VXLAN Mastery · Lesson 02 / 08

How VTEPs find each other and flood traffic

Lesson 1 showed how one frame crosses the fabric. But what about a broadcast, or a MAC nobody has learned yet? In this lesson you will see exactly how VXLAN copies traffic to many switches, what that costs, how loops are avoided, and why engineers invented EVPN.

📘 Builds on Lesson 1⏱ ~35 minutes🖱 7 interactive demos🧠 Quiz and checklist
0 of 7 chapters completeProgress is saved in this browser
1Starter⏱ 3 min

Quick recap from Lesson 1

Try to answer each question in your head first, then click to reveal. If one feels fuzzy, revisit that chapter of Lesson 1 before continuing.

What does a VTEP do?

It wraps local Ethernet frames in VXLAN/UDP/IP and sends them to a remote VTEP, and unwraps frames it receives. Hosts never see VXLAN.

What is a VNI and how big is it?

A 24-bit segment ID in the VXLAN header: 16,777,216 values. Example: VLAN 100 maps to VNI 10100.

How many bytes does VXLAN add, and why does it matter?

With an IPv4 underlay and no outer VLAN tag, the typical on-wire encapsulation adds 50 bytes: 14 outer Ethernet + 20 outer IPv4 + 8 UDP + 8 VXLAN (excluding FCS). IP MTU excludes the outer Ethernet header. For a 1,500-byte inner IP packet: 1,500 + 14 inner Ethernet + 8 VXLAN + 8 UDP + 20 outer IPv4 = 1,550-byte underlay IP packet. Its outer Ethernet frame is 1,564 bytes before FCS. Additional tags or options change the on-wire size.

What does the spine read to forward a VXLAN packet?

Only the outer IP header (destination = remote VTEP). It never looks at the inner MAC.

Today’s question: in Lesson 1, Leaf 1 “already knew” where Host B lived. What does a leaf do when it doesn’t know, or when the frame is a broadcast? The answer is called BUM handling.
2Starter⏱ 5 min

What is BUM traffic?

Broadcast, Unknown unicast and Multicast. In our plain VXLAN flood-and-learn example, these frames are sent through the VNI’s configured replication mechanism. The exact recipients depend on VNI membership, snooping and forwarding policy. Unknown-unicast flooding can also be disabled.

📣 Broadcast

Destination FF:FF:FF:FF:FF:FF. “Everyone.” Example: an ARP request.

❓ Unknown unicast

A normal unicast destination MAC with no forwarding entry. Our flood-and-learn example floods it; production fabrics may restrict or disable unknown-unicast flooding.

👥 Multicast

For example, IPv4 multicast MACs use 01:00:5E and IPv6 multicast MACs use 33:33 prefixes. Replication depends on the multicast and snooping configuration.

Try it: known unicast or BUM?

Score 0 / 0

Assume unknown-unicast flooding and broadcast/multicast replication are enabled, with no multicast snooping restrictions. Leaf 1’s MAC table (VNI 10100): 00:BB:BB:BB:BB:02 → VTEP 10.0.0.2 and 00:CC:CC:CC:CC:03 → VTEP 10.0.0.3. Decide how Leaf 1 treats each frame.

Pick an answer.
Plain English: known unicast has a specific destination VTEP. BUM needs a replication decision. In this simplified lab we flood to all remote members; real designs may prune multicast and restrict unknown unicast.
3Intermediate⏱ 6 min

Two ways to deliver BUM: ingress replication vs multicast

VXLAN needs a way to copy one BUM frame to many VTEPs. There are two classic methods. Play with the slider to see how they scale.

🧾 Ingress (head-end) replication

The source leaf makes a separate unicast VXLAN copy for every other VTEP in the VNI. The underlay needs no special features, only IP routing.

📡 Underlay multicast

The VNI is mapped to a multicast group. The source sends one packet; the underlay (PIM) copies it where paths split. Needs multicast routing in the underlay.

Compare the two methods

Copies leaving the source leaf
Source uplink load per 1500 B frame
What the underlay must support
Ingress replicationUnderlay multicast
Underlay needsPlain IP unicast routingPIM and multicast routing (RPs, groups)
Copies from sourceN − 11
Operational effortLow (but VTEP lists needed)Higher (multicast design and troubleshooting)
Common inEVPN fabrics (most popular today)Large flood-and-learn fabrics, some legacy designs
4Intermediate⏱ 4 min

“How does a leaf know who the other VTEPs are?”

With ingress replication a leaf needs to know the remote VTEPs that participate in a VNI. In this lesson’s static flood-list model, an operator configures each peer manually. Other control mechanisms, including EVPN, can distribute this information dynamically.

The cost of manual flood lists

Move the slider
Peer entries to configure (per VNI)
Devices to touch when adding ONE new leaf
…and for 100 VNIs
Common mistake: omitting a remote VTEP from a static replication list. BUM reachability can become asymmetric, and ARP resolution may fail across certain pairs. Exact symptoms depend on peer lists and any separately learned unicast paths.
What we want: an automated way to learn which remote VTEPs participate in a segment. EVPN Type 3 routes can provide the information used to construct ingress-replication lists. Preview in chapter 7.
5Intermediate⏱ 5 min

Split horizon: the rule that prevents loops

In this basic ingress-replication model, a receiving VTEP must not re-replicate a tunnel-received BUM frame back into the same VXLAN overlay. It may forward it to eligible local ports. This is a split-horizon rule. More advanced multihoming designs add other loop-prevention behaviors.

One broadcast, three leaves

A host on Leaf 1 sends one broadcast. Count the VXLAN copies on the underlay after each round.

Why this is dangerous: traditional Ethernet frames have no hop-count field equivalent to IP TTL. Incorrect tunnel re-flooding can cause duplicate traffic and storms. This simulator uses a simplified doubling model; an actual loop’s growth depends on topology and forwarding behavior.
Pro detail: what about multihomed hosts?

When a host connects to two leaves (e.g. EVPN multihoming or MLAG), extra rules (like designated forwarder election and local-bias filtering) are needed so only one leaf delivers BUM to the host. We cover this in later lessons.

6Mastery⏱ 4 min

Why flooding does not scale

In this deliberately worst-case flood-and-learn example, assume every host generates BUM traffic, all leaves participate in one VNI, and ingress replication sends it to every remote VTEP. Estimate the resulting load; real networks use pruning and other controls to reduce it.

BUM load calculator

Change the numbers
BUM frames per second (new)
Underlay packets per second (ingress repl.)
Approx. BUM deliveries per receiving host/s
Plain English: large shared Layer 2 domains can produce many unnecessary deliveries. Actual per-host traffic depends on where hosts sit, pruning, and whether the sender receives its own broadcast. This calculator is an intentionally simplified upper-bound illustration, not a capacity forecast.
7Mastery⏱ 6 min

Preview: how EVPN improves discovery and reduces flooding

EVPN uses MP-BGP as a control plane so leaves can advertise reachability. It can reduce—not always eliminate—data-plane flooding. In this lab we assume compatible EVPN configuration, Type 2 routes carrying valid IP-to-MAC bindings, and proxy ARP/ARP suppression enabled. Step through to distinguish control-plane BGP routes from Ethernet frames and VXLAN data traffic.

Type 3: Inclusive Multicast Ethernet Tag (IMET)

“My EVPN segment participates in BUM replication through this tunnel information.” With appropriate import policy and VXLAN mapping, receiving VTEPs use Type 3 information to build a flood/replication list.

Type 2: MAC/IP Advertisement

“MAC 00:BB…:02 is reachable through Leaf 2.” A Type 2 route may also include IP 192.168.10.12. When a valid IP-to-MAC binding is available and proxy ARP is configured, the leaf can answer matching ARP requests locally.

Watch Leaf 1 learn with EVPN (configured proxy ARP)

Simulation assumptions: valid route import/segment membership, working MP-BGP EVPN sessions, a learned MAC+IP Type 2 binding for Host B and proxy ARP enabled. The final step deliberately resets the state to a different, plain-VXLAN example.
BGP EVPN route (control plane)Ethernet frame on a host linkUnicast VXLAN / dataBUM flood copy
Routed underlay + BGP EVPN control plane Host links: ordinary Ethernet (VLAN 100) SpineBGP route reflector Leaf 1VTEP 10.0.0.1 Leaf 2VTEP 10.0.0.2 Leaf 3VTEP 10.0.0.3 Host A00:AA…:01 Host B00:BB…:02 Host C00:CC…:03

Packets in this step

ProblemFlood-and-learnWith EVPN
Find remote VTEPsStatic lists in this labType 3 information builds the replication list when configured
Learn remote MACsFrom data-plane learningType 2 routes can advertise MAC reachability before the first remote data frame
ARPFlooded when requiredProxy ARP can reply locally for known valid bindings; unknown requests may flood
Unknown unicastFlooded if allowedReduced when MAC routes are learned; otherwise policy determines flood/drop
What to remember: Type 3 gives reachability/replication information; Type 2 advertises a MAC and optionally an IP; proxy ARP only answers when enabled and a usable binding is present. Next: MP-BGP, route distinguishers, route targets and the EVPN route types.
✓Check yourself⏱ 5 min

Test what you learned

Eight questions with an explanation for every answer.

Mastery checklist

Tick each one when you can explain it in your own words.

Cheat sheet

  • BUM = Broadcast, Unknown unicast, Multicast
  • Ingress replication: N − 1 copies from the source leaf
  • Underlay multicast: 1 copy, needs PIM
  • Split horizon: tunnel-received BUM goes to local ports only
  • Ethernet has no TTL, so loops become storms
  • EVPN Type 3 = inclusive multicast membership and replication information
  • EVPN Type 2 = MAC advertisement, with optional IP binding
  • Proxy ARP = optional local reply for usable known bindings; unknown requests may flood

Mistakes to avoid

  • Missing a VTEP in a static flood list
  • Disabling split horizon
  • Mixing ingress and multicast modes between leaves
  • Assuming VXLAN or EVPN eliminates all flooding
  • Ignoring ARP load in very large VNIs
Up next, Lesson 03: EVPN control plane. MP-BGP, route distinguishers, route targets and the route types in detail.
Sources

Learn from the standards

See RFC 7348 (VXLAN), RFC 7432 (EVPN), RFC 8365 (EVPN/VXLAN), and RFC 9161 (Proxy ARP/ND).