About PVE cluster unicast limitations

[robhost]

Hi,

just wondering why the PVE Wiki mentions the unicast limit explicitly with "do not use it with more that 4 cluster nodes"?

An example in the corosync git repo (see here) has 16 nodes with unicast, can't believe they are wrong there?

Is there any proof that this 4 node limit does really exist? Any experience would be appreciated. If not, we should cut this sentence from the wiki page. Or does this only take affect in shared networks with no dedicated cluster hardware?

 

[dietmar]

Is there any proof that this 4 node limit does really exist?

Unicast leads to massive packet duplication. You need to send each packet 15 times if you have 19 nodes. You can do that, but IMHO it does not make sense...

 

[robhost]

Unicast leads to massive packet duplication. You need to send each packet 15 times if you have 19 nodes. You can do that, but IMHO it does not make sense...

Thats right - but this does not imply the limit to 4 nodes, right?

 

[robhost]

Unicast leads to massive packet duplication. You need to send each packet 15 times if you have 19 nodes. You can do that, but IMHO it does not make sense...

@dietmar we tried to change a cluster back to multicast in this way:

• deleted the HA groups and removed VMs from them
• changed corosync config /etc/pve (removed udpu) and increased config version
• after that, we restarted corosync on all nodes and pvecm told us everything was fine
• *then all nodes restarted* due to fencing :-( have you any idea why this happened? why did the fence trigger without having any active HA groups or VMS?