Data-centre networking
Design a leaf-spine (Clos) fabric with an IPv6 underlay. Run BGP over IPv6 link-local ("unnumbered") sessions carrying IPv4 and IPv6 (RFC 8950, RFC 7404). Address Neighbor Discovery scaling in large L2 domains (RFC 7342).
Take the quiz · pass mark 70% · 9 questions
Explanation
Modern data centres are routed leaf-spine (Clos) fabrics: every leaf connects to every spine, there is no spanning tree, and BGP runs on every link. IPv6 simplifies this.
The underlay and unnumbered BGP
Fabric links do not need globally unique addresses. RFC 7404 describes using only link-local addressing inside a network: every fabric interface already has an fe80::/10 link-local address, so you can run BGP "unnumbered" over link-local, with no per-link addressing plan to design, maintain or exhaust. The neighbor is discovered via RA on the link and BGP peers over the link-local next hop.
A single unnumbered BGP session can carry both families. RFC 8950 updates the mechanism for advertising IPv4 NLRI with an IPv6 next hop, so one IPv6 (link-local) BGP session distributes IPv4 and IPv6 routes together — you do not need a parallel IPv4 peering. This is the backbone of dual-stack fabrics: unnumbered IPv6 sessions, carrying everything.
Leaf-spine mechanics
With s spines and l leaves, each leaf peers with every spine, so there are l * s leaf-spine links/sessions, and a packet from one leaf to another has s equal-cost paths (one per spine) — ECMP spreads load and survives a spine loss. Keep the fabric a clean two-tier Clos and let BGP + ECMP do the work.
ND scaling in L2 domains (RFC 7342)
Where you still have large L2 tenant subnets (a big /64 with many hosts, or overlay segments), Neighbor Discovery and its IPv4 cousin ARP can stress the first-hop gateway's neighbor cache and generate multicast. RFC 7342 collects practices for scaling ARP and ND in data centres: ND/ARP suppression and proxy at the leaf (answer locally from a learned mapping instead of flooding), controlling unknown-unicast/multicast, and sizing neighbor tables. Combined with first-hop security (L5.03) and neighbor-cache-exhaustion limits (RFC 6583), this keeps large L2 domains stable. Prefer routing to the leaf (small subnets) where you can, and suppress/scale ND where you cannot.
ictNetworks scenario
ictNetworks' data centre is a two-tier Clos with unnumbered BGP over link-local (RFC 7404) on every fabric link, each session carrying IPv4 and IPv6 with an IPv6 next hop (RFC 8950), so there is no underlay addressing plan and no parallel IPv4 peering. Tenant subnets are routed to the leaf where possible; the remaining large L2 segments use leaf ND suppression and proxy (RFC 7342) plus neighbor-cache limits, and the GUA tenant space comes from its documentation allocation (2001:db8::/32 in the lab).
Drawing diagram…
Worked examples
Example 1: Sessions in a 2-spine, 4-leaf fabric
With 2 spines and 4 leaves, each leaf peers with each spine, giving 4 * 2 = 8 leaf-spine BGP sessions. Adding a spine adds one session per leaf.
Example 2: The unnumbered next hop
Unnumbered BGP peers over the interface's link-local address, which is drawn from fe80::/10. No globally unique per-link addressing is needed (RFC 7404).
Example 3: Equal-cost paths across the fabric
From one leaf to another in a 2-spine fabric there are 2 equal-cost paths (one via each spine). ECMP uses both and tolerates the loss of one spine, which is why spine count sets both capacity and resilience.
Lab
Goal: Build a 2-spine, 4-leaf fabric with unnumbered BGP and dual-stack tenant subnets.
You need: a containerlab topology with FRR on 2 spines and 4 leaves, dual-stack tenant hosts behind the leaves, and vtysh/tcpdump. (FRR syntax varies by version — check the docs.)
- Bring up the fabric links and confirm each interface has an
fe80::/10link-local address. - Configure unnumbered BGP over link-local on every leaf-spine link (interface peering), and confirm 8 sessions come up.
- Enable IPv4 NLRI with an IPv6 next hop (RFC 8950) so each session carries both families; advertise dual-stack tenant subnets.
- Verify end-to-end IPv4 and IPv6 between tenants on different leaves, and confirm ECMP uses both spines (and survives a spine down).
- On a large L2 tenant segment, enable ND suppression/proxy (RFC 7342) and show the reduction in flooded ND.
What to submit: the FRR configuration, the 8-session BGP state, evidence of dual-stack reachability and ECMP, and the ND-suppression result.
Summary
- Build routed leaf-spine fabrics; run BGP on every link.
- Use unnumbered BGP over link-local (RFC 7404) so there is no underlay addressing plan.
- One IPv6 session carries IPv4 and IPv6 via an IPv6 next hop (RFC 8950) — no parallel IPv4 peering.
- Scale ND in large L2 domains with suppression/proxy and neighbor-cache limits (RFC 7342); prefer routing to the leaf.
- Next lesson (L7.06): cloud and containers.
Quiz imported · 9 questions · pass mark 70%. Lab questions are not auto-marked.