Server 2019 ReFS and Windows Deduplication

[allankjaer]

In the what's new document for Veeam B&R 9.5 Update 4 is says:

Added experimental support for block cloning on deduplicated files for Windows Server 2019 ReFS. To enable this functionality,
create ReFSDedupeBlockClone (DWORD) value under HKEY_LOCAL_MACHINE\SOFTWARE\Veeam\Veeam Backup and Replication
registry key on the backup server.

But what value should the reg (DWORD) have?

[Gostev]

1

[kghague-ccx]

And this is set on each Repo server which has a deduped ReFS volume or only the primary backup server?

[Gostev]

This is only on the backup server.

[jrick]

I'm currently testing ReFS with Dedup on server 2019, I have added the ReFSDedupeBlockClone DWORD value to my backup server (Win server 2016).
My test job is configured to create a daily synth full but don't use the block clone feature.

An idea why?

[doktornotor]

Probably because it requires Server 2019?

[HannesK]

correct, that key is only working on Windows 2019

[DonZoomik]

On WS2019 I saw it already working on one very short chain. Wasn't particularly fast but there's a lot of IO load going on.

[DonZoomik]

I hit my first compact and it seems that fast clone is not fast on deduplicated ReFS. It's seems to just copy (not clone - free space on volume is reducing about as fast) blocks to new file and target file is not deduplicated (size on disk = size).
Currently I'd say that it's no faster (if not slower) than NTFS deduplication-based compact.

[DonZoomik]

I hit a bigger compact and it's just as slow as with NTFS deduplication, about 6 hours to compact ~2,5TB of VBKs.
Is this expected? ReFS volume was created on WS2019, chain was restarted using active full on WS2019 on ReFS (eg all data has always been on this ReFS volume) and job log shows that "[fast clone]" is being utilized.

[HannesK]

Hello,
if the job log says that "fast clone" is used, then it seems unexpected to me. Could you please open a support case and share the case number here?

Thanks,
Hannes

[DonZoomik]

#03408303

[j.forsythe]

Hi.
What is the file size limit for ReFS + DeDup at Server 2019?
Thank you!

[DonZoomik]

Only first 4TB of a file is processed by deduplication but there is no hard file size limit. 64TB volume limit places some practical limits though.

[YoMarK]

What is to be expected with ReFS 3.0(with dedup) on Windows 2019 compared to ReFS 2016/2019 without dedup in a case where you use per-VM backups and everything fast-clone/Syntectic full's?
As I expect offline duplication now also is able to dedup the whole volume instead of only the backup chain(fast clone), but does someone know how effective it really is in this scenario? Veeam: some lab test you are willing to share?

[Gostev]

Hi Mark, this was the last minute development and we did not have any chance to perform effectiveness testing (only reliability testing), which is why the feature is declared as experimental. Also, the idea was to hear the effectiveness results from the community, as lab environment don't have real-world data anyway. Thanks!

[DonZoomik]

Case resolved by support.
Slowness has be reproduced in lab and it's as slow as with NTFS deduplication.

[absentminded]

Our fast clone synthetic takes hours. How was your case resolved?

[DonZoomik]

It's wasn't resolved per se, just replicated in lab and confirmed to work... badly.
Summary is that all block clone benefits are lost if file is deduplicated. ReFS deduplication has no real integration with block clone and metadata operations become full read and write operations (copied from dedup store to new blocks). Also, written data is no longer deduplicated.
TL;DR: fast clone is slow with ReFS Deduplication.

[jorgedlcruz]

Hi guys,
But that will be expected I will say isn't? To obtain the best possible performance, at a logical level really, you will need to use the usual ReFS for your synthetic, might be weekly or so, and then once you know your synthetic file is created, that is the file you want to apply dedupe.

So, look at your Windows Deduplication schedule and make sure you are not doing dedupe on the files of the open chain which will be used to create the synthetic, as if you do the server will need to dedupe all that files before do the ReFS, which will end in that long time to do the fast-clone.

I see this ReFS 3.0 and Dedupe kind of what Exagrid does, you have a landing zone for your weekly, where you run your synthetic operations, and then after you apply your dedupication to the chains which are closed already. Of course all of this is DIY vs. ExaGrid which gives you all of this and more out of the box.

[absentminded]

I confirm we do dedup files older than one day. Have to try eight days.

[frankive]

soo... I should stick with server 2019 and Refs and not enabling deduplication?

[m.novelli]

Hi guys, slight off topic... does Veeam support restore of deduplicated files stored on a source VM Windows Server 2019 , ReFS 3.0 volume?

Do I need to have Veeam installed also on box with Windows Server 2019 with "Data Deduplication" feature installed?

Thanks,

Marco

[DonZoomik]

Mount server must have Windows Server 2019 and data deduplication feature installed.

[DonZoomik]

soo... I should stick with server 2019 and Refs and not enabling deduplication?

While I was running it, merge seemed a bit more stable than with NTFS deduplication - no other benefits. With NTFS, I got occasional transient errors during merge (probably due to file locking from deduplication engine).

[csydas]

Hi Veeam,

Does this key work if the Windows 2019 ReFS Volume is presented as CIFS? Or does Veeam need to have the volume added as a Windows Server?

[HannesK]

Hello,
yes it does fastclone via SMB / CIFS on a REFS dedupe filesystem. I just tested it in my lab.

Anyway, I'm not sure whether it makes any sense. SMB is the slowest way for backups and the comments say that ReFS dedupe & fastclone is also slow

Best regards,
Hannes

[sege]

As I understand the reg-key is to be set at the backup server, not the backup repo?
My case is that I have a Win2016 backup server and the repo server is a 2019 with ReFS.

Will that work?

[HannesK]

that's what the document says. So if it does not work, please open a support case and post the case number here for reference.

[matt.burnette]

Has anyone had any issues so far?
I thought about enabling it, though it looks like the only place that talks about block cloning with data dedup references Veeam.
I also haven't found where this is supported by Microsoft.
I also am unsure of the exact mechanisms that might be affected by having both on at the same time.

Does anyone have more information on how these can or can't be enabled at the same time?

I feel it could be a bad idea to have both at the same time since both are chunking data and replacing it with metadata/pointers.
How is it possible that these do not conflict with one another?
Is there any Veeam guidance on how they are testing it or what the best practices might be to enable this?