Summary

  • Petya Vasileva's RIPE 93 session on 26 October brought July ANRW research on combining path changes with delay, loss and throughput measurements from a mesh spanning more than 170 research sites.
  • In the team's Singapore hardware case, a displayed corridor's delay improved from 229.9 to 82.4 ms while loss rose from zero to 2.55% and throughput fell from 786.4 to 0.2 Mbps. The study treats that combination as a high-impact routing and performance event, not a victory for a faster path.

The network gave the operator a compliment and a bill. Delay on the illustrated SWT2_CPB → TAIWAN-LCG2 corridor fell sharply. Yet the same original ANRW slides show packet loss climbing to 2.55% and throughput almost disappearing. The path crossed commercial transit after a Singapore hardware problem, according to the researchers. A dashboard that celebrated only the lower latency would have sent the incident to the wrong queue with a smile.

That mismatch is the subject of Petya Vasileva's RIPE 93 session on 26 October. The team presented the method at the Applied Networking Research Workshop in July, so these are established findings carried to a new operator audience. Its perfSONAR mesh covers more than 170 WLCG/OSG sites in 42 countries and produces roughly four million tests a day. Those tests do not arrive together: the slides illustrate one-minute latency, ten-minute throughput and traceroutes separated by six to 24 hours. A route may change between observations; a slow transfer can reach the desk before the next path trace.

The method builds a recent baseline for each source-destination corridor, scores unusual traceroute paths using graph and structural features, and aligns the path event with measured delay, loss and throughput in a 30-minute window. The slides show thresholds used in its performance checks: one-way delay at least twice baseline, loss of at least 0.5%, or throughput at most 65% of baseline. Those settings turn separate instruments into a shared clock for a particular operator workflow.

The output has four trays. A route shift plus degraded performance is high impact. A route shift with no performance tests is a test outage — an evidence gap, not proof that users lost service. A changed route with no measured performance harm is a reroute only. Ambiguous cases wait for review. This separation matters because a changed traced path can make an alarming diagram without damaging a transfer, while missing tests can make an actual problem look wonderfully quiet.

The team also shows a UK–Netherlands submarine-cable fault appearing in its live triage capture hours before a network-operations ticket, with later ticket confirmation reported in the slides. The point is the handoff: a measurement can justify attention before a formal incident record catches up, but the label should carry the evidence that earned it. For operators, the governance question is who may turn a coloured cell into an urgent ticket and which measurements that person can inspect. A single green speed number did not save the Singapore transfer.

Original sources