Hi Andreas thank you for answering these questions. But I guess there is one more missing
element that I don't get it:
The backup device that you backup your MDS/MDT/MGS to : did you create it with mkfs.ext3
or mkfs.lustre ?
In other works can I do a backup using dd to an ext3 file system device or not?
Best Regards,
Amit
-----Original Message-----
From: Dilger, Andreas [mailto:andreas.dilger@intel.com]
Sent: Wednesday, April 10, 2013 4:44 PM
To: Pablo Escobar
Cc: Kit Westneat; Kumar, Amit; hpdd-discuss(a)lists.01.org
Subject: Re: [HPDD-discuss] Lustre Disk to Disk MDS Backup Question
On 2013/09/04 4:54 PM, "Pablo Escobar" <pescobar001(a)gmail.com> wrote:
could you post or link some more detailed technical docs about how you
snapshot and backup the mds?
My main doubt is that if I do an snapshot and backup of mds, for
example, on friday and during the weekend the client nodes modify the
filesytem and then I need to recover on monday. wouldn´t it be
inconsistent? would be possible to recover the data in this situation?
Is there any docs about this recover procedure?
Snapshots are done using normal LVM commans "lvcreate -s" to create a snapshot
of the MDT device. I'm sure there are equivalent dmadmin commands to do the same?
Then, simple "dd" to copy the MDT snapshot to a separate block device (actually,
two separate devices on alternate days, and one only manually).
The MDT is small, and the backup devices do not need to have good IOPS performance, just
simple 5400-RPM SATA is enough for streaming "dd"
copies. Since the MDT is a single point of failure for all of the data in the filesystem,
it makes sense to keep this backed up even if the data is too large to back up.
This is only intended for emergency recovery, since I also test experimental patches on my
home system at times, but it holds real data that the rest of my family cares about (music
collection, all of the TV we watch, etc).
Yes, if there is a recovery from the backup MDT copy, the newer files are lost, but that
is true of any backup system. In theory it would be possible to reconstruct the new files
from their OST objects using lfsck (in my case mostly just newly-recorded TV shows), but
that has not been a primary concern compared to being able to recover the rest of the
filesystem.
Cheers, Andreas
2013/4/10 Dilger, Andreas <andreas.dilger(a)intel.com>
On 2013/09/04 9:47 AM, "Kit Westneat" <kwestneat(a)ddn.com> wrote:
>You should do that in order to get a clean image. If you don't, then
>you are more likely to get an inconsistent image. Of course, an
>inconsistent image is better than no image, but ideally you'd want to get a clean
one.
Or, since this is a device mapper device, you can create a snapshot,
and then use "dd" from the MDS snapshot (which should be consistent).
That's what I do for my MDS backups, and it has saved me a couple of
times.
Cheers, Andreas
>On 04/09/2013 11:33 AM, Kumar, Amit wrote:
>> Hi Kit
>>
>> This is good to know. Do I still need to shutdown the lustre FS? And
>>Umount the devices?
>>
>> Thank you
>> Amit
>>
>> Sent from my iPhone
>>
>> On Apr 9, 2013, at 10:27 AM, "Kit Westneat"
>><kwestneat@ddn.com<mailto:kwestneat@ddn.com>> wrote:
>>
>> Hi Amit,
>>
>> If you are doing a disk to disk backup, and the size of the backup
>>disk is >= the main disk, I would just use dd and not worry about
>>formatting the disk beforehand. The dd will replicate the filesystem
>>structure exactly, so there's no need to mkfs. I would do something
>>like 'dd
>>if=/dev/dm-0 of=/dev/dm-2 bs=1M'.
>>
>> - Kit
>>
>> On 04/09/2013 02:40 AM, Kumar, Amit wrote:
>> I forgot to paste the link I referenced to perform backup: here it is:
>>http://wiki.lustre.org/manual/LustreManual18_HTML/BackupAndRestore.ht
>>ml Any thoughts on this will be very helpful!
>> Thank you,
>> Amit
>>
>> From:
>>hpdd-discuss-bounces@lists.01.org<mailto:hpdd-discuss-bounces@lists.0
>>1.o
>>r
>>g> [mailto:hpdd-discuss-bounces@lists.01.org] On Behalf Of Kumar,
>>g> Amit
>> Sent: Monday, April 08, 2013 12:47 PM
>> To: HPDD-discuss@lists.01.org<mailto:HPDD-discuss@lists.01.org>
>>(HPDD-discuss@ml01.01.org<mailto:HPDD-discuss@ml01.01.org>)
>> Subject: [HPDD-discuss] Lustre Disk to Disk MDS Backup Question
>>
>> Dear All,
>>
>> ³Version 1.8²
>>
>> We have both MGS & MDT mounted on two different block devices.
>> Now I would like to perform a disk to disk backup of both of these
>>block devices to another set of block devices.
>>
>> Although this link tells me how to perform that. There are some
>>missing things that I want to understand before I can perform this
>>backup
>>
>> We have two devices /dev/dm-1 & /dev/dm-0 mounted as /mgs and /mdt
>>respectively.
>>
>> Now I have two more devices /dev/dm-2 & /dev/dm-3 which I want to
>>mount it as /mgs-backup and /mdt-backup and then perform the disk to
>>disk backup. Because we are not backing up any of the metadata right now.
>>
>> Here is where I get lost: when I want to make the lustre file system
>>it asked me if the device is going to be an mgs or mdt or it will be
>>combined. Now that I already have an /mgs and /mdt partition on this
>>node what will happen if I make new block device to be an mgs and mdt
>>partition, will this conflict with the existing ??????
>> Or can I just make the new devices just an ext3 partition and still
>>be able to make a disk to disk copy of meta data that I can restore
>>when needed.
Cheers, Andreas
--
Andreas Dilger
Lustre Software Architect
Intel High Performance Data Division