Performance of Mirror Servers


Sponsor Ellen Zegura
ewz@cc.gatech.edu
216 GCATT
Area Networking

Problem
Server replication is one technique to improve scalability of a networked service. Server replication involves placing identical content servers ("mirrors") in geographically distributed locations in the Internet. Clients are directed to mirrors in an attempt to distribute load and improve client performance. A number of interesting research problems are suggested by a server replication approach. To help gain insight into the issue of performance across mirror servers, a group at CMU made a large number of measurements to mirrors (see application-aware anycast webpage). A paper describing their work can be found on the webpage, along with the raw data.

This project consists of obtaining the CMU datasets and investigating additional interesting questions (beyond those presented in the Infocom'99 paper). The dataset is very rich, and the Infocom paper had limited space to present results. There are quite a few interesting additional questions to ask. For example, are there a few servers that are generally preferred by nearly all clients? Or do client preferences tend to balance load across the servers? Once you understand the data, you should propose some of your own questions to examine.

Preparation for the project includes reading the Infocom'99 paper, and getting the data files. Please contact me to discuss ideas on other interesting information to extract from the data.