I'm not totally up to speed of how PBS functions. But I noticed that ZFS seems to be the preferred repository to use and was wondering if a specific record size or other tweaks should be considered when used in this case to combat fragmentation or write amplification. From what I understand changes are tracked in many chunk filed. They seem to be relatively small from what I'm reading.
Backup Repo ZFS record size considerations?
- Thread starter jayg30
- Start date
There are 4 MiB chunks for block-level backup and 64KiB to 4MiB for file level backup.
What really helps on a ZFS pool with spinners can be a mirrored (for safety!) special device on fast SSDs or NVMe devices:
https://pbs.proxmox.com/docs/sysadmin.html#zfs-special-device
Lot's of memory can also help, as then the ZFS ARC, or the kernel page cache for other filesystems, can host frequently used data in memory to accessed very fast.
Regarding record size, this depends a bit on what you plan to backup. If you only backup VMs on block level you could in fact use a larger record size, but you may not see a to big impact.
If you do both, file and block level, then the default 128 record size should be a OK tradeof, write amplification normally happens if a much smaller chunk of data than the record size is written, with the minimum of 64KiB chunks and 128KiB record size this shouldn't have such an impact either (as not all chunks are 64KiB).
What really helps on a ZFS pool with spinners can be a mirrored (for safety!) special device on fast SSDs or NVMe devices:
https://pbs.proxmox.com/docs/sysadmin.html#zfs-special-device
Lot's of memory can also help, as then the ZFS ARC, or the kernel page cache for other filesystems, can host frequently used data in memory to accessed very fast.
Regarding record size, this depends a bit on what you plan to backup. If you only backup VMs on block level you could in fact use a larger record size, but you may not see a to big impact.
If you do both, file and block level, then the default 128 record size should be a OK tradeof, write amplification normally happens if a much smaller chunk of data than the record size is written, with the minimum of 64KiB chunks and 128KiB record size this shouldn't have such an impact either (as not all chunks are 64KiB).
