[Linux-HA] heartbeat over ethernet with ipfail question

[Matthew Soffen]

On Tue, 2005-03-15 at 09:44, R.Schramp at telecom.tno.nl wrote:
> Hello All,
> 
> I have a heartbeat HA cluster built up off two PC's. Each is equiped
> with 2 Eth nics, with IPMI support (for stonith) (Sunfire V65)
> 
> On the nics i have configured ethernet bonding and connected both to
> two different ethernet switches.
> 
> When both ethernet switches are down, clearly both nodes can't see
> each other, but also stonith won't work. This will lead to both nodes
> becomming master :(.
> 
> I thought I could solve this with ipfail: when a node can't ping ANY
> of it's ping host it should never become master. However this appears
> not to be the case with the ipfail module.
> 
> Any suggestions how to configure this? 
> 
> Clearly a serial link would solve the situation but that would make
> the wiring mess even more complicated, so I prefer to see if there is
> a software way.
> 
> Also coupeling one of the Eth nic's via a cross could solve the thing
> but would remove the ethernet redundancy I have in place now.
> 
> best regards,
> 

The problem is that with only an Ethernet communication line, you have a
single point of failure ( the IP Stack ).  If something were to corrupt
the IP stack on one machine (say, the standby), then it wouldn't be able
to communicate to the master so it would take over.

Having 2 totally different lines of communication (i.e. a serial line )
you minimize that potential for problem because even if Ethernet croaks,
the serial line would still be alive and communicating ( it would still
see the master server ).

If that isn't possible, you might want to have a dedicated cross over
Ethernet connection between the 2 servers.  That can mitigate some
problems ( of the switches failing ).  You still need to deal with the
single point of failure though ( the IP stack ).

Matt