基于DRF的Kubernetes调度算法研究
A DRF-based Adaptive Scheduling Algorithm for Kubernetes
林雨昇 1王纯1
作者信息
- 1. 北京邮电大学计算机学院北京 100876
- 折叠
摘要
针对Kubernetes原生调度器在多维资源环境下缺乏全局感知能力，导致资源分配不均及I/O与网络带宽等隐性瓶颈无法识别的问题，本文提出了一种基于带权重主导资源公平的自适应调度算法。该算法突破了传统DRF仅考虑CPU与内存且权重静态的局限，引入了CPU、内存、I/O及网络带宽的多维资源模型，并设计了基于实时监控数据的动态权重调整机制。通过将资源紧张度指数反馈至调度决策层，算法能够自动识别集群中的"主导资源"并引导任务避开热点节点。实验结果表明，相较于原生Kubernetes调度器，该算法在内存资源利用率上提升了3.9%，负载标准差降低了19.3%，有效实现了多维资源的均衡调度与隐性瓶颈规避
Abstract
To address the lack of global perception in Kubernetes\' native scheduler under multi-dimensional resource environments-specifically, the resulting resource imbalance and inability to identify "invisible bottlenecks" such as I/O and network bandwidth-this paper proposes an Adaptive Scheduling Algorithm based on Weighted Dominant Resource Fairness (W-DRF). The algorithm overcomes the limitations of traditional DRF, which only considers CPU and memory with static weights. It introduces a multi-dimensional resource model encompassing CPU, memory, I/O, and network bandwidth, and designs a dynamic weight adjustment mechanism driven by real-time monitoring data. By feeding the Resource Tension Index back into the scheduling decision layer, the algorithm can automatically identify the "dominant resource" within the cluster and guide tasks away from hotspot nodes. Experimental results show that, compared to the native Kubernetes scheduler, the proposed algorithm improves memory resource utilization by 3.9% and reduces the load standard deviation by 19.3%, effectively achieving balanced scheduling of multi-dimensional resources and avoiding invisible bottlenecks.关键词
计算机应用技术/Kubernetes/资源调度/主导资源公平/负载均衡/Key words
Kubernetes/Dominant Resource Fairness/Dominant Resource Fairness/Load Balancing引用本文复制引用
林雨昇,王纯.基于DRF的Kubernetes调度算法研究[EB/OL].(2026-03-27)[2026-03-28].http://www.paper.edu.cn/releasepaper/content/202603-264.学科分类
计算技术、计算机技术
评论