Deep Q-Network-Assisted Energy-Aware Scheduling Framework (DQN-EAS) for Scientific Workflows

In modern computing environments, scientific workflows play a vital role in orchestrating complex computational tasks across diverse resources and domains. As high-performance computing (HPC) systems continue to evolve, the need for energy-efficient scheduling methods — without compromising task completion time — has become increasingly critical, especially in heterogeneous environments. This paper introduces the Deep Q-Network-Assisted Energy-Aware Scheduling (DQN-EAS) framework, which leverages deep reinforcement learning to balance multiple objectives — specifically, minimizing energy consumption and makespan — during workflow scheduling across heterogeneous HPC clusters. DQN-EAS is evaluated using benchmark workflows from Epigenomics, LIGO, Montage, and SIPHT, covering a wide range of task complexities. Experimental results demonstrate that DQN-EAS consistently achieves optimal or near-optimal makespan while significantly outperforming traditional approaches such as HEFT, NSGA-II, and A2C in terms of energy efficiency. Notably, DQN-EAS delivers energy savings of up to [Formula: see text] joules and reduces energy consumption by up to 20% compared to the best-performing baselines. Furthermore, under a weighted objective setting, DQN-EAS maintains a near-optimal makespan while achieving substantial energy reductions — validating its robustness and adaptability as a solution for energy-aware workflow scheduling in modern HPC systems.