Server allocation in a CDN

In this paper we consider the problem of data center allocation to user requests so as to optimize utility of users while satisfying server constraints. The utility or quality of service may be measured as either a required throughput rate or more abstractly in terms of a utility-per-unit of flow. While the typical assignment of a data center to a user is via a matching algorithm, like in Akamai, the use of optimal assignment as well as multiple access paths(as in smart routing in multihoming contexts), allows web services to choose multiple data centers as services. We model the server allocation problem as a graph optimization problem. We propose a ϵ-approximation optimal utility allocation to the fractional version of the problem in this bipartite network. A 1/2-approximation can be obtained from the fractional relaxation. We also design a greedy approach and use it as a basis for comparison. Finally, we present experimental results on an existing ISP network topology.