thx for the insights ... I am not 100% sure what your practical advice
is ...
could you sum that up to
"do rather not use btrfs/zfs for qcow images to begin with but any
other FS (ext4, xfs) with LVM?"
In my case it is not that I am running heavy workloads on such VM's.
Practical current usecase is setting up 2 VM's to practive for LPIC-1.
So they will be gone again sooner or later.
all best
Gunnar
On 29.04.25 23:17, František Ryšánek wrote:
Regarding QEMU guest HDD images on btrfs:
Both BTRFS and ZFS are COW-based filesystems and I've come across some
warning notes in the interwebs, that this COW mechanism results in
unpleasant and worsening fragmentation problem, if such FS is used as
underlying storage for files that get "overwritten in place" a lot -
such as guest VM hdd images or database extents. Fragmentation hurts
even on SSD's, because it multiplies the number of IOps required to
read the file. It is especially detrimental on spinning rust with its
slow heads (some 60-100 random IOps).
ZFS is namely good for storing backups = files that you store once,
but after that you leave them alone (there are no in-place rewrites).
For backups, the dedupe features of ZFS come in handy.
In btrfs, the COW can be suppressed, using the "nodatacow" mount
option (on a whole BTRFS volume), or you can apply it on individual
subvolumes using the chattr +C command. This will rid you of COW and
fragmentation for normal operation. But, if you try some snapshots,
COW will still activate, because there's no way around it for snapshots.
On a tangential note, the DeviceMapper subsystem in Linux, which is
the basis of Linux LVM, has a "snapshot-origin" option (an option to
dmsetup). Yes, DeviceMapper and LVM can do snapshots too, at the block
layer. By default, COW allocates new blocks for a write, and splices
them to the live "origin" file, whereas the snapshot file receives the
stale blocks just deallocated from the origin file...
So the "snapshot-origin" option turns this around, data are replaced
in-place in the origin file upon mod, and the snapshot file gets a few
extra blocks allocated, and the old data from the origin file gets
copied onto this new storage...
Have you checked Proxmox? It has a nice GUI, and uses the LVM for
storage (gives LV's to guests).
To save disk space on the hypervisor instance, you should aim to use
sparse image files or thin-provisione LVM. And, couple these to
DISCARD emulation in the guest (in the virtio storage device).
Frank
