Towards performance-centric fairness in datacenter networks

Fair bandwidth allocation in datacenter networks has been a focus of research recently, yet this has not received adequate attention in the context of private cloud, where link bandwidth is often shared among applications running data parallel frameworks, such as MapReduce. In this paper, we introduce a rigorous definition of performance-centric fairness, with the guiding principle that the performance of data parallel applications should be proportional to their weights. We investigate the problem of maximizing application performance while maintaining strict performance-centric fairness and present the inherent tradeoff between resource utilization and fairness. We then formulate the link bandwidth allocation problem with the objective of maximizing social welfare across all applications, so that resource utilization can be manipulated and improved by allowing a tunable degree of relaxation on performance-centric fairness. Based on dual based decomposition, we present a distributed algorithm to solve this problem, and evaluate its performance with extensive simulations.