A REINFORCEMENT LEARNING-BASED DYNAMIC CONTAINER ORCHESTRATION BASED ON THE APPLICATION PROFILE

INTRODUCTION: In the last years, cloud computing has gained popularity mainly because of its properties of being an elastic and on-demand service. The elasticity refers to the easiness of increasing or reducing components in a computer cluster, and on-demand services consist of paying for the physical hardware usage. OBJECIVES: Cloud computing providers extensively use virtual machines and containers as virtualization techniques. However, containers are lightweight and cheaper, in energetic terms, in contrast to virtual machines. Because of this, cloud providers tend to migrate their virtual machine-based infrastructures to containers, leading to more container management as they are operating system processes. MATERIALS AND METHOD: Container orchestration fulfills a fundamental role in managing multiple containers along a cluster. Despite containers being lightweight, they lack efficient management of resources, incurring SLA deviations. To minimize those deviations, container orchestration techniques based on reinforcement learning show a promising approach in a future scenario, as those algorithms learn through interaction with the environment. RESULTS: The current work proposes a Recurrent Neural Network-based technique to detect SLA deviations, enabling the use of a reinforcement learning-based solution for the dynamic allocation of containers based on the application profile in future scenarios. FINAL CONSIDERATIONS: The detection task used three Recurrent Neural Networks, achieving, on average, 91% of the true-positive rates.