[Linux-cluster] GNBD -> MD -> GFS?

[Jayson Vantuyl]

MD is a problem.  Since it's not cluster aware, you can have a  
situation where two atomic writes aren't actually atomic.

Specifically, there are times when a write to disk need to either  
happen or not happen.  Often this write spans more than a few  
blocks.  With clustered locking, you get a guarantee that two writers  
won't attempt to hit the same data at the same time.  However, with  
Linux MD, it's just not that simple.

Assume that you have blocks A and B that both must be written  
atomically.  If two nodes try to write them simultaneously, it is  
possible for block A to be a copy from one node but block B to be  
from a different one due to timing.  Things get even hairier when MD  
is rebuilding a RAID, as both machines will both attempt to remirror  
the system!  Madness ensues.

This is the reason that Redhat has put together the cmirror target.   
CMirror will do a simple RAID1 style DM target across a cluster.   
There is also CSnap for doing snapshotting.  There isn't anything  
like CRAID5.  Unfortunately, this is all a bit bleeding edge and not  
so integrated into the stack you want to use (CLVM would need to be  
in there for one).

On May 22, 2007, at 12:02 AM, Nathaniel Eliot wrote:

> I've been looking for a Linux clustered file system that can mirror/ 
> stripe
> across storage servers.  The end goal is to build a HA cluster from  
> just FOSS
> and commodity parts.  I've found two existing possibilities, each  
> with drawbacks:
>
> 1) GPFS can do redundant storage, but isn't FOSS.
> 2) DRBD + GNBD + GFS is free, but can only do two active nodes.
>
> I'm considering putting software RAID between GNBD and GFS on each  
> storage node.
>  The superblocks would be retained, so each MD would theoretically  
> have the same
> view of data.  GFS is cluster aware, and GNBD built for clusters;  
> the big
> question is, can they work sandwiching the non-cluster-aware MD?
>
> Any quick links/explanations as to why this won't work, and/or useful
> suggestions, would be appreciated.  Thanks,
>
> -- 
> Nathaniel Eliot
> T9 Productions
>
> --
> Linux-cluster mailing list
> Linux-cluster at redhat.com
> https://www.redhat.com/mailman/listinfo/linux-cluster



-- 
Jayson Vantuyl
Systems Architect
Engine Yard
jvantuyl at engineyard.com


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/linux-cluster/attachments/20070522/2ce0137f/attachment.htm>