Summary

  • On 26 October at RIPE 93, James Rice opened a discussion on layer-3 Internet exchanges: give each participant port a point-to-point link to exchange routers instead of relying on one shared layer-2 peering LAN.
  • His earlier RIPE 92 case study described traffic reaching the wrong ports or blackholing while route-server BGP remained established. The proposed product moves more forwarding control, and accountability for it, into the exchange.

A route server can say “we're connected” while a packet explores the building. James Rice's RIPE 93 session asks whether that mismatch is baked too deeply into the usual Internet exchange offer. His alternative is a layer-3 port: a point-to-point link from each participant to the exchange's routers, rather than another seat on a common layer-2 peering LAN. The exchange would do the forwarding. The sales brochure would need a sharper sentence about where a packet goes.

The concern has an operating history. In a RIPE 92 case study in May, Rice described a MAC-address change leaving neighbors with stale entries. Traffic for the old address could be flooded across the shared LAN; some member devices then forwarded it toward the intended recipient. The official transcript records his account of unexpected traffic, duplicate replies and BGP sessions that sometimes kept humming. A control-plane handshake was still taking place. The data plane was taking the scenic route.

Rice explained the persistent neighbor state through forward-progress confirmation in RFC 4861, section 7.3.1. The protocol allows reachability evidence from higher layers; in his account, traffic that eventually arrived could help keep a stale neighbor entry alive. The point for an operator is practical: “BGP up” and “only the intended port received customer traffic” are different test results. A route-server status page reports one of them.

The RIPE 93 abstract puts a new product on the table. Instead of asking participants to share a forwarding domain, an IX could route separately to each port. That could make participant isolation a feature with a test and a price. It also hands the exchange a bigger steering wheel. Its routers would choose and forward paths, maintain policy, handle failover and produce the evidence when a member reports a misdelivery. Participants would need to see how those decisions are made and what happens when they fail.

Rice asked which networks would prefer such a port and which exchanges would offer one. That question is the news; there is no new industry vote to report. The useful buying question is more exact: what does the port contract promise about delivery to the intended member, how can the customer test that promise, and who fixes a failure when the BGP session still looks beautifully green? “Connected” is a status. “Delivered to the right place” is a service.

Original sources