[ClusterLabs] [Slightly OT] OCFS2 over LVM

[Digimer]

On 24/08/15 07:55 AM, Jorge Fábregas wrote:
> On 08/24/2015 06:52 AM, Kai Dupke wrote:
>> Not sure what you want to run on top of your 2-node cluster, but OCFS2
>> is only needed when you need a shared file system.
> 
> This is for an application that manages the high-availability by itself
> (in an active/active fashion) and the only thing that's needed from the
> OS is a shared filesystem.  I quickly thought about NFS but then the
> reliability of the NFS server was questioned etc.  I could create an NFS
> cluster for that but that will be two more servers.  You get the idea.
> 
> I'm still googling "NFSv4 vs OCFS2"....  If anyone here have experience
> (going from one to the other) I'd like to hear it.
> 
> 
>> For plain failover with volumes managed by cLVM you don't need OCFS2
>> (and can save one level of complexity).
> 
> This is my first time using a cluster filesystem and indeed I get it:
> there's lots of things to be taken care of & many possible ways to break it.
> 
> Thanks,
> Jorge

Speaking from a gfs2 background, but assuming it's similar in concept to
ocfs2...

Cluster locking comes at a performance cost. All locks need to be
coordinated between the nodes, and that will always be slower that local
locking only. They are also far less commonly used than options like nfs.

Using a pair of nodes with a traditional file system exported by NFS and
made accessible by a floating (virtual) IP address gives you redundancy
without incurring the complexity and performance overhead of cluster
locking. Also, you won't need clvmd either. The trade-off through is
that if/when the primary fails, the nfs daemon will appear to restart to
the users and that may require a reconnection (not sure, I use nfs
sparingly).

Generally speaking, I recommend always avoiding cluster FSes unless
they're really required. I say this as a person who uses gfs2 in every
cluster I build, but I do so carefully and in limited uses. In my case,
gfs2 backs ISOs and XML definition files for VMs, things that change
rarely so cluster locking overhead is all but a non-issue, and I have to
have DLM for clustered LVM anyway, so I've already incurred the
complexity costs so hey, why not.

-- 
Digimer
Papers and Projects: https://alteeve.ca/w/
What if the cure for cancer is trapped in the mind of a person without
access to education?