Sharmen Akhter, Eui-Nam Huh, Journal of Internet Computing and Services, Vol. 27, No. 1, pp. 121-131, Feb. 2026
10.7472/jksii.2025.27.1.121, Full Text:  HTML
Keywords: Deep Learning, heterogeneous computing, CPU-GPU scheduling, Reinforcement Learning, inference optimization

Abstract

Hybrid CPU–GPU execution has become a practical solution for deep learning workloads constrained by limited GPU memory or heterogeneous compute environments. However, to the best of our knowledge, existing schedulers primarily optimize global objectives such as makespan or throughput while neglecting fine-grained temporal synchronization between CPU and GPU subtasks. This leads to significant idle gaps and transfer blocking, reducing the real parallel efficiency. In this paper, we introduce TempoRL, a temporal-aware reinforcement learning (RL)-based co-scheduler that directly minimizes CPU–GPU timing imbalance and maximizes data transfer overlap within intra-layer execution. TempoRL learns to select adaptive task-split ratios that minimize synchronization delay and hide transfer latency through concurrent computation. Through a lightweight environment simulation and empirical analysis, TempoRL demonstrates its ability to automatically converge toward near-optimal partitioning policies, achieving up to a 56.6%reduction in timing gap and notable improvements in end-to-end latency compared to heuristic and static scheduling baselines. The proposed framework provides a foundation for learning-based fine-grained hybrid scheduling, smoothing the path for more temporally efficient co-execution across heterogeneous processors.

Statistics
Cite this article