[Linux-cluster] Quorum disk over RAID software device

Thomas Meller thomas.meller at gmx.net
Thu Feb 11 14:26:46 UTC 2010

Thanks Gordan!

You put me to the right direction.

Yes, we are driving an active/passive solution.
Each node can see the storage LUNs of both nodes on startup.
We sort them in an init script within the initrd (remove and re-add) to get a well-defined order. Then we start multipathing via vpath. Afterwards, we start root-raid1 via MD, then we mount the root fs and go on booting. This is what we did for the last 4 years.

Driver-based multipathing was our first trial, before we found the current solution, but our colleagues rebelled. Therefore RDAC/vpath. Makes up for a big fat initrd, though.

But ok, techniques evolve, and today dm-multipath is a supported solution.

DRBD is unknown land for me. But it's really promising.
The drawback is a resync when one node boots and longer resync times.
(..if it works as I guess. link to documentation highly welcomed)

We will see how this biest works when FC-communication fails on one or both nodes.

> > I still have a third opportunity. I can use Qlocig's driver-based
> > multipathing and keep using host-based mirroring instead of
> > using dm-multipath, which currently prevents me from setting up
> > raid-devices as root-fs.
> I'm still not sure how all these relate in your setup. Are you saying 
> that you are using the qlogic multi-path driver pointing at two 
> different SANs while the SANs themselves are sorting out the synchronous 
> real-time mirroring between them?

No, both qlogic HBAs do see both Storage controllers via their own SAN.
There are two switched SANs, crossovered for fault-tolerance, but else autonomous. (don't ask how)
The driver does multipathing over both HBAs, so we see the same controllers/LUNs once instead of twice.

But we don't use this anymore. It worked ok, though.
Mirroring disk partitions from controllerA to controllerB was always done by MD-raid1. (simple&good)

> > Well, that will work, but is somewhat ugly.
> > 
> > So far, I had only a short glimpse on OSR. I think I will need
> > to dive deeper.
> It sounds like you'll need to add support for qlogic multi-path 
> proprietary stuff to OSR before it'll do exactly what you want, but 
> other than that, the idea behind it is to enable you have have a shared 
> rootfs on a suitable cluster file system (GFS, OCFS, GlusterFS, etc.). 

As far as I understood, no. I need to do multipathing and mirroring. Whether I do it by dm-multipath and MD-raid or by QLA-Multipath and DRBD, or any other combination, doesn't change the result. It only changes the way the whole thing behaves in case of a failure.

Actually, I hope to be able to start a DRBD device in the initrd and mount it as root. That is the first hurdle to be taken.
The rest is nice-to-have.

> It's generally useful when you need a big fat initrd
> to bootstrap things such as RHCS components, block device drivers 
> (e.g. DRBD), or file systems that need a fuller environment to start up 
> than a normal initrd (e.g. GlusterFS, things that need glibc, etc.).

Nice-to have. I am glad to know there is such a thing. When we started, there was nothing.

I will report back when I have some results.

Thanks for your help!


Jetzt kostenlos herunterladen: Internet Explorer 8 und Mozilla Firefox 3.5 -
sicherer, schneller und einfacher! http://portal.gmx.net/de/go/atbrowser

More information about the Linux-cluster mailing list