Candidate: Kaige Qu
Title: Multi-Timescale Dynamic Resource Provisioning for Services in SDN/NFV-Enabled Core Networks
Date: December 11, 2020
Time: 9:30 AM
Place: REMOTE ATTENDANCE
Supervisor(s): Zhuang, Weihua
The service-oriented fifth-generation (5G) and beyond core networks are featured by customized services with differentiated quality-of-service (QoS) requirements, which can be provisioned through network slicing enabled by the software defined networking (SDN) and network function virtualization (NFV) paradigms. A network slice supports a composite service via virtual network function (VNF) chaining, with dedicated packet processing functionality at each VNF. For each network slice with a target traffic load, the end-to-end (E2E) service delivery is enabled by VNF placement at NFV nodes (e.g., data centers and commodity servers) and traffic routing among corresponding NFV nodes, with static resource allocations. To provide continuous QoS performance guarantee over time, it is essential to develop dynamic resource provisioning schemes for services experiencing traffic dynamics in different time granularities during virtual network operation.
In this thesis, we focus on processing resources and investigate three problems on dynamic processing resource provisioning for delay-sensitive services in two different timescales.
In problem I, we investigate a dynamic flow migration problem, to accommodate the large-timescale changes in the average traffic rates with average end-to-end (E2E) delay guarantee, while addressing a trade-off between load balancing and flow migration overhead. We develop optimal formulations and efficient heuristic algorithms, based on a simplified M/M/1 queueing model with Poisson traffic arrivals. An equal-length decision epoch is used, under the assumption that the average traffic rate of the Poisson traffic is stable within a decision epoch but can change across different epochs.
Motivated by the limitations of Poisson traffic model, in problem II, we restrict to a local network scenario and study a dynamic VNF scaling problem based on a real-world traffic trace with nonstationary traffic statistics in large timescale. A change point detection driven traffic parameter learning and resource demand prediction scheme is proposed, based on which dynamic VNF migration decisions are made at variable-length decision epochs using deep reinforcement learning. An fBm traffic model is employed for each detected stationary traffic segment, based on properties of Gaussianity and self-similarity of the real-world traffic.
In Problem III, we focus on a sufficiently long time duration with given VNF placement and stationary traffic statistics, and study a delay-aware VNF scheduling problem in the presence of bursty and unpredictable small-timescale traffic dynamics, to achieve network utility maximization with minimum timely throughput guarantee for each service. Based on the Lyapunov optimization technique, an online distributed VNF scheduling algorithm is derived, which greedily schedules a VNF at each NFV node based on a weight incorporating the backpressure-based differential backlogs of required CPU cycles, the timely throughput performance, and the packet delay.