
correlated to newness, and most requests are for old objects.
The observations about the workload lead to an surprising conclusion: requests
for objects do not follow a Zipf distribution. This results from the
immutability of the multimedia files as the fact the clients only download a
file once. To some extent, Kazaa clients act like a Zipf workload going
through a proxy cache.
The authors use the observed results to help build a model that allows further
investigation. The model uses a Zipf distribution to determine the popularity
of objects, but then the clients obey a fetch-at-most once policy. A number
of observation result from this model. First the effectiveness of the system
diminishes as clients age since the most popular objects have already been
downloaded. New objects improvement performance since many clients will tend
to download the same new objects. The authors also find that there exists
significant untapped locality in the system.
I think many of the observations in this paper make sense given the nature of
the Kazaa system. I would question how well results for a University network
would extend to other scenarios. The population of college students may have
greater overlapping interests than general Kazaa users (though certainly
college students probably are a considerable portion of all users!). This
might skew the locality results. A trace outside UW would likely be outside
the scope of the paper, but I think it would provide a significant validation
of the conclusions. Overall, I think this paper provides an insightful look
at a growing phenomenon, but I wonder how universal those insights will prove.
This archive was generated by hypermail 2.1.6 : Mon Mar 08 2004 - 14:14:53 PST