MRLCC: an adaptive cloud task scheduling method based on meta reinforcement learning

Abstract

Abstract Task scheduling is a complex problem in cloud computing, and attracts many researchers’ interests. Recently, many deep reinforcement learning (DRL)-based methods have been proposed to learn the scheduling policy through interacting with the environment. However, most DRL methods focus on a specific environment, which may lead to a weak adaptability to new environments because they have low sample efficiency and require full retraining to learn updated policies for new environments. To overcome the weakness and reduce the time consumption of adapting to new environment, we propose a task scheduling method based on meta reinforcement learning called MRLCC. Through comparing MRLCC and baseline algorithms on the performance of shortening makespan in different environments, we can find that MRLCC is able to adapt to different environments quickly and has a high sample efficiency. Besides, the experimental results demonstrate that MRLCC can maintain a high utilization rate over all baseline algorithms after a few steps of gradient update.

Keywords

• Meta reinforcement learning
• Deep reinforcement learning
• Task scheduling
• Resource management