Hierarchical NFV Datacenters : Resource Allocation with Cost - Latency Tradeoff

Network Function Virtualization (NFV) allows datacenters to consolidate network appliance functions onto commodity servers and devices. Currently telecommunication carriers are re-architecting their central offices as NFV datacenters that, along with Software-Defined Networking (SDN), help network service providers to speed deployment and reduce cost. However, it is still unclear how a carrier network shall organize its NFV datacenter resources into a coherent service architecture to support global network functional demands. This work proposes a hierarchical NFV/SDN-integrated architecture in which datacenters are organized into a multi-tree overlay network to collaboratively process user traffic flows. The proposed architecture steers traffic to a nearby datacenter to optimize user-perceived service response time. Our experiment results reveal that: (1) the 3-tier architecture is favored over others as it strikes a good balance between centralized processing and edge computing and (2) the resource allocation should be decided based on traffic’s source-destination attributes. Our results indicate that when most traffic flows within the same edge datacenter, the strategy whereby resources are concentrated at the carrier’s bottom-tier datacenters is preferred, but when most traffic flows across a carrier network or across different carrier networks, a uniform distribution over the datacenters or over the tiers, respectively, stands out from others. As innovation of network technology accelerates, hardware-based network appliances rapidly reach end-of-life, becoming a primary source of expenditures of today’s carrier networks. Currently telecommunication carriers address this problem by adopting Network Function Virtualization (NFV) technology to re-architect their central offices as NFV datacenters [1]. Such NFV datacenters consolidate network appliance functions onto commodity servers and devices, which helps service providers to speed service deployment and reduce cost. NFV datacenters can be directed by programmable control planes with Software Defined Networking (SDN) protocol to intelligently steer user traffic flows to the best suited network function unit therein. At present it is still unclear how a carrier network shall organize its NFV datacenters into a coherent service architecture and how to deploy the datacenter resources to support global network functional demands. From one perspective, traditional system architects would like to squeeze as many resources into a fewer super datacenters as possible to achieve economy of scale. From another perspective, telecommunication carriers, when acting as service providers, may want to place sufficient datacenter resources at the Internet’s edge to improve user-perceived service response time [2]. This calls for a new NFV service architecture that can flexibly and efficiently support global network functional demands while accommodating a variety of user traffic patterns and timing constraints. In response to these challenges, our design philosophy is to divide resources of carrier networks among many NFV datacenters and connect them into a multi-tree overlay network. In the proposed architecture, most of the NFV datacenters act as edge datacenters, which are deployed at major subscriber access networks to reduce data transport latency and cross-network traffic. The remaining NFV datacenters are placed above the edge datacenters in the hierarchy to absorb transient traffic bursts. These NFV datacenters collaboratively process global network functional demands on the carrier networks and intelligently steer traffic flows to a nearby datacenter which yields the minimum expected user-perceived response time. This article introduces our efforts of designing such NFV network architecture and integrating hierarchical NFV datacenters with SDN-based control plane mechanisms. 1 Ying-Dar Lin and Chien-Ying Huang are with National Chiao Tung University. Yuan-Cheng Lai is with Taiwan University of Science and Technology. Correspondence: Chih-Chiang Wang is with National Kaohsiung University of Applied Sciences, E-mail: [email protected] Hierarchical NFV Datacenters: Resource Allocation with Cost-Latency Tradeoff