MCSG LAN fail over issue.

[Joe Short]

I have an MCSG (v11.16) cluster configured on 2 DL585 servers running RHEL AS 4 32 bit.
I have configured bonding for 2 of 3 NICs on each server, the third NIC is being used for dedicated heartbeat on a dedicated network.
With the package up and running on the promary server, I tested bonding by unplugging a NIC. The bond worked, and no fail over occured. however, when I unplugged the second NIC in the bond (only 2 NICs bonded) again no failover occurred. On HP-UX this would have triggered a failover of the package. When I completely disconnected all NICs on the primary server, the alternate server crashed and rebooted, but did not take the package.
Is this normal behavior on LINUX, or did I miss something? If so, what might I have missed?

[Steven E. Protter]

Shalom Joe,

You slightly misunderstand bonding in Linux. Unplugging one of the two bond cables does not result in a failure. Bonding defaults to active-passive on Linux and you can't force active-active except on intel nics.

So: If you unplug both cables, you should trigger a failover.

SEP

[Joe Short]

Steven,
That is my exact issue. Unplugging both cables does not trigger a failover. Based on my experience building clusters on HP-UX, I expected it would, but it did not. I am wondering if perhaps I am missing something, or there is another issue here.

[Steven E. Protter]

Ah!

Both cables and no failover.

What do the serviceguard logs say?

I would tend to think the failover LAN configuration is wrong or the NIC is bad or the card is plugged into the wrong lan.

My question now is can a NIC card be a failover LAN and a heartbeat LAN. My understanding of SG is that its either or, not both.

Please clarify.

SEP

[Joe Short]

If you define your NIC as HEARTBEAT_IP, both HB and Data can be used on that NIC. If it is defined as STATIONARY_IP, no HB is passed over it. What I have is 2 servers, each with 3 NICs. 2 NICs are in a bond, the third is on a separate network, that is simply between the 2 clustered servers. It is there to pass HB in the event the production network should go dark. If both bonded NICs on a server are unplugged, the cluster should respond by moving the package to the alternate server. In this case, that did not occur. My cluster config file is attached.

[Joe Short]

And the package log (I have a single package) and system log do not indicate anything out of the ordinary. I am wondering if the NODE_TIMEOUT parameter is set too high.

[Joe Short]

I had the bonding driver set incorrectly.
In /etc/modprobe.conf there is an options entry for the bond. it should read as follows

options bond0 miimon=100 mode=1

My was incorrect it read

options bond0 miimon=100 mode=0

Mode 1 is failover mode
Mode 0 is load balancing mode.