review of p2p workload

This paper measures a representative p2p file sharing system, Kazaa,
during a 200 day period, and then analyzes its user and object
characteristics. The most interesting observation in this paper is that
"fetch-at-most-once" behavior of Kazaa distinguishes itself with dynamic
web contents, and leads to non-zipf request distribution. Moreover, this
paper models this non-zipf behavior and explores the locality in caching
for Kazaa-like p2p file-sharing workload.

The model presented in this paper assumes that underlying popularity of
objects in such p2p file sharing system still obeys zipf's law.
Fetch-at-most-once prevents a client to achieve the same object again. So,
we wonder if the number of clients is huge; the impact of
"fetch-at-most-once" of a client may be invisible among a large amount of
requests from different clients. In this case, the request curve may be
still close to zipf's law.

The authors observe that users slow down as they age. Although they give
some explanation, I still feel hard to understand this behavior. If a user
likes to use Kazaa to download, his interests should not diminish with
time; his interests should be triggered by whether there are new favorite
files available in Kazaa. Since there are always new contents born in
Kazaa, I assume users will download different contents with time. The
results of this paper may indicate the abnormal behavior of users in
Kazaa. They maybe use Kazaa once and leave for a long time; next time,
they may use a new user id.

In brief, this paper provides insightful observation about the impact of
"fetch-at-most-once". It is helpful for designing the replication strategy
of a file storage system based on this non-zipf request distribution.
Received on Mon Nov 28 2005 - 10:45:40 EST