ACRE-DRL Framework Reduces Makespan in Heterogeneous Task Scheduling

Researchers have introduced a new deep reinforcement learning framework designed to improve the efficiency of task scheduling in heterogeneous distributed high-performance computing systems. The study, published in the journal Cluster Computing, presents a method called Adaptive Context and Relation Enhanced Deep Reinforcement Learning, or ACRE-DRL. This approach aims to address the challenges associated with scheduling multi-objective directed acyclic graph tasks, where the system must simultaneously manage task dependencies, communication costs, and dynamic processor loads. As the scale of tasks increases, the expanded scheduling space traditionally reduces exploration efficiency and training stability in reinforcement learning algorithms, a problem the new framework seeks to mitigate.

The ACRE-DRL framework uses relation-enhanced spatiotemporal embeddings to represent task dependencies, communication costs, and scheduling context. Additionally, the framework employs statistical-moment aggregation to capture variations in processor load.

To train the model effectively, the researchers developed an enhanced D3QN training scheme that combines noisy networks and prioritized experience replay. Furthermore, an adaptive reward mechanism balances makespan and load balance.

Experimental results demonstrate that ACRE-DRL outperforms the strongest baseline methods in the tested environments. On Random DAGs, the framework reduced the average makespan by 15.77% compared to the leading baseline. On the Montage workflow, the model achieved an 11.72% reduction in average makespan.

Beyond reducing the makespan, the study reports that ACRE-DRL improves parallel efficiency and resource utilization.

Subject of Research: ACRE-DRL: adaptive context and relation enhanced deep reinforcement learning for multi-objective dependent task scheduling in heterogeneous systems

Article Title: ACRE-DRL: adaptive context and relation enhanced deep reinforcement learning for multi-objective dependent task scheduling in heterogeneous systems

Article References: Wang, M., Lv, M., He, X., Liu, C., & Qiao, S. (2026). ACRE-DRL: adaptive context and relation enhanced deep reinforcement learning for multi-objective dependent task scheduling in heterogeneous systems. Cluster Computing, 29(16), Article 921. https://doi.org/10.1007/s10586-026-06736-z

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06736-z

Keywords: ACRE-DRL, adaptive, context, relation, enhanced, deep, reinforcement, learning, multi-objective, dependent, task, scheduling