Underlay multicast: the network makes the copies
With underlay multicast, the source leaf sends one VXLAN packet whose outer destination is a multicast group address. The underlay routers (spines) already know which leaves are interested in that group, so they replicate the packet at the branching points of a tree.
📻 Plain-English analogy
A radio station transmits once. Everyone tuned to that channel hears it. The station does not call each listener. Leaves “tune in” to a group (a channel) by joining it. The spine is the transmitter tower that repeats the signal to everyone tuned in.
🧩 In VXLAN terms
Each VNI is mapped to a group such as 239.1.1.100. Every VTEP that has the VNI joins that group. A BUM frame becomes one packet to 239.1.1.100 and the tree delivers it to all members.
The shape of the traffic
1 packet leaves the source, the spine copies itMapping VNIs to multicast groups
In this example each VNI is configured with an administratively scoped multicast group in 239.0.0.0/8; other address choices depend on network design and reserved-address rules. For each shared VNI, its group mapping must be consistent across participating VTEPs. You can give each VNI its own group, or let several VNIs share a group.
Try the mapping
See who receives the BUM frameVNI membership: 10100 = Leaf 1, 2, 3 · 20200 = Leaf 1, 3 · 30300 = Leaf 3, 4 · 40400 = Leaf 2, 4
One group per VNI
More selective delivery: only VTEPs joined to that group should receive its VXLAN BUM traffic when multicast forwarding state and filtering are correct. Costs more multicast state (one mroute entry per group) on the spines.
Shared groups
Saves spine state, but a leaf that joined the group for another VNI still receives the packet. It receives unnecessary multicast traffic and discards packets for unsupported VNIs (potentially in hardware). This can consume unnecessary bandwidth; the receiving device may discard the packet in hardware, so CPU cost depends on the implementation.
Before any frame moves: joining the group
Multicast delivery requires suitable forwarding state. A VTEP can signal group membership using IGMP toward an upstream multicast router; a PIM-capable leaf/router may instead send PIM joins. This specific walkthrough assumes the leaf VTEPs run PIM-SM and the spine is the Rendezvous Point (RP). Press Next step to see that particular design.
Build the multicast tree for VNI 10100
Packets in this step
Packet-by-packet: the same ARP broadcast, with multicast
The tree from chapter 3 is already built. Host A ARPs for Host B. Compare with Lesson 2A: how many packets leave Leaf 1 now?
Follow one broadcast through the multicast tree
Packets in this step
Inside the multicast packet
Compared with unicast VXLAN, the key changes are the outer multicast destination IP and corresponding next-hop/link-layer destination information; routing behavior is multicast rather than unicast. The outer destination IP is the group, and the outer destination MAC is the multicast MAC derived from it.
| Field | Value for VNI 10100, group 239.1.1.100 |
|---|---|
| Outer Ethernet dst (derived from the group) | 01:00:5E:01:01:64 |
| Outer Ethernet src (Leaf 1) | 0C:01:01:01:01:01 |
| Outer IP src (source VTEP) | 10.0.0.1 |
| Outer IP dst (group) | 239.1.1.100 |
| UDP src → dst | 52814 → 4789 |
| VNI | 10100 |
| Inner Ethernet dst → src | FF:FF:FF:FF:FF:FF ← 00:AA:AA:AA:AA:01 |
| Payload | ARP request |
Group IP → multicast MAC calculator
Type any group in 224.0.0.0–239.255.255.25501:00:5E + the low 23 bits of the group IP. The top bit of the second IP octet is dropped, so 32 different groups share one MAC. For example, “239.1.1.100” and “238.129.1.100” map to the same multicast Ethernet MAC because the top four IP bits and one additional bit are not represented.PIM designs and the comparison with ingress replication
PIM sparse mode (ASM)
In PIM-SM Any-Source Multicast (ASM), receivers establish shared-tree state toward an RP; first-hop routers can register new sources. This lab models ASM, not the only possible multicast design.
Anycast RP
Multiple RPs can provide RP redundancy with anycast addressing, but the exact failover and load distribution depend on topology and routing. Source-registration synchronization mechanisms (such as MSDP or PIM anycast-RP procedures) depend on the anycast-RP design.
Bidirectional PIM
Bidirectional PIM uses a shared bidirectional tree and avoids the same per-source forwarding state used by traditional shortest-path-tree switching. It can suit many-source multicast designs, but requires appropriate platform support.
Replication calculator
Change the numbers| Ingress replication | Underlay multicast | |
|---|---|---|
| Packets leaving the source per BUM frame | N − 1 | 1 |
| Who replicates | Source leaf | Spines (and any router on the tree) |
| Underlay requirement | Unicast routing only | Multicast forwarding and membership signaling (for example PIM-SM ASM and IGMP) |
| Extra state | Flood list per VNI on leaves | Multicast forwarding state on participating routers; (*,G) and/or (S,G) by design |
| Operations | Simple | More to design and troubleshoot |
| Best fit | Small to medium fabrics, WAN or multi-vendor | Large flat VNIs with heavy BUM, if multicast is already trusted |
Configuration and troubleshooting
Two parts: the VTEP maps a VNI to a group, and the underlay runs PIM so the group has a tree.
What’s wrong? Pick a symptom
Troubleshooting helperTest what you learned
Eight questions, each with an explanation.
Mastery checklist
Tick each one when you can explain it in your own words.
Cheat sheet
- Source sends 1 packet to the group address
- Each VNI maps to a group; same mapping everywhere
- Leaf VTEPs can signal membership using IGMP or PIM, depending on design; PIM routers build multicast forwarding state
- Outer dst MAC = 01:00:5E + low 23 bits of the group
- After RPF succeeds, a multicast router copies to eligible OIL interfaces, excluding incoming
- Shared groups cause wasted deliveries
Next
Next up is Lesson 03: the EVPN control plane (MP-BGP, route distinguishers, route targets and route types), which can also remove most flooding in the first place.