AI-Driven Predictive Modeling for System Performance and Resource Management in Microservice Architectures

Abstract

This study proposes a unified end-to-end framework for response-time prediction in microservice architectures. The framework begins with data-quality controls and normalization to align and denoise multi-source system metrics. It constructs spatiotemporal representations by combining multivariate feature representations with historical response signals. It employs Transformer-based cross-scale modeling to capture short-term fluctuations, long-term trends, and long-range service dependencies. The model is trained with a mean-squared-error(MSE) objective and robust regularization. We conduct systematic evaluations on a public microservice dataset under various conditions, including disk‑I/O throttling, container‑affinity changes, CPU‑quota and preemption‑intensity variations, as well as workload migration and traffic‑peak switching. The framework examines the dynamic changes of overall error and high-percentile error through continuous prediction. Results demonstrate that the method maintains low RMSE and MAE under complex runtime conditions, significantly suppresses P95 fluctuations, and preserves high goodness of fit. It remains robust to concept drift and resource‑policy changes, providing proactive and actionable signals for capacity planning, elastic scaling, and service‑level (SLO) assurance.