Abstract
Networks, the basis of the modern connected world, have evolved beyond the con- nectivity services. Network Functions (NFs) or traditionally the middleboxes are the basis of realizing different types of services such as security, optimization func- tions, and value added services. Typically, multiple NFs are chained together (also known as Service Function Chaining) to realize distinct network services, which are pivotal in providing the policy enforcement and performance in networks. Network Function Virtualization (NFV) is becoming more prevalent and enabling the soft- warized NFs to fast replace the traditional dedicated hardware based middleboxes in Communication Service Provider (CSP) networks. However, Virtualized Network Function (VNF) chains posit several systems and network level resource manage- ment and failure resiliency challenges: to ensure optimal resource utilization and performance at the system-level; and at the network-level to address optimal NF placement and routing for service chains, traffic engineering, and load balancing the traffic across Virtualized Network Function Instances (VNFIs); and to provide High Availability (HA), Fault Tolerance (FT) and Disaster Recovery (DR) guarantees. We begin by presenting NFVnice, a userspace NF scheduling framework for Service Function Chaining (SFC) to address the system-level resource utilization, per- formance, and scale challenges. NFVnice presents a novel rate-cost proportional scheduling and chain-aware backpressure mechanisms to optimize the resource uti- lization through judicious Central Processing Unit (CPU) allocation to NFs, and improve on the chain-wide performance. It also improves the scalability of NF de- ployment by allowing to efficiently multiplex multiple NFs on a single core. NFVnice achieves judicious resource utilization, consistently fair CPU allocation and provides 2x-400x gain in throughput across NF chains. Next, in order to address network-level challenges, specifically the orchestration and management of NFs and SFCs we develop DRENCH - a novel semi-distributed resource management framework to efficiently instantiate, place and relocate the net- work functions and to distribute traffic across the active NF instances to optimize both the utilization of network links and NFs. We model DRENCH as shadow-price based utilitarian market with Software Defined Networking (SDN) controller as a Market orchestrator to solve the Extended Network Utility Maximization (ENUM) problem. DRENCH results in better load balancing across Network Function In- stances (NFIs) and significantly lowers the Flow Completion Time (FCT) providing up to 1Ox lower FCT than the state-of-the-art solutions. We also present Neo-NSH, which extends on Network Service Header (NSH) to provide a simplified chain-wide steering framework. Neo-NSH leverages the SDN controller and discriminates the path-aware chain-wide transport at the control plane and service-aware but instance agnostic routing at the data plane. This separation presents two-fold benefits i) min- imize the path management complexity at the SDN controller ii) orders of magnitude reduction in the switch Ternary Content Addressable Memory (TCAM) rules; thus it enables for scalable, agile and flexible service function chaining. Finally, in order to achieve efficient NF migration and to address HA for NF chains, we present REINFORCE - an integrated framework to address failure resiliency for individual NF failures and global service chain-wide failures. REINFORCE presents a novel NF state replication strategy and distinct mechanisms to provide timely detection of NFs, hardware node (Virtualized Network Function Manager), and net- work link failures; and provides distinct failover mechanisms with strict correctness guarantees. NF state replication exploits the concept of external synchrony and rollback recovery to significantly reduce the amount of state transfer required to maintain consistent chain-wide state updates. Through the optimization techniques like opportunistic batching and multi-phase buffering, REINFORCE achieves very low latency (2 orders of magnitude lower latency) and less than 20% performance overheads. REINFORCE achieves NF failover within the same node in less than 100µseconds, incurring less than 1% performance overhead; and chain level failover across servers in a Local Area Network (LAN) within tens of milliseconds. In ad- dition, we present REARM, that adopts the concept of transient VNFs to migrate VNF within and across Data Centers (DCs) to facilitate HA in the event of disaster or power outages that frequent the Green Data Centers (GDCs). This dissertation combines abstract mathematical models to describe and derive NFV system behaviors, in order to design and develop system-level implementations for a set of working, ready-to-deploy NFV solutions. Our implementations have demonstrated their superior performance in addressing system-level performance, scale, and failure resiliency challenges. The proposed key solutions have been im- plemented on OpenNetVM, an open-source NFV framework, and are applicable to other NFV systems due to our generic design.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have