[linux-lvm] Snapshots (overheads and 2.6 kernel status)

[linux-lvm] Snapshots (overheads and 2.6 kernel status)

[C R Ritson]

Does anybody know how much overhead is involved in keeping a snapshot of
an active volume (for 24 hours)? I am currently running LVM1 (waiting
for snapshots in LVM2/device mapper) and am making a snapshot of my user
volumes overnight. These are left available for 24 hours in the hope
that users who damage a file will be able to get "yesterday's copy" from
the snapshot without us needing to restore from tape. There is obviously
some overhead involved, but does anyone know how much?

I guess that the worst case situation is when a file is updated for the
first time that day (an example might be a recompilation, particularly
if the underlying disk is nearly full, so that the blocks released by
removing the object files are quickly reused). In this case, the
original data blocks must be placed into the snapshot cache and the new
data blocks written to disk. Two writes for the price of one.

Perennial question - having had a look at the dm-devel and the lvm2
archives, I cannot find any comments about the status of snapshots in
the 2.6 kernel. Can anyone enlighten me?

Chris Ritson (Computing Officer and School Safety Officer)

Rm 618, Claremont Bridge,       EMAIL: C.R.Ritson@ncl.ac.uk
School of Computing Science,    PHONE: +44 191 222 8175
University of Newcastle,        FAX  : +44 191 222 8232
Newcastle on Tyne, UK NE1 7RU.  http://www.cs.ncl.ac.uk/

Re: [linux-lvm] Snapshots (overheads and 2.6 kernel status)

[Steve McIntyre]

On Mon, Nov 29, 2004 at 03:08:50PM -0000, C R Ritson wrote:
>Does anybody know how much overhead is involved in keeping a snapshot of
>an active volume (for 24 hours)? I am currently running LVM1 (waiting
>for snapshots in LVM2/device mapper) and am making a snapshot of my user
>volumes overnight. These are left available for 24 hours in the hope
>that users who damage a file will be able to get "yesterday's copy" from
>the snapshot without us needing to restore from tape. There is obviously
>some overhead involved, but does anyone know how much?

It can be quite significant, depending on your workload. I'm using
snapshots in a similar way, and I saw performance drop by a factor of
2 (or more) on our main NFS server. It's very useful to have the
snapshot, so we live with it.

-- 
Steve McIntyre, Plasmon                       smcintyre@software.plasmon.com
 
"C++ ate my sanity" -- Jon Rabone