topic Re: 2-node cluster with single network interface in Operating System - HP-UX

2-node cluster with single network interface

Guy Schoofs — Thu, 22 Dec 2005 15:46:36 GMT

Hi,

i need to configure a 2-node SG-cluster, but unfortunately the nodes only have 1 network interface (i know that this isn't exactly HA but i do not have a choice at this moment).
When i disconnect the network-cable on the node where my SG-package is running, the cluster starts reforming (because there's no heartbeat) but it's always the node where the package was already running, that gets the lock disk. (leaving me with an package without network)
I would expect that there is only 50% chance that the same node gets the package.
Is it not possible to configure SG so that it will not try to get the lock disk in case of a network-failure, leaving the lock-disk for the other node?
Does implementing a standby heartbeat over RS232 make any difference for the above scenario?

Thanks in advance for any response,

Re: 2-node cluster with single network interface

Geoff Wild — Thu, 22 Dec 2005 15:53:59 GMT

"RS-232 line prevents a TOC from the momentary loss of heartbeat on the LAN due to network saturation" from this:

http://docs.hp.com/en/B3936-90079/ch03s07.html#d0e4914

Also read on same page "Responses to Hardware Failures " - you can decide how you want a package to behave in a network failure - IE - stay up or come down...

Rgds...Geoff

Re: 2-node cluster with single network interface

Sanjay_6 — Thu, 22 Dec 2005 16:11:44 GMT

Hi Guy,

I do not think you can control which server gets the lock disk in the case of a tie breaker situation.

Hope this helps.

regds

Re: 2-node cluster with single network interface

Duncan Edmonstone — Thu, 22 Dec 2005 16:27:42 GMT

It's not really surprising that this is the case when you think about SG timings - I expect that the value for HEARTBEAT_INTERVAL + NODE_TIMEOUT adds up to more than NETWORK_POLLING_INTERVAL, which I think means you would expect the node that loses network connectivity to begin cluster reformation first.

Only way I can see around this would be to monitor the subnet in your package and to set NODE_FAIL_FAST_ENABLED to YES in your package conf file. That way on subnet failure the node should TOC, giving the other node a chance to grab the cluster lock...

... but of course you're running in an unsupported configuration - so really SG can choose to behave however it likes!

HTH

Duncan

Re: 2-node cluster with single network interface

melvyn burnard — Thu, 22 Dec 2005 17:16:35 GMT

Well as you say, this is an unsupported and non-HA configuration. However, if you really wish to run it like this, then use the Serial heartbeat between the nodes to allow the nodes to decide which one will TOC.

The serial heartbeat is NOT a full heartbeat, but it does allow communication so that the node that has LOST the network will be TOC'ed, and the node that still has a working network will remain running and take over the package if needed.