Abstract:

With the growing importance of MLOps methodologies in the deployment and maintenance processes of machine learning models, there is an increasing demand for efficient and scalable management of computational resources in such environments. Despite numerous studies on tools supporting MLOps, relatively limited attention has been devoted to the optimization of task and component allocation regarding available system resources. This article presents the concept of a method for optimizing component allocation in a system with MLOps architecture, using a mathematical model, a genetic algorithm, a MLOps environment simulator to evaluate solutions, and multi-criteria evaluation methods. The proposed approach incorporates critical infrastructure parameters, including CPU and GPU processing capabilities, memory usage, and network bandwidth, alongside the operational characteristics of MLOps-specific components.