Defrag and Compact on Large VM's

Curious what y’all are doing in this sort of situation. 

I have a client with a rather large VM (95 TB, consisting of VMDK’s at 100GB (OS), 30TB, 19TB, 20TB, 9TB and 17TB).  The repo (Synology NAS) is about 90TB of usable space with a second NAS at 24TB. The volumes are ISCSI presented as RDM’s to the virtual Veeam server running 2012 R2.  Volumes are formatted REFS.

The backup job for this server is split into three jobs...one for the OS disk and the 9TB and 30TB disks (active data) to the primary NAS, one for the 19TB and 20TB disks (Archive Data) to the Primary NAS, and one for the 17TB disk (Archive Data from an old location) to the secondary NAS).

Current backup jobs are forever forward incrementals keeping 31 days of restore points, but as such, we have defragments and compactions running, but they take a VERY long time (the first mentioned job has been running for almost 16 hours and is at 6%, and that’s with only 3 restore points due to an issue I fixed yesterday where the volumes were showing as RAW due to a bad patch that the assigned engineer just noticed after weeks of failures.  Given the size of the VM, we don’t have a ton of space to play with on either multiple fulls or with the compaction process.  I believe we had either forward or reverse incrementals running before (non-forever forward) but I think we ran into issues with filling up the repo.

Just curious what strategy you’d take to make this better.  We are using REFS, but it feels like we’re not necessarily getting a lot out of it in regards to block-cloning.  Otherwise I’d consider turning on synthetic fulls.  Are we at a limitation due to 2012 R2 perhaps?  Any other ideas on how to make this more efficient?

Ultimately, I’d like to get a new, purpose-built Veeam server on site with lots of local storage, but the NAS they’re currently using is only 2 or 3 years old, so could be a hard sell.  We can repurpose it for sure for a second copy or something like that as I don’t like using the NAS, for sure not with REFS on it.  Any ideas of what we can do for the time being?