[Lustre-discuss] Lustre Thumper Fault Tolerance

[Andreas Dilger]

On Mar 10, 2008  09:09 -0600, Colin Faber wrote:
> Is this true even in the case of mounting the OSS as a read only node?

Yes, definitely even a "read only" mount can cause serious corruption.
There are several issues involved, the most dangerous is that even for
read-only mounting the journal is replayed by the kernel or otherwise
the filesystem may appear to be corrupted.

In addition, there is the problem that (meta)data that is cached on the
read-only mounting node will become incorrect as the writing node is
changing the filesystem.  The ext3 filesystem is not cluster aware.

In order to prevent situations like this, the newer releases of ldiskfs
and e2fsprogs have an "mmp" (multi-mount protection) feature which will
prevent the filesystem to be mounted on another node if it is active
on one node (either mounted, or running e2fsck).

This will be enabled by default on newly-formatted filesystems which
are created with the "--failover" flag, and can also be enabled by
hand with "tune2fs -O mmp /dev/XXXX" (replace with MDT or OST device
names as appropriate).  This will prevent the filesystem from being
mounted or e2fsck'd by old kernels/e2fsprogs so it isn't enabled by
default on existing filesystems.

> Andreas Dilger wrote:
>> On Mar 07, 2008  00:04 +0530, Neeladri Bose wrote:
>>   
>>> To address the performance hit (whatever be the %age) if we setup DRDB in 
>>> active-passive mode across the 4500's but have the LustreFS points to 
>>> separate raid sets from the network across the DRDB pair of 4500's & thus 
>>> become an active-active solution which may actually increase the 
>>> throughput of the LustreFS.
>>>
>>> Can it be a possible scenario using DRDB on Linux with ext3 & LustreFS?
>>>     
>>
>> No, Lustre does not support active-active export of backing filesystems.
>> This doesn't work because the backing filesystems (ext3/ZFS) are not
>> themselves cluster-aware and mounting them on two nodes will quickly
>> lead to whole filesystem corruption.

Cheers, Andreas
--
Andreas Dilger
Sr. Staff Engineer, Lustre Group
Sun Microsystems of Canada, Inc.