Cost-Aware SFC Collaborative Scaling Based on Multi-Agent RL With Stackelberg Game

With the exponential growth of network computing demands, Service Function Chain (SFC) scaling can meet the evolving demands and provide more service functionalities, which is crucial for addressing the shortcomings of network resources. However, in multi-domain and heterogeneous edge networks, existing SFC scaling methods aim to reduce the additional costs of delays and resources during scaling, which ignore the resource redundancy and accumulation caused by the high and low pressure of virtual network load. Additionally, the inappropriateness of selection order of inter-domain transit nodes and intra-domain service nodes leads to frequent SFC scaling and placement. To tackle these challenges, we propose a relative-cost-aware SFC collaborative scaling and placement mechanism (SFC-CSP) based on Multi-Agent Reinforcement Learning (MARL) with Stackelberg game. Firstly, we introduce a priority-based VNFs scaling queue to reduce the times of frequent SFC scaling. Then, to alleviate the imbalance between delay, resource redundancy, and resource accumulation, we establish a relative cost-based multi-objective optimization function. The aim is to minimize the delay cost, the relative computing resource cost, the storage resource cost, and bandwidth resource cost. Furthermore, to reduce the impact of selection order for inter-domain transit node and intra-domain service node on asynchronous SFC placement, we design an SFC-CSP mechanism based on MARL with Stackelberg game, considering the interaction between node action selections. Experimental results demonstrate that our proposed method not only achieves higher SFC acceptance rates compared with other methods, but also performs well in reducing end-to-end delay of SFC and resource accumulation.