[Gluster-users] trouble combining nufa, distribute and replicate

[Jeff Darcy]

On 07/07/2010 07:13 AM, Matthias Munnich wrote:
> thanks a lot for your comment! It sounds like I hit a bug here.  Would 
> you still feel queasy if I added "option lookup-unhashed yes" to avoid dht?

I don't think that will really avoid dht, especially the
extended-attribute parts that are the likely cause of conflict with nufa.

> What I reported here was some initial testing.  In the end I like to use 
> glusterfs to provide a uniform name space for our O(20) workstations
> with lokal storage of 4 to 16TB.  The data should be mirrored (once)
> for reliability and stored locally were possible for speed.   I also
> would prefer not to glue together local disk using LVM or software
> raid to keep it easy to add/remove disks without having to expand a
> filesystem. 
> 
> Any hints how to  set this up  with glusterfs?

I think you can get "close enough" with just nufa on top of replicate.
If you have two disks per machine, no matter how many machines you have,
assign one to be active and one passive.  Pair each active to a passive
on another node however you like using cluster/replicate, then combine
all of those using cluster/nufa with local-volume-name pointing to the
replica pair where the active half is local.  If you had more than two
disks per machine, this wouldn't work quite as well but the differences
should be minor.  If that's not good enough, I just looked at the code
and it doesn't seem like it would be hard to make nufa recognize
multiple volumes (instead of just one) as local.  If you'd like, I could
try generating and submitting a patch for that.