[Linux-cluster] CLVM/GFS2 distributed locking

[Digimer]

On 12/30/2011 08:39 AM, Stevo Slavić wrote:
> Hello RedHat Linux cluster community,
> 
> I'm in process of configuring shared filesystem storage master/slave
> Apache ActiveMQ setup. For it to work, it requires reliable distributed
> locking - master is node that holds exclusive lock on a file on shared
> filesystem storage.
> 
> On RHEL (5.4), using CLVM with GFS2 is one of the options that should work.
> Third party configured the CLVM/GFS2. I'd like to make sure that
> distributed locking works OK.
> What are my options for verifying this?
> 
> Thanks in advance!
> 
> Regards,
> Stevo.

I've not had experience with Apache HA specifically, so I can only give
generalized comments.

DLM is very reliable... Red Hat's flagship clustered filesystem, as well
as it's cluster resource manager, rgmanager, both make extensive use of
DLM. If it wasn't reliable, there would be very big problems for Red Hat. ;)

The main issue with DLM is that you have to have tested and working
fencing. When a node is lost, fenced blocks DLM so anything using DLM
will also block until the fence action completes. This is by design
because you don't want nodes making changes while the cluster is in an
unknown state.

The other issue is performance; Your locks are, by necessity, travelling
over a network. You need to ensure that you network latency is as
minimal as possible. You will want to allocate time to tune DLM and GFS2
to make sure you've got decent performance. This is the nature of
clustered locking, more than anything specific to DLM.

For GFS2, one of the easiest performance wins is to set
'noatime,nodiratime' in the mount options to avoid requiring locks to
update the access times on files when you only read them.

-- 
Digimer
E-Mail:              digimer at alteeve.com
Freenode handle:     digimer
Papers and Projects: http://alteeve.com
Node Assassin:       http://nodeassassin.org
"omg my singularity battery is dead again.
stupid hawking radiation." - epitron