Policy-Based Reinforcement Learning for Online Container Scheduling in Meta Computing

Meta computing is a novel computing paradigm that harnesses the collective computing resources of the Internet, offering efficient, fault-tolerant, and personalized services while ensuring strong security and privacy. Nevertheless, overcoming the existing barriers to integrating network-wide computing power and unifying the entire network's computational resources remains a formidable challenge. Containers have gained significant popularity as a virtualization solution, holding promise in addressing the challenges presented by how to efficiently utilize resources in meta computing. Existing research has shown that task scheduling at the granularity of containers can substantially reduce the completion delay and resource consumption, making it more suitable for meeting real-time requirements. To minimize both the total task time and energy costs, we introduce an online container scheduling algorithm tailored for a small-scale meta computing framework, which formulates the online container scheduling problem to optimize the overall task utility. We adopt a policy gradient-based Reinforcement Learning (RL) algorithm that accounts for the unique characteristics of meta computing and expects to get a good performance. Experimental results validate that our RL-based algorithm outperforms other commonly used baseline algorithms.