Lore: A Learning-based Approach forWorkflow Scheduling in Clouds

The workflow scheduling problem is a critical challenge in clouds. Meticulously designed heuristics have attempted to address the intricate decision problem at a high cost. A more general approach is expected to handle different types of workflows and resource configurations. In this paper, a deep reinforcement Learning based apprOach for woRkflow schEduling (Lore) in clouds has been proposed to minimize the completion time of workflows. Moreover, Monte Carlo Tree Search and graph convolutional network are applied to improve performance further. Experimental results show that Lore outperforms the baselines, reducing average makespan by 2 - 10%, and enabling resource utilization increase by up to 20%.