A self-adaptive approach to job scheduling in cloud computing environments

Due to its convenience and flexible services, cloud users have drastically increased during the past decade. Manual configuration for the available resources makes the resource management process potentially error-prone. While optimal scheduling is an NP-complete problem, it becomes more complicated due to other factors such as resource dynamicity and on-demand consumer applications’ requirements. In this research, we have used deep reinforcement learning (DRL) as a sequential decision-making method for automatic resource management that changes its behavior to deal with environmental changes. The proposed approach uses the discrete soft actor-critic algorithm which is a model-free deep reinforcement learning algorithm. The proposed approach is compared to similar reinforcement learning-based automatic resource management researches using Google’s dataset. Results show that the proposed approach improves the slowdown and the balance of slowdown at least, 3 and 5 times in the left-bi-model, 4 and 3 times in the right-bi-model, 3 and 7 times in the normal-model, 4 and 2 times in the balanced-bi-model and 3 and 3 times using the Google's dataset.