[OpenAFS] looking for a dfs on linux

[Cees de Groot]

Octave <oles@ovh.net> said:
>Do you know if a file system (AFS ?) can do it ? Or works
>with failed servers ?
>
There is mirroring stuff available for Linux, based of the concept of network
block devices. DRBD (http://www.complang.tuwien.ac.at/reisner/drbd/) seems to
be the most popular. It's not RAID-1, because one of the servers is in
stand-by mode and you need to have heartbeat or something similar for
fail-over. The good thing is that you can mirror anything because the
network-mirrored device is just a block device; apart from performance
considerations, you'll be able to mirror databases, fileservers, and with
a bit of initrd hacking even the root filesystem.

AFS can replicate read-only volumes. By making read-write volumes small and
scattering them over multiple servers, plus replication of read-only volumes,
you won't have HA but at least a server going down will have a very low
impact (because only the set or R/W volumes living on that server will be
gone). AFS complements this with excellent backup tools for quick restoration
of the dead volumes on other servers. A sort of manual failover :-).

There's GFS, which lets you do what you want (pool a lot of servers for
HA), but it seems the original GFS people have gone commercial/closed
source with it (however, for an Enterprise Computing gadget it's not
extremely expensive, but the open source people are looking at their
legal options so it might not be a long-term thing) and I don't know
whether the Open Source fork of GFS is ready for prime-time. In any case,
this is a very interesting development, because GFS lets you mix-and-match
anything from network block devices to dual-attached fiberchannel drives.

Lessee - Intermezzo, which is CodaFS done right but it is in development and
personally I don't like the idea of having a filesystem done in Perl ;-).
Seriously: doing it in Perl is bound to make it more stable than in C, but I
really don't know whether Intermezzo is already usable. If it is, and lifts
the limitations from CodaFS, it's something worth looking at: not only
replication, but disconnected operation is also possible.

Apart from that, Compaq, IBM and Silicon Graphics are involved in a
pissing contest about who can open source the best clustering/HA solutions
for Linux.  It seems like some of the bewildering heap of products are
going to be part of Linux 2.5 - Compaq's CI, IBM's network locking stuff,
and GFS seem to be on the candidate list. With a bit of luck, SuSE 8.4
will have this all out-of-the-box ;-).

(discussion topic: what's the value of running AFS on top of DRBD or GFS? Is
it complementary or competing?)


-- 
Cees de Groot               http://www.cdegroot.com     <cg@cdegroot.com>
GnuPG 1024D/E0989E8B 0016 F679 F38D 5946 4ECD  1986 F303 937F E098 9E8B