[clug] GFS - add some more abstraction [SEC=PERSONAL]

Daniel Pittman daniel at rimspace.net
Tue Mar 30 19:11:03 MDT 2010

"Roppola, Antti" <Antti.Roppola at daff.gov.au> writes:

> So is it just me, or does GFS suffer from compulsive abstraction?

...just you, I think.

> For a shared filesystem it seems to want layers for clustering, LVM and a
> good fistful of services too.

For a multi-master shared file-system using a single underlying storage
device, it wants...

Which is a much harder job than just "shared file system", and gives you much
higher performance in return: you can get the full speed of your SAN or DAS
devices on each node, more or less, at the latency of SAS rather than IP.

The requirements are more or less the same as OCFS and friends, incidentally,
which are some of the few competitors in the space.

> It seems a bit of overkill if one just wants a shared filesystem on a small
> private network.

Absolutely.  I can't imagine why GFS would look attractive in that scenario.

> I find myself thinking fondly about NFS and wondering if GFS is really worth
> all that extra effort.

It really isn't.  Just use NFS, or CIFS, or some other network file system,
and you should be much happier.

> Has anyone else grappled with this?

*nod*  If multi-master and replicated is attractive, you might also find
glusterfs interesting to investigate — but it has another set of trade-offs
compared to just using a basic NFS type solution, and probably isn't worth it
for your needs.

> P.S. The GFS documentation also seems pretty patchy. There's a few behemoth
> documents that seem to assume you are building a data centre and not too
> much on why each layer is important.

I hate dealing with RedHat products because that is usually the state of the
documentation: almost all reference works, very few useful documents.

OTOH, if you are using GFS you *are* building a data centre (class solution),
so that assumption isn't too surprising.


Actually, the assumption that you know why distributed locking for your shared
storage is a pretty reasonably thing, too, since folks who need this would
already know why their cluster management solution needed locking. :)
✣ Daniel Pittman            ✉ daniel at rimspace.net            ☎ +61 401 155 707
               ♽ made with 100 percent post-consumer electrons

More information about the linux mailing list