[linux-lvm] relocate on write for snapshot

[linux-lvm] relocate on write for snapshot

[Ming Zhang]

Hi, I wonder if there are any thought about a relocate on write policy
for snapshot instead of copy on write policy used now?

instead of copy old one to snapshot, overwrite old one with new one, 2
writes and 1 reads. it is possible that write new data to a usused
location directly.

i know later remove a snaphot will be a little trouble, but there must
be some way to get around it.

just a rough thought, any comment?

ming

Re: [linux-lvm] relocate on write for snapshot

[Clint Byrum]

On Fri, 2004-11-19 at 11:39 -0500, Ming Zhang wrote:
> Hi, I wonder if there are any thought about a relocate on write policy
> for snapshot instead of copy on write policy used now?
> 
> instead of copy old one to snapshot, overwrite old one with new one, 2
> writes and 1 reads. it is possible that write new data to a usused
> location directly.
> 
> i know later remove a snaphot will be a little trouble, but there must
> be some way to get around it.
> 
> just a rough thought, any comment?
> 

I thought the way it worked now was for snapshot LV to be marked as the
"active destination" for all new writes to the logical volume. On write,
the data isn't copied, just the metadata is changed to reflect which
physical extent the block is now located on, and the old one is then
reallocated as part of the snapshot.

Am I wrong?

-- 
Clint Byrum <cbyrum@spamaps.org>

Re: [linux-lvm] relocate on write for snapshot

[Nils Juergens]

On Fri, 19.11.04, Ming Zhang <mingz@ele.uri.edu> wrote:

> instead of copy old one to snapshot, overwrite old one with new one, 2
> writes and 1 reads. it is possible that write new data to a usused
> location directly.

Since you have to delete the snapshot (and I don't think they live very long
on most systems) you don't gain anything (because you have copy the changes
back from the snapshot), but you make the cases of more than one snapshot
and (possibly) read-write snapshots a lot harder, maybe not in CPU or IO,
but certainly in code complexity, which is a thing you never do to gain
a bit of performance.

> i know later remove a snaphot will be a little trouble, but there must
> be some way to get around it.

Yes, you have to do the work you saved a couple of minutes earlier :)

just my EUR 0.02,
Nils

-- 
Nils Juergens  | nils@muon.de | icq 7090774
Having problems sending big files over the net?
Try out Efisto (http://efisto.org).

Re: [linux-lvm] relocate on write for snapshot

[Ming Zhang]

On Fri, 2004-11-19 at 12:19, Nils Juergens wrote:
> On Fri, 19.11.04, Ming Zhang <mingz@ele.uri.edu> wrote:
> 
> > instead of copy old one to snapshot, overwrite old one with new one, 2
> > writes and 1 reads. it is possible that write new data to a usused
> > location directly.
> 
> Since you have to delete the snapshot (and I don't think they live very long
> on most systems) you don't gain anything (because you have copy the changes
> back from the snapshot), but you make the cases of more than one snapshot
> and (possibly) read-write snapshots a lot harder, maybe not in CPU or IO,
> but certainly in code complexity, which is a thing you never do to gain
> a bit of performance.
> 
on some systems that mainly for disaster recovery purpose, there will be
a lot of snapshots available for a lv. this can make the recovery much
easier.

> > i know later remove a snaphot will be a little trouble, but there must
> > be some way to get around it.
> 
> Yes, you have to do the work you saved a couple of minutes earlier :)
> 
a dedicated snapshot merge process can solve this.

> just my EUR 0.02,
thanks. discussion always can make thing clarified.


> Nils