RFC 7348 • EVPN • AI / GPU CLUSTERS

VxLAN for Networking AI Era

Scale-out multi-tenant overlays, high-radix Clos topologies, and lossless RoCEv2 transport for ultra-large distributed AI training fabrics.

16.7M+ Virtual Networks (24-bit VNI)
4789 UDP IANA Standard Port
50 Bytes Standard IPv4 Overhead
Zero Loss RoCEv2 / PFC Alignment

Why VXLAN in Modern Architectures?

Overcoming Layer-2 limits with Layer-3 ECMP scale and dynamic BGP EVPN control planes.

16M Layer 2 Overlays

Breaks the 4,096 VLAN limitation by introducing a 24-bit VXLAN Network Identifier (VNI), enabling massive multi-tenancy in hyper-scale cloud & AI fabrics.

🌐

L3 ECMP Underlay

Runs transparently on routed IP networks using all available Equal-Cost Multi-Path (ECMP) links without spanning tree protocol (STP) blocked ports.

🤖

AI & GPU Cluster Ready

Seamlessly isolates AI tenants and distributes AllReduce / collective communication across multi-node GPU systems with RoCEv2 encapsulation.

AI ARCHITECTURE

VXLAN in the Era of AI & GPU Training

How modern AI data centers leverage EVPN-VXLAN with RDMA and RoCEv2 for high-throughput tensor parallelism.

1. RoCEv2 & RDMA Encapsulation

RDMA over Converged Ethernet v2 (RoCEv2) packets encapsulated within VXLAN tunnels allow seamless L2 adjacency across leaf-spine pods while maintaining sub-microsecond latency.

2. Flow Entropy & ECMP Hashing

The VXLAN outer UDP source port is calculated from an inner header hash (e.g. 5-tuple flow). This provides high entropy for spine switches to distribute AI elephant flows across all spine links.

3. BGP EVPN Control Plane

Eliminates flood-and-learn multicast reliance by using MP-BGP EVPN (RFC 8365) for host MAC/IP route advertisement and ARP suppression across GPU compute nodes.

4. Buffer Management & Congestion

Outer DSCP copying guarantees that Priority-based Flow Control (PFC) and Explicit Congestion Notification (ECN) marking remain intact across the underlay.

VXLAN Encapsulation Anatomy

Click on any header segment below to inspect byte boundaries, fields, and RFC requirements.

Outer Ethernet 14 Bytes
Outer IP 20 / 40 Bytes
Outer UDP 8 Bytes (Port 4789)
VXLAN Header 8 Bytes (24-bit VNI)
Inner L2 Frame Payload + Inner MAC

VXLAN Header (8 Bytes)

The VXLAN header contains flags, reserved fields, and the 24-bit VXLAN Network Identifier (VNI).

Bit Range Field Description
Bits 0-7 Flags (0x08) Bit 3 is I-flag (VNI valid). Remaining bits reserved and must be 0.
Bits 8-31 Reserved (24 bits) Must be set to 0 on transmit, ignored on receipt.
Bits 32-55 VNI (24 bits) VXLAN Network Identifier designating the virtual overlay segment.
Bits 56-63 Reserved (8 bits) Must be set to 0 on transmit.

Interactive MTU & Overhead Calculator

Calculate exact frame overhead and required underlay MTU for your cloud or AI fabric.

Standard Ethernet: 1500 | AI / RoCEv2 Jumbo Frame: 9000
Encapsulation Overhead: 50 Bytes
Minimum Underlay MTU: 1550 Bytes
Recommended Switch MTU: 1600 Bytes

Quick Specification Reference

Key parameters for network engineers and system administrators.

Core VXLAN Parameters

  • Standard: IETF RFC 7348
  • IANA Port: UDP 4789
  • Linux Kernel Default: UDP 8472
  • Identifier Size: 24-bit VNI (16,777,216 IDs)
  • Default Encapsulation: MAC-in-UDP

Ubuntu Linux Kernel Commands

# Add VXLAN interface in Ubuntu / Linux
ip link add name vxlan100 type vxlan \
    id 100 \
    dstport 4789 \
    local 192.168.10.1 \
    dev eth0

# Bring up interface
ip link set up dev vxlan100
ip addr add 10.0.100.1/24 dev vxlan100