Does forcing 4Kn on VM disks actually help?

[mercifultape]

Running PVE on hosts with 512/512 disks, storage is ZFS (ashift=12) and Ceph.
Guests use virtio-scsi-single, mostly Linux, some Windows.

I wonder if it's worth adding (or similiar):

args: -global scsi-hd.physical_block_size=4096 -global scsi-hd.logical_block_size=4096

For Linux it feels pointless - page cache is 4K so guests should already send 4K
writes regardless - except maybe for O_DIRECT workloads, but most DBs should use >= 4K pages anyway.

Questions:
1. On 512/512 disks with ZFS (ashift=12) / Ceph underneath, does this give any performance gain for Linux/Windows guests?
2. Does it change anything on native 4Kn disks?
3. Has anybody benchmarked it and can provide results?

Thanks.

 

[psalkiewicz]

Hi, In terms of performance I wouldn't expect any measurable gain.

The real-world gain for Linux is mostly at partition alignment and filesystem mkfs time — tools like parted, fdisk, and mkfs.xfs/mkfs.ext4 read the reported physical block size and align accordingly. If your logical block size is 512, some tools may create slightly suboptimal layouts, though most modern tools default to 1MB alignment anyway, making this moot in practice.

It's good for the sake of correctness but I wouldn't expect any significant improvement. It's more of an academic discussion than real world tweak

 

[stanleybaxter]

Hi, In terms of performance I wouldn't expect any measurable gain.

The real-world gain for Linux is mostly at partition alignment and filesystem mkfs time — tools like parted, fdisk, and mkfs.xfs/mkfs.ext4 read the reported physical block size and align accordingly. If your logical block size is 512, some tools may create slightly suboptimal layouts, though most modern tools default to 1MB alignment anyway, making this moot in practice.

It's good for the sake of correctness but I wouldn't expect any significant speed stars improvement. It's more of an academic discussion than real world tweak

That makes sense, thanks for the clear explanation. Sounds like it’s more about doing things “right” rather than expecting real performance gains, especially with modern tools already handling alignment well. Good to know it’s not something worth over-optimizing in most setups.