[pnfs] Linux pNFS road map

[J. Bruce Fields]

On Wed, Dec 12, 2007 at 12:35:55PM -0800, Marc Eshel wrote:
> I know that this what the protocol requires but why would you choose to do 
> i/o through the MDS when the performance is guarantied to be worst, even 
> for small files.

Sure.  But if you have some little client that needs access to some of
the same data but doesn't need performance, it'd be really convenient to
be able to just plug it in.

> I think that to make this really usable it will take much 
> more work and I believe that trying to extend GFS2 to support pNFS is much 
> better route to take. We can talk about it some more in tomorrow's status 
> meeting.

As far as I can tell, the reason we've got NFS in the kernel is that
that's where the filesystems are--it's hard to get sensible semantics
for sharing between remote clients and local applications otherwise.

So, I agree with Marc, for a cluster filesystem like GFS2 that's already
in the kernel and that might host applications and services other than
NFS, I think it makes sense to use the kernel nfs server.

spNFS still sounds interesting!  But if you don't care about local users
on the filesystem--if you're essentially doing the filesystem in
userspace--then why not just run the whole service in userspace?  Samba,
Apache, and others, seem to do that happily.

Jeff Garzik has some sort of userspace NFSv4 server.  I haven't looked
at it--my impression was it's pretty primitive--but it might be a start:

	http://linux.yyz.us/projects/nfsv4.html

Or maybe the newpynfs server could be used to bootstrap something.  Or
there's a guy whose given talks on an NFS server called "Ganesha", but I
don't know if he's released code.  Or it might be possible to just to
port some of the knfsd code to userspace.

--b.