Summary

  • Dmytro Shypovalov’s NANOG 98 prototype takes SR policy counters from routers, calculates traffic rates in a sampler and sends those rates to a path controller.
  • His slides show twelve active sampling policies in a demonstration. The same slides place a fuller production design in 2027, after sampling, controller agreement and rollback are worked through.

Segment Routing won a tidy bargain: transit routers no longer keep per-tunnel state. Then the bandwidth bill arrived. A router cannot adjust a reservation it does not track, and a path controller cannot react to traffic it never hears about. Dmytro Shypovalov put that missing conversation at the centre of his NANOG 98 presentation in Miami on 20 October.

His route around the gap starts with counters that routers already expose. A separate sampler fetches SR label-switched-path counters over gNMI, turns them into a traffic rate at an adjustment interval, and advertises the result toward a Path Computation Element through a BGP-LS SR-policy route. The controller can then recompute reservations and send an updated segment list. The router carries packets; the sampler supplies the measurement; the controller chooses the path.

The deck includes a sample policy showing 39.924 Gbps from six samples across 50.4 seconds. Another screen shows twelve active sampling policies and three automatic reservations around 30 Gbps. These are working demonstration screens. They establish that the loop can be assembled; they do not show how it behaves over months on an operator’s mixed network.

Shypovalov is candid about the weak joint. A rate, unlike a raw counter history, discards trends and context. In the first design the PCE does not choose the sampling method. The sampler does not bring interface or segment-identifier counters into the same decision. A burst, a stale sample and a change of destination can therefore ask for different policy moves while the controller sees a thinner picture than the network operator has.

His proposed next design puts raw telemetry into a store and lets the PCE sample it alongside other counters. The deck also describes multiple PCEs advertising distinguishable copies of a policy and shortest-path forwarding if all controllers fail. Those are design claims and operator checks to make, not a field-failure trace in the presentation. The final slide says he hopes the fuller version runs in production sometime in 2027.

The interesting line is between the twelve green policies on a demo screen and a control loop someone can trust at 3 a.m. Shypovalov has drawn it himself. A controller needs an agreed sampling rule, consistent results across controllers and a way to freeze or roll back a bad adjustment. Moving state out of the router did not abolish responsibility for the decision. It moved the decision to a place where its evidence can be inspected.

Sources: NANOG 98 session · Shypovalov’s original slides · RFC 8733 · RFC 9857.