DRS and Passive nodes

[whynotq]

This is a odd request and may be along the lines of other threads already running but there is enough unique about this to post it seperately. I have a customer who wants to configure a 3 node V3.5 cluster but in a N + 1 scenario. effectively they want to have an Active/Active/Passive config where the 3d node is in the cluster but not used unless one of the other hosts fails, this is so they don't "over subscribe" whilst deploying VMs. what they are looking to do is have all 3 nodes avaliable for HA but only the 2 Active nodes running DRS. any suggestions?

[MR-T · Accepted solution]

This isn't the best use of hardware, although I'm sure you already know that.

The admission control feature of HA will ensure you don't provision more machines than the cluster can handle in a failure situation.

I would always suggest using the 3 nodes which actually reduces the impact of a node failure due to less vm's being moved.

[MR-T]

If you really don't want machines running on the 3rd node, you'll need to place DRS in smanual mode.

So you'll not have any balancing.

If you place DRS in semi-automatic mode it will try and automatically newly provisioned VM's on the 3rd node.

[whynotq]

two posts in 13 minutes, you are either after extra points (which i doubt) or you were thinking it over whilst making a cup of tea... :smileygrin:

i was much of the same opinion as regards ESX taking care of the provisioning to make sure it doesn't over subscribe itself but they are thinking that ahving the spare capacity with a "passive" node would be a better use of resources for failure protection. They argue a good case but the complexity of trying to prevent DRS provisioning to the 3rd node whilst maintaining HA availability is going to see them going our route rather than theirs.

[RussH]

I can somewhat understand the passive node approach for some situations, particularly where reservations are used. If you enable "strict admission control" then VMWares over-conservative HA algorithm uses the VM with the largest reservation as the base to work out how many VMs can be powered on without violating HA constraints. If you use large reservations, this will invariably leave your clusters way under-utilised. If you turn off "strict admission control" then its down to you to work out how many VMs (with reservations) can be safely powered on without violating HA constraints, this is obviously risky.

Also, if you have 8 nodes in a cluster for example, then 1/8 capacity of each node is reserved for HA failover. If you use reservations, then the biggest VM that you can possibly create will be 1/8th the resource of the physical machine else it will not fit in case of a failover!

Neither of the "issues" I described above apply if you keep a node as a passive spare.

Like I say, most of these "issues" dont apply if you dont make extensive use of reservations.

Russ.