[openstack-dev] [cinder] LVM snapshot performance issue -- why isn't thin provisioning the default?

yang, xing xing.yang at emc.com
Tue Sep 15 20:56:19 UTC 2015


Hi Eric,

Regarding the default max_over_subscription_ratio, I initially set the
default to 1 while working on oversubscription, and changed it to 2 after
getting review comments.  After it was merged, I got feedback that 2 is
too small and 20 is more appropriated, so I changed it to 20.  So it looks
like we can¹t find a default value that makes everyone happy.

If we can decide what is the best default value for LVM, we can change the
default max_over_subscription_ratio, but we should also allow other
drivers to specify a different config option if a different default value
is more appropriate for them.

Thanks,
Xing


On 9/15/15, 1:38 PM, "Eric Harney" <eharney at redhat.com> wrote:

>On 09/15/2015 01:00 PM, Chris Friesen wrote:
>> I'm currently trying to work around an issue where activating LVM
>> snapshots created through cinder takes potentially a long time.
>> (Linearly related to the amount of data that differs between the
>> original volume and the snapshot.)  On one system I tested it took about
>> one minute per 25GB of data, so the worst-case boot delay can become
>> significant.
>> 
>> According to Zdenek Kabelac on the LVM mailing list, LVM snapshots were
>> not intended to be kept around indefinitely, they were supposed to be
>> used only until the backup was taken and then deleted.  He recommends
>> using thin provisioning for long-lived snapshots due to differences in
>> how the metadata is maintained.  (He also says he's heard reports of
>> volume activation taking half an hour, which is clearly crazy when
>> instances are waiting to access their volumes.)
>> 
>> Given the above, is there any reason why we couldn't make thin
>> provisioning the default?
>> 
>
>
>My intention is to move toward thin-provisioned LVM as the default -- it
>is definitely better suited to our use of LVM.  Previously this was less
>easy, since some older Ubuntu platforms didn't support it, but in
>Liberty we added the ability to specify lvm_type = "auto" [1] to use
>thin if it is supported on the platform.
>
>The other issue preventing using thin by default is that we default the
>max oversubscription ratio to 20.  IMO that isn't a safe thing to do for
>the reference implementation, since it means that people who deploy
>Cinder LVM on smaller storage configurations can easily fill up their
>volume group and have things grind to halt.  I think we want something
>closer to the semantics of thick LVM for the default case.
>
>We haven't thought through a reasonable migration strategy for how to
>handle that.  I'm not sure we can change the default oversubscription
>ratio without breaking deployments using other drivers.  (Maybe I'm
>wrong about this?)
>
>If we sort out that issue, I don't see any reason we can't switch over
>in Mitaka.
>
>[1] https://review.openstack.org/#/c/104653/
>
>__________________________________________________________________________
>OpenStack Development Mailing List (not for usage questions)
>Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
>http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev




More information about the OpenStack-dev mailing list