**************Revised**************
University of Houston
Department of Computer Science
In partial fulfillment of the Requirements for the Degree of
Master of Science
Yu-Ren Chung
will defend his thesis
Mirror Selection by Extraction of
Long-term Throughput from Short HTTP Transfers
Abstract
Mirrors of a web server host replicated materials at different
places around the world. They provide an opportunity for better performance
assuming one can choose a mirror that will deliver good performance. In this
paper, we first compare several simple mirror selection metrics, including
random selection, geographical distance, hop count, round-trip time, and
response time of a small document. We demonstrate that the response time is the best of
these metrics while geographical distance has no merit.
The key issue
underlying mirror selection is estimation of the long-term expected throughput
of a HTTP connection. We propose a set of heuristics to estimate the
throughput by closely monitoring a TCP transfer. The transfer is broken down into a
series of segments received every Round Trip Time. Heuristics tested include EWMA
(Exponential Weighted Moving Average), Skip K: ignoring the first K values in a series,
and Last K: using only the last K values. We investigate dynamic tuning of
heuristics based on analysis of the series and explore an Early Stop
heuristic to end processing when estimation is deemed
sufficiently accurate. We show that the better heuristics can estimate
more accurately and cheaply than average throughput for a short file transfer.
Date: Monday, April 30, 2001
Time: 11:30 AM
Place: 550-PGH
Faculty, students, and the general public are invited.
Thesis Advisor: Dr. Jaspal Subhlok