[Gluster-users] server side afr, workstation unify?

[Keith Freedman]

I'm not sure if anyones responded to you.  if not, it might be 
helpful to post your config files.

my comments are inline below

At 09:11 PM 8/4/2008, Wolfgang Pauli wrote:
>Hi,
>
>I am having trouble with the following setup, hope that somebody will be able
>to help.
>
>We have a couple of workstations and two file-servers. Let's just 
>keep it at 2
>workstations for now, because I can just add more clients later I guess. So
>the workstations are called echo and reality, and the servers are dream and
>neo.
>
>I am wondering whether it is possible to have a unified filesystem between
>echo and reality, using the nufa scheduler to keep things local as much as
>possible, and at the same time to have afr running on the servers neo and
>dream so that those guys each have a complete copy of the workstations (echo
>and reality).

When you mean, unified filesystem, you mean the same filesystem 
available on both clients?  and not a Union filesystem (whereby you 
combine multiple filesystems into one large virtual filesystem) via unify.

If so, this is kind of standard.   What I'd suggest is look in the 
wiki examples for the server side AFR, and then set up your clients 
using the example, my memory escapes me, sorry, but there was an 
example which used IP round-robin to do a kind of automated server 
failover, but this may not help with load-balancing.

So I suggest you set up the local cache translator to reduce your 
network overhead.
If you need better load balancing on the servers then you can use 
NUFA or something like that to help out.

>I can get it to work with only afr, or only unify, but combining the two
>somehow does not do it. The best I got so far was to have empty directories
>on the servers.

what're you trying to unify?  If you're servers are AFRing eachother, 
then this is one filesystem.
if you then want to unify multiple AFR bricks, you can do this, but 
again, I think you're better off having all the hard work done on the 
servers, and letting the clients just mount the unified volume from 
the servers.

I think it's possible to AFR the unify metadata nowadays, but I don't 
know for sure.

>Worst case is that I go with not doing the unified thing, but it 
>would be nice
>because it is so much faster and most of the time, files are just accessed
>locally anyway. With only client side afr, nfs seemed to be faster (with
>varying performance translators), and with server side afr, nfs and glusterfs
>where about the same speed - with the big advantage of not having a single
>point of failure. :)

Again, I think all your problems are solved by doing all the hard 
work on the servers and the clients just mount the unified brick from 
the server.

>Maybe this setup just doesn't make any sense at all ...
>
>Well, hope somebody has some suggestions!

I'm not sure if I pointed you in the right direction or down a path 
of despair.. hopefully the former.

Keith