Virtualization / 2026-04-06
ZFS or LVM-thin: the real difference on a single-disk box
Same SSD, two storage backends: snapshots, write amplification and memory use, plus when ZFS is the wrong choice.
The "ZFS or LVM-thin" choice during a Proxmox install is usually clicked without much thought. Three months later, changing it is no longer a click.
Memory is the first gate
The ZFS ARC takes a large slice of RAM by default. On a 16 GB box running two or three containers plus one VM, an ARC above 8 GB is normal, and everything else lives at the edge of swap. The symptom is "nothing is running, but it feels slow".
LVM-thin does not have this problem: it barely uses memory, and the price is no checksums, no compression, no dedup.
The rule is simple: under 16 GB of RAM with several workloads, do not pick ZFS. Not because ZFS is bad, but because you do not have the resources it assumes.
Write amplification
Copy-on-write does cost endurance on small random writes. Same consumer SSD, measured:
| Item | ZFS (default recordsize) | LVM-thin |
|---|---|---|
| 4K random write IOPS | ~32k | ~51k |
| Sequential write | about equal | about equal |
| Daily writes, same workload | 1.8x | 1x |
For container workloads — many small files plus constant logging — that gap widens. Consumer SSDs already have low TBW; 1.8x the writes means roughly half the useful life.
Snapshots mean different things
Both snapshot in seconds, but rollback granularity differs. LVM-thin snapshots are block level and consume space in the same volume group; when the VG fills, snapshots stop working. ZFS snapshots are filesystem level and can be shipped with send/receive to another machine.
If off-site backup is a hard requirement, zfs send is far less work than running vzdump and moving the archive.
What I actually pick
- Single disk, under 16 GB, containers only → LVM-thin
- Several disks, RAID-Z, incremental off-site sync → ZFS
- LVM-thin for the system disk, a separate ZFS pool for data → the common compromise
One warning: old snapshots do not migrate between backends. Change it at install time, not after the data is there.
Swap on ZFS is a misconfiguration
ZFS and swap compete for memory. Once paging starts, performance falls off a cliff, and you can end up with a system that is alive but whose IO is wedged. Either disable swap or set swappiness to 1:
sysctl -w vm.swappiness=1
echo "vm.swappiness=1" >> /etc/sysctl.d/99-zfs.conf
Capacity is reported two ways
zfs list shows usable space, zpool list shows raw capacity, and they differ by 20–30% because of parity and reservation. Planning against the zpool number overestimates what you have, and ZFS degrades noticeably past 80% full — so a quota calculated from raw capacity lands you in the slow zone.