VNF Dynamic Scaling and Deployment Algorithm Based on Traffic Prediction

NFV separates network functions from hardware-dependent middle boxes, which can significantly reduce costs and improve network management flexibility. It has been widely used in operator networks. However, due to traffic fluctuation in the network, using virtual network functions to provide flexible services is still challenging. In addition, most VNF scaling methods are passive in nature, which may cause high latency and fail to meet the QoS requirements of services. Therefore, this paper first proposes a GRU-based traffic prediction model and scales in/out VNF instances in advance based on the prediction result. Then we design a VNF buffering mechanism to avoid frequently releasing and creating VNF instances. Furthermore, based on the scaling results of VNF, we apply a DRL algorithm called A3C to train the agent and then obtain the optimal strategy of deploying new instances. Simulation results show that compared with other methods, the proposed proactive method can respond to traffic fluctuation in advance and reduce the total operating costs.