ONTAP Discussions

Does anyone use load sharing mirrors any more (ONTAP 9.1+) for SVM root vols?

netappz
5,120 Views

Hi All,

 

I have an inherited environment where there is a mish mash of load sharing SVM root vol mirrors - some setup and running to some of the nodes, some uniniatilized and some with none at all - set up by my predecessor (who no longer works for the company).

 

I am trying to bring some order and sense to this, but no one at my current employer can give me a straight answer whether they are needed or not, and I have a case open with NGS, but we seem to be going round in circles, with them just pointing me to documentation that contradicts each other and asking me to delete and re-create the mirrors.

 

What I would like to establish is if they are even required in the first place - my previous employer had 2 6node clusters and did not use LS mirrors at all. Another ex-colleague also says he does not use them, and they went out with 8.3x (I know volume LS mirrors are deprecated but these are SVM root volume LS mirrors).

 

 

In our current environment we are on ONTAP 9.1P9 and have a 12 node cluster and 45 SVMs (yes, really) so following what I can glean to be the best practices for this version we should have a LS per SVM root vol, so that would be approx 45 x 12 = 540 mirrors (which is simply ridiculous). We are working on a plan to get to 9.7 (and then possibly up from that to maybe 9.8 or 9.9) but this will take a few months at least due to hardware swapouts etc.

 

Can any  of the Netapp ninjas out there offer any advice on this ?, what do you do?

 

1 ACCEPTED SOLUTION

AlexDawson
4,965 Views

The aim for SVM rootvol LS mirrors is to enable a cluster to continue serving data if a HA pair goes offline when that pair is not holding epsilon but IS holding an/several SVM root vols, but not all of the data for that SMV, and also to enable path traversal operations in volumes that exist across multiple HA pairs without crossing the cluster interconnect, which may create performance issues in some scenarios.

 

It is absolutely not needed in a HA pair.. but.. for a 12 node cluster, their benefit is not zero. Multiply the CIs and divide the MTBF.. and suddenly the risk that the scenario proposed might happen becomes slightly more apparent.

 

However, if you don't have regular updates of that LS mirror, the potential exists to do something like create a new export, and then the nodes holding the LS destinations comes out of sync until this is updated.

 

I think your calculation is a little high - 12 nodes, but you only need one per HA pair, and not on the source pair, so you'd need 5 copies of each SVM root for 45 SVMs, so that's 225.. but as you rightly note, that's not a small number.

 

If your SVMs and their LIFs and their volumes exist only on a single HA pair at any time, then LS mirrors would provide little benefit.

 

Our current TR on NFS at https://www.netapp.com/media/10720-tr-4067.pdf, page 12, does recommend LS mirrors for SVM roots, and should be considered authoritative on the subject, but in the end it is a survivability and performance recommendation, not an operational requirement.

View solution in original post

7 REPLIES 7

cruxrealm
5,027 Views

As of 9.x ls mirror on root volume serves only for DR purposes.  We still use LS mirror on production NAS SVMs and create it only on its partner node.   You do not need to create mirrors on all nodes.  A single copy is enough (at least for our purpose).

 

For SVMs that have offsite BC/DR,  we do not create LS mirror,  since we snapmirror the data and root volume together.  

AlexDawson
4,966 Views

The aim for SVM rootvol LS mirrors is to enable a cluster to continue serving data if a HA pair goes offline when that pair is not holding epsilon but IS holding an/several SVM root vols, but not all of the data for that SMV, and also to enable path traversal operations in volumes that exist across multiple HA pairs without crossing the cluster interconnect, which may create performance issues in some scenarios.

 

It is absolutely not needed in a HA pair.. but.. for a 12 node cluster, their benefit is not zero. Multiply the CIs and divide the MTBF.. and suddenly the risk that the scenario proposed might happen becomes slightly more apparent.

 

However, if you don't have regular updates of that LS mirror, the potential exists to do something like create a new export, and then the nodes holding the LS destinations comes out of sync until this is updated.

 

I think your calculation is a little high - 12 nodes, but you only need one per HA pair, and not on the source pair, so you'd need 5 copies of each SVM root for 45 SVMs, so that's 225.. but as you rightly note, that's not a small number.

 

If your SVMs and their LIFs and their volumes exist only on a single HA pair at any time, then LS mirrors would provide little benefit.

 

Our current TR on NFS at https://www.netapp.com/media/10720-tr-4067.pdf, page 12, does recommend LS mirrors for SVM roots, and should be considered authoritative on the subject, but in the end it is a survivability and performance recommendation, not an operational requirement.

alessice
4,449 Views

Thanks, @AlexDawson for your explanation, so in an environment with only an HA Pair (like mine, with one FAS2720) is not needed to create an SVM rootvol LS mirror because in case all HA nodes go offline I have more big problems 🙂

 

Correct?

AlexDawson
4,352 Views

Yes, that's correct.

 

Hope this helps!

insert-name-here
4,416 Views

I would just like to add that SVM Rootvol LS mirror only supports CIFS or NFSv3 clients. NFSv4 does not use SVM Rootvol LS mirrors. From Page 52 of TR-4067.


"NFSv4.x operations do not leverage load-sharing mirror destinations when mounting volumes
and instead will use the source volume file handle. As a result, load-sharing mirrors don’t provide
protection against SVM root volume failures."

AndreMClark
3,564 Views

I have been working with ONTAP since the GX days and know the reasoning behind LSMs. As it is noted that this is a recommendation for SVM roots for NAS only, my ultimate question is this. Why isn't it part of the SVM creation workflow in System Manager? If you have a novice NetApp admin that doesn't read the best practices or likes to deal with the CLI, then they would never create any LSMs. 

AlexDawson
3,555 Views

And they'd almost certainly be fine if they didn't create LSMs.

 

Most clusters are 2 nodes only. For more, the chances of cluster partition are low anyway (due to the levels of redundancy), and for IO performance for namespaces, even then it's only if the amount of traffic warrants it. 

Public