Data Center Interconnect (DCI)
Domain 4 tests how you connect two independent fabrics without turning them into one fragile failure domain.
The main options
- Routed DCI with EVPN Type 5: each DC advertises tenant prefixes; traffic is VXLAN inside each DC and plain IP (or IPsec) between them. Preferred when tenants do not need L2 stretch.
- L2 stretch with Type 2 across the DCI, typically via VXLAN stitching on border leaves, so a subnet spans sites. This brings MAC mobility, duplicate-MAC detection, and BUM across the WAN.
- EVPN-MPLS DCI with VXLAN-to-MPLS interworking for providers already running MPLS.
Design rules
Keep RDs unique per site, keep the anycast gateway MAC identical if a subnet is stretched, keep the L3 VNI and RT mapping consistent at both borders, and use symmetric IRB so the VRF context survives the hop between DCs. Filter Type 2 routes at the border when only routed prefixes should cross. If two sites use overlapping tenant subnets, you need separate VRFs plus NAT, or Type 5 will simply not resolve.
Worked example
DC1 leaf advertises 10.10.10.0/24 as a Type 5 route with RT target:65000:10010 and L3 VNI 5010. The DC1 border re-originates it over the WAN to the DC2 border, which imports the RT into the matching VRF and re-advertises into the DC2 fabric. The packet path is leaf1, border1, border2, leaf2: two VXLAN segments joined by IP routing, not one end-to-end tunnel.
Traps
Assuming a VXLAN tunnel extends end-to-end across the WAN; reusing an RD at both sites; stretching L2 and then creating an anycast gateway with mismatched virtual MACs, which breaks ARP; forgetting WAN MTU when the payload is already encapsulated.