I'm trying to plan a production platform where we might want to try
for at least four nines, five if anybody wants to pay the price we'll
have to set...

So I want to be able to have two independent 'disk boxes' with
replicated content --- these could be SAN disk shelves or Linux hosts
acting as iSCSI targets, with internal hardware RAID in either case
--- and five to ten application hosts in a cluster using a shared
file system to access the disks.

It's not essential to have symmetrical read access to the two copies,
one could be a passive standby, but active-active is neater.

First question: Am I silly for wanting this?

Now add to this that I might be asked to provide business continuity
by replicating the file systems over a MAN...

When looking around for similar setups, the closest thing I found was
the ability to run GFS/OCFS2 over DRBD. But I don't really want my
disks replicated five to ten times, two should be enough.

What I really would like is:

1) Use LVM to make a mirrored VG out of the logical disks on the
separate SAN boxes, activate that VG on all the application servers
and run OCFS2 on that.

In theory, OCFS2 or GFS should be able to tolerate the write/read
order inconsistency that mirroring introduces, because they use
distributed locking to preserve order when it matters. But I haven't
read up on exactly what shared FSes expect from the underlying device,
so I could be wrong.

But if a write to one mirror fails, LVM would need to how to disable
it on all cluster hosts before it signals completion (of that write or
the next write barrier) to the shared FS instance on the local host.

Will CLVM in RedHat do that? Will it ever appear in a mainline kernel?

(The next half isn't really a ha-dev question, but I'm putting it in
for completeness).

2) Use DRBD on raw RAID disk between two Lunix boxes, export that
mirrored disk as a iSCSI target, and let the application boxes use
iSCSI multipathing under the shared FS.

A 2U Linux box with room for 6 Hotswap disks comes to about the same
as a single-controller iSCSI shelf with room for 14 disks, so it's not
quite as convenient as 1). And the iSCSI target code is not the most
mature, I've heard. And I can't see where LVM would fit in there at
all. And I've got two different sorts of Linux host now. So this is
clearly a fallback solution.

2a) Put the disks in two SAN boxes, let one Linux box own each disk
and do the same as above. Generates a lot of superfluous net traffic,
and the 'disk master' resources probably cannot share hosts with
anything else --- but they could still be 'regular' cluster hosts.
LVM is still a problem, and running iSCSI on DRBD on iSCSI is silly,
but is it viable?

And for completeness:

3) I can throw extra money at the thing and get SAN boxes that will do
the mirroring for me.

Regards,
Lars Mathiesen
Sifira A/S
_______________________________________________________
Linux-HA-Dev: [email protected]
http://lists.linux-ha.org/mailman/listinfo/linux-ha-dev
Home Page: http://linux-ha.org/

Reply via email to