Kubernetes 環境中節點退役流程對容器資源配置與排程策略之影響分析

Abstract

隨著雲端運算與微服務架構的普及，Kubernetes 已成為主流的容器編排平台。然而其預設調度策略未能充分考量節點硬體世代差異與工作負載特性，導致生產環境中常見資源配置效率低落與基礎設施成本偏高等問題。 為因應此一挑戰，本文以某企業部署於 AWS EKS 的一年運行快照為基礎，設計一組具實務代表性的模擬環境，並提出兩階段調度策略：「Guaranteed Spread & New-Node Priority Scheduling（GSNPS）」與「CPU-Resource Aligned Scheduling（CRAS）」。GSNPS 於初始 Pod 排程階段納入節點硬體世代與區域分佈考量，以提升部署的更新效益與容錯能力；CRAS 則於後處理階段應用模擬退火演算法（Simulated Annealing, SA），將 CPU 密集型工作負載重新配置至計算優化節點，以提升資源對位效率。 根據 24 小時模擬結果，本策略能有效促進老舊節點的退役，並提升整體資源使用的成本效益，顯示其在異質化 Kubernetes 叢集下具備實用潛力。

With the rise of cloud computing and microservice architectures, Kubernetes has become the standard platform for container orchestration. However, its default scheduler often overlooks factors such as hardware generation and workload-specific resource profiles, leading to suboptimal resource utilization and elevated infrastructure costs in long-running production clusters. This thesis presents a case study based on an anonymized one-year snapshot from an enterprise-scale AWS Elastic Kubernetes Service (EKS) deployment. Two scheduling strategies are proposed: Guaranteed Spread and New-Node Priority Scheduling (GSNPS), which prioritizes zone-level fault tolerance and favors newer, cost-efficient instances during pod placement; and CPU-Resource Aligned Scheduling (CRAS), a post-processing refinement that applies Simulated Annealing (SA) to reallocate CPU-intensive workloads to compute-optimized nodes. A 24-hour simulation demonstrates that the combined GSNPS and CRAS strategy significantly improves infrastructure efficiency. GSNPS facilitates the retirement of underutilized legacy nodes, while CRAS enhances resource alignment by matching workload profiles to node specializations. The results validate the effectiveness of this two-stage scheduling framework in promoting cost-efficient node usage and sustainable cluster operation in heterogeneous Kubernetes environments.