Approximate server selection algorithms in content distribution networks

Server selection is an important function in any replication-based infrastructure, aiming at redirecting client requests to the "best" server according to some predefined metrics. Previous research work has mainly focused on client-side redirection schemes, where the client is responsible for the server selection process. Furthermore, previous work has shown that client probing techniques perform significantly better in discovering the "best" server, compared to hop- or RTT-based schemes. Client probing, however, is not very scalable, since the number of clients and servers in the network will be very large. In this paper, we propose a novel technique to transform the server selection problem into a problem of optimal routing, which enables us to shift the redirection process from the client to the server-side. In particular, we consider the environment of a content distribution network (CDN), and propose a flexible framework that can be used to optimize the server selection process, according to various metrics and/or policies. Using trace-driven simulations, we show that the proposed method can improve significantly the response time of HTTP requests while keeping the control overhead at a very low level.