LXC backup slow vzdump to NFS storage

[foobar73]

pve-manager/7.1-11/8d529482 (running kernel: 5.13.19-6-pve)

LXC housed on ZFS

very small (8GB) LXC taking 1/2 an hour to backup to an NFS mount regardless of compression settings. It is a 10Gb/s wire to the NFS server.

It seems to be related to the tar command somehow, because a KVM backup is very fast.

To prove my theory, I ran vzdump with the stdout flag to tar -tv and it seems very, very slow.

Any ideas?

 

[Dunuin]

You could edit the /etc/vzdump.conf and change the "tmpdir" to a local SSD or even RAM disk. That way it will tar it locally and then send the archive to the NAS. By default vzdump would use your NFS share to store the temporary data.

 

[foobar73]

I forgot to say that I have two other identical hardware-wise nodes without this problem. The difference is those are running pve 6.3.

I tried the tmpdir option and nothing changed.

 

[foobar73]

I found the culprit but do not understand why it works, primarycache and secondarycache were both set to metadata. Setting them to all fixed it.

 

[Dunuin]

Is your NAS, PBS or PVE server using ZFS?

 

[foobar73]

yes, both are. PVE is ZFS, NAS is ZFS

 

[rvdmeer]

I found the culprit but do not understand why it works, primarycache and secondarycache were both set to metadata. Setting them to all fixed it.

Exactly the same here. Backup of a 670mb container took almost one hour. After changing primarycache and secondarycache to all it takes a couple of seconds (11 to be precise).

For anyone having the same:

$ zfs get primarycache
$ zfs set primarycache=all local-zfs
$ zfs set secondarycache=all local-zfs

Change local-zfs to your specific storagepool