DRL-Deploy:Adaptive Service Function Chains Deployment with Deep Reinforcement Learning

Abstract

With the development of software-defined networking (SDN) and network function virtualization (NFV), the service function chain (SFC) has become a popular paradigm for carrying and completing network services. In this new computing and networking paradigm, virtual network functions (VNFs) are deployed in software entities/virtual machines through physical device networks in a flexible manner to improve resource utilization and reduce management effectiveness. In this case, it is critical to effectively deploy SFCs within an acceptable time to improve quality of service, while meeting the constraints of the physical network. In this paper, we propose an adaptive deep reinforcement learning based method for the online deployment of SFC requests with different QoS requirements, called DRL-Deploy. DRL-Deploy integrates graph convolutional neural networks to effectively extract physical network features and then adopts a parallel method to improve training efficiency, which can converge to the best state. We compare with existing benchmarks, and extensive experiment results show that DRL-Deploy outperforms all others in terms of acceptance rate and long-term average revenue by 8.6% and 22.9%, respectively, while reducing long-term average cost by 36.4%.