Enhancing Networks via Virtualized Network Functions

Abstract

In an era of ubiquitous connectivity, various new applications, network protocols, and online services (e.g., cloud services, distributed machine learning, cryptocurrency) have been constantly creating, underpinning many of our daily activities. Emerging demands for networks have led to growing traffic volume and complexity of modern networks, which heavily rely on a wide spectrum of specialized network functions (e.g., Firewall, Load Balancer) for performance, security, etc. Although (virtual) network functions (VNFs) are widely deployed in networks, they are instantiated in an uncoordinated manner failing to meet growing demands of evolving networks. In this dissertation, we argue that networks equipped with VNFs can be designed in a fashion similar to how computer software is today programmed. By following the blueprint of joint design over VNFs, networks can be made more effective and efficient. We begin by presenting Durga, a system fusing wide area network (WAN) virtualization on gateway with local area network (LAN) virtualization technology. It seamlessly aggregates multiple WAN links into a (virtual) big pipe for better utilizing WAN links and also provides fast fail-over thus minimizing application performance degradation under WAN link failures. Without the support from LAN virtualization technology, existing solutions fail to provide high reliability and performance required by today’s enterprise applications. We then study a newly standardized protocol, Multipath TCP (MPTCP), adopted in Durga, showing the challenge of associating MPTCP subflows in network for the purpose of boosting throughput and enhancing security. Instead of designing a customized solution in every VNF to conquer this common challenge (making VNFs aware of MPTCP), we implement an online service named SAMPO to be readily integrated into VNFs. Following the same principle, we make an attempt to take consensus as a service in software-defined networks. We illustrate new network failure scenarios that are not explicitly handled by existing consensus algorithms such as Raft, thereby severely affecting their correct or efficient operations. Finally, we re-consider VNFs deployed in a network from the perspective of network administrators. A global view of deployed VNFs brings new opportunities for performance optimization over the network, and thus we explore parallelism in service function chains composing a sequence of VNFs that are typically traversed in-order by data flows.