Veeam server on VM with thin provisioned storage for backups

[NetFire-mdh]

Hey folks,
Another engineer and I are discussing how to deploy Veeam. We've previously used bare metal hardware for our Veeam hosts, but we've decided to change our platform up, and Veeam Backup and Replication will now be running in Hyper-V virtual machines.

I am proposing one of two solutions: we can either provide the raw disk to the Veeam server guest in Hyper-V, or we can utilize a thickly-provisioned VHDX for Veeam's backup storage.

The other engineer is proposing that we should thin provision VHDX's on the Hyper-V host for the Veeam guest's backup storage.

I believe that thin vs thick provisioned, performance will likely be similar or the same initially: this isn't something we can just easily take some benchmarks of. However, I also believe that over time, as multiple thinly-provisioned VHDX files compete for space on the physical volume and fragmentation occurs both inside those VHDXs (on the ReFS volume in the Veeam guest) and on the host system's NTFS filesystem where the VHDX files would exist, performance would slowly degrade over time until things were quite bad.

Can anyone weigh in on this, or on any of the potential caveats of either solution? Any data would be appreciated of course.

Much appreciated!

[HannesK]

Hello,
although you already decided for the more complex virtual setup, I will still recommend KISS (keep it stupid simple). Backup server as VM (only the management server) is okay for installations with geo-redundant storage with transparent failover. Agree if you have that and more than 1000 VMs (my rule of thumb is "one server for up to 1000 VMs"). For the repository I would go for a physical machine to avoid all the complexity of virtual filesystems etc.

I did not read anything about your storage. In most cases, the storage vendor recommends the setup with thin / thick.

Yes, with REFS fragmentation it can become slower over time. But no idea how to predict or really measure it. I prefer thick-provisioning in general, as it makes sure, that critical systems are not over-provisioned. I don't believe, that anyone can seriously predict the fragmentation performance reduction.

Best regards,
Hannes

[nmdange]

I would stick with vhdx rather than presenting the storage raw, because then you can storage migrate your refs volumes to new hardware without losing the block cloning space savings. I've been running this way with thin vhdx for over a year without any noticeable performance drops.