Storage Spaces Direct on Windows Server 2025: 2-Node Cluster Step-by-Step Guide

A 2-node S2D cluster is the smallest supported Storage Spaces Direct design, and the one where early decisions are hardest to undo. This guide helps you make those decisions first, then builds the cluster.

You will decide between hyperconverged and disaggregated deployment, choose a network and RDMA approach, pick a volume resiliency and a witness, and then work through prerequisites, networking, validation, enabling S2D and creating volumes. Each step explains the reasoning so you can adapt it to your hardware, and each choice is marked as reversible or not.

Decision 1: hyperconverged or disaggregated

In Windows Server, hyperconverged means Hyper-V VMs and S2D storage run on the same nodes. Disaggregated (Microsoft’s docs call it “converged”) means a storage cluster running S2D and a Scale-Out File Server (SOFS) serves SMB3 shares to a separate Hyper-V compute cluster.

Recommendation: choose hyperconverged unless one of these is true: compute and storage must scale independently, several compute clusters share the storage, separate teams own each tier, or per-core licensing on compute-heavy hosts is a concern.

Decision 2: network and RDMA

S2D uses SMB3, including SMB Direct and SMB Multichannel, between nodes. Microsoft requires at least 10 GbE and recommends RDMA. You make two sub-decisions: the RDMA protocol, and switchless versus switched.

Microsoft’s campus cluster documentation cites up to 30% CPU savings from RDMA.

What each choice locks in

• Both nodes must use the same RDMA protocol, and Network ATC requires identical adapter make, model and speed within an intent. Changing protocol or vendor later means new NICs.
• Some adapters support both iWARP and RoCE (Marvell FastLinQ and Intel E810 families are commonly cited), and the mode is set per adapter. Verify your exact model and driver.
• NICs must carry the Windows Server logo that matches your OS version.

Switchless versus switched

• Switchless: each node connects directly to the other. For two nodes, use at least two links, each on its own subnet. You save the switches, and RoCE needs no switch-side DCB, although hosts still need it. Management and compute traffic need separate NICs. Adding a third node requires a redesign.
• Switched: two top-of-rack switches for redundancy, and lossless DCB configuration for RoCE. It scales, and it lets you converge traffic on fewer ports.

Decision 3: resiliency and witness

Volume resiliency

Resiliency is chosen per volume, and you can mix types in one cluster. Volumes cannot be converted between resiliency types. You create a new volume and migrate the data.

Witness

A 2-node cluster needs a witness. Two votes alone cannot form a majority, so losing either node would stop the cluster. The witness also acts as the tie-breaker for the storage pool.

• A disk witness is not an option, because S2D has no shared disks.
• Never host the witness on a VM that runs on this cluster. That creates a circular dependency.
• With a witness you survive any single node or witness failure. Two simultaneous failures stop the cluster.
• The witness type is reversible at any time with Set-ClusterQuorum.

What is hard to reverse

Prerequisites and hardware

Servers

• Windows Server Datacenter edition on both nodes (S2D requires it). Identical model, firmware, CPU and drive layout.
• Memory for the OS and workloads, plus 4 GB of RAM per TB of cache-drive capacity for S2D metadata.
• Drives behind an HBA in pass-through mode (no RAID, no multipath), with a separate boot device.

Drives per node

• Keep the number of capacity drives a whole multiple of the number of cache drives.
• Choose cache drives with high write endurance (Microsoft suggests around 3 drive writes per day).
• Nested resiliency needs the four-capacity-drive minimum per server.
• Leave roughly one capacity drive per node unallocated as repair reserve.

Network, directory and Azure

• At least 10 GbE, two or more storage ports per node, symmetric NICs on both nodes.
• Domain-joined nodes, working DNS and time sync, static IPs, and an OU where the cluster can create its computer object.
• A storage account if you plan a cloud witness.
• OEM-supplied firmware and drivers for Windows Server 2025. Inbox NIC drivers are often wrong for RDMA.

Build steps

Examples use placeholder names S2D-N1, S2D-N2 and S2D-CL01.

Step 1: Prepare the operating system (both nodes)

# Same build and patch level on both nodes, then set a high-performance power plan
powercfg /setactive 8c5e7fda-e8bf-4a96-9a85-a6e23a8c635c

Mismatched patch levels and power-saving profiles cause validation warnings and uneven performance.

Step 2: Install the required features

$Nodes = "S2D-N1","S2D-N2"
Invoke-Command $Nodes {
  Install-WindowsFeature -Name Hyper-V, Failover-Clustering, FS-FileServer, `
    Data-Center-Bridging, NetworkATC, NetworkHUD, RSAT-Clustering-PowerShell, `
    Hyper-V-PowerShell -IncludeManagementTools -Restart
}

Omit Hyper-V on disaggregated storage nodes. NetworkATC and NetworkHUD are the Windows Server 2025 Network ATC features.

Step 3: Clean the drives

Invoke-Command $Nodes {
  Update-StorageProviderCache
  Get-StoragePool | ? IsPrimordial -eq $false | Set-StoragePool -IsReadOnly:$false -ErrorAction SilentlyContinue
  Get-StoragePool | ? IsPrimordial -eq $false | Get-VirtualDisk | Remove-VirtualDisk -Confirm:$false -ErrorAction SilentlyContinue
  Get-StoragePool | ? IsPrimordial -eq $false | Remove-StoragePool -Confirm:$false -ErrorAction SilentlyContinue
  Get-PhysicalDisk | Reset-PhysicalDisk -ErrorAction SilentlyContinue
  Get-Disk | ? Number -ne $null | ? IsBoot -ne $true | ? IsSystem -ne $true | ? PartitionStyle -ne RAW | % {
    $_ | Set-Disk -IsOffline:$false; $_ | Set-Disk -IsReadOnly:$false
    $_ | Clear-Disk -RemoveData -RemoveOEM -Confirm:$false
    $_ | Set-Disk -IsReadOnly:$true;  $_ | Set-Disk -IsOffline:$true
  }
}
Get-PhysicalDisk | Group-Object MediaType, BusType, CanPool | ft Count, Name

Every data drive should report CanPool = True. Anything else means leftover metadata, a RAID or multipath path, or an unsupported controller.

Step 4: Pre-validate and create the cluster

Test-Cluster -Node $Nodes -Include "Storage Spaces Direct","Inventory","System Configuration" -Verbose
New-Cluster -Name S2D-CL01 -Node $Nodes -NoStorage -StaticAddress 10.0.0.50

-NoStorage stops the cluster from claiming disks, because S2D must claim them later.

Step 5: Configure host networking

Path A: Network ATC (recommended on Windows Server 2025). It requires Datacenter and an existing cluster, because the intent is stored cluster-wide. Adapters in one intent must match in make, model and speed on every node.

# Management and compute converged on two ports
Add-NetIntent -Name MgmtCompute -Management -Compute -AdapterName "NIC1","NIC2"

# Storage on the dedicated RDMA ports.
# Switchless: turn off automatic storage IPs and assign your own per direct link.
$so = New-NetIntentStorageOverrides
$so.EnableAutomaticIPGeneration = $false
Add-NetIntent -Name Storage -Storage -AdapterName "SMB1","SMB2" -StorageOverrides $so

Get-NetIntentStatus

For a switchless layout, each direct link needs its own subnet, for example 10.71.1.0/24 and 10.71.2.0/24. Verify the override object names and storage VLAN behavior with Get-Help Add-NetIntent on your build and with your OEM’s switchless guide. For a switched design, leave automatic IPs on.

Path B: manual configuration. Do this before Step 4, because without ATC there is no cluster dependency.

New-VMSwitch -Name SETswitch -NetAdapterName "NIC1","NIC2" -EnableEmbeddedTeaming $true -AllowManagementOS $true
Enable-NetAdapterRdma -Name "SMB1","SMB2"

# RoCE only: host-side DCB. The switch configuration must match.
New-NetQosPolicy "SMB" -NetDirectPortMatchCondition 445 -PriorityValue8021Action 3
New-NetQosPolicy "Cluster" -Cluster -PriorityValue8021Action 7
New-NetQosPolicy "Default" -Default -PriorityValue8021Action 0
Enable-NetQosFlowControl -Priority 3
Disable-NetQosFlowControl -Priority 0,1,2,4,5,6,7
New-NetQosTrafficClass "SMB" -Priority 3 -BandwidthPercentage 50 -Algorithm ETS
New-NetQosTrafficClass "Cluster" -Priority 7 -BandwidthPercentage 1 -Algorithm ETS
Set-NetQosDcbxSetting -Willing $false -Confirm:$false
Enable-NetAdapterQos -Name "SMB1","SMB2"

For iWARP you can skip the DCB block. If your NIC supports both modes, set the RDMA mode explicitly. The property name varies by vendor.

Step 6: Run full validation

Test-Cluster -Node $Nodes -Include "Storage Spaces Direct","Inventory","Network","System Configuration" -Verbose
Get-NetAdapterRdma
Get-SmbClientNetworkInterface
Get-SmbMultichannelConnection      # run while copying a large file over the storage network

Read the report instead of only checking for green: every drive should be listed and symmetric across nodes, RDMA should be enabled on the storage NICs, and MTU should be consistent. A failure in the S2D or network sections must be fixed first. Enabling S2D on top of a bad network is the most common source of poor performance later.

Step 7: Configure the witness

# Cloud witness
Set-ClusterQuorum -CloudWitness -AccountName "<storageaccount>" -AccessKey "<key>"

# Or a file share witness (the cluster computer account needs Full Control on the share)
Set-ClusterQuorum -FileShareWitness "\\FS01\S2DWitness"

Get-ClusterQuorum

Do this before enabling S2D, so the storage pool has its tie-breaker from the start.

Step 8: Enable Storage Spaces Direct

Enable-ClusterStorageSpacesDirect -PoolFriendlyName "S2D Pool" -Confirm:$false -Verbose

Get-StoragePool S2D* | ft FriendlyName, HealthStatus, OperationalStatus
Get-PhysicalDisk | Group-Object MediaType, Usage, BusType | ft Count, Name
Get-StorageTier | ft FriendlyName, ResiliencySettingName, MediaType, NumberOfDataCopies

This one command creates the pool, claims the drives and binds the cache, so it is the point of no return. With a single media type and no cache wanted, add -CacheState Disabled. Check that cache devices appear as cache and capacity devices as capacity.

Step 9: Create volumes

Plan two volumes, one per node, so Cluster Shared Volume ownership is balanced.

# A) Two-way mirror (the 2-node default), 50% efficiency
New-Volume -FriendlyName "CSV01" -FileSystem CSVFS_ReFS -StoragePoolFriendlyName S2D* -Size 4TB -ResiliencySettingName Mirror

# B) Nested two-way mirror (all-flash), 25% efficiency
New-Volume -FriendlyName "CSV01" -FileSystem CSVFS_ReFS -StoragePoolFriendlyName S2D* `
  -StorageTierFriendlyNames NestedMirrorOnSSD -StorageTierSizes 2TB

# C) Nested mirror-accelerated parity (mirror size first, parity size second)
New-Volume -FriendlyName "CSV02" -FileSystem CSVFS_ReFS -StoragePoolFriendlyName S2D* `
  -StorageTierFriendlyNames NestedMirrorOnSSD,NestedParityOnSSD -StorageTierSizes 200GB,1.8TB

• Tier names: nested tier templates are created automatically on Windows Server 2022 and later, and only 2019 needed manual creation. Confirm names with Get-StorageTier, and use the OnHDD variants for HDD capacity.
• Capacity example: with 4 × 3.84 TB SSD per node, raw capacity is about 30.7 TB. After one drive per node of reserve, roughly 23 TB is allocatable. That gives about 11.5 TB usable with two-way mirror, about 5.8 TB with nested mirror, and about 8 to 9 TB with nested parity, depending on the mix.
• Thin provisioning: Windows Server 2025 adds thin volumes (-ProvisioningType Thin on mirror volumes). Older cmdlet documentation says tiered volumes must use fixed provisioning, and nested volumes use tiers, so verify and treat thin and nested as mutually exclusive until you confirm otherwise in a lab.

Step 10: Verify before production

Get-VirtualDisk | ft FriendlyName, ResiliencySettingName, HealthStatus
Get-ClusterSharedVolume
Get-StorageJob

Benchmark with DISKSPD or VMFleet, then run a controlled node-failure test while the cluster is still empty. Confirm VMs fail over, the witness behaves as expected and resync completes.

If you choose disaggregated

• Storage cluster: run the same steps 1 to 8, without the Hyper-V role.
• Scale-Out File Server: Add-ClusterScaleOutFileServerRole -Name SOFS01, then create volumes and SMB shares.
• Share permissions: grant the Hyper-V compute hosts’ computer accounts and your admin group full control on the shares and folders.
• Compute cluster: a separate Hyper-V failover cluster whose VM paths point at \\SOFS01\<share>.
• Network: a switched RDMA fabric between compute and storage nodes. Nested resiliency still works on a 2-node storage cluster.

Operations: patching a 2-node cluster

In a 2-node cluster, the surviving node holds the only copy of the data while its partner is down. Rebooting the second node while resync is still running can take volumes offline, so patch one node at a time and wait for storage jobs to finish.

Suspend-ClusterNode -Name S2D-N1 -Drain -Wait
Get-StorageFaultDomain -Type StorageScaleUnit | ? FriendlyName -eq "S2D-N1" | Enable-StorageMaintenanceMode
# patch and reboot the node
Get-StorageFaultDomain -Type StorageScaleUnit | ? FriendlyName -eq "S2D-N1" | Disable-StorageMaintenanceMode
Resume-ClusterNode -Name S2D-N1 -Failback Immediate
Get-StorageJob    # wait until empty before touching the next node

Cluster-Aware Updating can automate this sequence. Size CPU and memory so either node can run every VM, since a node failure puts the whole workload on one server.

Windows Server 2025 changes

Troubleshooting

Questions that decide your design

1. Scale: must compute and storage scale independently, or will this stay a 2-node pair?
2. Failure tolerance: must the cluster survive a node plus a drive failure, or is one failure at a time enough with good backups?
3. Drives and capacity: what is your drive layout, and how much usable capacity do you need after 25% nested efficiency?
4. Fabric: do you already run a DCB-capable network or have RoCE experience?
5. Witness: can both nodes reach Azure over HTTPS, or do you need a file share witness on an independent machine?
6. Growth: will you add a third node within the hardware lifetime? That rules out nested resiliency and affects switchless.
7. Security baseline: does it require SMB signing or encryption, and have you measured the effect on RDMA throughput?
8. Drivers: does your OEM provide Windows Server 2025 drivers and firmware for your NIC and HBA?
9. Headroom: can either node run every VM on its own?

If you also manage Windows Server, Hyper-V and hybrid Azure environments, more practical guides are on VMoreCloud.

Frequently asked questions

Sources and further reading

• Microsoft Learn: Deploy Storage Spaces Direct on Windows Server
• Microsoft Learn: Storage Spaces Direct hardware requirements
• Microsoft Learn: Nested resiliency for Storage Spaces Direct
• Microsoft Learn: Create volumes on Azure Local and Windows Server clusters
• Microsoft Learn: Deploy host networking with Network ATC
• Microsoft Learn: Add-NetIntent
• Microsoft Learn: What’s new in Windows Server 2025
• Microsoft Learn: Create a Storage Spaces Direct campus cluster
• Microsoft Learn: Storage Spaces Direct overview

This guide is based on Microsoft documentation and community sources available on October 28, 2026. Cmdlet parameters, hardware support and defaults can change with updates, and some details depend on your OEM. Items marked verify were not confirmed in Microsoft documentation. Test the full build, including a node-failure test, in a lab before production, and keep tested backups.

👤 About the Author

Khurram Shahzad

Hybrid Cloud & Virtualization Engineer

Hybrid Cloud & Virtualization Engineer specializing in VMware (VCF, vSAN, Aria), Microsoft Azure, AWS, Windows Server, Microsoft 365, Entra ID, Intune, Linux, Nutanix, Proxmox, networking, backup & disaster recovery, Cisco and FortiGate. I share practical technical guides, infrastructure projects and insights to help IT professionals build and manage modern IT environments through VMoreCloud.

🖥️ VMware🔷 Azure☁️ AWS🪟 Windows Server🐧 Linux🗄️ Virtualization🛡️ Cybersecurity🔗 Networking💾 Backup & DR✖️ Nutanix🟠 Proxmox

📄

120+

Articles Published

👥

10K+

Readers

🎓

Learning & Sharing

A Stronger Tech Community

View all articles →

“Learn  |  Build  |  Share”

☁️ VMoreCloud|Tech Knowledge for a Better Tomorrow

IT  •  Cloud  •  Virtualization  •  Security  •  Open Source