[Linux-cluster] Startup and Shutdown of a Virtual/Physical cluster

[Paolo Marini]

Thanks for the answer, I suspected that failover domains could do the
trick, I will try this solution.

Anyway, shutdown is for me still unclear.

starting up a cluster involves starting the cman daemon, clvmd, gfs,
mounting the gfs filesystems and starting rgmanager. Stopping cman does
not work on a cluster, because the other nodes still think that the
shutting down node is up, and they end fencing it if it is no more
responding. In fact, the shutdown procedures of a single node involves the
removeal of the node from both the fencing and cman agents (fence_tool
remove and cman_tool leave remove commands).

So stopping the daemons is not enough. Beside this, the ordered shutdown
of a physical/virtual cluster requires the shutdown of the virtual first
then of the physical cluster.

Paolo

> Hi Paolo,
>
>
>> When the physical cluster starts up, and the nodes are operational, the
>> virtual nodes are started in an unordered mode, e.g. the allocation to
>> the
>> physical nodes is not predetermined. If the physical cluster nodes come
>> up
>> at different times, the first operational node of the physical cluster
>> tries to bring up all the virtual guests, even if the available memory
>> is
>> not sufficient. This causes some instabilities at system startup, and
>> requires some manual intervention in order to distribute the xen guests
>> among the physical cluster nodes.
>>
>> Is there some way to prevent this kind of behaviour ?
>>
>
> yes there is. Basically what you want to do is the same thing that is
> done with nfs services in the nfs cookbook.
> you basically create three failoverdomains, one for each node, with
> ordered=1, then put your nodes in with a priority of 2 except one node
> which gets priority of 1. every domain then has to have a different node
> with the priority of 1.
> then all you have to do is put the vms in the failoverdomain with the
> node you want it to be started on automagically, so in case you have
> failoverdomain1 with node2 having priority 1, you put
> domain="failoverdomain1" in your <vm> definition.
> this will also do some migration magic if your cluster is up. so if a
> node has to take over a vm that's in a failoverdomain with a failed node
> having the highest priority, it will automatically migrate the vm back,
> once the higher priority node has become available again.
>
>
>> Another aspect is related to system shutdown. My system is powered by an
>> UPS, but once in a year due to maintenance on the power supply it may
>> happen that there is need to shutdown the system, or in case of blackout
>> with a long time. How is it possible to completely shutdown the cluster
>> by
>> UPS command ?
>
> i believe shutting down rgmanager on every node will do the trick. or
> else, shutting down cman should work too :)
>
> enjoy,
> johannes
>
> --
> Linux-cluster mailing list
> Linux-cluster at redhat.com
> https://www.redhat.com/mailman/listinfo/linux-cluster
>