VirtPhy: A Fully Programmable Infrastructure for Efficient NFV in Small Data Centers Cristina K. Dominicini, Gilmar L. Vassoler Federal Institute of Education, Science and Technology of Esp´ırito Santo (IFES), Brazil Email: {cristina.dominicini, gilmarvassoler}@ifes.edu.br Moises R. N. Ribeiro, Magnos Martinello Federal University of Esp´ırito Santo (UFES), Brazil Email: moises@ele.ufes.br, magnos@inf.ufes.br Abstract—This paper proposes, implements, and evaluates VirtPhy, a novel approach to build small-scale data centers net- work infrastructure for network functions virtualisation (NFV). VirtPhy is fully programmable to respond to NFV stringent demands related to throughput, latency, cost, rapid innovation, and efficient orchestration. A 8-node testbed demonstrates its key enabling mechanisms, which are programmed to jointly make the most of the available physical infrastructure. I. I NTRODUCTION Network functions virtualisation (NFV) has attracted a lot of attention in recent years as a potential solution for reducing costs and providing scalable network services [1]. Emerging trends in telecommunication networks, such as Fog Computing and Internet of Things (IoT), portend new types of services that require the geographical distribution of small-scale data centers (SSDCs) that will provide virtualised network func- tions (VNFs) at the edge of current networks [2]. In these new scenarios, the SSDs have to reduce costs while still being able to efficiently manage its computing and network infrastructure to meet demanding bandwidth and latency requirements of highly dynamic service requests. Traditional data center infrastructures have been designed to serve a scenario with a much larger number of servers and their network infrastructure is not though for complex service chaining interconnections between VNF elements. Moreover, their network reconfiguration mechanisms are limited, and packet routing/forwarding is not optimized. In this scenario, it is hard to build a cohesive solution between the underlying network topology and VNF orchestration decisions. To tackle these issues, we investigate the integration of NFV and software-defined networking (SDN) in SSDCs and take the virtualisation of network functions to a next level by virtualis- ing not only middleboxes, but also the network elements that interconnect servers. To this end, we propose VirtPhy, a fully programmable NFV infrastructure based on a server-centric data center architecture that aims to provide mechanisms for efficient NFV in SSDCs, considering scenarios of up to 512 commercial off-the-shelf (COTS) servers. The remainder of this paper is structured as follows. In Section II, we describe the current network infrastructures for NFV and their respective challenges. Then, Section III de- scribes our proposal to address the challenges and the VirtPhy architecture. Then, we describe our testbed implementation in Section IV. Section V contains the evaluation of VirtPhy, VNF VNF VNF VNF VNF VNF VNF VNF Computing Networking Network-centric Topology (a) VNF VNF Computing and Networking VNF VNF VNF VNF VNF VNF Server-centric Topology (b) Fig. 1. Comparison between data center network architectures: (a) traditional network-centric networks, and (b) server-centric networks. followed by the related works in Section VI. Finally, we outline the conclusions and future works in Section VII. II. NETWORK I NFRASTRUCTURES FOR NFV Traditional data center infrastructures for NFV are based on network-centric topologies, as seen in Fig.1(a). In two-tier and three-tier tree data center architectures, the servers hosting the VNFs are the leaf nodes and the network elements are responsible for routing/forwarding NFV traffic between server nodes. This approach presents several drawbacks: (i) The traffic forwarding is managed exclusively by expensive core, aggregation and edge network elements with high throughput capacity to avoid over-subscription; (ii) The network solutions are tied to proprietary hardware and software with limited room for innovation; (iii) The physical network infrastructure offers no extra degree of freedom to the NFV Orchestrator, which is tied to costly VNF migration operations when an overload event is detected; (iv) The routing is typically done in layer 3 using IP routing tables and ARP flooding, increasing network latency; and (v) The Spanning Tree (STP) protocol, used to build a logical topology that avoids loops, limits the bisection bandwidth, impacting on the network performance. One could argue that some of these problems can be solved by using SDN-enabled switches. This approach allows layer 2 forwarding with the use of flow tables. Besides, the NFV Orchestrator can interact with an SDN controller with a view of the whole network, such that the Orchestrator efficiently decides where to place VNFs in the network and the SDN Con- troller steers flows through service chains. On the other hand, this management flexibility takes its toll: (i) The costs can increase with the investment in expensive OpenFlow-enabled