[nycbug-talk] Approach for the NFS cluster

[Steven Kreuzer]

On May 19, 2009, at 11:15 AM, George Rosamond wrote:

> Aleksandar Kacanski wrote:
>> Hi, I am interested if anyone implemented NFS in HA configuration and
>> what are folks experiences with different scenarios. I am interested
>> in pNFS, but there is DRBD (harbeat) or just plain cluster of nfs
>> servers without sharing metadata. My requirement is that I need to
>> make service available and I can't have one point of failure...
>>
>> Regards, --Aleksandar (Sasha) Kacanski
>
> AFAIK, carp was developed for file mounts. . . I think SMB though.
>
> Worth the pursuit.

The biggest issue with high availability nfs is that you need to share
a common pool of storage between two servers.

Take a look at geom_mirror and ggate{c,d} and carp.

One the backup machine, you would export the storage as
a block device using ggated to the primary server.

On the primary server, you would mount the block device using
ggatec and then use geom_mirror to mirror writes to the
primary and backup server.

Export the volumes via nfs on each host and configure a carp
interface that clients will connect to.

This is an extremely oversimplified explanation of how you could provide
HA NFS. There are alot of things to take into consideration such as how
you would fail back etc

If this is mission critical, this is definitely an instance where I  
would
solve this problem by throwing money at it. I would suggest you look at
Isilon, NetApp and Sun, all who do exactly what you want very well.

--
Steven Kreuzer
http://www.exit2shell.com/~skreuzer