Dedup only saving 5% on 750gb of windows VM's

[wally_kroeker]

I currently have two ESX 3.5 hosts connected to a 2050 running 7.3.1.1 through iscsi.  I recently had to create a brand new lun and migrate all of my VM's to it becasue the vmfs file system became corrupted.  On my old lun dedup was saving me 65% of my space but on the new one I am only saving 5%.  It is a 1 TB volume with space settings set to volume, fractional space reservation set to 0% and snapshot reserve set to 0%.  On the Lun space reservation is turned off.  I have set dedup to run between 3 and 6 am everyday and it now seems to take 6 min to run every night.  I can't figure out why this would have changed.  Any help would be apreciated.

[wally_kroeker]

I think I have solved my problem when all of my vm's most of which are 2003 servers where out of alignment.  when I did the alignment they started deduping.  that being said I think the reason this fixed it is becasue dedup is esentially seeing the alignment as new data.  I think the reason this all started is becasue I turned on dedup with the windows app "system manager"  the default action of this app is to not scan old data but just new.  Keith Asen from netapp said that if I had kicked off dedup with sis start -s then it would have worked from the start.  THanks for all the help everyone.

[radek_kubka]

To make our discussion complete...

It looks like we were (sort of) reinventing the wheel - this is what I just found in TR-3749 on page 69:

12.4 THE IMPACT OF PARTITION MISALIGNMENT
[...]
In addition to the negative performance impact storage savings with NetApp data deduplication will be
negatively impacted, reducing the total amount of storage savings.

Nice one

Regards,

Radek

[radek_kubka]

On my old lun dedup was saving me 65% of my space but on the new one I am only saving 5%.

Isn't your block alignment screwed up by any chance? That may explain poor dedupe ratio (I think)

[eric_barlier]

I have not heard that block align. issues will impact on dedupde savings. Dedupe is at WAFL level and the block un-alignment issue is when you put

a file system on top of WAFL and its not aligned = impact on perf.

Its an interesting thought though. Does anyone hold conclusive information about this? Is Radek onto something here?

Cheers,
Eric

[scottgelb]

I'm interested in hearing the official answer on this too... in early asis presentations I remember seeing something about fixed 4k blocks but could be wrong...then also heard misalignment didn't hurt deduplication but haven't heard confirmation from an engineering resource.

[radek_kubka]

Guys,

This is where I am coming from, using purely common sense:

Scenario A, three identical VMs, but no identical blocks within each VM, all properly aligned, so NTFS blocks are mapped identically to WAFL blocks.

Scenario B, same VMs, vm1 properly aligned, vm2 has 1k offset & vm2 has 2k offset, so NTFS blocks from each VM are mapped in a different way to WAFL blocks.

Scenario A should give 3:1 de-dupe savings, whilst scenario B will give no savings, as NTFS blocks will be "chopped" from a different starting point in the case of each VM (unless purely by chance some aligned blocks will get identical 'twins' from miss-aligned ones).

Am I missing something?

Regards,
Radek

[eric_barlier]

Radek,

Interesting thought, it certainly warrants investigation. Can anyone from NetApp engineering or other shed any light on this please?

If I was a betting man (never been to Vegas, its not my gig) I d put money on bad MBR align issue not having impact on dedupe savings.

I d still like to know for sure.

Cheers,

Eric

Message was edited by: eric barlier

[renault]

All,

Reading this topic I think two possible reasons of poor dedup performance

- The directories of your VMs are compressed. Compressed data cannot be deduplicated

- Use of crypt engine like decru. Crypted data cannot be deduplicated.

My VM are not block aligned and we get 50% of dedup perf.

You may also zeroes the free space of your VMs

And if you have only 2 ESX host, you shoud try NFS storage. But it's an other subject.

Just my 2 cents

ML

[keitha]

I have been working with Wally on this issue and am stumped. Here is an update and a couple of things that still baffle me. First we have checked and rechecked the LUN and volume reservation settings. They are correct. Second although I agree alignment is likely an issue, these same VMs were deduping at nearly 70% beore Wally copied them to this new LUN. Same VMs and the VMware copy process shouldn't have affected the dedupe rate.

Now the really odd bit. As we deploy new VMs or align the old once they seem to dedupe at the right ratio, last I spoke to Wally the ratio was 15% and climbing. Perhaps the copy process did mess something up in the alignment process. Anyhow it does seem to be fixing itself as we align the VMs correctly using MBRALIGN.  Thanks to everyone for the suggestions.

Keith

[eric_barlier]

Hi keith,

I reckon how you did the copy will be a factor here. Say if you were to use NDMPCOPY that will nicely restripe your block layout. This is good

if your volume is fragremented, NDMPCOPY a volume is a known workaround for fixing defrag issues. however as the blocks are re-striped I think

it will undo what dedupe did as the block layout will have changed. Does that makes sense?

Interesting topic still as Mark-Laurent DeLaruelles statement opposes what you have observed.

Cheers,

Eric

[wally_kroeker]

Interesting.  I just started reading about alignment and how it affects performance.  Can someone point me to an easy tutorial on how to check alignment and how to fix it if it is off?  I had someone tell me that when I create the lun and choose vmware as the type that will take care of the WAFL to vmfs alighnment.

[keitha]

Wally,

Included in the Host Utilities Kit are mbrscan and mbralign. mbrscan accepts the path to a vmdk file as input and tells you if the NTFS filesystem in the vmdk is properly aligned with the VMFS filesystem. The VM can be running during the scan. mbralign accepts the same parameter and will fix any alignment issues with the VM. The VM does need to be off for the align process though and you do need at least enough space in the volume for a copy of the VMDK. The end result is a properly aligned VM.

Alignment certainly helps performance and load on the storage array but I would be surprised if it effected dedupe this much as I would expect all the VMs to be equally misaligned thus dedupe well. Especially if they deduped well on a different volume.

Keith

[satheesh1]

I was under the impression we need to shutdown VM, when we do mbralign. That's the mbralign documentation says. Please correct me if I'm wrong.

Satty

[keitha]

You are correct, the VM does need to be shutoff for mbralign. Reread my post, I might have said it in an odd way but I did say the VM has to be off for mbralign.

Keith

[eric_barlier]

I reckon you need to set vol guarantee to none. Can you look in old autosupports to set what the old volume used to have? you find old asupsin /etc/log/ folder on controller.

If you are worried about writes to your LUN you can then use "vol autosize" command so that the volume will grow when

the LUN needs space. We use this config and it works like a charm. This config forces free blocks back to

aggr. level = easier to monitor/monitor.

So long you dont use snapshots you dont have to worry about "snap autodelete" once/if the volume reaches its max size.

Cheers,

Eric

[radek_kubka]

I reckon you need to set vol guarantee to none.

That's bad idea, as vol guarantee set to none enforces 100% fractional reserve (can't be changed).

Vol guarantee set to volume (plus vol auto-size & snap auto-delete, plus FR of 0%) & thinly provisioned LUNs within that volume is a way to go!

Regards,

Radek

[eric_barlier]

Hi Radek,

By setting it to none you can start monitoring your space at aggregate level as opposed to have to monitor X numbers of volumes. It obviousily a lot

easier and allows for cap. planning to be facilitated at the same time.

However for some customers setting vol guarantee to none is not good of course. I just think any statement that categorically states

its not good to set it to none will not be right 100% of the time. it does work for us, in fact I love it! :-0

Cheers,
Eric

Message was edited by: eric barlier

[radek_kubka]

I just think any statement that categorically states
its not good to set it to none will not be right 100% of the time. it does work for us, in fact I love it! :-0

Hi Eric,

Yes, yes - point taken.

What I was trying to stress is the fact that 100% fractional reserve kicks in when vol guarantee is set to none & there is no way to dodge this. Setting FR to 0% (or some small number when SnapManager products are in place) is now a part of official NetApp best practice, so you see where I am coming from...

Having said that, I fully appreciate 100% FR may still be a good thing for some people in some environments. And I also admit FR in itself is still a very hairy topic! (see my post about FR & reallocation: http://communities.netapp.com/thread/4431?tstart=0)

NB this is all irrelevant if there are no LUNs in the volume in question.

Regards,
Radek

[eric_barlier]

Hi Radek,

Good points. I was never a fan of FR 100%.. I believe "vol guarantee none" should only be used when you  fully thin provision by turning

LUN space reservation off as well, and you have <vol autosize on> with <snap autodelete> as well.

thanks for sharing your valid points with us.

Eric

[lfreeman]

Yes you should of course see 65% savings on the new LUN just like you did on the old LUN.  Your LUN settings seem correct.  Question, did you run the "sis on" command on the volume before you did the migration, or run "sis start -s" on the volume after the migration?  You'll have to have done one or the other to get the full dedupe savings.

DrDedupe.