Storage Networking: SANs Expert(s):
10 July 2002
Snapshot copy can be done in a number of ways. There are software
solutions that work at the host system level to provide "snapshot"(copy on
write) functionality such as Veritas' or Compaq's Virtual Replicator, or Sun's
Instant Image. Then there are firmware-based solutions done at the storage
subsystem level. Your choice of which to use would depend on the number of host
systems needing the functionality. Host-based solutions are licensed per server,
where firmware-based solutions are priced per storage subsystem.
If you only have a few servers that need snapshot functionality, then using
host-based software may be the cheaper way to go.
There are two types of "snapshots" available. Complete copy, a clone or exact
duplicate of the original volume, and metadata-based copy that uses
"copy-on-write" functionality. A metadata copy is very fast as it just copies
the "pointers" to where your data is stored whereas the clone or complete copy
copies the entire LUN to a separate physical disk. Metadata copies track which
blocks on the disks are changed as you write to the original volume and does a
"copy-on-write" to a designated pool of storage set aside for the snapshot.
Copy-on-write moves the original data block to the snapshot before a write is
allowed to that block. This keeps the snapshot data consistent in time with the
exact time the snapshot was taken. Snapshot volume read requests to unchanged
data blocks are redirected back to the original volume while read requests to
data blocks that have been changed are directed to the "copied" blocks in the
Metadata snapshots have an impact on performance on the original volume while
they exist since write requests to the original volume must be "copied out" to
the snapshot before the write is allowed.
Clone copies are more highly available since they are exact duplicates of the
original volume that now reside on a separate physical disk. A clone can be made
available by "splitting" a mirrorset volume into two JBODs or copied internally
inside the storage array from one RAIDset to another. Check with your storage
vendor for how this is performed.
If the mirrorset is already in place before the clone is taken, then the
clone-based image will happen very quickly. If the original volume was a JBOD
prior to issuing the clone command, then it will take time for the volumes to
get in sync before the clone can be split off.
As for your database, the most common way of creating database clones or
snapshots is to "quiese" the database prior to creating the snapshot. By "quiese"
I mean stopping all writes to disk and flushing your file system cache. Doing
this, will give you a transactionally consistent image that can be remounted
without file system checks or database consistency checks.
Some vendors (like HDS) have firmware-based software that will flush host memory
down to disk and freeze, then thaw I/O to disk as the snapshot is taken. This
means you can keep the database (or mail system) up and running while the
snapshot is taken. Check with your storage vendor to see if this capability is
currently available on their arrays.
Q: You explained why metadata snapshots have an impact on performance on the
original volume, since write requests to the original volume must be "copied
out" to the snapshot before the write is allowed. Isn't there any storage vendor
that offers a snapshot feature based on a different principle such as one that
acknowledges the write once it has reached the cache and only later proceeds to
perform the additional operations while relying on dual-write or other,
This question posed on 22 November 2002
The answer to your question is yes, but I have not found this to be the case in
software-based solutions. The HDS 9000 storage arrays use a cascaded snapshot
approach where writes to the original volume is given an I/O complete back to
the application immediately. The cascaded writes to the snapshot volume happen
in ASYNC mode using dual writes from the cache in background mode. This
alleviates many of the performance issues associated with writes to the primary
volume when linked to an image copy. This is similar to metadata snapshooting,
although there is an actual complete physical copy of the data. It's just as
fast as creating a metadata copy of the disk from a storage pool but you also
get the added benefit of RAID protected snapshots.
In the original question, I was speaking of metadata copy using software, not
All Rights Reserved, Copyright 2000 - 2003, TechTarget