SONiC
Module 4 · EVPN/VXLAN in SONiC

Symmetric and Asymmetric IRB

35 minLesson

Both IRB models solved the same problem differently. This lesson contrasts them at the level a SONiC operator needs — what each requires on every VTEP, how the L3VNI does or does not appear, and the operational tradeoffs — with a hard caveat: identify what your specific lab image supports before assuming either workflow.

Prerequisites: Lessons 4.1, 4.3, 4.4

Learning objectives

  • Contrast asymmetric and symmetric IRB in terms of what each VTEP must hold and configure.
  • Explain the role of the L3VNI and type-5 routes in the symmetric model.
  • Enumerate the operational tradeoffs (scale, config symmetry, MAC/ARP scale) of each.
  • State the verification step to confirm which model your image supports before designing a workflow.
Why this matters

Choosing an IRB model dictates your config template, your scaling limits, and your troubleshooting commands. Picking one your image does not fully support is a costly late discovery.

Inter-subnet forwarding — the thing tenants actually ask for — lives or dies on the IRB model, so this is where design meets daily operations.

Analogy — Two ways terminals hand off between tenant labels
In the analogyIn SONiC
The origin terminal must stock every destination tenant's labelsAsymmetric IRB
A shared transit label (L3VNI) both terminals understand for cross-tenant handoffSymmetric IRB
The neutral transit-lane label used only between terminalsL3VNI
What it actually is: Asymmetric IRB: the ingress VTEP routes into the destination subnet and then bridges in the *destination* VNI, so every VTEP must be provisioned with every VNI it might talk to (route-in, bridge-out; the return path uses different VNIs, hence 'asymmetric'). Symmetric IRB: both ingress and egress route through a common L3VNI using type-5 routes (route-in, route-across-L3VNI, route-out), so a VTEP only needs the VNIs of locally attached subnets plus the shared L3VNI — the forward and return paths are symmetric.
Where the analogy stops working

The label analogy hides the real cost driver: asymmetric IRB forces every VTEP to hold ARP/ND and MAC state for *all* subnets it routes to, which is the scaling wall. It also hides that whether symmetric IRB, asymmetric IRB, or only L2 (no IRB) is available is entirely image/version dependent in SONiC — the analogy makes both look equally on-the-shelf when your image may stock only one.

Read this before you design anything

Do not assume your SONiC image supports symmetric IRB, asymmetric IRB, or L3VNI routing at all. Support and defaults vary by version, distribution, and ASIC/SAI. Confirm on the actual lab image first: check show evpn vni for L2 vs L3 VNIs, look for L3VNI/VRF-VNI support in the image's docs, and validate a routed cross-subnet flow in a lab before committing a design. Treat everything below as the model to verify, not a guaranteed workflow.

Asymmetric vs symmetric IRB

AspectAsymmetric IRBSymmetric IRB
Forwarding modelRoute-in, bridge-out in destination VNIRoute-in, route-across L3VNI, route-out
VNIs per VTEPMust configure *every* VNI it talks toOnly locally attached VNIs + shared L3VNI
Route type usedRelies on type-2 MAC/IP across all VNIstype-5 IP prefixes over the L3VNI
State/scale costARP/MAC for all remote subnets → scaling wallOnly local host state + prefixes → scales better
Config symmetryAsymmetric VNI use forward vs returnSymmetric; easier templating
Typical modern choiceSimpler mentally, limited scalePreferred at scale — if the image supports L3VNI

What the L3VNI actually is

In the symmetric model the L3VNI is a per-VRF transit VNI that carries routed traffic between VTEPs. type-5 routes advertise tenant prefixes tagged with the VRF's route target and the L3VNI; the ingress VTEP routes the packet into the L3VNI, ships it, and the egress VTEP routes it out into the destination subnet. No VTEP needs to know the destination subnet's L2VNI unless a host is locally attached there.

This is why symmetric IRB scales: a VTEP holds host/ARP state only for subnets with local hosts, plus routed prefixes — not the full L2 state of every subnet in the fabric.

Confirming what your image exposes

FRR / vtyshIllustrative output
show evpn vni                 # look for VNIs typed as 'L3' (VRF) vs 'L2'
show evpn vni detail          # VRF binding, L3VNI, router MAC per VNI
show bgp l2vpn evpn route type prefix   # are type-5 routes present at all?
vtysh -c 'show version'       # FRR version — feature availability varies
show version                  # SONiC image/build — record it in the lab log

Run these on the actual lab image before choosing a model. The presence/labels here tell you what is realistically available.

Operational tradeoffs at a glance

DimensionAsymmetricSymmetric
VNI provisioningAll VNIs on all routing VTEPsLocal VNIs + one L3VNI per VRF
ARP/ND scaleHigh (all remote subnets)Lower (local subnets only)
Troubleshooting focustype-2 across many VNIstype-5 + L3VNI + router MAC
Failure blast radiusVNI sprawl → more to misconfigureTighter, but L3VNI is a shared dependency
Image dependencyVerify L2 IRB supportVerify L3VNI/type-5 support — often the gating factor
Troubleshooting implications
  • For a symmetric-IRB fabric, verify the router MAC and L3VNI as first-class objects, not afterthoughts.
  • For asymmetric IRB, confirm the destination VNI exists on the ingress VTEP — a missing VNI there is the classic break.
  • Always record the image/build; an inter-subnet workflow that 'worked in class' may not exist on a different image.

Key takeaways

  • Asymmetric = route-in/bridge-out, every VNI everywhere; symmetric = route via shared L3VNI with type-5.
  • Symmetric scales better but depends on L3VNI/type-5 support in your image.
  • IRB model dictates config template, scale, and troubleshooting commands.
  • Confirm your specific image's IRB support before designing any inter-subnet workflow.

Tested against: SONiC 202305 / 202311 lineage · Upstream (community) unless noted · SONiC-VS (virtual) baseline · last reviewed 2026-07. Exact commands and behavior can vary by SONiC release, image, hardware platform, and enterprise distribution.

Knowledge check

Q1

You are designing inter-subnet forwarding and want the symmetric IRB model so VTEPs only hold local subnet state plus a shared L3VNI. Before committing the design, what is the single most important thing to confirm?