Analysis of Resource Management Methods Based on Reinforcement Learning

As the scale of service-based applications rapidly grows over recent years, tremendous amount of user data is being generated on a daily basis, which needs to be processed by computing jobs. Distributed computing frameworks are extensively applied to efficiently process large-scale data using finite resources, which has consequently placed resource management at the center of attention for many researchers. Traditional heuristic-based resource management algorithms are widely used in the industry, while often require experts with rich experience to design and tune rules, which is usually a time- consuming process and difficult to be generalized to computing jobs with distinct natures and scales. With the immense successes of reinforcement learning (RL) in the fields of games, auto- driving, and robotics, researchers begin to model and learn the task of resource management through the perspectives of RL, which has been proven to outperform conventional methods by experimental results. In this paper, we aim to summarize the relevant background, introduce both the heuristic-based and RL- based algorithms and propose a few areas of improvement for future work to come.