EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Multi-Objective Task Scheduler Using Deep Reinforcement Learning in Heterogeneous Kubernetes

Wei-Chun Chen, Chih-Yu Wen

IEEE Access · 2026 · 14 · 166402-166418

DOI: 10.1109/access.2026.3747297 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Wei-Chun Chen und 1 weiteren Autor:innen. Die Arbeit nennt 49 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Science and Technology Council (NSTC) of Taiwan, Smart Sustainable New Agriculture Research Center (SMARTer) at NSTC of Taiwan.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The rise of containerized workloads in cloud-native systems demands intelligent, adaptive task scheduling. Traditional Kubernetes schedulers rely on static heuristics, failing to manage heterogeneous hardware effectively, which causes suboptimal resource utilization and system imbalance. To address this, we propose EPPO-MOHA, a multi-objective deep reinforcement learning (DRL) scheduling algorithm tailored for heterogeneous Kubernetes environments. We formulate scheduling as a Markov decision process (MDP) using a heterogeneous-aware state representation that incorporates node-level capacities. Additionally, a multi-objective reward mechanism jointly optimizes resource utilization, load balancing, and response time. Experiments under varying workload scales demonstrate that EPPO-MOHA consistently outperforms existing heuristic and DRL-based schedulers. In a 1000-task scenario, EPPO-MOHA outperforms the best baselines by improving average resource utilization by $\approx 1.78$ % and reducing average load imbalance by 11.8%, while achieving a 100% scheduling success rate and an average makespan of 1291 s. When scaling to 1750 tasks, it demonstrates superior robustness, exceeding leading methods in average resource utilization by over 0.78% and reducing average load imbalance by 20.7%, while maintaining a 100% success rate, the shortest average makespan of 2043 s, and competitive response times. These results confirm that EPPO-MOHA effectively balances conflicting objectives and ensures robust performance across varying workload scales in heterogeneous environments.

Förderung

  • National Science and Technology Council (NSTC) of Taiwan
  • Smart Sustainable New Agriculture Research Center (SMARTer) at NSTC of Taiwan

Zitieren

APA

Chen, W., & Wen, C. (2026). Multi-Objective Task Scheduler Using Deep Reinforcement Learning in Heterogeneous Kubernetes. IEEE Access, 14, 166402-166418. https://doi.org/10.1109/access.2026.3747297

Vancouver

Chen W, Wen C. Multi-Objective Task Scheduler Using Deep Reinforcement Learning in Heterogeneous Kubernetes. IEEE Access. 2026;14:166402-166418. doi:10.1109/access.2026.3747297

BibTeX
@article{Chen2026,
  author = {Chen, Wei-Chun and Wen, Chih-Yu},
  title = {Multi-Objective Task Scheduler Using Deep Reinforcement Learning in Heterogeneous Kubernetes},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {166402--166418},
  doi = {10.1109/access.2026.3747297}
}
Kostenlos starten →