[linux-lvm] Re: write performance with active snapshot

[Peter Daum]

Hi,

Larry Dickson wrote:
> My guess is that you are getting the typical seek overhead. Have you 
> tried making a volume group out of two separate RAID arrays (driving 
> different spindles), and using lvdisplay --maps to make sure the parent 
> volume is on one array, the snapshot(s) on the other?

That was my suspicion, too (although I could not imagine such an extreme
impact). Just for testing I added a single disk to the same volume group
and put the snapshot onto that disk - amazingly it made hardly any
difference (Actually, I'm almost glad about that, because the combination
of a 12-disk-array with a single disk would be under almost all other
aspects foolish).

One thing that does improve the performance a little (actually by 100%,
which in this case meens still pretty lousy 16 MB/sec) is to increase
the chunk size to 512kb. (I don't know yet, how this might affect
performance when dealing with many small files) ...

Regards,
                     Peter


> On 11/9/08, *Peter Daum* <gator_ml at yahoo.de <mailto:gator_ml at yahoo.de>> 
> wrote:
> 
>     Hi,
> 
>     for an application I am just working on it looks like lvm snapshots
>     would
>     be just what I need as far as functionality is concerned. Unfortunately,
>     I am experiencing such a massive degradation in performance, that the
>     result is almost useless.
> 
>     I'm working on a fairly fast machine (Quadcore, 8GB RAM) with a big
>     hardware RAID array and lvm2 (Debian Lenny; Linux 2.6.26-1-amd64;
>     LVM version:2.02.39 (2008-06-27)
>     Library version: 1.02.27 (2008-06-25)
>     Driver version:  4.13.0)
> 
>     Sequentially writing to a file (ext3) on a logical volume, I get a
>     sustained performance of ~ 250 MB/sec. When I create a snapshot
>     volume, the write throughput drops to 7-8 MB/secs (on the original
>     volume; writing to the snapshot I see a significant degradation,
>     but not nearly, as bad; read performance is o.k.).Is this "normal"
>     or is there anything I can do to about it?
> 
>     I looked in this list and searched the WWW but couldn't find any
>     concrete information on the performance impact of snapshots
>     (except http://www.nikhef.nl/~dennisvd/lvmcrap.html).
>     It seems like write performance should probably be less then 1/3
>     of the original throughput, because every write to the source
>     volume causes 3 I/O operations plus some overhead for meta data.
>     More difficult to estimate would be the time lost by additional
>     head movements. Still, a throughput degradation by a factor of 30
>     seems pretty extreme.
> 
>     Any ideas?
> 
>     Regards,
>                             Peter Daum