On Wed, Jul 05, 2006 at 10:34:41PM +0200, Petter Reinholdtsen wrote:
> Well, I do not even know what multipath _is_, nor why it is important.
> If that is representative, I suspect the people interested in
> multipath have some work to do to raise the awareness of the problem.
>
> This email is a very good start, but it seem to assume that everyone
> know what multipart is and why it is important. Is multipath machines
> as common as the ppc64 machines, or is the problem affecting a lot of
> users?
Let me give a brief explanation of multipath.
Let's say you want a bunch of disk space. A whole lot -- maybe
terabytes worth. So you buy a SAN, which is a device that might have
dozens or hundreds of disks in it. And you can hook multiple servers to
this SAN. So you have a SAN controller, a bunch of disks in whatever
RAID configurations you like hooked to it, a fibre channel switch, and
each server hooked to the FC switch.
Suddenly you have a lot of really important single points of failure
that could take down not just one but many servers -- the FC switch, the
SAN controller, the FC cables, etc.
So the solution is to build two distinct I/O paths for any server to
reach the disks. The SAN will have two controllers (each with access to
disk enclosures). You'll have two FC switches, one controller cabled to
each. And each server will have two FC links, one to each switch.
Now, when you bring up this system, Linux will assign *two* /dev/sdx
devices for each RAID LUN (basically looks like a disk). At any given
time, exactly one will be readable and useful. That is, the disk can be
probed on both controllers, but only one path will support I/O at any
given time.
Adding to the complexity, which one to use can vary while the system
runs. For instance, if a SAN controller dies, everybody switches over
to the backup path.
The multipath-tools package is the userland support necessary to make
all this work in a sane fashion. It uses the dm-multipath kernel module
to do that.
But it's got some problems:
1) It doesn't properly scan partition tables in multipath devices
2) It doesn't integrate with initramfs, so it's not possible to boot
off a multipath device unless more work is done
3) Some other general bugs and issues
BTW, multipath is often called MPIO (MultiPath I/O)
> > I am gravely concerned, though, about the lack of attention this
> > package is receiving. Does anyone intend to give it some TLC
> > anytime soon?
>
> Perhaps you could give it some tender loving care, and talk to the
> people maintaining the affected packages using IRC and email, and
> hopefully get them to realize why they should fix it in time for
> etch. :)
That's what I intend to do. It's maintained by the LVM folks, though,
and seems to be tied reasonably closely to that somehow. I'm not as
familiar with all this as they are. But it seems like the package is
not really being looked after, given its bug reports.
I have already uploaded multipath-tools-initramfs to Incoming, which
simply installs initramfs hooks and scripts to make it possible to boot
from multipath. We are successfully using it with these scripts at our
site.
> I suspect you might wait in wane if you expect someone else to do
> it. :)
I understand. I'm just trying to figure out if there are interested
parties out here to pitch in, if the LVM folks have plans for it, etc.
I'm brand-new at this and wouldn't be at all surprised if someone else
was more capable at it than I am.
-- John