VXLAN EVPN Là Gì? Khi Nào Nên Dùng Trong Thực Tế

VXLAN EVPN là gì và khi nào nên triển khai? Đội ngũ chia sẻ kinh nghiệm thực chiến về kiến trúc overlay, route type, lỗi MTU và cách chọn đúng thiết kế.

If you run a data center with a few dozen racks, you have probably faced this scenario: a VLAN needs to stretch into another rack, then another building, and every expansion adds another broadcast domain. We have hit this wall many times in our own lab and field work. So what is VXLAN EVPN, and when should you actually use it? The honest answer is that it exists to remove a structural bottleneck, not to add complexity for its own sake.

What VXLAN EVPN actually solves

VXLAN is an overlay technology. It wraps Layer 2 frames, and optionally Layer 3 packets, inside UDP and carries them across an ordinary IP network. Think of it as a tunnel that lets two distant segments behave as one. EVPN is the control plane. It uses BGP to advertise which MAC and IP addresses sit behind which VTEP, so every switch in the fabric knows where traffic should go.

The two work as a pair. VXLAN moves the data, while EVPN tells the devices where that data belongs. Without the control plane, you are back to guessing, and guessing is exactly what caused the problems described above.

In our experience there are three clear signals that point toward this design. The first is having many VLANs or tenants that must stay isolated from each other. The second is needing to stretch Layer 2 across several racks or data centers without dragging every broadcast along. The third is a genuine multi-tenancy requirement, where each customer or environment needs its own routing space. If none of these apply, you probably do not need the extra machinery.

The opposite case is worth stating plainly. A small network with one simple Layer 2 domain and a handful of switches gains nothing from EVPN. We once watched a team build a full EVPN fabric for an office of fifty people. Operations became noticeably harder than running a pair of switches with ordinary VLANs, and nobody gained anything for the trouble.

Why EVPN replaced flood-and-learn

Early VXLAN deployments relied on flood-and-learn to discover MAC addresses. When a switch did not know the destination MAC, it flooded the frame everywhere. That works on a small lab bench. On a large fabric, every new address triggers a burst of broadcast and unknown-unicast traffic across the underlay. That eats bandwidth, and it wears out both the devices and the engineers who troubleshoot them.

EVPN changes the approach by letting BGP share the information first. Leaf switches announce the MAC and IP addresses behind them using BGP route type 2, the MAC/IP advertisement. Route type 3, inclusive multicast, tells peers which VTEPs need to receive BUM traffic (broadcast, unknown-unicast and multicast) for a given VNI. Once that information sits in the BGP table, leaves no longer need to discover the network by flooding it.

The standards behind this are RFC 7432, which defines EVPN over BGP MPLS, and RFC 8365, which applies EVPN to NVO3 architectures such as VXLAN. When you need a reference for a technical claim, those two documents are the most reliable starting point.

Separating overlay from underlay

People new to this often blur these two layers, so we usually describe them as distinct roles. The underlay is the routed IP network beneath everything, typically running eBGP, OSPF or IS-IS. Its only job is to make sure every VTEP loopback can reach every other. It does not need to know anything about customer VLANs or MAC addresses, and that ignorance is a feature rather than a gap.

The overlay is the VXLAN tunnel running on top of that underlay, between VTEPs, together with the EVPN control plane riding on iBGP or eBGP. Tenants, VLANs and VRFs all live here. Keeping these roles separate in your documentation and your change process prevents a lot of confusion later, especially when a fault report lands and someone must decide which layer to inspect first. For teams building out the wider network side of this, the engineering resources at iServerHosting cover the infrastructure foundations that sit underneath a fabric like this.

MTU is the detail most often overlooked. VXLAN adds roughly 50 bytes of overhead to each packet. If the underlay stays at a standard 1500-byte MTU, full-sized frames from hosts get fragmented or dropped. The correct practice is to raise the MTU across the whole underlay path to 9216, or at minimum to a value large enough to carry the extra headers.

Core building blocks: VTEP, VNI and EVPN route types

A VTEP, or VXLAN Tunnel Endpoint, is usually the leaf switch itself. It encapsulates and decapsulates VXLAN traffic as frames enter and leave the fabric. A VNI, the VXLAN Network Identifier, names a Layer 2 segment in the overlay. It is usually mapped to a VLAN on the customer side and to a VRF on the Layer 3 side.

You will meet three EVPN route types most often. Type 2 carries host MAC and IP addresses for bridging and routing. Type 3 advertises which VTEPs participate in a VNI so BUM traffic can be replicated correctly. Type 5 carries external IP prefixes, commonly used for routing between subnets and outside the fabric. Knowing which route type you are looking at makes the BGP table far less intimidating.

Integrated Routing and Bridging forces a choice between symmetric and asymmetric models. Asymmetric is simpler conceptually but requires every leaf to carry every VNI for every subnet, which inflates state. Symmetric is more flexible and scales better, but it depends on hardware capability and software release. Verify your vendor’s documentation before committing to either.

A minimal leaf configuration

The example below is a stripped-down NX-OS style snippet, included to illustrate the configuration blocks rather than to serve as a deployable template. Syntax varies by vendor and release, so always check the official reference for your platform.

feature nv overlay
feature vn-segment-vlan-based
feature bgp
feature interface-vlan

! Loopback used as the VTEP source address
interface loopback0
  ip address 10.0.0.11/32

! NVE is the VXLAN tunnel interface
interface nve1
  no shutdown
  source-interface loopback0
  host-reachability protocol bgp
  member vni 10100
    suppress-arp
    ingress-replication protocol bgp

! Map VLAN to VNI
vlan 100
  vn-segment 10100

! BGP session carrying the EVPN address family
router bgp 65001
  neighbor 10.0.0.1
    remote-as 65001
    update-source loopback0
    address-family l2vpn evpn
      send-community extended

Loopback0 is the VTEP’s identity and must be reachable from every other leaf. The NVE interface performs the encapsulation. The suppress-arp option reduces ARP broadcasts by letting the leaf answer locally once it knows the MAC and IP binding. Finally, the l2vpn evpn address family is where BGP carries the EVPN routes, so forgetting it is a classic reason a fabric looks configured but learns nothing.

Design mistakes we keep seeing

The most common error is forgetting the MTU on the underlay. The network appears healthy under small pings, then drops packets during large file transfers. Operators lose hours chasing what looks like an application problem, when the real fault sits in a single interface setting.

A second error is duplicate anycast gateway addresses or mismatched route targets. When two leaves advertise the same MAC and IP pair without consistent route target configuration, forwarding tables oscillate and traffic bounces between paths. Route distinguishers should also be unique per VTEP, and it is worth checking them early.

A third is inconsistent handling of ARP suppression and BUM replication across leaves. If one leaf uses ingress replication while another expects multicast, BUM traffic silently disappears for some hosts. Nothing logs an error, which makes it especially frustrating to find.

The last is placing route reflectors poorly or running only one. A single RR failure can take down the entire EVPN control plane. We recommend at least two, positioned so they do not depend on any one leaf for their own reachability.

Platform differences and a pre-deployment checklist

Juniper and Cisco both implement EVPN, but configuration structure and command names differ. The table below summarises the main deployment steps without quoting performance figures, since those depend heavily on hardware and release.

Step Juniper (Junos) Cisco (NX-OS)
Enable the overlay EVPN and VXLAN under the routing instance feature nv overlay
Create the VTEP VTEP interface with a source loopback nve1 interface with source-interface
Map VNIs vni-options under the VLAN member vni under nve1
BGP EVPN session family evpn signaling address-family l2vpn evpn

Before moving to production, we walk through a short checklist. Verify MTU end to end, and confirm every VTEP loopback is reachable. Check that RDs and route targets are unique, and make sure the BGP EVPN sessions are fully established. Finally, compare the MAC table against expectations to confirm addresses are learned where they should be. Teams with many devices benefit from validating configuration automatically before it reaches the leaves. Our write-up on CI/CD for Juniper & Cisco Hybrid Fabrics covers how to test configuration changes before they touch production switches, a habit that pays for itself quickly.

VXLAN EVPN compared with other options

Compared with traditional Layer 2 extension over MPLS, VXLAN EVPN is simpler to run on a plain IP underlay and does not demand deep MPLS expertise. MPLS has a longer history in carrier networks and may suit you if that infrastructure already exists and your team knows it well. Proprietary vendor fabrics integrate tightly and deploy quickly at first, but they tie you to a single supplier for years, a cost that rarely appears on the first quote.

When we help teams decide, we ask three questions. How large will the network become, and does it need to grow quickly? Which technologies does the operations team already know? Is multi-vendor interoperability required? If the answers point to a large network, a team comfortable with BGP, and a mixed-vendor environment, EVPN is usually a sound choice. If they point the other way, a simpler design is often the wiser investment. For broader context on how choices like this fit into wider technology planning, GoAcademy offers useful background that complements the engineering side of a rollout.

Frequently asked questions

Is VXLAN EVPN worth it if I only have a few VLANs?

Usually not. A handful of VLANs in a simple Layer 2 domain works fine with conventional configuration. The value appears as scale and tenant count grow, because the operational cost of flood-and-learn becomes significant at that point, and the control plane starts paying back its complexity.

Does every switch support EVPN?

No. Support depends on the ASIC and the software version. Features such as symmetric IRB appear only on certain platforms, and some models support the control plane but not every feature you might want. Check the vendor compatibility matrix before buying or deploying, and ask for it in writing.

Can I use eBGP for the underlay and iBGP for the EVPN overlay?

Yes, and this is a fairly common design. eBGP in the underlay gives each leaf its own AS and scales well, while iBGP in the overlay simplifies route reflector management. You must still configure next-hop behaviour and route targets carefully to avoid routing surprises, particularly when both layers carry BGP sessions.

What is the most common mistake beginners make?

From what we have seen, it is failing to raise the MTU on the underlay. The fault is hard to spot because the network still passes small packets, and it often surfaces only once real workloads start moving. Our practical advice is to check MTU first in every deployment, before touching anything else.

If you are preparing a VXLAN EVPN rollout, start by drawing the underlay and overlay separately, and set the MTU correctly from day one. Validate on a single pair of leaves before expanding to the full fabric. Taking small steps will save you a great many late nights, and it gives your team the confidence to grow the design as the real requirements become clear.

Related topics