MA3C: A Multi-Agent A2C Scheduling Approach for Real-Time Heterogeneous Serverless Edge-Fog Continuum

As the integration of serverless architecture and the edge-fog continuum continues to evolve, the need for efficient and effective scheduling methods becomes increasingly critical. Traditional scheduling approaches often struggle to meet the unique challenges posed by these dynamic and heterogeneous environments. Deep Reinforcement Learning (DRL), particularly the Asynchronous Actor-Critic (A2C) algorithm, offers a promising solution to these challenges. Our proposed method, MA3C (Multi-Agent Actor-Critic Continuum), utilizes a centralized value network (critic) in the fog node, while equipping each edge node with a decentralized policy network (actor). This design enables agents to optimize their scheduling decisions not only based on their own experiences but also by learning from the experiences of others. Simulation results demonstrate the effectiveness of MA3C, consistently outperforming alternative methods in terms of response time, failure rates, and overall reward, even as the complexity increases with higher node counts and request arrival rates. Specifically, MA3C achieves a 14% improvement in total performance and a 10% enhancement in response time compared to other policies. This study paves the way for new advancements in scheduling methodologies for the serverless edge-fog continuum, providing a robust and scalable solution for future research.