network · deep · draft

EVPN-VXLAN leaf-spine overlay

An L3 Clos underlay with EVPN as the BGP control plane and VXLAN as the NVO data plane — multi-tenant L2/L3 segments, distributed IRB, and scalable MAC/IP learning without STP-stretched VLANs.

In one sentence. Pure IP Clos moves packets well but does not by itself provide multi-tenant L2 segments, flexible subnet placement, or scalable MAC learning across a large fabric.

Why it exists

Classic stretched VLAN + STP and data-plane-only VXLAN learning hit multihoming, load-balancing, and provisioning limits. EVPN (RFC 7432 lineage) plus EVPN-NVO (RFC 8365) replace flood-and-learn overlays with BGP-advertised MAC/IP state over NVO tunnels — commonly VXLAN (RFC 7348). This page teaches fabric placement and IRB modes; for EVPN route types, NLRI formats, and VXLAN header fields see Protocol Atlas at https://protocols.alexflux.com — do not treat Architecture Atlas as a protocol encyclopedia.

Visual walkthrough

Walkthrough

Intra-subnet L2 over VXLAN

Same-subnet hosts on different leaves bridge on an L2 VNI. Spines carry underlay IP between VTEP loopbacks; they need not terminate EVPN in an ERB design.

Step 1 / 3
Spine 1 · underlayunderlay transitSpine 2 · underlayLeaf 1 · VTEPVTEP / NVELeaf 2 · VTEPLeaf 3 · VTEPHost A · subnet 10Host B · subnet 20Host C · subnet 10

Leaves are VTEPs (NVEs): they encapsulate/decapsulate VXLAN and speak EVPN. Spines provide underlay reachability between VTEP loopbacks — lean IP in many ERB designs.

Walkthrough

Asymmetric IRB (inter-subnet)

Ingress leaf routes into the destination L2 VNI; the egress leaf only bridges. RFC 9135 warns that participating PEs need remote host ARP/MAC state and MAC-VRF/IRB for all relevant subnets.

Step 1 / 3
Spine 1 · underlaySpine 2 · underlayLeaf 1 · VTEPLeaf 2 · VTEPLeaf 3 · VTEPHost A · subnet 10subnet 10Host B · subnet 20subnet 20Host C · subnet 10

Host A (subnet 10) reaches Host B (subnet 20). Asymmetric IRB: the ingress PE routes the packet into the destination subnet’s L2 VNI so the egress PE can bridge only.

Walkthrough

Symmetric IRB (inter-subnet)

Ingress routes into a tenant L3 VNI; egress routes then bridges. Uses Router’s MAC extended community and a dedicated L3 VNI per tenant VRF (RFC 9135 / Juniper Type-2 symmetric).

Step 1 / 3
Spine 1 · underlaySpine 2 · underlayLeaf 1 · VTEPIRB + L3 VNILeaf 2 · VTEPIRB + L3 VNILeaf 3 · VTEPHost A · subnet 10Host B · subnet 20Host C · subnet 10

Same Host A → Host B intent, symmetric IRB: both ingress and egress PEs perform routing. The tunnel is IP-VRF to IP-VRF via a tenant L3 VNI.

Control vs data plane

Control plane

MP-BGP EVPN address family between VTEPs (often via spine route-reflectors). Underlay IGP/BGP provides loopback reachability so outer VXLAN packets can ECMP across spines. Arista AVD documents underlay/overlay pairings (eBGP/eBGP, OSPF/iBGP, ISIS/eBGP, and others). Leaf VTEPs advertise Type-2 MAC/IP, Type-3 IMET, Type-5 prefixes, and related routes — protocol message anatomy belongs on Protocol Atlas.

Data plane

VTEPs encapsulate Ethernet into VXLAN (UDP/IP) toward remote VTEP IPs. L2 VNIs bridge MAC-VRFs; L3 VNIs carry tenant IP-VRF traffic in symmetric IRB. In Juniper ERB examples, spines stay lean IP while leaves terminate L2 + L3 VXLAN (IRB); spine-routed / centralized IRB is a different placement with different scale trade-offs and is not asserted as a universal default.

Request / packet path

North–south

Endpoint → leaf VTEP → underlay (often via spine) → border/gateway leaf → external networks or DCI. Border participation in EVPN vs pure IP handoff is design-specific.

East–west

Intra-subnet: bridged on an L2 VNI end-to-end between VTEPs. Inter-subnet asymmetric IRB: ingress routes into the destination L2 VNI; egress bridges only. Inter-subnet symmetric IRB: ingress routes into a tenant L3 VNI; egress routes then bridges (RFC 9135). Walkthroughs compare both modes side by side; a dedicated /compare page can come later.

Scaling & math

Fabric bandwidth oversubscription remains an underlay Clos property — the overlay does not remove it (see spine-leaf / Clos). Asymmetric IRB scale caution from RFC 9135: remote host ARP/MAC state and MAC-VRF/IRB for all relevant subnets on participating PEs. Symmetric IRB uses Router’s MAC extended community and a dedicated L3 VNI per tenant VRF. Platform max VNIs / VRFs / MACs are datasheet-specific (unknown as a universal table here).

When it breaks

  • Endpoints on a single-homed leaf lose overlay connectivity; dual-homed hosts may survive via ESI multihoming.

    Cause. VTEP leaf failure or isolation.

    Mitigation. Dual-home servers / ESI multihoming where availability requires it; verify remaining VTEP EVPN state.

  • New MAC/IP learning stalls; existing data-plane tunnels may continue until state times out.

    Cause. EVPN route-reflector / controller set unavailable.

    Mitigation. Redundant RRs; monitor EVPN session health separately from underlay forwarding.

  • Overlay tunnels fail between partitioned VTEPs; intra-partition traffic may continue.

    Cause. Underlay IP reachability loss between VTEP loopbacks.

    Mitigation. Treat underlay Clos health as a hard dependency; repair leaf–spine ECMP and loopback advertisement first.

  • Cross-tenant connectivity or unexpected broadcast domain merge.

    Cause. VNI mis-mapping or route-target leak between tenants.

    Mitigation. Audit RT import/export and VNI assignments; treat as configuration blast radius.

  • Flooding storms or excess BUM within an L2 VNI.

    Cause. BUM handling mis-design (ingress replication / underlay multicast / IMET issues).

    Mitigation. Review Type-3 IMET and BUM design against RFC 8365 / platform guidance; limit L2 stretch scope.

Misconceptions

  • “Asymmetric and symmetric IRB are interchangeable defaults.” — RFC 9135 defines different tunnel semantics and scale properties; asymmetric holds more remote host/IRB state on PEs.
  • “VXLAN alone provides multi-tenant fabric control.” — RFC 7348 is the data-plane encapsulation; scalable MAC/IP learning and multihoming are EVPN control-plane concerns (RFC 7432 / 8365).
  • “Spines must always run EVPN.” — Juniper ERB examples keep spines as lean underlay IP with leaf VTEPs; other designs place IRB on spines — placement is design-specific, not universal.

Reference expression

Primary teaching sources: RFC 7348 (VXLAN), RFC 7432 (EVPN), RFC 8365 (EVPN NVO), RFC 9135 (IRB), Juniper ERB anycast / symmetric IRB docs, and Arista AVD eos_designs EVPN underlay/overlay pairings. Short architecture excerpts only — follow the RFCs and vendor guides for configuration. Protocol-level deep dives: https://protocols.alexflux.com. review: draft.

Standards & sources

Known unknowns

  • Exact “default” IRB mode and spine EVPN participation are vendor/product-specific — not asserted as universal.
  • Platform maximum VNIs, VRFs, and MAC/ARP scale are datasheet-specific and not tabulated here.

Check yourself

  1. What problem does EVPN-VXLAN primarily add on top of a pure IP Clos?

    • It removes the need for underlay ECMP
    • Multi-tenant L2/L3 segments and scalable MAC/IP learning over NVO tunnels
    • It replaces BGP with STP
    • It guarantees 1:1 oversubscription

    Answer: Multi-tenant L2/L3 segments and scalable MAC/IP learning over NVO tunnels. IP Clos is underlay reachability; EVPN-VXLAN adds overlay L2/L3 and control-plane learning.

  2. In asymmetric IRB inter-subnet forwarding, what does the egress PE typically do?

    • Route into a tenant L3 VNI
    • Bridge only after ingress routed into the destination L2 VNI
    • Drop all inter-subnet traffic
    • Terminate VXLAN on the spine only

    Answer: Bridge only after ingress routed into the destination L2 VNI. RFC 9135: ingress routes to dest L2 VNI; egress bridges.

  3. In symmetric IRB, inter-subnet traffic between PEs uses which tunnel context?

    • Only the destination L2 VNI with no routing on egress
    • Tenant L3 VNI (IP-VRF to IP-VRF); both PEs route
    • Raw Ethernet without VXLAN
    • STP blocked links only

    Answer: Tenant L3 VNI (IP-VRF to IP-VRF); both PEs route. Symmetric IRB uses L3 VNI between IP-VRFs; egress also routes.

  4. RFC 9135’s scale caution for asymmetric IRB highlights…

    • That spines must host all IRBs
    • Remote host ARP/MAC state and IRB/MAC-VRF for all relevant subnets on participating PEs
    • That VXLAN UDP ports are optional
    • That underlay ECMP is forbidden

    Answer: Remote host ARP/MAC state and IRB/MAC-VRF for all relevant subnets on participating PEs. Asymmetric designs push more host/IRB state onto PEs.

  5. In many ERB designs, where does IRB typically sit?

    • Only on WAN firewalls
    • On leaf VTEPs (distributed), with lean IP spines
    • Only in the public cloud control plane
    • On every host NIC mandatorily

    Answer: On leaf VTEPs (distributed), with lean IP spines. Juniper ERB examples: leaves do L2+L3 VXLAN; spines lean underlay.

  6. Does the overlay remove Clos oversubscription?

    • Yes — VXLAN is always 1:1
    • No — bandwidth oversubscription remains an underlay Clos planning property
    • Yes — if you use asymmetric IRB
    • Only when Type-5 routes are present

    Answer: No — bandwidth oversubscription remains an underlay Clos planning property. Overlay semantics ≠ fabric capacity ratios.

  7. Where should you go for EVPN route-type and VXLAN header deep dives?

    • Invent packet captures on this page
    • Protocol Atlas (https://protocols.alexflux.com) for protocol depth; this page for fabric architecture
    • Anonymous forums only
    • Unlabeled AI summaries

    Answer: Protocol Atlas (https://protocols.alexflux.com) for protocol depth; this page for fabric architecture. SPEC Phase 4: cross-link Protocol Atlas; do not copy protocol pages here.

  8. What provides underlay reachability between VTEP loopbacks?

    • STP root priority alone
    • Underlay IGP/BGP (Clos ECMP) independent of EVPN MAC learning
    • Only Type-2 routes without IP underlay
    • DNS anycast

    Answer: Underlay IGP/BGP (Clos ECMP) independent of EVPN MAC learning. Overlay tunnels require underlay IP between VTEPs.