TCPIP setup for high-availability Itanium cluster

[Paul Jerrom]

Hi all,
Greetings from sunny New Zealand.
I'm looking for suggestions and input for a Multi-Site cluster configuration. There are three nodes; 2xRX2620s, each has 4 (yes, four - we didn't realize when ordering that they came with 2 on-board!) Gigabit ethernet ports; and an AlphaServer DS10. DS10 is there as a vote provider/tie-breaker only, all three servers are in different buildings. There are two networks between the servers; a public one (.10 subnet), and a private one. I will be using volume shadowing between the two rx2620s, so I ideally want as much cluster and volshad traffic as possible to go over the private network, but if necessary to fail over to the public. Multi-site clustering is required as this is a real 24x7 operation, and I can't let anything as minor as a plane crash, meteorite strike or civil unrest stop production.
I've been looking at load brokering, failSAFE etc, but just wondered if anyone has set up a similar environment, and how they set up the IP addresses (the number of addresses is not an issue, by the way, I have my own subnet to play with), cluster aliases, public and private addresses etc.
VMS, TCP/IP versions all the latest; applications will include RDB and MessageQ.
[BTW I've already read Matt Muggeridge's excellent paper on high availability TCPIP.]
Thanks in advance.
Paul Jerrom.

[Jan van den Ende]

Paul,

forget about IP clustering, as that is just a crude Unixian failover with ALL cluster connection to exactly ONE node, and failover only upon node failure.
Use DNS round-robin or (preferably) metric+loadbroker.

Have each pair of nodes connected to the others over at least two GEOGRAPHICALLY INDEPENDENT network lines.

For 24 x 7, or even 24 x 365.25 operation, you probably also need some way of rolling upgrade of your applics.

We implement this by having a separate service name for each application, which is devided by round-robin over every node (normally 4 for us) that offers that service.
Planned upgrades are done by taking one node for that service out of the round-robin.
User sessions are limited to 10 hours, so after that the node is free of that app.
On THAT node we perform the upgrade, and the verification. If all is OK, we move the app over the the new-node version. Depending on the acceptability of simultanuous running of the versions, we do or do not kill the user session on the old version. After all old-version sessions are gone, we re-install round-robin.
This allows for NO (simultanuous multiversion allowed) or minimal (break running sessions & restart for no multiversion) applic interrupion.

hth,

Proost.

Have one on me.

jpe

[Andy Bustamante]

Hello from San Diego.

You can use SCACP to set priority for cluster traffic. You don't need to configure anything, out of the box this just works for a network interface cluster.

For the two networks, you should make sure that there is redundant physical networking equipment supporting connectivity. Two VLANs on the same switch are not redundant.

I would consider using LAN failover on your public network, you can combine LAN failover and failSAFE IP. I'd want to have a "service" address or addresses and a dedicated management address for each system.

As Jan says pass on cluster alias.

What sort of storage are you planning?


Andy