[Openstack-operators] Avoiding Cinder stampedes with RBD? (Kilo)

Ned Rhudy (BLOOMBERG/ 731 LEX) erhudy at bloomberg.net
Tue Mar 22 13:25:39 UTC 2016


We have a situation where tenant A is trying to launch large numbers of instances from a single RBD volume snapshot in Cinder (e.g., 40 instances at once). We made an unrelated change recently to enable rbd_flatten_volume_from_snapshot by default in order to save tenants who create large chains of volumes and snapshots until they run out of quota and then can't figure out how to unwind the chain, because the relationships are not trivially traceable. This change appears to be causing tenant A significant heartache now, because when he launches some large number of instances at once, most of his instance launches time out, presumably because Cinder is stampeding onto Ceph and trying to create flattened RBD images for every single instance simultaneously.

My question is, is there anything I can do to stop this stampeding? A review of cinder.conf options for Kilo didn't point out any obvious setting that could be adjusted to force Cinder to serialize its operations here, but maybe I missed something.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-operators/attachments/20160322/537ca3ba/attachment.html>


More information about the OpenStack-operators mailing list