What to Check When Planning a Hyper-V Failover Cluster

Before building a Hyper-V failover cluster: matching hardware and firmware, SAN or S2D, redundant networking, quorum and witness placement, N+1 capacity and validation tests.
Technical summary
Technology
Windows Server, Hyper-V, Failover Clustering, CSV, S2D
Topic
Virtualization infrastructure
Level
Intermediate–advanced
Prerequisites
  • Active Directory domain
  • Server, storage and network inventory
In this article9 sections

A Hyper-V failover cluster restarts virtual machines on another node when a physical server fails and makes planned maintenance possible without downtime through live migration. But the protection a cluster provides is decided at the design stage, not in the setup wizard. The wrong storage, nodes connected to a single switch or capacity that hasn’t been calculated make the cluster useless when a failure happens. The points below need to be settled before ordering and installation. Plan traffic separation on Hyper-V together with the enterprise VLAN and network segmentation plan on the physical switches.

1. What is protected and what isn’t?

A cluster protects against hardware and operating system-level failures: a server shutting down, a node being taken down for maintenance, a power supply or motherboard failing. Deleted data, file corruption, ransomware or a bad update, however, affect all nodes at once through shared storage.

A cluster is an availability solution. It is not a backup.

That is why a cluster project should be budgeted together with a separate backup and restore plan. A common situation is that no money is left for backup after the cluster investment; that only moves the risk instead of reducing it.

2. Hardware compatibility and consistency

  • Nodes should be the same model, the same CPU family and the same firmware version as far as possible.
  • Live migration between different CPU generations may require processor compatibility mode.
  • Server, network card and storage components must be supported for the Windows Server version that will be used.

Firmware and driver differences show up as warnings in the cluster validation test and are often ignored. Then a network card behaves differently on one node and live migration fails now and then. Before installation, bring BIOS, RAID/HBA, NIC firmware and drivers to the same level on all nodes and document that level. Nodes should run the same Windows Server version, at the same update level and in the same Active Directory domain.

3. Storage model

Virtual machine disks are kept on storage that all nodes can access. Common options:

  • Cluster Shared Volumes (CSV) on shared SAN (iSCSI, Fibre Channel or SAS)
  • Storage Spaces Direct (S2D): software-defined storage built from the nodes’ local disks; hardware and network requirements are stricter

The storage itself must not be a single point of failure either: dual controllers, multipath (MPIO) connectivity and a capacity growth plan should be considered.

Operational capability matters as much as budget in the choice. On the SAN side, the problem is usually entry-level single-controller devices and MPIO not being configured: there are two cables, but the operating system sees the same LUN as two separate disks. On the S2D side, you need hardware the vendor has validated for S2D, RDMA-capable network cards and the right resiliency (mirror, parity) for the number of nodes. Two-node S2D is possible, but while one node is in maintenance the other runs on a single copy; keep that window short.

4. Network design

Different types of traffic in a cluster should be separated, with redundant connections planned for each:

  • Management and cluster communication
  • Virtual machine (user) traffic
  • Live migration
  • Storage traffic (for iSCSI or S2D)

With Switch Embedded Teaming (SET), VLANs and QoS these traffic types can also be separated over the same physical links; however, separate and sufficient bandwidth is recommended for storage traffic.

On the physical side the rule is simple: each node should be connected to two separate switches. A single switch is a single point of failure. If iSCSI is used, two paths are built on two separate subnets and MPIO balances them; MPIO is preferred over NIC teaming for iSCSI. If jumbo frames are enabled on the switches, verify end to end with ping -f -l that the same MTU is set everywhere; half-configured jumbo frames are a common cause of hard-to-find performance problems.

5. Quorum and witness

When communication between nodes is lost, the cluster decides by vote which side keeps running. Especially in two-node setups, a witness (disk, file share or cloud witness) should be defined. It is important that the witness does not share a point of failure with the cluster.

A common mistake is placing the file share witness on a virtual machine that runs on the cluster itself. When the cluster has a problem, the witness goes with it. Place the witness outside the cluster: on a separate physical server, at a separate location, or as a cloud witness in Azure. You can view the current configuration with Get-ClusterQuorum and change it with Set-ClusterQuorum.

6. Capacity headroom (N+1)

When a node goes offline, its virtual machines must be able to run on the remaining nodes. CPU and memory capacity should be calculated for this scenario; otherwise some virtual machines can’t start when a failure occurs.

An example: in a two-node cluster with 256 GB of memory per node, the total memory of all virtual machines must not exceed the usable memory of a single node (after the host’s own needs). In practice, each node in a two-node setup runs at most half full. When memory runs out, Hyper-V won’t start the machines. So define VM priorities: when memory is tight, domain controllers and business applications start first and test machines wait.

7. Validation and testing

  • A cluster validation test (Validate Cluster) should be run before installation and after every major change.
  • Planned node draining and live migration should be tested.
  • An unexpected node shutdown should be tried in a controlled window.
  • Applying updates node by node with Cluster-Aware Updating should be planned.

You can run the validation test in PowerShell with Test-Cluster; keep the report it produces. Microsoft support asks for a clean validation report when you open a case. Read the warnings one by one: a storage or network item left at “warning” becomes the problem itself at the first real failure. In the unexpected shutdown test, check not only that VMs start on the other node but that the applications work for users.

Checklist

  • Hardware compatibility and firmware versions aligned
  • Storage model and redundancy defined
  • Network traffic separated, connections redundant
  • Quorum and witness configured
  • N+1 capacity calculated
  • Validation test clean, failover tests done
  • A separate backup and restore test plan in place

Frequently asked questions

Does a failover cluster replace backup?

No. A cluster keeps virtual machines running on another node when a server fails. Deleted data, corrupted files or ransomware show up the same way on every node, so a separate, tested backup is needed.

Are Hyper-V Replica and a failover cluster the same thing?

No. A failover cluster provides automatic takeover between nodes that share the same data. Hyper-V Replica copies a virtual machine to another server asynchronously; failover is planned or started manually, and it is not an automatic high-availability cluster. Replica doesn’t replace backup either.

Why does a two-node cluster need a witness?

When one of the two nodes loses contact, a third vote is needed to decide which side keeps running. That vote comes from a disk, file share or cloud witness.

Support

Contact us for support or a discovery session on this topic.

Contact form
[email protected]

Eretek System Support

Server installation, configuration, incident response and scheduled maintenance. Active Directory, virtualization and backup are managed with the same discipline.

Knowledge Center

Other articles