PBS with HDDs

[benoitc]

The doc and anyone of this forum is advising to use SSDs for PBS but I didn't found a performance doc that can show the differences.

I have a cluster of 3 machines (3x 2T SAS HDD) that I want to backup. Would it be enough fast to use a DL380 witjh 6x4T disks in 1+0 zfs raid? What kind of performances / issues should I expect from it?

 

[Dunuin]

The doc and anyone of this forum is advising to use SSDs for PBS but I didn't found a performance doc that can show the differences.

I have a cluster of 3 machines (3x 2T SAS HDD) that I want to backup. Would it be enough fast to use a DL380 witjh 6x4T disks in 1+0 zfs raid? What kind of performances / issues should I expect from it?

PBS stores everything as small chunk files of max 4MB in size. Usually its more like 2MB per chunk. So if you backup 12 TB of guests that will result in roughly 6 million chunk files. PBS cleans up data by running a GC task that has to read and write the atime of each of those chunk files. So a daily GC task would create atleast 12 million IO (6 million times reading atime + 6 million times writing new atime). Now lets say your HDDs got a IOPS performance of 100 IOPS. With 6 disks in raid10 that would result in 300 IOPS. Divide 12 million IO by 300 IOPS and you get 40,000 seconds or over 11 hours. So yes, might work but won't be fast...a SSD could do that GC task in a few minutes.
At least use ZFS with two enterprise SSDs in mirror as special devices to store your metadata. A HDD-only datastore is a terrible idea.

 

[jdancer]

Granted that flash is the way to go but I do backup two Ceph clusters with a Dell R200 and Dell R620 using SAS drives.

These Dells are decommissioned and still functional so they make great PBS servers.

Not the fastest but it does backup/restore just fine.

PBS benchmarks https://forum.proxmox.com/threads/how-fast-is-your-backup-datastore-benchmark-tool.72750/