Data Center Routing and IP Fabric Design
Domain 3 - Data Center Routing and IP Fabric Design
This domain tests the underlay and overlay routing design for a scalable IP fabric: addressing, BGP and IGP choice, ECMP, route reflection, and failure behavior. Juniper designs commonly use a Layer 3 leaf-spine underlay with BGP and an EVPN overlay.
Core rules
- Use /31 or /127 point-to-point links and a unique loopback per device; loopbacks are the BGP router ID and the stable endpoint for overlay peering.
- Underlay options: eBGP with unique leaf ASNs, OSPF, or IS-IS. If eBGP spines must re-advertise leaf loopbacks to other leaves, enable advertise-external on the spines or use an IGP. Do not build a full iBGP mesh in the underlay.
- Overlay: iBGP EVPN between leaf loopbacks, with spines as route reflectors. Two RRs minimum for redundancy.
- ECMP requires equal-cost paths and BGP multipath; Junos installs multiple next hops when multipath is configured and the paths are equal.
- Summarize only at safe boundaries; a summary that blackholes a failed component is worse than no summary.
Worked example
A fabric has four spines and 32 leaves. Each leaf uses eBGP AS 65001–65032; spines use AS 65100. Each leaf advertises its loopback /32. On the spines, configure advertise-external so a route learned from leaf 1 is sent to leaf 2. Each leaf sees four equal-cost paths to every remote loopback and installs them with multipath. The EVPN overlay runs iBGP between leaf loopbacks, with two spines acting as RRs. If one spine fails, BGP withdraws that path and ECMP converges to three paths; the overlay session to the remaining RR stays up.
Traps
- eBGP split horizon: a spine does not re-advertise eBGP routes without advertise-external.
- Forgetting next-hop resolution or loopback reachability.
- Placing both RRs in the same failure domain.
- Assuming ECMP works without multipath configuration.
- Using a single ASN everywhere and creating AS-path loop drops.