A Flexible Deployment Scheme for Virtual Network Function Based on Reinforcement Learning

Network function virtualization (NFV) technology is widely used in network slicing in 5G networks and traffic processing in wide area network. However, with the increasing of service requests during the life cycle of a virtual network function (VNF), how to flexibly deploy the VNF becomes a key problem to make maximum use of the limited capacity of the physical network resources, and meanwhile satisfy the requirements of quality of service (QoS) in NFV scenario. In this paper, aiming to solve whether and how to scale VNF on demand, we formulated this problem as a non-convex linear mathematical optimization model where the optimization goal is to minimize the delay and energy consumption of the service function chain (SFC). Specifically, we propose a VNF flexible deployment scheme based on Reinforcement Learning (RL). Moreover, we train the agent by interacting with the physical network environment and take action according to the state of physical node to find the optimal physical resource allocation strategy of the VNF scaling. In addition, the state space, action space and the reward function are defined as available resource, migration or scaling decision and the reciprocal of total cost respectively. Extensive simulation results demonstrate that the proposed algorithm outperforms the comparison algorithm in terms of reducing the delay and increasing the ratio of successful scaling request.