\section{Conclusion}
\label{section:conclusion}

We have presented measurements of the performance of replicated web
servers which have ramifications for server selection system designs.
We have found that performance observed by a given client can vary
widely from mirror to mirror.  However, the set of servers that a
client must visit to achieve good performance is fairly small.  Once a
client has found a good server, neither time scale over which the
server has been good nor moderate changes in server performance are
good indicators about when to begin searching for a new good server.
However, a large performance drop does signal a client that a server's
rank is likely to have changed.  To further substantiate and expand
our conclusions, future work includes collecting longer traces, trying
other mirror sets, and exploring shorter time scales.

The data collected for this study is available on the World Wide Web
at http://\D www.cs.cmu.edu/\~{ }acm/\D research/\D anycast.html.

\section{Acknowledgments}

We would like to thank the client sites (Berkeley, Georgia Tech., ISI,
the University of Kentucky, the University of Massachusetts, the
University of Texas, the University of Virginia, and Washington
University in St. Louis), who gave us guest accounts and tolerated our
experiments.  We would also like to thank the server administrators
for their support and cooperation.

% don't forget to put Bruce
% Lowekamp and Dave Eckhardt into final version, assuming the % paper
% gets accepted