Dynamic Server Selection in the Internet

Our second Brown Bag Lunch of the semester will feature Professor Mark Crovella. He will discuss one of several interesting topics that he is working on.

Abstract

To provide scalability, services on the Internet increasingly need to be replicated. For example, popular WWW documents might be copied and placed on multiple servers; this balances the load on individual servers and decreases the distance between servers and clients. Unfortunately, once services are replicated, clients need a method for selecting which server to use -- the "server selection" problem. In this talk I'll outline a new approach that Bob Carter and I are taking to solving the server selection problem in the Internet. Previous approaches have required the client to select a single server, in advance (static selection). In this talk I'll show that there are significant advantages to deferring the selection decision until the instant the service is needed. Using measurements and experiments performed on the Internet, I'll show what sorts of metrics are best for dynamic selection strategies, and why dynamic selection strategies always outperform static strategies.