[Linux-cluster] Virtual ips 'drifting' from one node to another

[Eric Kerin]

Chris Harms wrote:
> We went with 2 APC remote power switches.  The ports can be grouped 
> together so you can turn off multiple ports at the same time to combat 
> redundant power supplies keeping the machine on.  Just remember to 
> have the cluster use the same IP address to connect to the APC device 
> for each node so they can't both be fenced simultaneously.  The APC 
> software only allows one session at a time so the second node will be 
> blocked from accessing it on the same IP.
>
Be careful with port groups.  In the quick bit of testing I did with my 
7920s, if I unplugged the network cable on the 2nd PDU (the one that my 
nodes were not accessing for control)  it said that it turned off the 
port, but the 2nd PDU still had both ports on (for obvious reasons)  
Which would quickly lead to a corrupted filesystem...

I personally use this config for my nodes (this allows for the cluster 
to know if fencing failed on either controller, and not continue in that 
case):
<device name="left" port="2" option="off"/>
<device name="right" port="2" option="off"/>
<device name="left" port="2" option="on"/>
<device name="right" port="2" option="on"/>

Thanks,
Eric Kerin
eric at bootseg.com