EVRM: Elastic Virtual Resource Management framework for cloud virtual instances

• Published in: Future generation computer systems 2025-04, Vol.165, p.107569, Article 107569
• Main Authors: Wang, Desheng, Li, Yiting, Zhang, Weizhe, Yu, Zhiji, Tian, Yu-Chu, Li, Keqin
• Format: Article
• Language: English
• Subjects: Deep reinforcement learning; Elastic resource management; Multiple virtual resources; OpenStack
• ISSN: 0167-739X
• DOI: 10.1016/j.future.2024.107569

As cloud demand for computation and network resources fluctuates, effective resource management becomes essential for optimizing allocation and enhancing performance in virtualization-based applications. Current methods struggle to efficiently schedule multiple virtual resources for dynamic workloads. To address this, we propose a self-adaptive elastic virtual resource management (EVRM) framework that comprises a monitor, analyzer, planner, and executor, enabling dynamic scheduling of CPU, memory, and bandwidth for virtual instances. Central to EVRM is a resource management model employing a novel deep reinforcement learning approach, the deep deterministic policy gradient-based resource allocation (DDPG-RA), which coordinates resource allocation by automatically exploring optimization policies and learning complex relationships between resource allocation and performance. Additionally, DDPG-RA features an action refinement algorithm to derive multiple resource allocations from its outputs. Experimental results using OpenStack demonstrate that EVRM significantly enhances performance, achieving approximately 52.87% faster benchmark completion times and a 41.37% reduction in average time under both light and heavy loads, outperforming three competing approaches while optimizing physical resource utilization. •Build a novel model for elastic virtual resource management.•Propose a self-adaptive algorithm to quantitatively calculate target resource.•Investigate the key factors of virtual instance running performance.•The EVRM improves the resource utilization and virtual instance running performance.