[Linux-cluster] Fencing required for node failover

[Digimer]

On 12/29/2011 05:21 PM, Achint Mehta wrote:
> Thanks for the explanation.
> 
> It appears that until fence daemon performs the fencing activity the
> recovery will not be made.
> So I would have to setup fencing.
> 
> 1. I am not sure what kind of fencing device to add for simple servers.

IPMI is popular, if your server supports it. Big vendors have their own
IPMI-like interfaces you can also use for fencing; HP use iLO, IBM uses
RSA, Dell uses DRAC, etc.

These all rely on the host server having power (though are otherwise
totally independent and can be used even when the server is shut off).
If the host node's power supply (or power feed) fails, then so do these.

Alternatively, or ideally, as a backup would be to use switched PDUs.
These are power bars that have network connections. You can turn off and
on individual outlets on these devices to cut the power to nodes.

Personally, I use IPMI for my primary fence device and a PDU as a backup.

> 2. Is this fencing device supposed to be a dedicated machine outside of
> the cluster.

Either/or. Devices which can confirm a node's state are best, like IPMI,
but as I mentioned, they suffer from sharing a power source. Switched
PDUs are totally independent, but they can't confirm a node is truly off.

> 3. Also, would fencing be successful if the node  to be fenced is not
> reachable on the network.

Provided the fence device is reachable, then the state of the node does
not matter.

I cover exactly how to implement fence devices on RHEL 6.2 here;

https://alteeve.com/w/2-Node_Red_Hat_KVM_Cluster_Tutorial#Defining_Fence_Devices

It includes examples using IPMI, APC switched PDUs and HP's iLO.

-- 
Digimer
E-Mail:              digimer at alteeve.com
Freenode handle:     digimer
Papers and Projects: http://alteeve.com
Node Assassin:       http://nodeassassin.org
"omg my singularity battery is dead again.
stupid hawking radiation." - epitron