Round Robin policy with only certain HBAs?

[jedijeff]

We are moving from an Hitachi 9985 which is active/active to a Compellent which is active/passive.

I was wondering if I have 4 hba's and I have 2 going to 1 controller, and 2 to the other controller---could I somehow get a Round Robin to only work with each pair of HBAs?

Or as I think about this more, if the Compellent supports ALUA, then I could simply set RR on all of them and the Compellent should report back what controller the volume is attached too and the esx will RR between the 2 HBAs that are on that contoller?

Thanks!

[mcowger]

Its actually even easier than this.

The path selection policy (MRU, RR, Fixed) can be set on a per LUN basis. So just set RR on all your 9985 LUNs, and MRU (or whatever Compellant recommends - they may recommend RR) on your Compellant LUNs.

--Matt

VCP, vExpert, Unix Geek, Storage Nerd

[jedijeff]

I may not have been totally clear. Lets say I have a volume that is currently owned by ControllerA. ControllerB will take it over should ControllerA fail---so this is your typical active/passive setup.

So i a simple case I have 2 hba's on 2 fabrics--so 1 hba has a connection to ControllerA, the other to ControllerB. At this point I would do MRU.

But to get some extra io out of the esx host I put 2 more hba's...so 2 hba's on fabricA and 2 on fabricB. That means I now have 2 hba's to talk to ControllerA which currently owns the volume.

How could you tell esx to round robin between those 2 only--and ignore the other 2 which are on a controller that does not own the lun.

I maybe not understanding all the intelligence that is going on behind the scenes and that esx can somehow figure that out. I really am missing my active/active san and i have not even decommited yet :smileyhappy:

And lastly a very stupid question: in any round robin setup,,,how is there not any disk corruption? I assume will a loaded host it is entirely possible that at T=0 esx sends 1000iops down hbaA,,,at T=1 he sends 1000iops down hbaB. But due to latency or any zillion numbers of variables,,,the iops sents at T=1 reach the storage array before the 1000ips at T=0?

I guess I have always been very nervous about round robin---though I read it is very safe and actually supported on any array(according to vmware tid)

Thanks!

[mcowger]

The exact same way. Stop thinking about this on an HBA level (because thats now how ESX thinks about it).

If you set LUN A (whether it has 2, 4, or 16 paths) to be accessed under the MRU PSP, it will be accessed using only the MRU method. If you REALLY want to force it, find all your Compellant LUNs and set them all to 'Fixed' woth your chosen path as the primary.

As far as round robin safety....order of operations is guarenteed by the protocol, because each frame has an ID. Just like IP, the frames can come in slightly out of order and the array will handle it correctly.

--Matt

VCP, vExpert, Unix Geek, Storage Nerd

[jedijeff]

Thanks!

[iforbes]

Does Compellent support ALUA? I can't find an answer anywhere.

Thanks

[mcowger]

I'm 95% certain they do. That being said, the best answer would come from your Compellant rep.

(full disclosure, I work for a Compellant competitor)

--Matt

VCP, VCDX #52, Unix Geek, Storage Nerd