|dc.description.abstract||The paradigm of network function virtualization (NFV) with the support of software-defined networking has emerged as a prominent approach to foster innovation in the networking field and reduce the complexity involved in managing modern-day conventional networks.
Before NFV, functions, which can manipulate the packet header and context of traffic flow, used to be implemented at fixed locations in the network substrate inside proprietary physical devices (called middlewares). With NFV, such functions are softwarized and virtualized. As such, they can be deployed in commodity servers as demanded. Hence, the provisioning of a network service becomes more agile and abstract, thereby giving rise to the next-generation service-customized networks which have the potential to meet new demands and use cases.
In this thesis, we focus on three complementary research problems essential to the orchestration and provisioning of NFV-enabled multicast network services. An NFV-enabled multicast service connects a source with a set of destinations. It specifies a set of NFs that should be executed at the chosen routes from the source to the destinations, with some resources and ordering relationships that should be satisfied in wired core networks.
In Problem I, we investigate a static joint traffic routing and virtual NF placement framework for accommodating multicast services over the network substrate. We develop optimal formulations and efficient heuristic algorithms that jointly handle the static embedding of one or multiple service requests over the network substrate with single-path and multipath routing.
In Problem II, we study the online orchestration of NFV-enabled network services. We consider both unicast and multicast NFV-enabled services with mandatory and best-effort NF types. Mandatory NFs are strictly necessary for the correctness of a network service, whereas best-effort NFs are preferable yet not necessary. Correspondingly, we propose a primal-dual based online approximation algorithm that allocates both processing and transmission resources to maximize a profit function that is proportional to the throughput. The online algorithm resembles a joint admission mechanism and an online composition, routing, and NF placement framework. In the core network, traffic patterns exhibit time-varying characteristics that can be cumbersome to model. Therefore, in Problem III, we develop a dynamic provisioning approach to allocate processing and transmission resources based on the traffic pattern of the embedded network service using deep reinforcement learning (RL). Notably, we devise a model-assisted exploration procedure to improve the efficiency and consistency of the deep RL algorithm.||en