PhD Position: Frugal Distributed Training with Volatile Resources

This PhD position, based at the Inria Centre at Université Côte d'Azur in Sophia Antipolis, is part of the Inria–Hivenet Challenge Cupseli initiative. The successful candidate will be employed by Hivenet (soon to become Antimatter) and hosted primarily by the NEO project-team, working in close collaboration with the ARGO project-team and Hivenet engineers. The research focuses on designing novel distributed training algorithms that can operate efficiently in environments where computing and storage resources are shared for limited, unpredictable periods. The candidate will model resource availability patterns, develop algorithms that exploit these patterns, provide theoretical convergence guarantees, and evaluate the methods on both simulations and the Hivenet platform. The work bridges machine learning, distributed optimization, federated learning, and sustainability considerations.

• Develop new distributed training algorithms tailored to environments with volatile, heterogeneous computing resources
• Model resource availability patterns using real-world traces and measurements from the Hivenet platform
• Design algorithms that decide resource activation, data transfer strategies, and local computation budgets
• Provide theoretical convergence and training-time guarantees for the proposed methods under intermittent availability
• Investigate frugal and sustainable training strategies, including carbon-aware client and time-slot scheduling
• Evaluate proposed algorithms through simulation on standard ML datasets and controlled availability models
• Test the most promising algorithms in conditions approximating the Hivenet production system
• Optionally contribute to Master-level teaching and Master student supervision if interested
• Collaborate with Hivenet engineers and researchers from both NEO and ARGO teams