Windows Server 2025 Storage Spaces: Parity Interleave hard-locked to >=64KB on Fixed Provisioning (Breaks ReFS 64KB AUS Alignment)

Hi everyone, novice sysadmin here looking into this in my homelab during my free time from work

I wanted to share a critical undocumented change (or potential API bug) I’ve encountered while deploying a cold data pool on Windows Server 2025, which completely breaks traditional ReFS performance optimizations that worked flawlessly in Windows Server 2022.

The Goal:

Deploying a 5-disk mechanical drive array in RAID 5 (Parity) using Storage Spaces.

(Before the Ryan Reynolds “But why?” meme drops in the comments: Yes, I know spinning rust and Storage Spaces parity isn’t the most glamorous setup for a homelab. However, this is the exact stack our organization is thoroughly used to at work, and I wanted to replicate it here to squeeze every bit of performance out of it!)

To prevent Read-Modify-Write (RMW) penalties and Write Amplification, the objective is to align the Stripe Data Width perfectly with the maximum ReFS Allocation Unit Size (64KB).

With 5 columns, we have 4 data columns (5 - 1 parity). In Windows Server 2022, setting -Interleave 16KB resulted in:

4 columns x 16KB interleave = 64KB stripe data width

This matched the 64KB ReFS AUS beautifully, providing atomic writes and native performance on spinning rust.

The “You shall not pass (this command, I might add)” problem in Windows Server 2025:

When executing the exact same deployment command in WS 2025, the StorageWMI provider throws a Not Supported error:

New-VirtualDisk -FriendlyName “RAID 5 Cold Data” -StoragePoolFriendlyName “RAID 5 Constellation” -Interleave 16KB -UseMaximumSize -ResiliencySettingName Parity -NumberOfColumns 5

New-VirtualDisk : Not Supported
Activity ID: {45a547e2-fce9-0006-8a48-a545e9fcdc01}
At line:1 char:1

  • New-VirtualDisk -FriendlyName “RAID 5 Cold Data” -StoragePoolFriendly …
   + CategoryInfo          : InvalidOperation: (StorageWMI:ROOT/Microsoft/...SFT_StoragePool) [New-VirtualDisk], CimException
   + FullyQualifiedErrorId : StorageWMI 1,New-VirtualDisk 

The Inconsistency: Fixed vs. Thin

After extensive testing, I found a bizarre discrepancy in the API validation logic:

  • If you create the Virtual Disk as -ProvisioningType Fixed (or leave it default/maximum size), it strictly rejects any Interleave value below 64KB.
  • However, if you explicitly change the command to -ProvisioningType Thin, the API happily accepts -Interleave 16KB (or lower) without throwing any errors.

This proves that the underlying storage architecture is technically capable of handling a 16KB interleave on the parity engine, but Windows Server 2025 is enforcing an undocumented API restriction specifically on Fixed disks. Forcing a minimum 64KB interleave on a 5-column fixed layout increases the Stripe Data Width to 256KB, resulting in heavy RMW performance degradation for standard 64KB ReFS clusters.

How to help:

I have already filed a high-priority feedback report directly to the Microsoft Core Storage engineering team. My organization is seriously considering dropping Windows Server entirely for our data storage if this isn’t addressed, as we cannot deploy our planned ReFS architecture.

If you are running Windows Server 2025 in your lab or enterprise environment, could you please try to replicate this behavior? If you want this feature back, please upvote my feedback report on the Feedback Hub using this link: https://aka.ms/AA11tvzr (Note: You need a Windows machine with Feedback Hub installed to open this link).

Has anyone found a clean registry bypass or workaround for this validation lock on W. S. 2025 Storage Pools?

P.S: To all the Proxmox + ZFS on TrueNAS enthusiasts out there: if “Microslop” (we all know we hate them for many reasons, and more) decides not to fix this issue, we will fully transition to ZFS in the end and become “part of the ship, part of the crew”.

Thanks in advance!

1 Like

I stopped using ReFS after the last time it ate my data. I’m glad to see that there are still improvements being made to it though, it does feel like there are not many good alternatives to ZFS these days and a monoculture is never good for the ecosystem, even with software.

While I use Storage Spaces, I stayed with NTFS. Has not given me any issues in the… gosh darn it I had this setup for 8 years already??

I recommend to stop using ReFS. I learned it hard way. Some files silently changed to random sequence, there is no checkdisk that actually works on ReFS, is not recoverable when things go wrong…

Thanks for the feedback, everyone and apologies for the late response.

I totally get the historical distrust toward ReFS (we’ve all heard or lived through the early Windows Server 2016 horror stories when implemented on production servers it broke and corrupted all info on them). However, we didn’t pick Windows and ReFS blindly. In our organization, we rely on Windows Server because it’s the operating ecosystem we are most accustomed for, and we specifically chose ReFS to benefit from its core enterprise advantages: we need real-time data integrity streams (-Enable=true on and -Enforce=false, we don’t want to delete completely “bad files” if flagged before verification as Enforce=true does) and native scrubbing to combat silent data corruption (bit rot), which is something NTFS simply does not provide on its own.

Furthermore, we planned our topology with modern backup standards. For instance, tools like Veeam heavily recommend formatting repositories with an ReFS 64KB Allocation Unit Size to leverage Fast Clone / Block Cloning technology, as they show on their forums, IIRC. This allows for near-instantaneous synthetic full backups and massive space savings.

However, when Microsoft has probably introduced an arbitrary API block that breaks mechanical drive alignments for a 64KB cluster layout, they weren’t just messing with a homelab setup; they were breaking a fully intentional engineering architecture designed to protect production data while squeezing performance out of spinning rust.

The architectural inconsistency is what bothers me the most. The storage engine can clearly handle a 16KB interleave on this 5-column parity layout because it works flawlessly when using -ProvisioningType Thin as I have shown previously. Hard-locking -ProvisioningType Fixed to a minimum of 64KB interleave is just a lazy StorageWMI validation block that forces a 256KB Stripe Width onto a 64KB file system (as ReFS max AUS is 64KB), which tanks performance with heavy Read-Modify-Write (RMW) cycles and forces users to use the “Mirror-accelerated parity” setup (which can be avoided if it worked as in Windows Server 2022).

Bumping this to see if anyone running Windows Server 2025 has found a way to bypass this validation lock via registry tweaks or direct CIM/WMI object manipulation?

Cheers!

I got to agree with you, there seems to be no documentation other than this seemingly forced imposition when mixing -ProvisioningType Fixed and -Interleave 16KB.

This seems to have been identified back in late 2025 as well, with no answer; but I’m guessing you’re already aware of this.

But for context, the issue is not with ReFS, but with Storage Spaces itself. Not sure why this is an issue now, or what was the need to put this change.

Now if you want to get “creative”, you can create the storage pool in WS2022, and just move the disks to WS2025 as is, without any encryption. That SHOULD work in theory. (i.e. I have moved my Storage Space around servers with no issue and no need to reconfigure anything like this, between 2022 and 2025)

It will work until it doesn’t, after some update. Not a great plan

been there with the nic teaming feature, took them a while to fully remove all loopholes to enable it manually on win Pro, but eventually it stopped working alltogether

Fair point!

However, I may state that NIC Teaming is still a thing on Windows Server? Not quite potato-potato, but we’ll see, this is a test environment IIRC.

Final Update:

Hello there once again,

After one month digging information regarding this issue, it looks like the mystery is officially solved, and, as suspected, Microsoft has left us with a broken storage stack in Windows Server 2025. TL;DR: do not use Windows Server 2025 if you strive for storage optimization, it is completely broken and they will not fix it any time soon. The last version with ReFS optimization is Windows Server 2022.

Searching through community deployments and Microsoft Q&A tickets, I have confirmed that multiple sysadmins are running into this exact same issue, as teltersat has shown, among others. Anyone trying to deploy a traditional 5-column Parity virtual disk with -ProvisioningType Fixed and a low interleave (like 16KB) is getting slapped in the face with the exact same New-VirtualDisk : Not Supported API block. The second they switch to Thin or raise it to 64KB, the validation passes and the virtual disk is created.

To make matters worse, a quick Google search has shown telemetry from other enterprise users and revealed that Windows Server 2025’s storage subsystem is facing severe regressions. The redirection routine (RaidpAdapterRedirectDpcRoutine inside storport.sys) is currently consuming up to 1.70x more CPU overhead compared to Windows Server 2022 under identical workloads, and recent cumulative updates have done absolutely nothing to address it, as shown in this Microsoft question, Windows Server 2025 storage performance issue - Microsoft Q&A.

Basically, Microsoft has crippled the classic and rock-solid spinning rust optimizations to heavily push their new Azure Stack HCI / Mirror-Accelerated Parity agendas, leaving standard local storage deployments out in the cold. Since their corporate support has remained completely silent after weeks of high-priority escalation, my organization is officially pulling the plug on Windows Server for this project. We refuse to compromise data integrity on NTFS or suffer performance degradation on unaligned ReFS. At most, we will roll back and deploy a WS 2022 instance, but management has lost faith in the ecosystem given the circumstances.

To all the Proxmox + ZFS on TrueNAS enthusiasts here: Microsoft didn’t give us a choice. I’ll pass the final report to my organization and we’ll begin wiping the drives, migrating the whole backup and render ecosystem to Proxmox VE.

“Part of the ship, part of the crew.” See you guys on the ZFS side!

I’ll leave this thread here as a warning for anyone else planning their WS 2025 deployments. Thank you so much to everyone who chipped in and helped validate this mess!

Cheers!

1 Like

Im confused on this, as I have never seen 64KB be the maximum size the allocation unit size can be set to. The AUS can go up into the gigabyte range from my experience with it, in fact I had a hard time getting it as low as I wanted and something as low as 64KB was rejected for me as too small.

My interleaves are usually 128KB because of the issues getting AUS small enough. These are some of the flags I use for example:

 -NumberOfColumns 5 -ProvisioningType Fixed -AllocationUnitSize 1024MB -Interleave 128KB -WriteCacheSize 2GB

I did however notice significant regressions in Storage Spaces allowed command flags as you have said on Windows 11 Workstation/Server 2025. It doesnt allow nearly as much as Windows 10 Workstation/Server 2022. It would not accept my New-VirtualDisk command, and gave the same “Not Supported” error as you have seen from some of your commands. I had to go make the array on a Win 10 Workstation PC instead and then move the drives over. I only use fixed provisioning, and likely could have gotten around the problems if I have tried thin provision, but I never thought of that.