Mark,
An "AFS/DFS-aware tar" might seem speedy but I believe
you really need to run the backup at the volume or fileset
level for the simple reason that fileservers organise
data in volumes/filesets. Thus, when it comes to recovering
all the volumes from a failed hard disk on a fileserver it
is straightforward to get those volumes off the backup device
onto the recovered disk.
How would you recover the contents of a failed hard disk
on an AFS fileserver if you only have the directory/file
level backup in a tar archive? Where is the mapping between
filenames and disk/partition?
I would also question if the tar type of backup will scale.
Remember that all access to filesystem data via tar requires
data to be fetched via the AFS Cache Manager. So, the entire cell
will be "scanned" through the Cache. Can this be as efficient
as backing up by volume (with no CM operations involved)?
The other advantage of the volume backup is that all the
data is backed up. However, with "file level" backups (eg tar)
you may be denied access to sub-directories by ACL controls.
If you are looking for ways to speed up backup consider options like:
a) have a backup device per fileserver (use butc)
this could be a stacking tape device auto cycling tapes
b) use high speed backup devices (eg DLT beats writeable CD)
c) use high speed connections (eg ultra SCSI, SSA, ATM)
d) make your AFS database servers pure
(eg only run AFS db processing, nothing else)
e) make your AFS database servers fast (CPU clock speed, local disk)
e) subset your volumes to backup only data that needs to be backed up
--
paul http://acm.org/~mpb
intergalactical brotherhood of magicians
incandescent burning machinery
incredible bureaucracy machine
ikkle bikkle microprocessors
immutable borg monster
invisible blue monkeys
interesting big mess
i bought manhattan
i blame microsoft
i buy microsoft
i bash meanies
Mark Plaksin <[EMAIL PROTECTED]> wrote on [EMAIL PROTECTED]
on 23 Jun 1999 17:06:39 -0400
>This brings up an interesting point....
>
>Does everybody using AFS or DFS notice that it takes *forever* to backup
>lots of volumes (filesets in DFS-speak)? We're using DFS and the
>per-fileset overhead is very high. Tar'ing the data is super-speedy but
>using the DFS backup system takes ages.
>
>I've complained to IBM (or DCE/DFS vendor) that the backup system is so
>slow that it's broken but they weren't impressed :)
>
>We'd really love to see a AFS/DFS-aware tar! I bet that would be speedy
>and would probably meet our needs just fine.
