Re: [Nbd] OCFS2 cluster, HA and nbd-server

[Wouter Verhelst <w@...112...>]

On Thu, Mar 01, 2012 at 10:35:38AM +0100, Daniel Schwager wrote:
> Hi,
> 
> we would like to setup a OCFS2 (Oracle Cluster File System 2) 
> cluster with 2 nodes, using a SAN as a shared storage.
> Each of the nodes has access to the cluster filesystem. 
> 
> Now  the question: Can we setup the running ndb-server 
> processes to be  high-available (HA-linux, ...) ? 
> Of course, a ndb-client is connected to the nbd-server.
>
> ** Szenario1: Failure of one storage node:
> Is there a chance to restart a nbd-server autoamtically (HA) on another
> clusternode
> without problems concerning the connection (means nbd-client does
> not get in trouble) ? Of cource, the client must be configure for
> write-through.
>
> ** Szeario2: Planed migration of all running nbd-servers to another
> cluster node.
> If szenario1 cannot be managed, is there a chance to migrate the
> ndb-servers
> to another node - of course, both nodes has no failure. May we have to
> use LVS
> for TCP session migration ?

There's been talk of a kernel patch that would block writes (rather than
start issuing EIO or similar) until the process that had originally set
up the connection would exit. This would allow an nbd-client to restart
the same connection, or to connect to a fallback server. The former is
implemented in nbd-client (the -persist option), the latter isn't.

I don't know whether that has actually made it into the kernel, though.
You may need to talk to Paul about that.

Note also that until recently (i.e. before nbd 3.0), nbd-client
erroneously would let the child live after a fork() call, rather than
the parent; this would mean that with that version of nbd-client, this
patch wouldn't even work in the first place.

-- 
The volume of a pizza of thickness a and radius z can be described by
the following formula:

pi zz a