Ask Slashdot: Spanning Networked Disks?

Andy McCoy wrote in with this doosy! He asks: "Of the network filesystem implementations out there (NFS, AFS, Coda, SMB, NCP are the ones I'm aware of), is there any way to set up a pseudo-RAID network filesystem. I realize it would not be RAID per se, but a filesystem that a) would be distributed across multiple disks, and appear as 1 filesystem, b) would provide redundency between the servers (eg - 1 goes down, one or more of the other servers has a copy of that information, and is still available), and c) is efficient about bandwidth concerns?" Click below for more.

"The target situation is a group of *nix boxes which are set up as workstations, that also have big disks, or extra disks, that can be used for home directories, or other shared file mountings. The filesystem would merge all these available disks into one big 'logical' filesystem that the client would see, and create redundant backups behind-the-scenes between the disks.

A couple methods I have seen / heard of / thought of include:

- automount / NIS(+)
Set up the NIS auto_(home|software) maps to list individual system's exports, and to NFS mount directly from that system to the mount point. The problems with this scheme are that there is no redundancy, no transparency, and requiring NIS or something similar, and the delay associated with automounting.

- NFS re-export
Mounting all the small drives under one subdirectory on a server, and then re-exporting that directory to all the clients, with a bit of monitoring software to keep redundancy. The major (major) problems with this method are the non-transparency of it, and the fact that every file access is doubled (C1->S->C2->S->C1), which slows down already-slow NFS. I would consider a somewhat alternative to replace NFS with Coda, but the same bandwidth and transparency issues exist.

- NFS re-export with host masquerading
Mainly the same as above, except that the master NFS server happens to be smart enough to tell one of it's client disks to masquerade as itself when it sends back the response to the file access. The upside is the transparency increases and the bandwidth decreases (C1->S->C2->C1), but the complexity of this masqerading increases tremendously, and may cause interesting quirks with routings (since you are masquerading as a different server :)

I know Andrew FS and Coda are highly advanced, but I have not seen anything that indicates this sort of capability (spanning networked disks) even though they have the redundancy taken care of. (Unfortunately I have found very little info on AFS, since Coda has supposedly superseded it entirely)"