We propose a control-theoretic framework for cluster energy budgets that coordinates RAPL, vertical pod scaling, and admission control as coupled levers. Evaluated on a 64-node Grid’5000 cluster under mixed latency-sensitive and batch workloads.
@article{romial2026reactive,
title = "Reactive control of energy constraints in Kubernetes clusters",
author = "Romial, M. W. and Obama, C. and Pastor, J. and Lebre, A.",
journal = "ACM TOCS (under review)",
year = "2026"
}
We introduce AdaptRAPL, a scheduler extension that re-negotiates per-package power budgets every 250 ms based on observed SLO slack. On SPEC Cloud workloads, we achieve 19 % energy savings with < 2 % tail-latency regression.
@inproceedings{romial2025adapt,
title = "Energy-proportional scheduling via adaptive RAPL budgets",
author = "Romial, M. W. and Pastor, J. and Lebre, A.",
booktitle = "UCC '25",
year = "2025"
}
@inproceedings{romial2025multilever,
title = "Multi-lever re-adaptation of SLA under a power cap",
author = "Romial, M. W. and Lebre, A.",
booktitle = "BDCAT '25",
year = "2025"
}
@inproceedings{romial2024position,
title = "Kubernetes needs an energy descriptor",
author = "Romial, M. W. and Lebre, A.",
booktitle = "HotCarbon '24",
year = "2024"
}