[Linux-cluster] Private Network for all cluster comms?

[Digimer]

On 11/19/2010 02:46 PM, Colin Simpson wrote:
> I'm looking at implementing a RHEL 6 based cluster using DRBD. For this
> I plan to have a dedicated fast bonded network between the two nodes for
> the backend DRBD Storage Network (SN), well a 10Gb with a 1Gb failover.

A good plan

> Most examples I've seen of doing this seem to use the SN as ringnumber 1
> in totem, does it make more sense to be ringnumber 0. I assume this
> could be accomplished by setting the hostnames associated with SN
> network IPs as the node names in cluster.conf. And put the altname in as
> the hostname associated with the main network IPs.

I don't think it matters much, assuming that your cluster communication 
is also on a private network. I'd sort of argue that the cluster comm 
network would be quieter and thus less likely to suffer latency which 
you might see during periods of high disk I/O on your storage network.

> My reasons for thinking of doing this are that my SN network will be
> back to back wired so more reliable than a switch based network but more
> importantly it's performance will be higher for GFS2 locking etc

For this reason, it is a good one to use, regardless of whether it is 
ring 0 or 1.

> Does this make sense to do (anyone else do this)?

I do this.

> Does this make management harder at all say from a central off cluster
> luci, that the node names are not themselves are not resolveable from
> the main network DNS (the altname's will be). Any other nastyness that
> might arise?

When a ring fails, for any reason, it's recovery must be performed manually.

> One thing I don't like is the machine hostnames will not be their
> network DNS names. Would the cluster suite be happy if the hostname of
> the node matches the "altname" but not the "clusternode name"?

Cluster comm will, iirc, happen on the network resolved by `uname -n`. 
I'm not certain of this though, as it may be tied to the interface of 
the active ring.

> As a quick aside, I've only played with clustering on RHEL5 before so on
> RHEL6 does corosync.conf get generated from cluster.conf. If not, what
> would happen if I hand craft a corosync.conf for doing my config above ?

In RHCS3, corosync.conf should not, and is not used. Everything comes 
from cluster.conf.

> Would this upset things with regards using keeping things the right way
> around in cluster.conf with regards "clusternode name" and "altname"?

My only real concern is that your SN could get too noisy. Personally, 
I'd advice making it ring 0 and leaving your names alone. That said, 
either *should* work.

-- 
Digimer
E-Mail: digimer at alteeve.com
AN!Whitepapers: http://alteeve.com
Node Assassin:  http://nodeassassin.org