The next hop still needs a reachable link neighbor

After selecting a route, the sender determines the relevant next-hop address. For an on-link destination this may be the destination itself. For a remote destination it is generally the local router selected by the route. IPv4 ARP on Ethernet asks which link-layer address corresponds to that local IPv4 address. The request commonly uses link-layer broadcast; the resulting frame to the router still contains the original remote IP destination. Routers repeat local forwarding decisions on subsequent links.

Do not expect the remote server's Ethernet address to appear in the client's local neighbor cache. Link-layer addresses are meaningful on their link, and a router replaces the link envelope as it forwards. A subnet-mask error can make a client incorrectly treat a remote destination as on-link and send unanswered ARP requests instead of using its gateway. Comparing the selected route, chosen next hop, and actual ARP target often exposes that configuration error faster than inspecting application retries.

IPv6 Neighbor Discovery uses ICMPv6 messages and performs more than address resolution. Neighbor Solicitation and Advertisement help discover link-layer addresses; Router Advertisements support router and prefix discovery; Neighbor Unreachability Detection checks the ability to reach neighbors. Resolution commonly uses solicited-node multicast rather than an IPv4-style broadcast. IPv6 has no broadcast address type. A firewall policy that indiscriminately blocks ICMPv6 can therefore remove mechanisms the network needs for ordinary operation.

A cached neighbor entry can be usable while no longer recently confirmed. In the Neighbor Discovery state machine, stale does not simply mean broken. Traffic can trigger transitions through delay and probe states before a neighbor is declared unreachable. Conversely, an entry that once resolved does not certify that the remote service is healthy now. Inspect timestamps, interface state, link segmentation, and solicitation responses. Cached evidence is about a particular adjacency and time, not the complete end-to-end path.

While a neighbor is unresolved, an implementation may retain a bounded number of packets and eventually drop them if resolution fails. A stream of new unknown neighbors can therefore consume resources even when the route table is stable. Observe unresolved-neighbor counts and drops, not just interface bandwidth. A capture should answer whether solicitations leave the correct interface and whether replies return on the expected link. An empty application response alone cannot distinguish these possibilities.

In the queue lab, interpret arrival rate as work awaiting local resolution and service rate as completed resolution work under a deliberately simplified aggregate model. For one fresh run, press Reset, set both λ=3 and μ=2, predict growing backlog, then Play or Step. For a separate draining case, press Reset again, set both λ=2 and μ=3, predict the time to drain the initial four items, then Play or Step. Parameter edits pause but retain elapsed time and recompute from the initial four; they do not switch rates on accumulated backlog. The model begins with four work items and has no ARP packets, per-neighbor cache states, timers, or multicast. Its purpose is to make the resource consequence of stalled adjacency work visible, while packet captures remain the evidence for the actual protocol fault.