Energy-Efficient Deployment of Applications in the Edge-Network-Cloud Continuum M/F

Job Description

Thesis Subject

Context and Motivation

The deployment of AI applications is undergoing a paradigm shift with the advent of 5G/6G networks, the Internet of Things (IoT), and edge computing. This evolution enables services to be deployed across the edge-network-cloud continuum [1], leveraging heterogeneous resources from edge devices (e.g., smartphones, microcontrollers) to cloud data centers [2,3,4]. This new paradigm addresses critical challenges such as latency, bandwidth, and energy efficiency, but also introduces new complexities in resource allocation, model deployment, and energy optimization.

At the same time, AI models, especially deep neural networks, are becoming increasingly complex, with energy consumption and carbon footprint emerging as major concerns. For instance, training a single large language model can emit as much CO₂ as five cars in their lifetime, and inference tasks contribute significantly to energy use. The CARECloud project (PEPR CLOUD) explicitly targets reducing the environmental impact of cloud infrastructures, making this PhD topic highly relevant to national and global sustainability goals.

Scientific Objectives
This thesis aims to develop novel methods for deploying AI applications in the edge-network-cloud continuum, with a focus on reducing energy consumption while maintaining model performance. The research will address some of the following challenges:

* Energy-Aware Deployment Strategies
- Model Compression: Investigate techniques such as quantization, pruning, and knowledge distillation to reduce the computational and memory footprint of deep learning models without sacrificing accuracy [7,9,13].
- Cascade Systems: Explore early-exit architectures and multi-stage inference to dynamically select the most appropriate model (from lightweight to heavyweight) based on real-time constraints (e.g., battery level, network latency) [10,11].
- Federated Learning: Study federated learning (FL) as a means to distribute AI training and inference across edge devices, reducing the need for data centralization and lowering energy costs associated with data transfer and cloud compute. FL allows models to be trained locally on devices, with only model updates (not raw data) being communicated, thus improving energy efficiency and privacy [14].
- Resource-Aware Scheduling: Design algorithms to optimize task placement (edge vs. cloud) and scheduling policies for AI workloads, balancing latency, energy, and accuracy [13].

* Trade-offs Between Efficiency and Performance
- Quantitative Analysis: Measure the energy consumption of AI workloads across different deployment scenarios (edge, network, cloud) and model compression techniques.
- Adaptive Configurations: Develop adjustable models that can be reconfigured on-the-fly to adapt to varying environmental and resource constraints.

* Environmental Impact
- Carbon Footprint Modeling: Extend existing frameworks to estimate the CO₂ emissions of AI deployments, accounting for both compute and network energy use.
- Optimization for Sustainability: Propose green AI deployment strategies that minimize energy use and carbon emissions, in line with the CARECloud project's objectives.

Research activities:

1. Analyze energy consumption of AI deployments in the edge-network-cloud continuum.
2. Design algorithmic methods for energy-efficient scheduling of AI workloads.
3. Investigate trade-offs between energy efficiency and model accuracy in compression techniques.
4. Develop adaptive deployment frameworks using cascade systems and early-exit models.
5. Evaluate environmental impact of proposed methods using lifecycle assessment tools.


Competence:

The ideal candidate should have:
* Knowledge of machine learning, especially neural networks or graph neural network or federated learning.
* Strong mathematical and algorithmic background (optimization, probability, linear algebra).
* Programming expertise in Python, with experience in PyTorch or TensorFlow.
* Familiarity with networking and edge computing (e.g., MEC, IoT, 5G/6G).
* Analytical skills for designing and evaluating optimization algorithms.
Fluency in English (essential for scientific communication and collaboration).

References

[1] H. Hua, Y. Li, T. Wang, N. Dong, W. Li, and J. Cao, “Edge comput- ing with artificial intelligence: A machine learning perspective,” ACM Computing Surveys, vol. 55, no. 9, pp. 1–35, 2023.
[2] S. Wang, T. Tuor, T. Salonidis, K. K. Leung, C. Makaya, T. He, and K. Chan, “When edge meets learning: Adaptive control for resource- constrained distributed machine learning,” in IEEE INFOCOM 2018- IEEE conference on computer communications. IEEE, 2018, pp. 63– 71.
[3] G. Drainakis, P. Pantazopoulos, K. V. Katsaros, V. Sourlas, and A. Amdi- tis, “On the distribution of ml workloads to the network edge and beyond,” in IEEE INFOCOM 2021-IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), 2021, pp. 1–6.
[4] W. Gao, Q. Hu, Z. Ye, P. Sun, X. Wang, Y. Luo, T. Zhang, and Y. Wen, “Deep learning workload scheduling in gpu datacenters: Taxonomy, challenges and vision,” arXiv preprint arXiv:2205.11913, 2022.
[5] J. Lin, W.-M. Chen, J. Cohn, C. Gan, and S. Han, “Mcunet: Tiny deep learning on iot devices,” in Annual Conference on Neural Information Processing Systems (NeurIPS), 2020.
[6] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mobilenetv2: Inverted residuals and linear bottlenecks,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 4510–4520.
[7] Y. Cheng, D. Wang, P. Zhou, and T. Zhang, “Model compression and acceleration for deep neural networks: The principles, progress, and challenges,” IEEE Signal Processing Magazine, vol. 35, no. 1, pp. 126– 136, 2018.
[8] J. Yu and T. Huang, “Autoslim: Towards one-shot architecture search for channel numbers,” 2019. [Online]. Available: https://arxiv.org/abs/1903.11728
[9] H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han, “Once-for-all: Train one network and specialize it for efficient deployment,” in International Conference on Learning Representations, 2020. [Online]. Available: https://openreview.net/forum?id=HylxE1HKwS
[10] Viola, P., & Jones, M. (2001, December). Rapid object detection using a boosted cascade of simple features. In Proceedings of the 2001 IEEE computer society conference on computer vision and pattern recognition. CVPR 2001 (Vol. 1, pp. I-I). Ieee.
[11] Rabanser, S., Rauschmayr, N., Kulshrestha, A., Poklukar, P., Jitkrittum, W., Augenstein, S., ... & Tombari, F. (2025). Gatekeeper: Improving model cascades through confidence tuning. NeurIPS 2025.
[12] Natale, E., Ferré, D., Giambartolomei, G., Giroire, F., & Mallmann-Trenn, F. (2024). On the sparsity of the strong lottery ticket hypothesis. Advances in Neural Information Processing Systems, 37, 40565-40592.
[13] Barros, T. D. S., Giroire, F., Aparicio-Pardo, R., Perennes, S., & Natale, E. (2024, May). Scheduling with fully compressible tasks: Application to deep learning inference with neural network compression. In 2024 IEEE 24th International Symposium on Cluster, Cloud and Internet Computing (CCGrid) (pp. 327-336). IEEE.
[14] Savazzi, S., Rampa, V., Kianoush, S., & Bennis, M. (2022). An energy and carbon footprint analysis of distributed and federated learning. IEEE Transactions on Green Communications and Networking, 7(1), 248-264.

Your Work Environment

The thesis will be carried out as part of the PEPR CLOUD project CARECloud (Comprendre, Améliorer, Réduire les impacts Environnementaux du Cloud computing). Cloud computing and its many variations offer users considerable computing and storage capacities. The maturity of virtualization techniques has enabled the emergence of complex virtualized infrastructures, capable of rapidly deploying and reconfiguring virtual and elastic resources, in increasingly distributed infrastructures. This transparent resource management gives users the illusion of access to flexible, unlimited and virtually immaterial resources. However, the power consumption of these clouds is very real and a cause for concern, as are their overall greenhouse gas (GHG) emissions and the consumption of critical raw materials used in their manufacture. At a time when climate change is becoming more visible and impressive every year, with serious consequences for people and the planet on a global scale, all sectors (transport, construction, agriculture, industry, etc.) must contribute to the effort to reduce GHG emissions. Clouds, despite their ability to optimize processes in other sectors (transport, energy, agriculture), are no exception to this observation: the increasing slope of their greenhouse gas emissions must be reversed, or their potential benefits in other sectors will be wiped out. This is why the CARECloud project aims to drastically reduce the environmental impact of cloud infrastructures.

Compensation and benefits

Compensation

2300 € gross monthly

Annual leave and RTT

44 jours

Remote Working practice and compensation

Pratique et indemnisation du TT

Transport

Prise en charge à 75% du coût et forfait mobilité durable jusqu’à 300€

About the offer

About the CNRS

The CNRS is a major player in fundamental research on a global scale. The CNRS is the only French organization active in all scientific fields. Its unique position as a multi-specialist allows it to bring together different disciplines to address the most important challenges of the contemporary world, in connection with the actors of change.

CNRS

The research professions