Putting two switches in the core doesn’t mean redundancy; how those two switches appear to the access layer determines the outcome. With the wrong choice, either half the capacity is blocked by STP or a software upgrade blacks out the whole building. Below we compare three approaches that make a port channel spanning two devices possible (stack, MLAG and Cisco vPC) through the control plane difference, and address split-brain scenarios, upgrade downtime and selection criteria as they play out in the field.
The goal of device redundancy: active-active uplinks without STP
In a classic design, an access switch connects to two separate distribution switches with two uplinks, and STP blocks one of them. Half the capacity sits idle, and convergence in a failure is left to STP; you also need to plan root and gateway placement separately for each VLAN in the VLAN and routing design.
The goal is for the access switch to see its uplinks to the two distribution devices as a single LACP port channel. The access switch sees a single LACP system ID opposite it; the two members of the port channel terminate in different physical chassis. The result:
- Both uplinks carry traffic, and STP doesn’t block any port.
- When a distribution device goes down, the port channel loses one member; STP doesn’t recalculate.
- The root and gateway alignment problem largely disappears.
Three architectures deliver this result. The fundamental difference between them is whether the two devices run a single control plane or two separate control planes. How the port channel itself is built is explained in the LACP and EtherChannel configuration guide.
Stack: a single control plane, advantages and limits
In a stack, multiple physical switches work like a single logical switch. One member is in the active role (master with some vendors) and runs all control plane functions (STP, LACP, routing protocols, management); one member stays standby, keeping state synchronized. One configuration file, one management IP, monitoring like a single device.
There are two forms:
- Backplane stack: Common on access switches (such as the Cisco StackWise family). Connected in a ring with special stack cables; members are usually in the same or nearby cabinets.
- Virtual stack over Ethernet: On core and distribution-class devices, two chassis are turned into a single logical switch over regular high-speed ports. At Cisco, this approach is called VSS on older platforms and StackWise Virtual on the Catalyst 9000 family. The two chassis can be in different rooms.
The advantage is clear: operation is simple, a port channel spanning both members needs no special configuration, there’s a single router on the L3 side and no HSRP/VRRP is needed. The limitation comes from the same place. There’s a single control plane. A software bug or a faulty configuration change affects all members at once; a software upgrade also usually requires the whole stack to restart together.
One brain, one failure domain.
MLAG and vPC: separate control planes
In the MLAG (Multi-Chassis Link Aggregation) approach, the two switches remain independent devices: each has its own configuration, its own control plane and its own management address. The two devices synchronize data such as the MAC table, ARP/ND information and port-channel state over a peer link, and present a common LACP system ID to the downstream device.
MLAG isn’t an IEEE standard; each vendor offers its own implementation under a different name. In the Cisco Nexus family it’s vPC (virtual Port Channel); other vendors use names such as MLAG, MC-LAG, VSX or VLT. The concepts are similar; behavioral details (which side suspends its ports on a peer-link failure, consistency checks, L3 behavior) vary by vendor. Both peers must be from the same vendor and on compatible software versions.
A basic vPC configuration on Cisco Nexus looks like this:
feature lacp
feature vpc
!
vpc domain 10
role priority 100
peer-keepalive destination 192.168.255.2 source 192.168.255.1 vrf management
peer-gateway
auto-recovery
!
interface port-channel1
description vPC-PEER-LINK
switchport mode trunk
vpc peer-link
!
interface port-channel20
description ACCESS-SW-B2K1
switchport mode trunk
vpc 20
!
interface Ethernet1/20
channel-group 20 mode active
The vpc 20 number must be the same on both Nexus switches; vPC pairs the port channels on the two chassis through this number. Monitor consistency checks with show vpc consistency-parameters; a mismatch in Type 1 parameters can cause the relevant vPC to be suspended. A cause you may encounter is a change being made on only one peer and forgotten on the other.
Peer link, keepalive and split-brain scenarios
In a separate-control-plane architecture, the real design question is what happens when the two devices lose each other. Two separate links answer this question:
- Peer link: Carries state synchronization and, when needed, data traffic. It should be built as a port channel with at least two members, preferably with ports on different modules.
- Peer keepalive: Answers only the question “is the other device alive?” It should run over a path independent of the peer link, typically the management port or a separate L3 link.
| Status | Cisco vPC behavior | Effect |
|---|---|---|
| Peer link down, keepalive working | The secondary suspends vPC member ports and vPC VLAN SVIs | Traffic continues through the primary |
| Keepalive down, peer link working | No change; a warning is raised | Service isn’t affected; risk increases |
| Both down | Both devices may consider themselves primary | Split-brain: dual active, inconsistent forwarding |
| The primary device shut down completely | Because the keepalive was also cut along with the peer link, the secondary treats the other side as out of service and becomes operational primary | vPCs run through the secondary |
The equivalent risk on the stack side is the dual-active state: when the virtual stack link breaks, both chassis can stay active with the same configuration and the same IP addresses. Cisco StackWise Virtual detects this through a dedicated dual-active detection (DAD) link, or with ePAgP over port channels built with access switches, and puts one chassis into recovery mode by shutting its ports. A virtual stack isn’t put into production without DAD configured.
A dual-active core is worse than no core.
Downtime differences during software upgrades
The most concrete operational difference between the architectures shows up on upgrade night:
- Backplane stack: On most platforms, an upgrade completes with all members restarting together. During this time, all access connected to the stack is cut. Some platforms offer fast reload or staged upgrade methods that shorten the downtime; verify whether they’re supported in the release notes.
- Virtual stack (StackWise Virtual, VSS): Although there’s a single control plane, there are upgrade methods in which the two chassis restart in turn; support depends on platform and version.
- MLAG/vPC: Because the devices are independent, while one is upgraded the other carries all port channels over a single leg. Access switches only lose bandwidth. Check the vendor’s compatibility document that running the two peers on different versions during the transition is supported.
Upgrade frequency and the acceptable downtime window are set by organizational policy; the process itself is explained in the firmware update policy article.
Comparison table
| Criterion | Stack | Virtual stack | MLAG / vPC |
|---|---|---|---|
| Control plane | Single | Single | Two separate |
| Management | Single device | Single device | Two devices, consistency required |
| L3 gateway | Single router | Single router | HSRP/VRRP or equivalent |
| Physical distance | Short (stack cable) | Ethernet distance | Ethernet distance |
| Upgrade downtime | Usually full | Platform-dependent | Staged, low |
| Impact of a software bug | All members | Both chassis | Usually a single device |
| Configuration complexity | Low | Intermediate | High |
Selection criteria by scale and budget
When deciding, answer these questions in order:
- Is there a downtime window for planned maintenance? If a few minutes of downtime at night or on the weekend is acceptable, a stack or virtual stack is enough. If tolerance for downtime is very low (production, hospitals, 24/7 operations), MLAG/vPC gains weight in the enterprise network infrastructure.
- How many devices can the team manage? For a small IT team, a single configuration file is a big convenience. In MLAG, keeping the configuration of two devices consistent at all times requires discipline and preferably automation.
- Where will the devices sit? Placing the two core devices in different server rooms provides physical redundancy; in that case a backplane stack isn’t suitable.
- Where is the L3 design? If the core also does inter-VLAN routing, a stack has a single gateway; with MLAG, gateway redundancy is set up separately for each VLAN. Details are in the inter-VLAN routing article.
- What is at the access layer? Access switches usually remain stacks; MLAG/vPC mostly makes sense at the distribution and core layers.
The question of how many devices will be at which layer comes before this decision; the three-tier and collapsed core architecture comparison covers this step. STP stays on in all three architectures; for guard settings, see the STP, RSTP and MSTP guide. For platform-specific behavior, the vendor’s current configuration guide is authoritative; you can reach Cisco documents from the Cisco support portal.
To evaluate the redundancy model of your existing core and its fit with your growth plan, see our enterprise network consulting page.
Frequently asked questions
What happens if a stack cable breaks?
In a stack cabled as a ring, a single cable break doesn’t split the stack; traffic continues through the other direction of the ring, and only stack bandwidth drops. A break at two points splits the stack in two; without dual-active detection, both parts can stay active with the same configuration.
Can the vPC peer link and keepalive run over the same cable?
They shouldn’t. The keepalive’s job is to tell whether the other device is still running when the peer link breaks. If the two run over the same physical path, a single failure cuts both and creates split-brain risk. The keepalive is usually carried over the management port or a separate L3 link.
Does a small campus need MLAG?
In a single-building setup with a few access switches, a two-member stack is often enough and simpler to operate. MLAG’s benefit is that devices can be upgraded independently and a single software bug doesn’t take down both devices; this need comes to the fore in environments with low tolerance for downtime.
Why is peer-gateway used in vPC?
Some devices (for example some NAS units and load balancers) send replies not to the HSRP/VRRP virtual MAC but to the source router MAC of the incoming packet. If this packet lands on the other vPC peer, it has to be carried over the peer link and may be dropped under vPC rules. peer-gateway makes each peer also route locally the packets arriving for the other side’s router MAC, eliminating the problem.