We have conducted experiments, using content servers that are topologically located in North America and Europe. The clients were located in the USA at the University of Maryland and the University of Pittsburgh, the Technion - Israel Institute of Technology, and the University of Passau in Germany. At each of these sites, we deployed four to eight clients. Multiple clients scattered within a client AS were needed to verify if clients within a client AS exhibit similar performance.
A full description of our experiments is given in a forthcoming paper [5]. There are two main conclusions of our empirical analysis. First, there is a positive impact of the client AS on coverage. Clients within a client AS exhibit similar latency distributions. Second, aggregate LPs can be constructed to provide good coverage for client-server pairs.
Couple of notes are in order at this point. First, many ISPs (service providers) operate a backbone that spans a wide geographic area. We expect that in this situation, the client AS will have little significant impact on the aggregate LPs. In an ongoing experiment, we are testing clients within a campus AS that is geographically distributed over several tens of miles.
Second, we note that individual LPs are designed to reflect some network and server workload, as well as other factors that impact latency, e.g., timeout. However, LPs may not adequately capture additional factors that will have an impact on coverage. e.g., time intervals when a content server is unavailable, replication of content as in content delivery networks, the presence of proxy servers that cache objects to reduce latency, etc. These factors are captured in what is termed resource profile and client profile in our project. Joint together, these three profiles are expected to provide better service for data delivery in Wide Area Applications.