HGATS: A Heterogeneous GPU-aware Adaptive Training Scheduler for Large-scale Deep Learning toward BIM-driven Multi-objective Building Energy Retrofit Prediction
Fundamental Scientific Reports in Multidisciplinary Areas · 2026
DOI: 10.66238/fsrma112 ↗Kurz gefasst
Fachartikel in der Zeitschrift Fundamental Scientific Reports in Multidisciplinary Areas, 2026. Verfasst von Hao Hu Bowling und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The increasing adoption of Building Information Modeling (BIM) and multi-objective deep learning models for building energy retrofit assessment has substantially increased the computational demand of intelligent building analytics, making efficient large-scale model training on heterogeneous Graphics Processing Unit (GPU) clusters a critical challenge. Existing GPU scheduling strategies primarily rely on static allocation policies or homogeneous resource assumptions, which are unable to effectively balance training efficiency, resource utilization, and scheduling adaptability for computationally intensive BIM-driven prediction tasks. To address this issue, this paper proposes Heterogeneous GPU-aware Adaptive Training Scheduler (HGATS), an adaptive resource scheduling framework for large-scale deep learning training toward BIM-driven multi-objective building energy retrofit prediction. HGATS jointly models training task characteristics and heterogeneous GPU resource states through a Task Profiling Encoder (TPE) and a Heterogeneous Resource State Encoder (HRSE), while a Cross-domain Scheduling Fusion Module (CSFM) dynamically matches task requirements with heterogeneous computing resources using a cross-attention mechanism. In addition, a Dynamic Adaptive Scheduling Optimizer (DASO) continuously updates scheduling decisions according to online training feedback to improve resource allocation efficiency. A BIM-driven multi-task prediction model simultaneously estimating energy saving ratio, discounted payback period, operational carbon reduction, and life-cycle carbon benefit is employed as the representative large-scale training workload. Experimental results demonstrate that HGATS reduces end-to-end training time by 28.8% and improves GPU utilization by 26.3% compared with state-of-the-art scheduling methods, while maintaining comparable prediction accuracy across all BIM-driven retrofit tasks. These results verify that HGATS provides an effective and scalable infrastructure-level solution for adaptive heterogeneous GPU resource scheduling in large-scale deep learning systems for intelligent building energy retrofit applications.
Zitieren
APA
Bowling, H. H., Xie, R., & Wang, Y. (2026). HGATS: A Heterogeneous GPU-aware Adaptive Training Scheduler for Large-scale Deep Learning toward BIM-driven Multi-objective Building Energy Retrofit Prediction. Fundamental Scientific Reports in Multidisciplinary Areas. https://doi.org/10.66238/fsrma112
Vancouver
Bowling HH, Xie R, Wang Y. HGATS: A Heterogeneous GPU-aware Adaptive Training Scheduler for Large-scale Deep Learning toward BIM-driven Multi-objective Building Energy Retrofit Prediction. Fundamental Scientific Reports in Multidisciplinary Areas. 2026. doi:10.66238/fsrma112
BibTeX
@article{Bowling2026,
author = {Bowling, Hao Hu and Xie, Ruoxi and Wang, Yihan},
title = {HGATS: A Heterogeneous GPU-aware Adaptive Training Scheduler for Large-scale Deep Learning toward BIM-driven Multi-objective Building Energy Retrofit Prediction},
journal = {Fundamental Scientific Reports in Multidisciplinary Areas},
year = {2026},
doi = {10.66238/fsrma112}
}
EUVIMED