Result: Performance of server selection algorithms for content replication networks

In this paper, we investigate the problem of optimal server selection in content replication networks, such as peer-to-peer (P2P) and content delivery networks (CDNs). While a number of server selection policies have been proposed or implemented, understanding of the theoretical performance limits of server selection and the relative performance of existing policies remains limited. In this paper, we introduce a mathematical framework, based on the M/G/1 Processor Sharing queueing model, and derive closed-form expressions for the optimal server access probabilities and the optimal average delay. We also analyze the performance of two general server selection policies, referred to as EQ-DELAY and EQ-LOAD, that characterize a wide range of existing algorithms. We prove that the average delay achieved by these policies can theoretically be as much as N times larger than the optimal delay, where N is the total number of servers in the system. Furthermore, simulation results obtained using our M/G/1-PS workload model and the ProWGen Web workload generator show that the optimal policy can reduce the average delay of requests by as much as 30% as compared to EQ-LOAD and EQ-DELAY, in realistic scenarios. They also show that the optimal policy compares favorably to the other policies in terms of fairness and sensitivity to traffic parameters.