Summary

  • RFC 2103 divided mobility into two independent effects: an endpoint’s locator might change, and the network topology might change. Either could occur without the other.
  • A Dynamic Association Module could maintain a fresh endpoint–locator binding when topology remained stable. Once a network moved, maps and routers also had to participate; the binding alone could no longer describe reachability.

The most important row in RFC 2103’s mobility matrix is the one that looks impossible at first glance. The locator does not change. The topology does.

Imagine a network node moving and acquiring different neighbours while remaining inside the same enclosing Nimrod node. Its endpoints may retain the same locators. A lookup can therefore return a current, perfectly correct endpoint–locator association. Yet packets entering the region still need a new account of connectivity. The association is true and the route derived from an old map can still be wrong.

That distinction marked the limit of a clean modular idea.

A useful abstraction—with a precise edge

RFC 1992 gave Nimrod topology-insensitive endpoint identifiers and topology-sensitive locators. Endpoint identifiers were stable; locators expressed where nodes sat in a hierarchical map. A change of provider illustrated the consequence: an organisation could not simply carry an old locator into a different part of the hierarchy without flattening the routing system.

RFC 2103 turned that separation into a mobility model. It described a Dynamic Association Module, or DAM, whose job was to maintain the changing relationship between endpoints and locators. The DAM was not a prescribed server and not a renamed DNS. It was an abstraction for comparing designs. Its function might live in a database, a home representative or distributed router state.

If a host crossed from one node to another while the map stayed unchanged, the abstraction worked neatly. The DAM recorded the new locator. Nimrod could be notified or could ask for the association after detecting that the old destination was unreachable. Identity remained stable; location changed; the routing system continued to operate over the same graph.

RFC 2103 then separated four cases:

  1. neither locator nor topology changes;
  2. the locator changes but topology does not;
  3. topology changes but the locator does not;
  4. both topology and locator change.

Only the second case belongs wholly to the appealing story of rebinding. In the third, the DAM has nothing to update even though Nimrod’s map must change. In the fourth, both modules have work to do.

Network mobility made the abstraction leak

The paper’s examples were organisations relocating and wireless networks moving between cells in trains, aircraft, cars and ships. A node that changed neighbours could alter the graph. A node moving outside its enclosing node could also force new locators on the devices it carried.

RFC 2103 warned that Nimrod’s ordinary topology-update mechanisms were likely to be optimized for relatively slow change. Fast-moving networks might need specific updates sent to appropriate node representatives so that packets entering a node could follow the new topology. The document did not specify that interaction. It recorded the requirement and the architectural consequence: network mobility might require closer coupling between the DAM and Nimrod routing or routers than endpoint mobility did.

This is not a minor implementation note. It changes what a successful observation proves. A fresh endpoint–locator record answers one question: which locator is currently associated with this endpoint? It does not answer whether the node’s neighbours changed, whether every relevant map received that change, whether a route was recalculated, or whether an installed path remained usable.

Three places to put the intelligence

RFC 2103 classified solutions by where the association was maintained and where remapping occurred.

A centralized database queried by the source was simple and offered a direct route. It also created a single failure point and was judged inadequate as Nimrod’s long-term scalable answer.

A distributed scheme with remapping at a home representative resembled Mobile IP. The source could keep sending toward the home location; the representative redirected traffic to the current location. This avoided changing sources and ordinary routers, confined much control information, and could conceal the mobile endpoint’s location from correspondents. It also produced triangular routing and dependence on home-representative reachability. The paper called this or a modification of it a viable first cut, while explicitly questioning future scalability.

The third approach placed state in routers. That increased complexity and involved more Nimrod entities, but it was better suited to cases in which topology changed. Border routers could use recent mapping state near the region where movement occurred. The promised path could become more direct within the limits of the hierarchy, but the cost was architectural involvement precisely where the clean DAM boundary had tried to avoid it.

Mobile IP was a mechanism, not a verdict

The paper then adapted the mechanism of RFC 2002. A mobile host discovered a foreign agent or acquired a transient locator, sent a registration request and caused a home agent to cache a mobility binding. Packets sent toward the home identity could be intercepted, wrapped and delivered toward the foreign agent, which removed the wrapper and consulted its visitor list.

Every verb in that sequence matters. Registration could be granted while later forwarding failed. A cached foreign-agent locator could become stale. A packet could reach the foreign agent while the visitor-list entry was missing, in which case RFC 2103 said it would be discarded and an error returned. An indirect path could work without being optimal. Authentication of a request did not decide whether the user was entitled to join the visited network.

The paper also separated offline mobility, where a session is torn down during a move, from online mobility, where it stays up. It considered online support desirable. It did not demonstrate that the first-cut design preserved a particular transport session. Nor did it measure scalability. Its abstract limit was sobering: any tracking system with response time o fails when attachment changes arrive faster than 1/o. No deployment benchmark supplied the value of o.

The historical boundary

The RFC Editor record and IETF Datatracker record preserve RFC 2103 as an Informational document from February 1997. It says directly that it is not a protocol specification. It leaves detailed authentication, deregistration, fast topology updates, predictive routing and the DAM–router interaction unresolved.

That incompleteness is the historical point. Nimrod tried to keep identity, location and routing in separate conceptual layers. RFC 2103 showed exactly where the layers stopped aligning. Moving an endpoint could be described by a new association. Moving a network changed the graph that made the association useful.

Lu Heng’s Running-Code Primacy and Minimum Initial Specification provide a disclosed modern lens, not historical evidence. A document can define a thin interface and leave future choices local. Operational truth still begins only when implementations update state, routers accept it, paths work and operators observe the result. His Reality Layers language sharpens the evidentiary divide: a symbolic record about a locator cannot substitute for executable connectivity.

The defensible conclusion is narrow. A current binding is evidence of a current binding. To claim mobility worked, the record must also preserve the topology revision, update propagation, route decision, forwarding state, authorization, session observation and delivery outcome that the binding did not contain.

Sources