An LLM-Driven Digital TWin for Data Center Monitoring

Abstract

Modern data centers generate telemetry beyond human processing abilities, while existing tools require correlating metrics across disparate interfaces. This study explores whether a digital twin (DT) with a large language model (LLM) can integrate physical systems, live metrics, and three-dimensional visualization via natural-language interface. Prior work towards a Computing Infrastructure Digital Twin established a CD center in Omniverse and Prometheus metric pipeline. However, the two were disconnected: there was no mapping of #D objects to physical location. To create an AI-driven DT, we validated each virtual object against physical data center and recorded its corresponding scene geometry. Then, using Model Context protocol, we given an LLM agent 19 tools covering PostgreSQL analytics, forecasting, and reversible control of the 3D scene. End users can now pose natural-language questions, receive live diagnoses, and trigger Omniverse visualizations. Across four evaluated LLMs, tool-calling tasks achieved success rates of 75% - 85% with a median response time of 39.1s (<6s on DT infrastructure). Database-query tasks achieve a 100% success rate. Furthermore, forecasting tools, tested using nine weeks of historical data, achieved a mean absolute error of 0.022% for seven-day-ahead predictions. The system also found 108 CPU/network anomalies exceeding three standard deviations from their hourly norms. These results demonstrate that LLM-enables DTs can give relevant information to operators, suggesting potential for better cloud-infrastructure supervision. Future work will incorporate deep-learning to improve predictions and error detection for automating security actions.