If live migration takes longer than expected, CSV latency rises during the day, or the host drops off management when a single cable is pulled, the problem is usually not in the hardware but in the virtual network design. A Hyper-V host carries management, cluster communication, live migration, storage and VM traffic over the same physical cards; how this traffic is separated determines whether the host stays up when a NIC fails. Below we build a converged virtual network with Switch Embedded Teaming (SET), with PowerShell examples and physical switch requirements.
Write down the traffic first
Before touching a cable, put on paper which traffic flows on each host in the server and system infrastructure, and which is latency-sensitive and which is bandwidth-sensitive:
| Traffic | Feature | Typical separation |
|---|---|---|
| Management | Low bandwidth, high availability | Separate VLAN, host vNIC |
| Cluster (heartbeat, CSV redirection) | Latency-sensitive | Separate VLAN, at least two paths |
| Live migration | Short bursts of high bandwidth | Separate VLAN, SMB/RDMA if possible |
| Storage (SMB3 / iSCSI) | Latency and bandwidth critical | Separate VLAN; separate physical NICs preferred for iSCSI |
| VM traffic | Variable, many VLANs | Trunk port, VLAN tag per VM |
In the 1 GbE era, a separate pair of physical NICs was used for each traffic type; that’s why hosts had six to eight ports. With 10/25 GbE, the common model is to combine two (sometimes four) fast ports under a single SET switch and separate traffic with host vNICs, VLANs and QoS. iSCSI is the exception: since it is designed to run multipath with MPIO, it is usually left outside the team on separate physical ports. Details are in the Hyper-V shared storage choice article.
Moving from LBFO to SET
LBFO (NIC Teaming), introduced with Windows Server 2012, creates a team at the operating system level, and the virtual switch used to be built on top of this team. SET, from Windows Server 2016 on, moves the teaming function directly inside the Hyper-V virtual switch. The differences for the decision:
| Feature | LBFO + vSwitch | SET |
|---|---|---|
| Teaming mode | Switch independent, static, LACP | Switch independent only |
| RDMA on host vNICs | None | Yes |
| NIC compatibility | Different models can be mixed | Same model, driver and firmware |
| Maximum members | 32 | 8 |
| Microsoft’s direction for Hyper-V | Being deprecated | Recommended method |
In Windows Server 2022, creating a Hyper-V virtual switch on an LBFO team is blocked by default; Hyper-V Manager doesn’t offer this option. When upgrading LBFO-based hosts, put rebuilding the virtual switch as SET into the plan. Do the change after putting the host in maintenance mode and draining the VMs, because when the vSwitch is deleted, the connected VM adapters drop off the network. Don’t skip the order.
Creating SET and load balancing modes
SET has two load balancing algorithms:
- HyperVPort: Each VM adapter and host vNIC is bound to one physical port. A VM’s traffic leaves through one port; the physical switch sees the MAC address on a single port. This is the default on Windows Server 2019 and later.
- Dynamic: Outbound traffic is distributed across ports per flow, and inbound traffic is mapped using HyperVPort logic. It can distribute the outbound load better in environments with a few high-traffic VMs.
In most environments the default HyperVPort is the right choice. Since SET works only switch independent, no port-channel or LACP is configured on the physical switch; you connect the two ports to two separate physical switches and get switch redundancy.
# Create a SET switch with two 25 GbE ports
New-VMSwitch -Name "SETswitch" `
-NetAdapterName "SLOT2-P1","SLOT2-P2" `
-EnableEmbeddedTeaming $true `
-AllowManagementOS $false `
-MinimumBandwidthMode Weight
# Check / set the load balancing algorithm
Get-VMSwitchTeam -Name "SETswitch" | Format-List
Set-VMSwitchTeam -Name "SETswitch" -LoadBalancingAlgorithm HyperVPort
-MinimumBandwidthMode is set only when the switch is created. Changing it later means rebuilding the switch. If you’ll give weights to host vNICs, choose Weight from the start.
Host vNICs and VLAN configuration
The host operating system reaches the network through virtual adapters (host vNICs) attached to the SET switch. Open a separate vNIC for each traffic type and give each its own VLAN. The VLAN numbers below are examples; change them to fit your own addressing plan. Treat traffic separation on Hyper-V together with the enterprise VLAN and network segmentation plan on the physical switches.
# Host vNICs
Add-VMNetworkAdapter -ManagementOS -SwitchName "SETswitch" -Name "Management"
Add-VMNetworkAdapter -ManagementOS -SwitchName "SETswitch" -Name "Cluster"
Add-VMNetworkAdapter -ManagementOS -SwitchName "SETswitch" -Name "LiveMigration"
# VLAN tags (single VLAN in access mode)
Set-VMNetworkAdapterVlan -ManagementOS -VMNetworkAdapterName "Management" -Access -VlanId 10
Set-VMNetworkAdapterVlan -ManagementOS -VMNetworkAdapterName "Cluster" -Access -VlanId 20
Set-VMNetworkAdapterVlan -ManagementOS -VMNetworkAdapterName "LiveMigration" -Access -VlanId 30
# Bandwidth weights (relative values)
Set-VMNetworkAdapter -ManagementOS -Name "Management" -MinimumBandwidthWeight 10
Set-VMNetworkAdapter -ManagementOS -Name "Cluster" -MinimumBandwidthWeight 10
Set-VMNetworkAdapter -ManagementOS -Name "LiveMigration" -MinimumBandwidthWeight 30
# VLAN on a VM adapter
Set-VMNetworkAdapterVlan -VMName "APP01" -Access -VlanId 110
Weights aren’t an upper limit; they are the relative share guaranteed during congestion. When the link is idle, every vNIC can use all the bandwidth. For unassigned VM adapters, the switch’s DefaultFlowMinimumBandwidthWeight value is used; designs that forget to leave a share for VM traffic come back as user complaints during heavy live migration.
The management vNIC gets the default gateway and DNS. No gateway is defined on the cluster and live migration vNICs, and DNS registration is turned off. This is exactly the common mistake: a gateway is entered on every vNIC, the host acquires several default routes and management traffic leaves through an unexpected interface. Then a ticket is opened saying “RDP drops now and then.”
Set-DnsClient -InterfaceAlias "vEthernet (Cluster)" -RegisterThisConnectionsAddress $false
Set-DnsClient -InterfaceAlias "vEthernet (LiveMigration)" -RegisterThisConnectionsAddress $false
RDMA and SMB Direct
RDMA lets SMB traffic move directly between memories while barely using the processor (SMB Direct). In a Hyper-V environment, SMB3 storage, Storage Spaces Direct and SMB-based live migration benefit from this directly. SET is the only teaming method that supports RDMA on host vNICs.
There are two common RDMA types. iWARP runs over TCP, and a lossless network configuration isn’t mandatory. RoCEv2 runs over UDP and requires end-to-end DCB (PFC and ETS) configuration. With RoCE, the switch and host DCB settings must match exactly; a wrong priority number on one side produces no error message, it just silently lowers performance. This is the hardest failure to find.
# Two RDMA-enabled host vNICs for SMB
Add-VMNetworkAdapter -ManagementOS -SwitchName "SETswitch" -Name "SMB1"
Add-VMNetworkAdapter -ManagementOS -SwitchName "SETswitch" -Name "SMB2"
Set-VMNetworkAdapterVlan -ManagementOS -VMNetworkAdapterName "SMB1" -Access -VlanId 711
Set-VMNetworkAdapterVlan -ManagementOS -VMNetworkAdapterName "SMB2" -Access -VlanId 712
Enable-NetAdapterRdma -Name "vEthernet (SMB1)","vEthernet (SMB2)"
# Pin each SMB vNIC to a separate physical port
Set-VMNetworkAdapterTeamMapping -ManagementOS -VMNetworkAdapterName "SMB1" -PhysicalNetAdapterName "SLOT2-P1"
Set-VMNetworkAdapterTeamMapping -ManagementOS -VMNetworkAdapterName "SMB2" -PhysicalNetAdapterName "SLOT2-P2"
# DCB example for RoCEv2 (priority 3 = SMB)
Install-WindowsFeature Data-Center-Bridging
New-NetQosPolicy "SMB" -NetDirectPortMatchCondition 445 -PriorityValue8021Action 3
Enable-NetQosFlowControl -Priority 3
Disable-NetQosFlowControl -Priority 0,1,2,4,5,6,7
New-NetQosTrafficClass "SMB" -Priority 3 -BandwidthPercentage 50 -Algorithm ETS
Set-NetQosDcbxSetting -Willing $false
Enable-NetAdapterQos -Name "SLOT2-P1","SLOT2-P2"
Because RDMA traffic bypasses the virtual switch’s software path, vSwitch weights don’t limit SMB Direct traffic; the DCB traffic class sets this traffic’s share. The ETS bandwidth value in the example is a configuration example; choose it based on the traffic mix in your environment. Microsoft’s reference document is the RDMA and Switch Embedded Teaming page.
The physical switch side
However clean the host side is, the design isn’t finished until the switch port on the other end is verified. For consistency with the VLAN plan on the campus side, the VLAN design in enterprise networks article is a complement.
| Requirement | Why |
|---|---|
| Trunk port with an explicitly defined allowed VLAN list | Host vNIC and VM VLANs arrive tagged |
| No port-channel / LACP | SET works only switch independent |
Edge port behavior (e.g. spanning-tree portfast trunk on Cisco) |
Avoids STP delay when the link comes up |
| The same MTU end to end | Jumbo frame inconsistency causes live migration and SMB errors |
| PFC and ETS for RoCE, with the same priority as the host | Lossless RDMA traffic |
If jumbo frames are used, the adapter’s “Jumbo Packet” advanced property on the Windows side (9014 bytes on most drivers) and the MTU on the switch side are set together. Enabling jumbo only on the storage and live migration VLANs and keeping the standard MTU for management and VM traffic makes troubleshooting easier. If the MTU is missing on a single switch in the middle of the path, large packets drop and small ones pass; ping works but live migration hangs. An end-to-end test is a must.
Verification before adding to the cluster
Run the following checks on each host and add the output to the change record:
# Team members and algorithm
Get-VMSwitchTeam -Name "SETswitch" | Format-List Name, NetAdapterInterfaceDescription, TeamingMode, LoadBalancingAlgorithm
# Host vNICs, VLANs and physical mapping
Get-VMNetworkAdapter -ManagementOS | Format-Table Name, SwitchName, MacAddress
Get-VMNetworkAdapterVlan -ManagementOS
Get-VMNetworkAdapterTeamMapping -ManagementOS
# RDMA and SMB Multichannel
Get-NetAdapterRdma | Format-Table Name, Enabled
Get-SmbClientNetworkInterface
Get-SmbMultichannelConnection
# End-to-end jumbo frame test (8972-byte payload for a 9014-byte adapter setting)
ping 10.0.71.12 -f -l 8972
Checklist:
- All SET members are the same model, driver and firmware version.
- Each host vNIC is on the right VLAN; the default gateway is only on the management vNIC.
- When a physical cable is pulled, management access and VM traffic continue without interruption. Do this test by actually pulling the cable, not on paper.
- If RDMA is used, RDMA-capable connections appear in the
Get-SmbMultichannelConnectionoutput. - In the Failover Cluster, the roles of the cluster networks (cluster only, cluster and client, none) are assigned correctly, and the networks live migration will use are defined.
To diagnose network-related errors during live migration, see the Hyper-V Live Migration errors article; for the overall cluster design, see the failover cluster planning checklist. For a review of your current host network configuration, there is our virtualization and data management service page.
Frequently asked questions
Should I configure LACP on the physical switch when using SET?
No. SET works only in switch independent mode; the physical ports aren’t put in a port-channel or LACP group, and each is configured as an independent trunk port. This lets you connect the two members to two separate physical switches.
Can I add NICs of different brands or speeds to SET?
Microsoft requires SET members to be the same vendor, model, driver and firmware version. A team built with different cards is unsupported, and unpredictable behavior can be seen in features such as RDMA.
Should I choose HyperVPort or Dynamic?
On Windows Server 2019 and later, the default HyperVPort is the right choice for most environments; each VM is bound to one port, and there is no MAC movement on the physical switch. If outbound traffic from a small number of VMs saturates a single port, Dynamic can be considered.
Should I also carry iSCSI traffic over SET?
Technically, iSCSI over host vNICs is possible. But since iSCSI redundancy is provided by MPIO, running it on separate physical ports outside the team is the common and recommended approach. Also check the storage vendor’s supported configuration.