Towards efficient service-level statistical multiplexing and resource allocation for inter-DC WAN

Abstract

Inter-datacenter (Inter-DC) WAN is shared by multiple services that generate a large amount of traffic on aggregate. Thus, efficiently multiplexing the traffic to improve network utilization while ensuring the requested quality-of-service for each service is highly challenging. To address this problem, we present a framework, called Statistical Network Abstraction (SNA), that is a new approach for managing bandwidth in inter-DC WAN to guarantee quality of service (QoS) and improve network utilization. In SNA, we first represent the amount of network resources and desired QoS level for each service in a statistical fashion. We also propose a centralized controller in SNA which performs (i) admission control (AdCon) and (ii) network provisioning (NetProv), by formulating and solving a min-cost flow problem with stochastic QoS guarantee and outputting a network embedding of the aggregate traffic demand onto the physical network. SNA differs from the earlier guaranteeing fixed bandwidth approaches in that service-level statistical multiplexing is realized without need of partitioning physical network resources for each service. We evaluate SNA in a real testbed and large-scale simulations, demonstrating that more service requests are accepted and higher QoS levels are provided compared to previous deterministic bandwidth request model.