Energy-Aware Scheduling of Large-Scale Deep Learning Training

Read full text Summarise with VIRAT
Authors: Nathalia Wolf, Luce Brotcorne, Víctor Bucarey L.
Cited by: 0
References: 29

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Deep learning (DL) training workloads impose high and increasingly variable energy demands on GPU-equipped data centers. Existing scheduling methods rarely account for the long runtimes, low preemption tolerance, and spatial-temporal variability of renewable energy that characterize large-scale DL workloads. This study proposes a scheduling framework for geo-distributed data centers that maximizes the use of green energy. This framework integrates a mixed-integer linear programming model, a rolling-horizon optimization scheme, and a heuristic-column-generation matheuristic to ensure scalability. Short-term renewable availability forecasts are incorporated using a decision-centric evaluation that measures their impact on optimization quality.

Fields of Study & Subjects

Cloud Computing and Resource Management Energy Efficiency in Computing Smart Grid Energy Management

Publication details

Journal / source
HAL
Publication date
18 March 2026
Language
EN
Document type
Preprint