On Thu, 19 Sep 2013, Fabio M. Di Nitto wrote:
On 09/19/2013 11:26 AM, David Lang wrote:
On Wed, 18 Sep 2013, Digimer wrote:
On 18/09/13 23:16, Fabio M. Di Nitto wrote:
On 09/19/2013 12:47 AM, David Lang wrote:
I'm trying to get a cluster up and running without using multicast,
I've
made my cluster.conf be the following, but the systems are still
sending
to the multicast address and not seeing each other as a result.
Is there something that I did wrong in creating the cman segment of the
file? unforutunantly the cluster.conf man page just referrs to the
corosync.conf man page, but the two files use different config styles.
If not, what else do I need to do to disable multicast and just use
udpu?
<cman two_node="1" expected_votes="1">
<totem vsftype="none" token="5000"
token_retransmits_before_loss_const="10" join="60" consensus="4800"
rrp_mode="none" transport="udpu">
<interface ringnumber="0" bindnetaddr="10.1.18.0" mcastport="5405"
ttl="1" >
<member memberaddr="10.1.18.177" />
<member memberaddr="10.1.18.178" />
</interface>
</totem>
</cman>
You don't need all of that...
<cman two_node="1" expected_votes="1" transport="udpu"/>
There is no need to specify anything else. memberaddresses et all will
be determined by the node names.
Fabio
To add to what Fabio said;
You've not setup fencing. This is not support and, when you use
rgmanager, clvmd and/or gfs2, the first time a fence is called your
cluster will block.
When a node stops responding, the other node will call fenced to eject
it from the cluster. One of the first things fenced does is inform dlm,
which stops giving out locks until fenced tells it that the node is
gone. If the node can't be fenced, it will obviously never successfully
be fenced, so dlm will never start offering locks. This leaves
rgmanager, cman and gfs2 locked up (by design).
In my case the nodes have no shared storage. I'm using
pacemaker/corosync to move an IP from one box to another (and in one
case, I'm moving a dummy resource that between two alerting boxes, where
both boxes see all logs and calculate alerts, but I want to have only
the active box send out the alert)
In all these cases, split-brain situations are annoying, but not critical
If both alerting boxes send an alert, I get identical alerts.
If both boxes have the same IP, it's not great, but since either one
will respond, the impact consists of any TCP connections being broken
each time the ARP race winner changes for a source box or gateway (and
since most cases involve UDP traffic, there is no impact at all in those
cases)
This is about as simple a use case as you can get :-)
If you are running only a pool of VIPs, with no fencing, then you want
to consider making your life simpler with keepalived instead of
pcmk+corosync.
Thanks, I'll look into it. for all these two machine clusters, what I really
want to use is heartbeat with v1 style configs, they were really trivial to deal
with (I've had that on 100+ clusters, some going back to heartbeat 0.4 days :-)
But since that's no longer an option, I figured it was time to bite the bullet
and move to pacemaker, and since RHEL is pushing pacemaker/corosync, that's what
we setup.
David Lang
_______________________________________________
Openais mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/openais
