Routing Anomalies & Performance Degradation in R&E Networks: Aligning Path Changes with Measured Impact

Large R&E networks are full of partial signals: traceroutes show path changes, OWAMP tests show delay and loss, iperf3 shows throughput - each running at its own cadence, inspected with its own tool. When a data transfer slows down, the real question is simple but hard to answer fast: did the path change, and does that explain it?

We built a routing-aware incident triage tool on top of perfSONAR data from the WLCG/OSG mesh (170+ sites, ~4M tests/day). For each monitored corridor (source-destination pair), it detects routing shifts using graph-based path embeddings scored against that corridor's own recent baseline, then aligns them with delay, loss, and throughput evidence. Finally, each pair is assigned to an operator-facing class - confirmed impact, routing change with no measurable effect, tests interrupted, or needs review - so operators get a prioritized list.

We will walk through how the routing representation and correlation stage work, then through real incidents: a router migration, a hardware failure that rerouted traffic through commercial transit, and a submarine cable fault the tool caught live, hours before the NOC ticket was filed. The method is perfSONAR-native: no BGP feeds, no additional probing, no labelled training data, and directly reusable on any comparable measurement mesh.