KubeAI-An AI-Powered Pod Placement Scheduler for Kubernetes
IEEE Access · 2026 · 14 · 80268-80279
DOI: 10.1109/access.2026.3696055 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von H. Dhanyatha und 7 weiteren Autor:innen. Die Arbeit nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Efficient pod scheduling is vital for low latency, balanced resource use, and reliable performance in Kubernetes clusters that handle dynamic microservice workloads. Current methods, including the standard heuristic-based Kubernetes scheduler and reinforcement learning-based schedulers like Deep Reinforcement Scheduler (DRS), mainly depend on static rules or infrastructure-level metrics. This reliance limits their ability to respond to real-time changes in workloads and service interaction patterns. This paper introduces KubeAI, a pod placement scheduler that considers observability and is based on Proximal Policy Optimization (PPO). KubeAI integrates real-time telemetry from OpenTelemetry and the Kubernetes Metrics Server. This telemetry includes CPU and memory usage, disk I/O pressure, network latency, node temperature and inter-pod communication frequency. The scheduling policy is applied using a Flask-based Kubernetes scheduler extender. This setup allows for real-time inference, smooth integration with the native Kubernetes scheduler and flexible updates to policies without changing core scheduling components.Tests on a real Kubernetes cluster show that KubeAI reduces pod scheduling latency significantly compared to the default heuristic scheduler. In compute-intensive situations, it has latency levels comparable to DRS. The results also show that KubeAI makes better use of CPU and memory resources and reduces scheduling failures by up to 35% relative to the heuristic scheduler by using infrastructure-level metrics for observability-aware reinforcement learning. KubeAI also reduces time taken for inter-service communication by about 15% compared to heuristic and DRS schedulers by incorporating service interaction patterns into pod placement decisions. These findings indicate that integrating reinforcement learning with comprehensive infrastructure-level and service-level observability enables more efficient and reliable Kubernetes scheduling.
Zitieren
APA
Dhanyatha, H., Lavanya, M., Lal, R. K., Srivatsha, N., Gadad, V., Rani, K. R. U., Goswami, S., & Mehra, H. (2026). KubeAI-An AI-Powered Pod Placement Scheduler for Kubernetes. IEEE Access, 14, 80268-80279. https://doi.org/10.1109/access.2026.3696055
Vancouver
Dhanyatha H, Lavanya M, Lal RK, Srivatsha N, Gadad V, Rani KRU, et al. KubeAI-An AI-Powered Pod Placement Scheduler for Kubernetes. IEEE Access. 2026;14:80268-80279. doi:10.1109/access.2026.3696055
BibTeX
@article{Dhanyatha2026,
author = {Dhanyatha, H. and Lavanya, M. and Lal, Rishabh Kumar and Srivatsha, N. and Gadad, Veena and Rani, K. R. Usha and Goswami, Subrata and Mehra, Harsh},
title = {KubeAI-An AI-Powered Pod Placement Scheduler for Kubernetes},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {80268--80279},
doi = {10.1109/access.2026.3696055}
}
EUVIMED