Layer 2 and Layer 3 Underlay and Overlay Networking
Domain 3 covers the two layers beneath every tenant: the IP underlay that makes VTEP loopbacks reachable with ECMP, and the BGP EVPN overlay that distributes MAC/IP and prefix reachability.
Underlay rules
- eBGP leaf-to-spine, one session per link,
family inet unicast, loopback /32 advertised through an export policy. multipath multiple-asso equal-cost paths install; confirm withshow route <vtep>showing two next hops.- BFD on every fabric link so a dead link is removed in milliseconds rather than at BGP hold time.
- Jumbo MTU 9000 on fabric links, at least 1550 on host-facing VTEP ports.
Overlay rules
iBGP with the spines as route reflectors is the common design: every leaf must receive every other leaf's Type 2 and Type 3 routes, so the RR group needs clients and a cluster ID. family evpn signaling must be negotiated. MAC learning is control-plane; the data plane is not how MACs are discovered. The L3 overlay uses Type 5 routes in a VRF with its own L3 VNI and vrf-target.
Worked example
Leaf1's loopback is 10.0.0.1/32, Leaf2's is 10.0.0.2/32. Leaf1 advertises its /32 to spine1 and spine2; each spine selects it best and re-advertises; Leaf2 receives two paths and, with multipath multiple-as, installs both as the VXLAN outer destination. Remove the export policy on one spine and the path count drops to one - the fastest proof that the underlay is symmetric.
Traps
Full-mesh iBGP because the reflector was skipped; assuming show bgp summary up means forwarding works (check show route forwarding-table for the VTEP prefix); hash asymmetry pinning one spine link near saturation; MTU left at 1500 on a transit link so large frames die silently in the middle of the fabric.