Optimal snapshot size

[Brandon Poyner]

We use the VxFS snapshot feature during our nightly dataprotector backup. Last night the snapshot file system ran out of space, causing the snapshot to become disabled. How would I go about determining either the optimal snapshot size or the true contents of the snapshot (those files that differ from the source)? Thanks.

msgcnt 1 vxfs: mesg 028: vx_snap_alloc - /dev/eva2/dnx_snap snapshot file system out of space
msgcnt 2 vxfs: mesg 032: vx_disable - /dev/eva2/dnx_snap snapshot file system disabled

[Alzhy]

Realistically, you should prepare for any eventuality that you snapshot may become as large as your source filesystem. More so to mitigate the risks during backups - make sure during your snapshot "window" that no significant filesystem activity (changes) occur. If they do or has the potential -- then prepapre your snapshots such that they have the potential to grow as big as the original...

[Ivan Ferreira]

If you do incremental backups, a little more of the size of the backed up data should be the size of your snapshot. If you don't do incremental backups, you can do test incremental backups to get statistics about the amount of modified data.

[A. Clay Stephenson · Accepted solution]

The tools supplied directly from Veritas have the ability to actually examine the snapshot buffer usage; the tools supplied in the OEM'ed version of vxfs do not have that ability. The size of the snapshot buffer depends upon two factors: 1) The activity of the original filesystem during the snapshot; 2) the duration of the snapshot. In principle, you could do a block by block read of each file in the two filesystems and gather your data but that seems impractical.

In practice, I have never needed a snapshot buffer larger than about 25% of the original but my backups are finished in about 4 hours and 15% is more typical. Note that only the first update/write of an original block need be written to the snapshot buffer so that the maximum buffer size would be 100% of the original.

One other "gotcha" that is far from obvious is that the logical device that houses the snapshot buffer should be mirrored or otherwise highly available because if the snapshot buffer becomes unavailable, the original filesystem will hang.