LINA: A Fair Link-Grained Inter-Datacenter Traffic Scheduling Method With Deadline Guarantee

Abstract

Inter-datacenter networks take the role of connecting edge clusters and datacenters distributed globally. Nowadays, inter-datacenter network topologies are becoming more and more complicated with multiple links on each path, multiple paths between two datacenters, and overlaps between different paths. Hence, a link-grained scheduling method is necessary for making routing decisions and bandwidth allocation strategies when scheduling inter-datacenter transfers. For inter-datacenter transfers, deadline guarantee and fairness are essential and conflict requirements. Unfair preemptive scheduling methods ensure deadlines but cause starvation and service interference problems, while fairly sharing bandwidth among transfers ensures fairness but causes deadlines missed. However, a link-grained scheduling method simultaneously ensuring these two conflicting objectives is still missing. In this paper, a centralized controller, called LINA, is proposed to find fair link-grained bandwidth allocation strategies, which ensure transfers can compete fairly with each other for bandwidth on links and guarantee their deadlines in the meanwhile. First, we formulate the competition among transfers as a bottleneck routing game and prove the existence and non-uniqueness of its Nash Equilibrium, which is also the optimal bandwidth allocation strategy. Then, we propose a deep reinforcement learning based algorithm to derive it. Finally, experimental results show that LINA achieves superior performance than state-of-the-art methods in terms of fairness by 50%.