A Data-Driven Framework for Shared Bottleneck Detection across SD-WAN Overlays

Software-Defined Wide Area Network (SD-WAN) solutions rely on multiple overlays to enhance resilience, bandwidth utilization, and routing flexibility. However, because overlays are decoupled from the underlying physical network (underlay), shared underlay links among distinct overlays can remain hidden from the SD-WAN controller, possibly leading to congestion upon their concurrent utilization. Correlation of one-way delay (OWD) measurements across overlays has been proposed to detect shared bottlenecks without explicit underlay knowledge, yet the accuracy of this approach depends critically on two overlooked parameters: the volume of probing traffic and the correlation threshold. If not properly tuned, these parameters can lead to missed detections, false positives, or unnecessary network overhead. In this work, we address these challenges by introducing a traffic load calibration framework combined with a data-driven correlation threshold selection method. We evaluate our approach in an emulated real-world WAN scenario with multiple overlays, showing that it accurately identifies shared underlay resources while minimizing probing-induced congestion. Our results demonstrate that proper calibration significantly improves bottleneck detection effectiveness in SD-WAN.