Finding heavy writers

[#1]

Our environment is VMware 5.0 linked to a cluster of NetApp FAS heads by NFS. For a number of reasons we are trying to determine which VM's are the ones doing most of the writes.

I am assuming that the NFS subsystem on the NetApp has the data that I need since every write is passed through it at the NFS File level. Is there a way to look at this? Is there another way to figure this out?

I don't care about reads, only about writes.

Thanks

[#2]

Do you mean that your VMWare datastores are NFS, and the guests are writing to their "local" VMDK filesystem?  Or that your VM guests are mounting NFS from NetApp?

You can get per-client NFS usage from NetApp, but if your datastores are NFS, all you'll see is ESX host access, not individual VM access.  That being said, the vSphere client gives you pretty good per-guest IO information.

To get per-host NFS stats, you need to enable the nfs.per_client_stats.enable option, then check out nfsstat -h.

Bill

[#3 · Accepted solution]

Bill has several great points, so I'm just going to add to them:

• The controllers themselves can only "see" things from a volume level for the most part, so watching the performance counters in vSphere will let you see which VM's are doing which I/O, but you'll need to sort of map that back to which datastores they all live on (and isn't granular enough for what you're looking for unless you only have one VM per datastore)
• OnCommand Balance (paid-for software that you can get a demo from your account team), which will give you awesome end-to-end analysis and details like you're looking for, all the way through to the VM application(s) if you add in credentials for the VM(s) themselves - it can figure out which VM's are the hottest by pulling that info from vCenter, and mapping everything back to the controller for you
• Depending on your version of ONTAP, you can try using Performance Advisor (free), which is bundled with OnCommand Unified Manager 5.x - but again this will only show you the volume level (and again, isn't granular enough for what you're looking for unless you only have one VM per datastore)

If you're up for a demo, I would highly recommend Balance, as it traces things like victims and bullies, has a great interface, and is really granular; because it does look end to end (since it talks to vCenter as well as the controllers).

[#4]

So it turns out that we have a copy of OnCommand Balance. I guess this is what the new guy gets for reaching out of the team.

Thank you all for your help.