[Ocfs2-users] Multiple clusters per node?

[Sunil Mushran]

Each volume has a different dlm domain. Only nodes that mount that
volume, join that domain.

Søren Kröger wrote:
> Hi Sunil
>
> Will the mastery involve all nodes in the cluster, regardless which 
> filesystem they have mounted? Or will it only involve those nodes 
> which have mounted the particular filesystem?
>
> Søren
>
> On 28/04/2009, at 22.21, Sunil Mushran <sunil.mushran at oracle.com> wrote:
>
>> Søren Kröger wrote:
>>> I'm trying to split up our big OCFS2 filesystem into 3 separate 
>>> LUN's, since there are only a limited amount of nodes which need 
>>> access to the different parts of the OCFS2 filesystem.
>>> One "Master" server with RW access should still be able to mount all 
>>> 3 OCFS2 LUN's, all others would only mount one of the 3 LUN's in RO 
>>> mode.
>>
>> In a distributed lock manager, no one node is the master node for
>> all resources. Mastery of resources is distributed amongst all the
>> nodes irrespective of the fact that a node may have mounted ro.
>>
>> Note that the locks on a resource are only created if a node actually
>> reads that inode. So it could be that you have three nodes reading
>> different directory trees that have little in common with each other.
>> In that case, each node will master the resources and the only lock
>> on each resource will be the local one.
>>
>>> What i want is:
>>> - Lowering the impact, when one node crashes and causes deadlocks on 
>>> the filesystem for other nodes
>>
>> When a node crashes, the surviving nodes will recover it. Why should
>> there be any deadlocks?
>>
>>> - Lowering the impact, when we have to resize a OCFS2 volume
>>
>> ocfs2 has online file system resize. However we require a cluster
>> volume manager. This should be working in SLES11 when they release
>> the HA extension package. We hope to have this running with (RH)EL6.
>>
>>> - Lowering the mastery time
>>
>> Again, ro nodes have as much a say in mastery as rw.
>>
>>> Is there one master per filesystem or one master per cluster?
>>> Would 3 separate filesystems under the same cluster be "separated" 
>>> enough to achieve?