Traditional network topology discovery methods rely heavily on specific protocols or are limited to single subnets, making them inadequate for large-scale, protocol-heterogeneous, and dynamic data center networks. To address this, we propose a framework that discovers physical device connections without active probing or protocol dependence. Our approach passively analyzes network traffic to form local topology fragments, which are then merged and refined into a complete map through a three-stage process: (1) topology prediction based on flow-data five-tuples, (2) topology optimization using delay covariance matrices to resolve anonymous routers, and (3) topology repair via a Deep Convolutional Generative Adversarial Network to recover missing links. The key innovations include a distributed and scalable discovery procedure and the theoretically-grounded application of image inpainting techniques to network graph completion. In simulations using NS3 and real topology image datasets, our method achieves over 95% accuracy and an F1-score above 0.93 in networks of up to 6000 devices, outperforming SNMP-based discovery in scalability while introducing minimal overhead. This work provides a practical, non-intrusive solution for topology discovery in modern data centers. • Protocol-agnostic framework discovers physical device connections via passive flow analysis. . • Distributed monitoring infers local topologies without active probing or centralized control. • Delay covariance matrices optimize topologies by resolving anonymous router structures. • Deep learning model repairs missing links using image inpainting on topology graphs. • Demonstrates scalability and effectiveness in large-scale data center environments.
Wang et al. (Sun,) studied this question.