On Mon, Apr 30, 2012 at 02:20:31PM +0200, Frank Van Damme wrote:
> Hi.
>
> I am in the process of setting up a small 2-node HA cluster (for NFS,
> active/passive) with a shared disk for storage. Because Corosync and
> pacemaker look nice and good, I am trying to make it work with this
> combination instead of cman, which is what clvm is compiled for in Debian.
>
> There are a couple of howto's on the web on recompiling the lvm package
> with Openais or Corosync , so that's what I did. However, I'm still
> confused about the relationship between the cluster manager and the lock
> manager. If I understood right, this is the correct order in which
> things will be started:
>
> * start Corosync (or openais, if clvm is compiled with openais support
> instead)
> * you need to run some kind of resource manager
> * the resource manager (pacemaker) must have a resource defined that is
> the dlm controld - I tried starting dlm_controld at the command line,
> this seems to make clvm happy but unfortunately that command has no man
> page.
> * then, a bunch of files appear in /sys/kernel/config/dlm
>
> Actually, that controld should be started by the cluster resource
> manager is also something I just read on the web, and it sounds kinda
> odd to me. After all, you need to run an instance on every cluster node
> that accesses shared storage, right?
>
>
> [[I also tried the same thing but with a clvm compiled for openais, I
> read it uses another lock manager ("Lck") but I'm not sure if and how
> this one is actually cluster-aware since it doesn't seem to rely on a
> resource defined in the cluster.]]
>
> [[Another thing that I find rather strange is that corosync is started
> waaaay at the end of the boot process, and clvm is started at the
> beginning, even if I changed the dependency of 'cman' to 'corosync' in
> clvm's init script. (clvm is started in rcS.d). But that's got more to
> do with lsb init scripting than with clustering, I suppose.]]
>
>
> If anyone could make me wiser or more educated about any of this, I'd be
> grateful :-)Hello, IIRC a distributed lock manager (DLM for short) is needed whenever an active/active resource is used, e.g. a shared filesystem. OCFS2 and GFS2 implement their own DLMs. For an active/passive cluster, you'd need to define your resources and a floating IP with stickyness WRT to those resources. A DRBD in master/slave would be more appropriate IMHO. Others might have other experiences/opinions or corrections. -- Primary key fingerprint: AD8F BDC0 5A2C FD5F A179 60E7 F79B AB04 5299 EC56
signature.asc
Description: Digital signature
