SCOT: Scalable Online Transformation of Container Systems with Reinforcement Learning

• Xiang Zheng,
• Jing Zhang,
• Yichen Lu,
• Jianguo Qian,
• Sheng Zhang,

Vol. 19, No. 9, September 30, 2025
10.3837/tiis.2025.09.005, Download Paper (Free):

• Container Scaling; Large-Scale Data Processing; Multi-Agent Reinforcement Learning; Power Dispatch Automation; Time-Series Forecasting

Abstract

Power dispatch automation systems face significant challenges in managing containerized applications, particularly during sudden load fluctuations and complex grid events that demand rapid and reliable resource scaling. Traditional orchestration methods often fail to ensure both system stability and efficient resource utilization across highly dynamic grid services. To address these limitations, we present SCOT, a scalable container orchestration system specifically designed for power grid environments. SCOT introduces three core innovations: an end-to-end fusion architecture that integrates a Transformer-based model for precise workload forecasting, a multi-agent reinforcement learning (MARL) framework for coordinated container-level scheduling, and a Lamarckian learning mechanism that enables continuous self-adaptation by retaining effective policies from past scenarios. In addition, SCOT embeds grid-specific constraints and dynamic service priorities into its decision-making process, significantly enhancing its responsiveness and robustness under volatile conditions. Extensive evaluation using real-world grid workload traces demonstrates SCOT’s effectiveness: it achieves a 14% improvement in resource utilization over Dynamic Scalable Task Scheduling (DSTS), reaches 88.7% accuracy in predicting and handling demand spikes, and limits SLA violations to just 1.2%. SCOT also exhibits strong responsiveness, provisioning 128 containers in 4.9 seconds—2.8× faster than DSTS—and successfully orchestrating 2,500 containers across 1,000 VMs within 3.5 seconds, while reducing post-surge stabilization time by a factor of 3.

Statistics

Cite this article