Abstract
Resource interference in co-located cloud environments poses significant challenges to efficient workload management and system performance. Traditional scheduling approaches often fail to adequately address the complex interactions between tasks sharing physical resources, leading to performance degradation and quality of service violations. This paper proposes a novel Multi-Agent Reinforcement Learning (MARL) framework for interference-aware task scheduling in co-located cloud workloads. The framework employs a hierarchical multi-agent architecture where distributed agents collaborate to optimize task placement while explicitly modeling and mitigating resource interference effects. Through centralized training and decentralized execution, agents learn coordinated scheduling policies that balance resource utilization, minimize job completion time, and reduce interference-induced performance degradation. Experimental results on a simulated 1000-server cloud cluster demonstrate that the proposed approach achieves 32% reduction in average job completion time compared to Dominant Resource Fairness scheduling, while maintaining 84% average resource utilization and effectively predicting interference patterns with 8.3% mean error.

This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright (c) 2025 Jianfeng Luo, Zhaoyu Qin, Mikkel Sørensen (Author)