Some of my profiling made it seem like the VHD backends were doing their low-level I/O with much bigger blocks than QCOW2 ... would that depend on the cluster size?

@olivierlambert said:

Might be interesting to test a different cluster size and see the impact

I tested it with a cluster size of 2 megabytes, and nothing changed

https://docs.xcp-ng.org/storage/qcow2_faq/#can-we-change-the-qcow2-cluster-size

I have performed some quick I/O benchmarks comparing the new QCOW2 format against the traditional VHD. In my tests, QCOW2 appears significantly slower than VHD on my hardware.

Test Environment
Hypervisor: Dell PowerEdge R420

CPU: Intel Xeon E5-2470 v2

Storage: Intel SSDPELKX010T8 NVMe

VM OS: Debian 13

VM Specs: 2 vCPUs, 1GB RAM

Setup: One 10GB VHD and one 10GB QCOW2 disk, both pre-filled from /dev/random.

Methodology
I used a custom test suite available here: https://vm01.unsoft.hu/~ventura/fio/fio_test_20250408.tar.gz

I also ran a simplefio loop with the following results:
VHD:

root@Debian-13-CloudInit-20250810:/mnt/vhd# for mode in read write; do for jobs in 1 16; do for bs in 4 64; do for t in "" rand; do printf "%2i qd %2ik % 4s " $jobs $bs $t; fio --name=random-write --rw=$t$mode --bs=${bs}k --numjobs=1 --size=1g --iodepth=$jobs --runtime=10 --time_based --direct=1 --ioengine=libaio|grep -e BW -e runt ; done; done; done; done
 1 qd  4k        read: IOPS=9625, BW=37.6MiB/s (39.4MB/s)(376MiB/10001msec)
 1 qd  4k rand   read: IOPS=5414, BW=21.2MiB/s (22.2MB/s)(212MiB/10001msec)
 1 qd 64k        read: IOPS=2657, BW=166MiB/s (174MB/s)(1661MiB/10001msec)
 1 qd 64k rand   read: IOPS=2575, BW=161MiB/s (169MB/s)(1610MiB/10001msec)
16 qd  4k        read: IOPS=45.7k, BW=178MiB/s (187MB/s)(1785MiB/10001msec)
16 qd  4k rand   read: IOPS=45.9k, BW=179MiB/s (188MB/s)(1794MiB/10001msec)
16 qd 64k        read: IOPS=16.7k, BW=1041MiB/s (1092MB/s)(10.2GiB/10001msec)
16 qd 64k rand   read: IOPS=16.7k, BW=1042MiB/s (1093MB/s)(10.2GiB/10001msec)
 1 qd  4k        write: IOPS=8842, BW=34.5MiB/s (36.2MB/s)(345MiB/10001msec); 0 zone resets
 1 qd  4k rand   write: IOPS=8880, BW=34.7MiB/s (36.4MB/s)(347MiB/10001msec); 0 zone resets
 1 qd 64k        write: IOPS=6095, BW=381MiB/s (399MB/s)(3810MiB/10001msec); 0 zone resets
 1 qd 64k rand   write: IOPS=6006, BW=375MiB/s (394MB/s)(3755MiB/10001msec); 0 zone resets
16 qd  4k        write: IOPS=49.3k, BW=193MiB/s (202MB/s)(1928MiB/10001msec); 0 zone resets
16 qd  4k rand   write: IOPS=47.3k, BW=185MiB/s (194MB/s)(1848MiB/10001msec); 0 zone resets
16 qd 64k        write: IOPS=14.3k, BW=891MiB/s (934MB/s)(8910MiB/10001msec); 0 zone resets
16 qd 64k rand   write: IOPS=15.5k, BW=966MiB/s (1013MB/s)(9663MiB/10001msec); 0 zone resets

QCOW2

root@Debian-13-CloudInit-20250810:/mnt/qcow2# for mode in read write; do for jobs in 1 16; do for bs in 4 64; do for t in "" rand; do printf "%2i qd %2ik % 4s " $jobs $bs $t; fio --name=random-write --rw=$t$mode --bs=${bs}k --numjobs=1 --size=1g --iodepth=$jobs --runtime=10 --time_based --direct=1 --ioengine=libaio|grep -e BW -e runt ; done; done; done; done
 1 qd  4k        read: IOPS=5866, BW=22.9MiB/s (24.0MB/s)(229MiB/10001msec)
 1 qd  4k rand   read: IOPS=4000, BW=15.6MiB/s (16.4MB/s)(156MiB/10001msec)
 1 qd 64k        read: IOPS=2229, BW=139MiB/s (146MB/s)(1394MiB/10001msec)
 1 qd 64k rand   read: IOPS=2161, BW=135MiB/s (142MB/s)(1351MiB/10001msec)
16 qd  4k        read: IOPS=16.9k, BW=66.2MiB/s (69.4MB/s)(662MiB/10001msec)
16 qd  4k rand   read: IOPS=17.6k, BW=68.8MiB/s (72.1MB/s)(688MiB/10001msec)
16 qd 64k        read: IOPS=7244, BW=453MiB/s (475MB/s)(4529MiB/10002msec)
16 qd 64k rand   read: IOPS=6994, BW=437MiB/s (458MB/s)(4372MiB/10002msec)
 1 qd  4k        write: IOPS=5551, BW=21.7MiB/s (22.7MB/s)(217MiB/10001msec); 0 zone resets
 1 qd  4k rand   write: IOPS=5159, BW=20.2MiB/s (21.1MB/s)(202MiB/10001msec); 0 zone resets
 1 qd 64k        write: IOPS=4024, BW=252MiB/s (264MB/s)(2515MiB/10001msec); 0 zone resets
 1 qd 64k rand   write: IOPS=4027, BW=252MiB/s (264MB/s)(2517MiB/10001msec); 0 zone resets
16 qd  4k        write: IOPS=14.5k, BW=56.8MiB/s (59.6MB/s)(568MiB/10002msec); 0 zone resets
16 qd  4k rand   write: IOPS=14.0k, BW=54.7MiB/s (57.4MB/s)(547MiB/10001msec); 0 zone resets
16 qd 64k        write: IOPS=6360, BW=398MiB/s (417MB/s)(3976MiB/10002msec); 0 zone resets
16 qd 64k rand   write: IOPS=6090, BW=381MiB/s (399MB/s)(3807MiB/10002msec); 0 zone resets

I would be interested to know if I'm overlooking something, or if the qcow2 format simply provides lower performance compared to VHD for the time being?

https://xcp-ng.org/blog/2026/05/05/qcow2-is-now-ga-in-xcp-ng/

You’ll have about three days to share your feedback if you’d like to be part of this final sprint .

Details at https://xcp-ng.org/forum/post/104961

In the XO5 storage UI, the disks appear, but their names and descriptions are lost, and even though they are currently attached to running VMs, the system doesn't recognize this:

If I try to assign a new name to one of the VDIs as an experiment, I get a "VDI does not exist" error, even though xe vdi-list does show all of these VDIs.

On the other hand, I have running VMs where QCOW2 VDIs are attached and mounted, but the xe vm-disk-list command doesn't show them.

I see that there's a new batch of updates from a few days ago — any chance they will address this?

@stormi said:

This is not exposed in Xen Orchestra yet.

When will it likely be available in the GUI?

QCOW2 FAQ

What storage space available do I need to have on my SR to have large QCOW2 disks to support snapshots?

Depending on a thin or thick allocated SR type, the answer is the same as VHD.
A thin allocated is almost free, just a bit of data for the metadata of a few new VDI.

For thick allocated, you need the space for the base copy, the snapshot and the active disk.

Must I create new SRs to create large disks?

No. Most existing SR will support QCOW2. LinstorSR and SMBSR (for VDI) does not support QCOW2.

Can we have multiples different type of VDIs (VHD and QCOW2) on the same SR?

Yes, it’s supported, any existing SR (unless unsupported e.g. linstor) will be able to create QCOW2 beside VHD after installing the new sm package

What happen in Live migration scenarios?

preferred-image-formats on the PBD of the master of a SR will choose the destination format in case of a migration.

Can we create QCOW2 VDI from XO?

XO hasn’t yet added the possibility to choose the image format at the VDI creation.
But if you try to create a VDI bigger than 2TiB on a SR without any preferred image formats configuration or if preferred image formats contains QCOW2, it will create a QCOW2.

Can we change the cluster size?

Yes, on File based SR, you can create a QCOW2 with a different cluster size with the command:

qemu-img create -f qcow2 -o cluster_size=2M $(uuidgen).qcow2 10G
xe sr-scan uuid=<SR UUID> # to introduce it in the XAPI

The qemu-img command will print the name, the VDI is <VDI UUI>.qcow2 from the output.

We have not exposed the cluster size in any API call, which would allow you to create these VDIs more easily.

Can you create a SR which only ever manages QCOW2 disks? How?

Yes, you can by setting the preferred-image-formats parameter to only qcow2.

Can you convert an existing SR so that it only manages QCOW2 disks? If so, and it had VHDs, what happens to them?

You can modify a SR to manage QCOW2 by modifying the preferred-image-formats parameter of the PBD’s device-config.

Modifying the PBD necessitates to delete it and recreate it with the new parameter. This implies stopping access to all VDIs of the SR on the master (you can for shared SR migrate all VMs with VDIs on other hosts in the pool and temporarily stop the PBD of the master to recreate it, the parameter only need to be set on the PBD of the master).

If the SR had VHDs, they will continue to exist and be usable but won’t be automatically transformed in QCOW2.

Can I resize my VDI above 2 TiB?

A disk in VHD format can’t be resized above 2 TiB, no automatic format change is implemented.
It is technically possible to resize above 2 TiB following a migration that would have transferred the VDI to QCOW2.

Is there any thing to do to enable the new feature?

Installing updated packages that supports QCOW2 is enough to enable the new feature (packages: xapi, sm, blktap). Creating a VDI bigger than 2 TiB in XO will create a QCOW2 VDI instead of failing.

Can I create QCOW2 disks lesser than 2 TiB?

Yes, but you need to create it manually while setting sm-config:image-format=qcow2 or configure preferred image formats on the SR.

Is QCOW2 format the default format now? Is it the best practice?

We kept VHD as the default format in order to limit the impact on production. In the future, QCOW2 will become the default image format for new disks, and VHD progressively deprecated.

What’s the maximum disk size?

The current limit is set to 16 TiB. It’s not a technical limit, it’s a limit that we corresponds to what we tested. We will raise it progressively in the future.

We’ll be able to go up to 64 TiB before meeting a new technical limit related to live migration support, that we will adress at this point.

The theoretical maximum is even higher. We’re not limited by the image format anymore.

Can I import without modification my KVM QCOW2 disk in XCP-ng?

No. You can import them, but they need to be configured to boot with the drivers like in this documentation: https://docs.xcp-ng.org/installation/migrate-to-xcp-ng/#-from-kvm-libvirt
You can just skip the conversion to VHD.

So it should work depending on different configuration.