Desarrollo de un sistema de auto-configuración de cómputo en la nube para reducción de costos

Abstract

This paper addresses the implementation of an intelligent autoscaling system for cloud applications using reinforcement learning (RL). The primary objective is to design and evaluate a model that optimizes resource allocation in AWS through hybrid training, combining simulated and real-world metrics. The hypothesis is that this approach will improve efficiency and reduce operational costs compared to traditional dynamic scaling strategies. A deployment flow was implemented in AWS, starting in Lambda and scaling to Fargate and EC2 based on the workload. Services such as Amazon CloudWatch for monitoring, AWS ECR for container management, API Gateway, and an Application Load Balancer were used to create a global endpoint that unifies the URLs of each service. An RL agent was implemented that, based on traffic data and resource consumption, learned optimal scaling policies. The results showed that the RL model was able to reduce response times and improve resource utilization by 30% compared to static scaling strategies. Keywords: autoscaling, reinforcement learning, cloud computing, AWS.