Dynamic Resource Scaling for VNF Over Nonstationary Traffic: A Learning Approach

Software defined networking (SDN) and network function virtualization (NFV) are key enablers for service-level customized network slicing in fifth generation (5G) core networks. Network slices are required to be isolated from each other in terms of service performance with traffic load fluctuations. In this article, the virtual network function (VNF) scalability issue is studied to meet the quality-of-service (QoS) requirement in the presence of nonstationary traffic, through joint VNF migration and resource scaling. A traffic parameter learning method based on change point detection and Gaussian process regression (GPR) is proposed, to learn traffic parameters in a fractional Brownian motion (fBm) traffic model for each stationary traffic segment within a nonstationary traffic trace. Then, the time-varying VNF resource demand is predicted from the learned traffic parameters based on an fBm resource provisioning model. With the detected change points and predicted resource demands, a VNF migration problem is formulated as a Markov decision process (MDP) with variable-length decision epochs, to maximize the long-term reward integrating load balancing, migration cost, and resource overloading penalty. A penalty-aware deep Q-learning algorithm is proposed to incorporate awareness of resource overloading penalty, with improved performance over benchmarks in terms of training loss reduction and cumulative reward maximization.

I. Introduction

The service-oriented fifth-generation (5G) core networks are featured by customized services with differentiated multi-dimensional performance requirements, which can be provisioned through network slicing enabled by the promising software defined networking (SDN) and network function virtualization (NFV) paradigms [1]–[4]. A network slice supports a composite service via virtual network function (VNF) chaining, with dedicated packet processing functionality at each VNF, such as intrusion detection system (IDS) and firewall. Packets processed by a VNF are transmitted to next VNF in the same VNF chain for further processing, generating traffic between consecutive VNFs, i.e., inter-VNF subflows. During the planning of network slices, VNFs are placed at NFV-enabled commodity servers or data centers, referred to as NFV nodes, and inter-VNF subflows are routed over physical paths between the locations of corresponding VNFs [5]. The VNFs and subflows are allocated with static amounts of processing and transmission resources, respectively, according to the estimation of long-term resource demands [5], [6]. During the operation of network slices, traffic arrivals of each service fluctuate over time, possibly leading to a mismatch between traffic load and resource availability, which is detrimental to service performance and resource utilization. With the flexibility provided by SDN and NFV, it is possible to migrate VNFs among several candidate NFV nodes in a neighborhood, with elastic processing resource allocation for consistent quality-of-service (QoS) guarantee in the presence of traffic fluctuations [7], [8]. Accordingly, the subflows are rerouted over alternative physical paths to the new VNF locations.