On Wed, 14 Aug 2002, Greg Freemyer wrote:
> >> The normal way we setup something like this is to use a shared scsi
> >> or fiberchannel disk setup, that way you no longer have a singe
> >> point of failure. If you are putting the disks into one of the
> >> two computers then if that system goes down you are dead in the
> >> water. The other aspect of the above setup is using something
> >> like failsafe, the two nodes monitor each other, and can (if
> >> setup correctly) shoot the other node down and take over the
> >> filesystem if it detects problems. failsafe itself is opensource,
> >> not sure if the components of it which fail over the filesystem
> >> to the other node are. Of course all this assumes using the fs
> >> via NFS - the second node takes over the ip address of the failed
> >> node.
> >> A quick look at drbd on the net seems to show that it is designed
> >> for these sorts of setups, and it has links to all the high
> >> availability stuff for linux.
> >> That aside, there should be no problem doing this, provided you
> >> make sure the unmount from one system is complete before attempting
> >> to mount on the other system. Mounting a filesystem which is
> >> already mounted elsewhere is not a good thing, the second system
> >> will think the fs needs recovery running on it.
> >> Steve
> I don't think drbd is designed for use in a shared SCSI environment.
> It is more of a RAID 1 driver where the 2 halves of the mirror are on the 2
> different servers using internal disks.
> I believe it has support for ordered writes, but I for one would not simply
> assume XFS and drbd are compatible.
I've asked a similar question to Dan Yocum (very active on this front)
some time ago, and AFAICT unfortunately XFS and DRDB don't play well
together because DRBD lies on the assumption of constant chunk I/O (ala
ext2/3), while XFS uses variable chunks.
IIRC, addressing this problem it's on the drbd TODO list (not in the