Skip to Main Content
Server selection plays an essential role in content replication networks, such as peer-to-peer (P2P) and content delivery networks (CDNs). In this paper, we perform an analytical investigation of the strengths and weaknesses of existing server selection policies, based initially on an M/G/1 processor sharing (PS) queueing-theoretic model. We develop a theoretical benchmark to evaluate the performance of two general server selection policies, referred to as EQ_DELAY and EQ_LOAD, which characterize a wide range of existing server selection algorithms. We find that EQ_LOAD achieves an average delay always higher than or equal to that of EQ_DELAY. A key theoretical result of this paper is that in an N-server system, the worst case ratio between the average delay of EQ_DELAY or EQ_LOAD and the minimal average delay (obtained from the benchmark) is precisely N. We constructively show how this worst case scenario can arise in highly heterogeneous systems. This result, when interpreted in the context of selfish routing, means that the price of anarchy in unbounded delay networks depends on the topology, and can potentially be very large. Our analytical findings are extended in asymptotic regimes to the G/G/1 first-come first-serve and multi-class M/G/1-PS models and supported by simulations run for various arrival and service processes, scheduling disciplines, and workload exhibiting temporal locality. These results indicate that our analysis is applicable to realistic scenarios.
Date of Publication: Dec. 2008