One of the prominent challenges in deploying Wide Area Applications (WAA) is scalable performance management. The unpredictable behavior of WAN calls for models to predict the end-to-end latency between a client and server. Early research in this area presents the concept of Latency Profiles (iLPs) as a tool to capture the changing latencies experienced by clients when connecting to a server. It also presents a technique to group iLPs into Aggregate Latency Profiles (aLPs), study the relationships between iLPs using concepts of Mutual Information and Correlation and managing a large collection of iLPs using Relevance Networks.

Present research in this area, which is also presented in this thesis, deals with a new method of latency prediction using peers, apart from using aLPs as done earlier. The method involves identifying a group of peer clients experiencing similar latencies to servers and building vector of confidence values in peer clients for each client (for each server). These confidence vectors can be used for latency prediction.

The success of this research on scalable performance monitoring must be validated against thousands of iLPs. A new experiment on Planet-Lab is designed for this purpose. PlanetLab is a globally distributed wide area testbed for deploying newtork services at the Internet scale. The experimentation on Planet-lab involves deploying clients and server written in Python. This ensures a better control over the working of clients and servers and also takes care of other details like recording the processor load at both the client and servers and obtaining AS-level BGP paths between clients and servers. The additional parameters of processor loads and BGP paths help to better analyse the relationships between iLPs.