A Hybrid UDE+NN Approach for Dynamic Performance Modeling in Microservices

Abstract

Microservice architectures have become a cornerstone of modern cloud-based systems due to their scalability, modularity, and flexibility. However, managing the performance of such distributed systems in dynamic environments presents significant challenges. Traditional performance models, such as fluid models based on queuing theory, often fail to capture the nonlinear and dynamic interactions between microservices, especially under fluctuating load conditions. In this paper, we propose a hybrid modeling approach that integrates Universal Differential Equations (UDEs) with Neural Networks (NNs) to enhance the accuracy and flexibility of microservice performance predictions. The UDE+NN model combines the interpretability and efficiency of fluid models with the adaptive learning capabilities of neural networks, capturing unmodeled system dynamics and improving the prediction of key performance metrics, including queue lengths and response times. Through extensive simulations, we demonstrate that the hybrid model significantly outperforms traditional fluid models, particularly in high-load and variable-traffic scenarios. Furthermore, the UDE+NN model enables real-time optimization of load balancing strategies, leading to better resource allocation and reduced operational costs. This work provides a robust framework for real-time performance management of microservice architectures, offering enhanced adaptability and predictive accuracy.