[Scottish] RAID Setup

[Colin McKinnon]

On Thursday 10 April 2008 10:59, Phillip Bennett wrote:
>
> What do you mean by 'fencing solution'?  I'm looking to only store data
> there that can be read/written to by multiple servers.  Shouldn't the
> filesystem (GFS) take care of this?  The data doesn't have to be fenced, it
> just has to be available.
>
Any cluster system needs to be able to determine where control resides if only 
a portion of the nodes are available - this is fencing. Have a read off the 
stuff at the Linux-HA project, and google for Stonith.

Definitely read the GFS docs - 
http://www.redhat.com/docs/manuals/csgfs/admin-guide/s1-ov-perform.html

Because you are in the unusual situation of having a machine with guaranteed 
availability within the cluster network (although not behaving as an 
equivalent cluster node) you could easily script the 'manual' fencing 
solution for GFS - but RTFM first.

Note that cluster filesystems are designed for low-level concurrency and may 
acheive this by trading off performance. A quick google suggests there are 
comparisons of NFS, ext3 and GFS available.

> And when you say OLTP, I presume you mean Online Transaction Processing?
> I'm just figuring out your terms and reading up first, so I might not be on
> your wavelength here...  It's just straight data files.  No databases or
> anything...

Yes, OK - so RAID 5 makes sense for a small network then (assuming that 
latency is not an issue).

>
> I also understand what you are saying in your last paragraph.  I'm not very
> articulate, as you can see!  I was meaning that you can't access a SCSI
> drive/device independent of the machine that's hosting it.  

Yes you can - you just need to set the id of the other adaptor to something 
other than 0 - people have been doing shared disks this way for years.

iSCSI will happily work with mutliple host accessing the same disk.

In both cases, though, you still need a cluster-capable filesystem if the 
disks are trully shared (as opposed to merely occupying the same bus with a 
failover option - but we're back to fencing again here).

>
> I have thought about this and I see what you mean about pushing the RAID
> over the network (I think).  I was wondering if it would actually work!
> After reading about DRDB though, this is exactly what I'm trying to
> achieve. Would there be any significant gain in using something like DRBD
> over a software RAID1 on the host?  Apart from the fact, it is DESIGNED to
> do what I want, of course. :)
>

I'd always seen DRBD as something more of a realtime backup solution - and I 
can't envisage how you would use it from multiple hosts without a clusterFS 
again.

HTH

C.