Summary

  • A LACNIC 46 presentation models origin-to-destination traffic across 930 IP/MPLS tunnels using 128 directed link series from ANTEL’s international network.
  • Its 2025 dataset contains 261,248 five-minute link observations; the slides report a concentration result but no numeric error score for the estimated traffic.

ANTEL’s international network has 31 nodes. Each ordered pair makes 930 possible tunnels; the measurement set has 128 directed physical links. That is the arithmetic behind a LACNIC 46 study trying to infer where traffic travels from what each link carries.

Claudio Risso, Pablo Cuello and Carlos Martínez presented the work to the Internet Measurement Working Group on 21 October, using ANTEL’s international network as it stood in May 2025. The question is how much traffic moves from each node to every other node—the origin-destination, or O/D, matrix. In IP/MPLS, those flows travel over label-switched paths (LSPs). The presenters say LSP statistics are not standardised and can be less dependable than physical-interface counters, so the model starts with the link readings.

The dataset records 2,041 samples per directed link at five-minute intervals: 261,248 observations across 128 links. A routing model maps each tunnel’s path through those links; a second model uses the network’s daily traffic rhythm to fill points between 51 reference times. Together they produce 47,430 parameters to fit. That is 5.5 link observations per parameter.

A ratio of 5.5 link observations per parameter is not an accuracy score. The presentation says the constrained optimisation problem has multiple solutions and iterative results can lean toward their starting values. The team tackles the fit in sequence, first giving the strongest tunnels attention, then adding the rest. Twenty-five LSPs—2.69% of the 930—carry 60% of total traffic. The slides show an average fitted profile for this group, but attach no numeric error measure to it.

A modified method estimates only tunnels whose traffic is missing. The available LSP counters are used to identify the busiest tunnels; the authors say the revised fit is substantially better. Their black-box route also claims two practical advantages: fewer observations and independence from whether the network runs native IP, MPLS or Segment Routing. That is the part an operator can price—if the fit earns a score first.

The deck calls the results promising and says work continues. It publishes no residual, percentage-error or held-out score, and excludes temporary topology changes during link failures and restoration. A measured-versus-inferred score on unseen periods and reroute cases would show whether the estimate belongs in capacity and failure planning or stays in the research notebook.

The result is a useful measurement path, not yet an operator’s forecast. It turns ordinary link series into a candidate view of traffic between nodes; the missing number is how wrong that view can be.

Sources: LACNIC 46 agenda record and the original IMWG presentation.