HA Failover and resources

[Richard Leishman]

Hello,

I have been running Proxmox for roughly two years now in production, know it quite well and working very well. We are currently putting together a new improved cluster of 5 nodes together with HA and fencing which is all up and running perfect in our office test rack.

Is there a way to change how rgmanager/proxmox decides which machine to failover to? I understand that it goes off the failover group and priority but can it be configured to work based on resources?

All our nodes have 64gb ram, some of our vm's have as much as 32gb of ram assigned to them and in the past I have accidentally dished out ram that the server does not have available which causes the host and all vm's to run dog slow and even at times freeze.

At the same time I have also been messing around with a small 3 node XenServer cluster/NFS shared storage which was a dream to setup (one click ha setup), supports auto failover based on available resources which works amazingly. But what I don't like is that XenCenter is windows only (I use mac and the ported openXenManager is so buggy) and the hassle it would be to migrate from Proxmox to XenServer.

Any advice would be amazing,

Thanks
Richard

 

[t.lamprecht]

Use fail-over domains which can be used to bind services to (subset) of nodes? See https://fedorahosted.org/cluster/wiki/FailoverDomains
or:

man rgmanager

Search for fail-over domain.

eg:

# define some failover domains
<failoverdomains>
# we won't a restricted domain (service only runs in the defined nodes
   <failoverdomain name="my_domain" nofailback="0" ordered="0" restricted="1">
      <failoverdomainnode name="pve1"  priority="1"/>
      <failoverdomainnode name="pve2" priority="1"/>
   </failoverdomain>
</failoverdomains>
# bind the domain and service
     <pvevm domain="my_domain" autostart="1" vmid="100" recovery="relocate"/>

note this is only an example, look up in the docs what you need for yourself, don't only copy paste

Use the same way to edit your config as for the fencing.

btw. with PVE4 (beta, not ready for production yet) you have groups to assign resources to node(s) in the UI, which makes those thing a lot easier to setup.