基於NSGA-II之Kubernetes自動擴縮的動態最佳化

Abstract

在現代雲端原生架構中，Kubernetes 已成為主流的容器編排平台，其中的自動擴縮機制對於控制基礎設施成本與維持服務穩定性具有關鍵作用。然而，Kubernetes 內建的 Horizontal Pod Autoscaler與 Vertical Pod Autoscaler 雖提供自動化的調整能力，但皆主要依賴靜態閾值與單一指標，如 CPU 或記憶體使用率，未能涵蓋服務品質指標，且兩者無法同時應用於同一部署，限制了彈性配置的可能性。實務上，若未能正確設定 Kubernetes關鍵參數，如資源配置、目標使用率、副本數範圍等，將可能導致資源浪費、效能下降，甚至系統不穩定。而在初期部署階段，由於工作負載特性尚未明朗，此類錯誤設定的風險亦相對提高。 本研究提出一套結合非支配排序基因演算法（NSGA-II）與動態選擇機制的自動擴縮參數最佳化框架。我們將autoscaling的參數建模成多目標最佳化問題，考量三項關鍵指標：資源成本（Resource Cost）、容器重啟率（Restart Rate）、請求失敗比例（Failure Ratio），分別對應基礎設施支出、系統穩定性與使用者服務品質。透過 NSGA-II 的演化式搜尋能力，本研究能有效探索複雜參數空間，產生兼顧多面向的高品質配置集合。其次，我們提出了一套動態選擇模組，可以根據即時指標與歷史績效，在執行期自動選擇最適合的參數組，以因應工作負載的動態變化。 實驗結果顯示，我們提出的框架在多種指標下均展現出顯著成效。首先，透過 NSGA-II 所產生的最佳參數組合，已可觀察到資源成本、系統穩定性與服務品質三者間的平衡趨勢，並優於原生 HPA 與固定參數設定。進一步地，動態選擇模組的整體表現明顯優於所有對照組方法，有效降低了總體資源使用成本，同時展現出高度的穩定性與彈性。整體而言，本研究提出的自動擴縮參數最佳化框架在 Kubernetes 環境下具備穩定、具彈性的自動化調控能力，為未來自動擴縮決策設計提供具體可行的方向與參考依據。

In modern cloud-native architectures, Kubernetes has emerged as the de facto standard for container orchestration. Among its core functionalities, autoscaling plays a vital role in controlling infrastructure costs and ensuring service stability. While Kubernetes offers Horizontal Pod Autoscaler (HPA) and Vertical Pod Autoscaler (VPA) as built-in solutions, both rely primarily on static thresholds and single-dimensional metrics—such as CPU or memory utilization—without accounting for quality of service. Moreover, HPA and VPA cannot be applied concurrently to the same deployment, limiting the flexibility of scaling strategies. In practice, misconfigured parameters can result in resource waste, degraded performance, or instability—especially during early deployment stages when workload characteristics remain unclear. This study proposes an optimization framework that combines the Non-dominated Sorting Genetic Algorithm II (NSGA-II) with a dynamic runtime selection mechanism for autoscaling parameter tuning in Kubernetes. We formulate the autoscaling configuration problem as a multi-objective optimization task, targeting three core objectives: resource cost, restart rate, and failure ratio—representing infrastructure cost, system stability, and service quality, respectively. Leveraging the evolutionary search capabilities of NSGA-II, our approach effectively explores the high-dimensional configuration space and identifies a set of well-balanced parameter combinations. To address workload variability, we further design a dynamic selection module that automatically chooses the most suitable configuration during runtime based on real-time metrics and historical performance. Experimental results show that our proposed framework delivers substantial improvements across multiple evaluation criteria. First, the configurations generated by NSGA-II demonstrate a clear trade-off balance among cost, stability, and service quality, outperforming both default HPA and fixed-parameter baselines. Furthermore, the dynamic selection mechanism consistently achieves the best overall performance, reducing resource costs while maintaining high resilience and adaptability under fluctuating workloads. Overall, this research presents a practical, extensible solution for dynamic autoscaling in Kubernetes, offering concrete guidance for future autoscaling strategies and system-level automation design.