How can I configure redundant LANs in a 2 node MC/SG cluster ?

[Lothar Krueler]

We`re using HP-UX 11 and Service Guard A.11.08 on two L2000. There are 4 LAN-cards in each node and we`ve got two HUBs and one crossed twisted-pair cable.
Our applications aren`t really online-switchable, so I want to prevent our packages as good as possible from beeing switched. I connected the crossed twisted pair cable to both LAN1 for heartbeat. The other LANs are connected to the HUBs:
LAN0 (data) to the one HUB, LAN2 and LAN3 (both blank/standby) to the other HUB. Each HUB ist connected to the main-network. If I cut (one or both) HUBs from the network the cluster is still alive in cause of the direct connected LAN1 (heartbeat), but I cannot access one node by the other node. The packages themselves donot have needed remote-access to other packages residing on the other node. If I reconnect the HUBs everything works again.
In case of single LANcard failure, HUB failure or whole network-failure I want all my packages keep running. There are lot of batch-jobs, which also should be running even the whole network fails and no user can access.
What`s the best way of using the LAN-cards and HUBs ? How do I have to do the connections and what entries are necassary in my cluster conf-files?
Thanks in advance,
Lothar

[John Palmer]

Your standby LAN details should be configured in your cluster configuration file.

cmquerycl automatically detects standby LAN cards and works out what can connect to what by using the 'linkloop' command.

Have you rerun 'cmquerycl' since cabling all the LAN cards?

The LAN/HUB configuration that you have described sounds OK to me.

[John Palmer]

Lothar,

further to my previous reply, standby LAN entries in the cluster configuration file just consist of records such as:-

NETWORK_INTERFACE lan0

with no subsequent HEARTBEAT_IP or STATIONARY_IP records.

Regards,
John

[Stefan Farrelly]

Your cluster config file (.ascii) in /etc/cmcluster should have entries in it similar to these; (We have 3 lancards on each of our SG nodes, lan3 is the standby, we use the other 2 for heatbeat)

NODE_NAME hal
NETWORK_INTERFACE lan2
HEARTBEAT_IP 10.40.225.146
NETWORK_INTERFACE lan3
NETWORK_INTERFACE lan1
HEARTBEAT_IP 192.1.1.1

And in your /etc/cmcluster//.ascii file by default network card switching should be enabled;
NET_SWITCHING_ENABLED YES

Now the tricky stuff is in your /etc/cmcluster//.cntl file where you can specify all sorts of stuff for network switching. Theres documentation inside this file, take a look at it.

[R.Suresh]

Hi,

Simple,

You have 4 lan interfaces in each server.
configure
one for hear beat
another for Network communication.

And leave the other two without configuring,

When you configure M/C SG it will search
for all available un-configured lan interfaces & use them as stand-by.

Refer the manual "Managing M/C service Gaurd"

Thanks
R.Suresh.

[melvyn burnard]

Well reading the question, I sort of gather that you are using MC/SG, but do not want htet application to switch??
If so, why use MC/ServiceGuard?

Correct me if I am wrong, but you seem to want to the application to STAY on the node, even if the network dies? You may be better off removing MC/SG, and just using the networking software known as APA or Auto Port Aggregation which would allow you to put all four networks into one "channel", and allow better throughput, as well as give you the redundancy of lan failure coverage.

[Lothar Krueler]

Hi there,
thank you all for your responses.

John , Stefan :
Sorry, I didn`t mention : If I cut off LAN0, SG uses LAN2 for data.
I`ve 4 NETWORK_INTERFACE entries : lan0 as stationary, lan1 as heartbeat, lan2 and lan3 blank. Would you suggest to using heartbeat instead of stationary in my case?
I`m afraid that the data might use the direct connection between the nodes in normal case and not the connections through the HUBs.

Is it necassary to enable netswitching in every package ?

Melvyn :
I don`t know APA. I don`t want SG to AUTOMATICALY switch, that means automatic package-shutdown if really required is OK. But before running again, the data must be checked for consistence, what the apps cannot do themselves. So we have to prevent switching as good a s possible. I think, every admin likes his apps on that node he defined and wants them stay there as long as possible.
We use SG for manual switching, if service is required and for the use of redundant interfaces etc. So we`ve only the downtime informing the users to do a new login and switching the apps manually .
So the applications (this means our production-processes) are available while one node is beeing serviced. Well, I wish our apps were online-switchable, but most of them aren`t yet, perhaps some of them will become in future.
If APA also can do this, tell me more about this please.

[John Palmer · Accepted solution]

Lothar,

There should be no problem in defining your normal data connection for HEARTBEAT as well as the dedicated LAN. This will provide resilience against the loss of the dedicated LAN.

If you have defined the IP address that the users use to connect to the service as STATIONARY then it will not switch - see man cmquerycl. Could this be your problem? You need to define a 'floating' IP address in your package control script and get your users to connect to that.

User data will not use your dedicated LAN unless you set up routing to do so.

You need to set NET_SWITCHING_ENABLED to YES for the package if you want to enable the use of a standby LAN card. You definitely want to set this to YES.

Hope this helps,
John

[Lothar Krueler]

Hi John,
seems that is it. Nearly every package has NET_SWITCH_ENABLED set to NO.
(One of my problems surely is the complete understanding of english texts.)
If I understood right, I should toggle it to YES and configure one lancard to a dedicated HEARTBEAT_IP and an other lancardcard instead of STATIONARY_IP to HEARTBEAT too . The other 2 lancards should be left blank as standby.
Am I right in understanding "floating IPs" as the different IPs I assigned to the packages ? The users are connecting to the package-IPs, not to the node`s.

I`ll try and test this on a sunday. When I`m ready I`ll set the magic icon.
Lothar :-)

[John Palmer]

Hi Lothar,

Yes - by 'floating' IP I mean the package IP that moves with the package.

Any LAN card which is configured (ie not a standby card) can be configured for hearbeat, this gives greater resilience.

cmquerycl will normally define your server
IP addresses as HEARTBEAT_IP where it detects that that LAN card can 'see' the other nodes. I suspect that 'HEARTBEAT_IP' addresses switch to a standby LAN in the event of failure whereas STATIONARY_IP addresses don't.

Regards,
John