Avaliação de plataformas de contêineres AWS para offloading computacional de aprendizado de máquina em dispositivos de baixo poder computacional

Abstract

The growing demand for Machine Learning (ML) applications on devices with limited computing power poses significant performance and energy consumption challenges. This work addresses this issue through computational offloading, evaluating the feasibility and efficiency of cloud container platforms for executing ML inference tasks. The main objective is to compare two Amazon Web Services (AWS) approaches: the serverless AWS ECS Fargate service and running containers on AWS ECS managed instances with EC2. For the analysis, a prototype image classification application is developed using the MobileNetV3Large model, and systematic experiments are conducted to collect metrics on latency, total processing time, and financial cost. The results demonstrate that, although the EC2 approach presents marginally superior performance in terms of latency, Fargate stands out as the more cost-effective option with lower operational complexity. It is concluded that Fargate represents a robust and financially viable solution for ML offloading scenarios, offering an advantageous trade-off between cost, simplicity, and adequate performance for most applications.