Skip to main content

PhD Candidate · Computer Science

Menra Romial

I work on making cloud infrastructures consume less energy, without making them slower.

I am a PhD candidate at IMT Atlantique and Inria, in the STACK team of the LS2N laboratory in Nantes, France. My thesis looks at power management across the whole cloud stack: from the hardware mechanisms that let a processor limit its own power, up to the orchestration layer where Kubernetes decides where and how workloads run.

At the hardware level, I study how Intel RAPL and Turbo Boost really interact on production servers, and build models that predict their behaviour. At the orchestration level, I design Kubernetes components that treat power as a first-class resource. This work relies on large-scale experiments on the Grid'5000 testbed and is carried out partly with Orange Research.

Before the PhD, I trained as a computer engineer at ENSPY in Yaoundé, Cameroon, and spent two years building and running production platforms as a developer and team lead. I am also involved in C5IN, a research and innovation network on cloud, edge and IoT in Cameroon.

Read my research Download CV
AffiliationsIMT AtlantiqueInriaLS2N

Featured work

Slowdown and energy ratio as the power cap decreases
Below about 50% of TDP, the slowdown outweighs the power saved and the energy ratio climbs above 1.
Euro-Par 2026Best Paper Award nominee

Understanding Power Limiting Mechanisms in Modern Processors: A Deep Dive Into Intel RAPL and Turbo Boost Dynamics

Power capping is the standard way to keep servers within an electrical budget. We measured how Intel's RAPL limits and Turbo Boost actually behave together on four generations of Xeon processors, and what that means for energy.

PDFDOISlides
  1. Few knobs matter. Turbo behaviour is driven by the long-term window and the two power limits; the short-term window is negligible.
  2. Turbo is predictable. A closed-form model gives how long a processor holds its boost, matching measurements almost exactly.
  3. There is a floor. Capping below half of TDP systematically costs energy instead of saving it.

Research

From the processor to the cluster

01

Hardware power limiting

How processors cap their own power. I study Intel RAPL domains, their power limits and time windows, and how they interact with Turbo Boost on real servers.

02

Modelling power behaviour

Turning measurements into predictions: analytical models of how long a processor holds its boost, and where power capping stops paying off in energy.

cluster power budget120 W68 W93 W
03

Energy-aware orchestration

Making power a first-class resource in Kubernetes: node-level capping, automatic server profiling and scheduling that follows real energy use.

News

Latest

All news →
  1. Talk

    Presented our work on Intel RAPL and Turbo Boost at Euro-Par 2026 in Pisa, Italy.

  2. Award

    Our Euro-Par 2026 paper was nominated for the Best Paper Award.

  3. Paper

    Paper accepted at Euro-Par 2026, the 32nd International European Conference on Parallel and Distributed Computing (acceptance rate 26.2%).

  4. Team

    Welcoming Clement Obama as M2 research intern, working on reactive control of energy constraints in Kubernetes clusters.

  5. Service

    Student volunteer at UCC and BDCAT 2025 in Nantes.

Publications

Peer-reviewed papers

Figure from Understanding Power Limiting Mechanisms in Modern Processors: A Deep Dive Into Intel RAPL and Turbo Boost Dynamics
Euro-Par 2026Best Paper Award nominee

Understanding Power Limiting Mechanisms in Modern Processors: A Deep Dive Into Intel RAPL and Turbo Boost Dynamics

Romial Menra, Guillaume Rosinosky, Remous-Aris Koutsiamanis, Sébastien Bolle, Jean-Marc Menaud

Euro-Par 2026: Parallel Processing, Lecture Notes in Computer Science, Springer, pp. 135-149

Takeaway Only the long-term window and the two power limits drive turbo behaviour; a closed-form model predicts turbo duration on four Intel Xeon generations; capping below half of TDP wastes energy instead of saving it.

PDFDOISlides
COMPAS2024
COMPAS 2024

Intégration de l'Aspect Énergétique dans Kubernetes

Romial Menra, Remous-Aris Koutsiamanis, Jean-Marc Menaud

COMPAS 2024, Conférence francophone en Parallélisme, Architecture et Système, Nantes, France

Takeaway Brings the server’s real power draw, measured through RAPL, into Kubernetes scheduling decisions, so that containers are placed according to actual energy use rather than requested resources.

PDF
All publications →

Talks

Recent presentations

Aug 2026

Understanding Power Limiting Mechanisms in Modern Processors

Euro-Par 2026 · Pisa, Italy

Slides →

Jul 2024

Intégration de l'Aspect Énergétique dans Kubernetes

COMPAS 2024 · Nantes, France

Paper →
All talks →