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

Manual configuration of available resources in the data center, as well as manual decision-making for customers' requests, makes the resource management process potentially error-prone. Therefore, the resource manager should make intelligent decisions for assigning available resources to existing requests to ensure scalable and efficient on-demand resource provisioning. Cloud job scheduling mechanisms aim to allocate the resources to users' submitted jobs optimally, yet optimal scheduling is an NP-complete problem. To address these challenges, many researchers have tried to tackle the job scheduling problem by proposing automatic solutions using Reinforcement Learning (RL) methods. Unfortunately, most of these methods ignore fair response time to all the incoming jobs with the proper utilization of data center resources. Tn this research, we use deep RL as a sequential decision-making method for automatic resource management that changes its behavior to deal with environmental changes. The approach uses the discrete soft-actor-critic algorithm. Tt has efficient sampling and stable learning convergence, as well as a precise adjustment of learning hyperparameters. Results show that compared to DeepRM and DeepScheduler, our approach improves slowdown and the balance of slowdown by at least three times using Google's dataset.