Resource Management Scheduling-Based on Proximal Policy Optimization

Job scheduling solutions based on traditional heuristics are being severely challenged by the uncertainty and complexity of data center environments. There is an urgent need for data centers to adopt new techniques to optimize job scheduling. Job scheduling is a part of combinatorial optimization, most of which are considered as NP-hard problems. Previous work has effectively demonstrated that reinforcement learning (RL) is significantly effective in solving NP-hard problems. In this paper, we propose a reinforcement learning-based scheduling algorithm RMP to effectively solve different resource management problems. The model combines job scheduling with resource management system optimization, captures resource management models using convolutional neural networks, and makes scheduling decisions using proximal policy optimization (PPO). The results show that our proposed algorithm has faster convergence and better scheduling efficiency in terms of job slowdown compared with current deep reinforcement learning (DRL) algorithms DeepRM and other heuristics algorithms.