Adaptive Service Function Chain Orchestration Algorithm Based on Deep Reinforcement Learning

In traditional networks, service function chains consist of multiple network service devices that data flows traverse from source to destination, which leads to unnecessary resource waste due to issues like fixed network topology and deployments. SDN/NFV breakthroughs in decoupled control and hardware virtualization provide server virtualization techniques that lift limitations of traditional network bottom layer resources. This enables possibilities for reducing link latency and resource overhead to achieve adaptive orchestration.We propose DeepSFC, a distributed self-learning service chain orchestration method based on DRL. After offline training, the intelligent agents apply online with minimal overhead. Without requiring expert knowledge, they can make orchestration decisions by periodically observing environmental monitoring information. Each agent only observes its direct neighbors rather than requiring global knowledge. Compared to state-of-the-art heuristic and DRL algorithms, DeepSFC significantly improves throughput on real-world network topology and traffic traces (11% and 15%), while each online decision takes less time (1 ms).