Performance- and energy-aware architecture for containerized cloud–edge continuum systems

Abstract

By 2030, the electricity demand of the Information Technology (IT) sector is projected to reach approximately 3,200 TWh, primarily due to the rapid growth of Artificial Intelligence (AI) and cloud-native services. This rapid growth establishes energy efficiency as a fundamental operational objective for cloud providers. However, achieving energy-efficient operation in cloud-native environments remains challenging due to a multi-tenancy architectural model, heterogeneous microservices, and highly dynamic workloads, where each service exhibits distinct performance characteristics and strict Service Level Agreements (SLAs). To avoid violating performance requirements, cloud providers often operate compute nodes at maximum performance levels, leading to significant energy inefficiencies. Modern processors provide advanced power management mechanisms, including Dynamic Voltage and Frequency Scaling (DVFS), uncore frequency scaling, and CPU idle state (C-state) control, enabling runtime adjustment of the performance–power trade-off. Despite their potential, effectively exploiting these mechanisms in cloud-native environments remains difficult due to the complex relationship between hardware control parameters, workload behavior, and application-level performance. Existing power governors, such as intel_pstate and schedutil, rely primarily on hardware-level utilization metrics and operate without awareness of microservice-level performance requirements. Consequently, they are unable to ensure energy efficient operation while maintaining strict performance compliance. This thesis addresses this limitation by developing performance-aware power management approaches tailored for cloud-native microservices. The proposed methods coordinate CPU core frequency, uncore frequency, and idle state control based on runtime workload characteristics and performance constraints. To achieve this, the thesis investigates and designs intelligent control strategies, including heuristic-based techniques and reinforcement learning–based controllers, capable of dynamically adapting to workload variability, multi-tenant interference, and evolving microservice performance requirements. The proposed approaches are implemented and evaluated on a real Kubernetes-based cloud-native testbed using representative microservice workloads and fine-grained power telemetry. Experimental results demonstrate that the proposed solutions significantly reduce system-level power consumption while maintaining strict performance requirement compliance across diverse deployment scenarios and workload conditions. These results validate the effectiveness of workload-aware power control in improving energy efficiency without compromising application performance. This thesis contributes novel control strategies, system implementations, and experimental insights for energy-efficient cloud-native computing. The proposed solutions provide practical and deployable mechanisms to improve the energy proportionality of modern cloud infrastructures, supporting the development of sustainable and performance-aware cloud systems.

Date	9 May 2026
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Nadjia Kara (Supervisor) & Abdelouahed Gherbi (Co-supervisor)