Time needed for take a snapshot
Hi list,
just for confirmation: is the time needed to take a snapshot constant, or does it depend on how many chunks are mapped in/by the specific thin pool and volume? In other workds, is snapshot create time O(1) or O(n)?
Thanks.
Signature · 5 lines
Danti Gionatan Supporto Tecnico Assyoma S.r.l. - www.assyoma.it email: [email protected] - [email protected] GPG public key ID: FF5F32A8
Dne 19. 01. 20 v 13:59 Gionatan Danti napsal(a):
Quoted 2 lines
Hi list,
just for confirmation: is the time needed to take a snapshot constant, or does it depend on how many chunks are mapped in/by the specific thin pool and volume? In other workds, is snapshot create time O(1) or O(n)?
Hi
Depends on what you are asking - in lvm2 - unfortunately some parts of the code are sometimes even O(n^2) - so the more LVs are in VG - the 'heavier' the operation gets.
So having thousands of LVs in a single VG will become probably your bottleneck.
The bTree mapping in kernel will probably be less problematic here.
Regards
Zdenek
Il 20-01-2020 10:22 Zdenek Kabelac ha scritto:
Quoted 1 line
So having thousands of LVs in a single VG will become probably your bottleneck.
Hi Zdenek, I was thinking more about having few LVs, but with different amount of data/mapping.
For example, is a very fragmented volume (ie: one written randomically) significantly slower to snapshot than an almost empty volume? I fully expect some small difference; however, if an empty volume take 0.2s and a fragmented one 20s, this would surely be significant.
Note that I never had such a slow snapshot, rather, even on aged and big volumes, it always take <1s. However, other experiences are welcome.
Thanks.
Signature · 5 lines
Danti Gionatan Supporto Tecnico Assyoma S.r.l. - www.assyoma.it email: [email protected] - [email protected] GPG public key ID: FF5F32A8
Dne 20. 01. 20 v 11:36 Gionatan Danti napsal(a):
Quoted 8 lines
Il 20-01-2020 10:22 Zdenek Kabelac ha scritto:
Quoted 1 line
So having� thousands of LVs in a single VG will become probably your bottleneck.
Hi Zdenek, I was thinking more about having few LVs, but with different amount of data/mapping.
For example, is a very fragmented volume (ie: one written randomically) significantly slower to snapshot than an almost empty volume? I fully expect some small difference; however, if an empty volume take 0.2s and a fragmented one 20s, this would surely be significant.
Note that I never had such a slow snapshot, rather, even on aged and big volumes, it always take <1s. However, other experiences are welcome.
Yep - kernel metadata 'per thin LV' are reasonably small - so even for big thin devices it still should fit within your time boundaries.
(effectively thin snapshot just increases 'mapping' sharing between origin and its snapshot - so the time needed depends on how many bTree nodes needs to be updated - so if you would manage to create heavily fragmented multi TiB thinLV, the time depends on speed of you metadata device - as long as device is fast (i.e. >= SSD) - operation should be quick. But ATM there is no scientific proof for the worst case scenario.
Regards
Zdenek
Il 20-01-2020 15:40 Zdenek Kabelac ha scritto:
Quoted 9 lines
Yep - kernel metadata 'per thin LV' are reasonably small - so even for big thin devices it still should fit within your time boundaries. (effectively thin snapshot just increases 'mapping' sharing between origin and its snapshot - so the time needed depends on how many bTree nodes needs to be updated - so if you would manage to create heavily fragmented multi TiB thinLV, the time depends on speed of you metadata device - as long as device is fast (i.e. >= SSD) - operation should be quick.
But ATM there is no scientific proof for the worst case scenario.
Well, I'll do some tests and update here if finding anything interesting. Thanks.
Signature · 5 lines
Danti Gionatan Supporto Tecnico Assyoma S.r.l. - www.assyoma.it email: [email protected] - [email protected] GPG public key ID: FF5F32A8