Data and Computing Resource Integrated Allocation System for Deep Learning Workload Scheduling

Siya Xu

doi:10.1109/tcc.2026.3731759

Data and Computing Resource Integrated Allocation System for Deep Learning Workload Scheduling

2026-09-07

0

PRE

AI

DOI：10.1109/tcc.2026.3731759

Original

Original request for help

Save

Abstract

En 中文

Deep learning workloads deployed in cloud environments increasingly rely on geographically distributed data resources and cost-aware scheduling. However, existing cloud resource management approaches primarily focus on computing infrastructure allocation while treating data as a static input, overlooking the strong coupling between data characteristics and computing resources during model training. This limitation often leads to suboptimal matching among datasets, training tasks, and heterogeneous cloud resources, thereby reducing training efficiency and model effectiveness. To address this challenge, this paper proposes a data and computing resource integrated allocation system for deep learning workload scheduling in cloud computing environments. The proposed framework extends conventional cloud resource orchestration by treating data as a first-class schedulable resource alongside computing resources. A unified representation model is developed to characterize heterogeneous data resources through task relevance, data quality, and pricing attributes, while heterogeneous computing resources are represented according to application-domain-specific processing capabilities and performance characteristics. To enable scalable decision-making in large resource spaces, a factorized Q-network-based scheduler(F-DQN Scheduler) with hierarchical action decomposition is adopted to coordinate data and computing resource selection. Experimental results demonstrate that explicitly incorporating data characteristics into cloud resource allocation decisions significantly improves resource matching quality, scheduling efficiency, and overall deep learning workload performance compared with conventional compute-centric scheduling approaches.

Keywords:

data resources allocation

heterogeneous computing resources

computing network

deep learning workload scheduling

Data and Computing Resource Integrated Allocation System for Deep Learning Workload Scheduling

Abstract

Journal

Organization

Cited Papers