Allocation unit size is chosen once, at format time, and then silently affects every read and write for the life of the volume. On a Windows VPS running SQL Server, the default 4 KB NTFS cluster forces most database I/O into read-modify-write operations that trickle across multiple clusters — and on network-backed or thin-provisioned virtual disks that is measurable. This guide covers the cluster sizes that matter, the 64 KB SQL Server sweet spot, its interaction with physical sector and VMFS/block-storage geometry, and how to fix an existing volume without reinstalling.
What Allocation Unit Size Actually Controls
NTFS divides the volume into clusters (“allocation units”) and allocates file space in whole clusters. A 4 KB cluster means a 60 KB SQL Server read is served by 15 separate cluster reads; a 64 KB cluster means one. The difference is not about wasting space for small files — SQL data files are huge and sequential — it is about how many distinct I/O operations the filesystem and the virtual disk have to perform per logical request, and how many page fragments the storage layer must reassemble.
ReFS behaves differently: it uses a fixed 64 KB block and cannot be formatted with another allocation unit size through the Format dialog, which is one reason DB volumes usually stay on NTFS.
Recommended Sizes by Workload
| Workload | File system | Allocation unit | Rationale |
|---|---|---|---|
| SQL Server data (.mdf/.ndf) | NTFS | 64 KB | Matches SQL Server’s 64 KB extent; one extent read = one cluster |
| SQL Server log (.ldf) | NTFS | 64 KB | Sequential writes; fewer clusters per log flush |
| SQL Server backup target | NTFS | 64 KB | Streaming sequential I/O; also better for compression |
| IIS content root (wwwroot) | NTFS | 4 KB (default) | Many small files; large clusters waste space |
| Windows system volume (C:) | NTFS | 4 KB (default) | Never change this; updates and WinSxS assume it |
| Backup/archive volumes with large files | NTFS | 64 KB | Fewer metadata updates per GB |
The rule of thumb: 64 KB for anything dominated by large sequential or extent-aligned I/O, 4 KB for volumes that hold tens of thousands of small files. A single volume holding both is a compromise; on anything larger than a few hundred GB, split the database files onto their own VHDX with 64 KB clusters.
Aligning With the Virtual Disk and Physical Sector
The cluster size only helps if it lands on an aligned boundary of the underlying storage. NTFS formats created on Windows Server 2012 and later are 1 MB-aligned by default, which is correct for virtually all modern virtual disks. Verify rather than assume:
# Physical sector size and logical sector size of the disk
Get-Disk | Format-Table Number,FriendlyName,PartitionStyle,BusType,LogicalSectorSize,PhysicalSectorSize
# Allocation unit size per volume
Get-Volume | Format-Table DriveLetter,FileSystem,AllocationUnitSize,SizeRemaining
# Free space and cluster count sanity check
fsutil fsinfo ntfsinfo D:
On a VPS, PhysicalSectorSize is often reported as 4096 even when the backing storage uses 512e emulation. A 64 KB cluster is an integer multiple of every one of these, so there is little to tune. What matters is avoiding a volume offset of, say, 32 KB against a 64 KB stripe — which happens when a partition was created by an old tool.
Formatting a New Data Volume Correctly
Attach the VHDX or second data disk, initialise it as GPT, then format with an explicit allocation unit size. Do not use the Format dialog’s default and then try to fix it later:
# Initialise and format a new data disk as 64 KB NTFS
Initialize-Disk -Number 1 -PartitionStyle GPT
New-Partition -DiskNumber 1 -UseMaximumSize -DriveLetter D
Format-Volume -DriveLetter D -FileSystem NTFS -AllocationUnitSize 65536 `
-NewFileSystemLabel 'SQLData' -Confirm:$false
# Confirm the result
Get-Volume -DriveLetter D | Format-List DriveLetter,FileSystem,AllocationUnitSize
For an existing volume, Format-Volume with -AllocationUnitSize will refuse to reformat without confirmation. If the volume holds data, the only safe conversion path is a new volume plus a file copy. That is the usual reason this decision should be made when the VPS is provisioned, not after launch. The Windows VPS hosting plans and specifications page is where you size the data disk before the first boot.
Converting an Existing Volume Without Reinstalling Windows
There is no supported in-place command to change NTFS allocation unit size — fsutil, chkdsk, and defrag will not do it. The practical procedure on a VPS is a second data disk and a cutover:
- Attach a new virtual disk of equivalent or larger size and format it with 64 KB clusters.
- Put the database in read-only, or better, stop the SQL Server service to guarantee a consistent copy.
- Copy with robocopy /MIR /COPYALL /J to preserve ACLs and use unbuffered I/O.
- Detach the old volume or rename it, then mount the new one at the original drive letter.
- Start SQL Server, run DBCC CHECKDB, and take a fresh full backup before reopening to clients.
Expect 20–60 minutes of downtime for a mid-sized database. Schedule the cutover; do not improvise it during business hours.
When 64 KB Hurts: The Counter-Arguments
Large clusters are not free. A 64 KB cluster means a 1 KB file consumes 64 KB, so a volume full of small files — a multi-million-file log directory or a git working tree — can lose 20–40% of its capacity to slack space. In those cases, use the default 4 KB and accept the extra clusters per I/O. Also note that Windows Server deduplication on NTFS requires the volume to use 4 KB clusters; if you rely on dedup for a file server role, you cannot adopt 64 KB there.
SQL Server does not care about slack because its files are pre-allocated and large, but it does care about the instant file initialisation setting and the 64 KB extent size — which is exactly why the alignment matters. A 64 KB cluster that is misaligned with the underlying storage can be worse than a well-aligned 4 KB cluster, so verify with the commands above before assuming the bigger number is better.
Verifying the Gain Instead of Trusting the Theory
Measure before and after with \PhysicalDisk(*)\Avg. Disk sec/Transfer and SQL Server’s own \SQLServer:Databases(*)\Log Write Waits. On a spinning-backed virtual disk the gain is usually visible as lower average transfer latency; on NVMe it may be within noise, because the device’s own queue depth hides cluster-count differences. The right conclusion is the measured one, not the textbook one — and on a small VPS, the config effort should follow the bottleneck. If I/O latency remains high after correct alignment, the constraint is the host’s storage tier, and the entry-level options compared on the Windows VPS hosting plans and specifications page explain which tiers provide NVMe-backed volumes.


