You should also make sure that your base images are in raw format. The rbd import from your output indicates that your base images are downloaded from glance to local storage on the compute node (to
Code:
/var/lib/nova/instances/_base/
), which is then converted to raw and then nova tries to create a new rbd object within the pool. Also check your disk space on compute nodes because this would fill up your disk if all the base images reside on your compute nodes. I would also recommend to disable image caching if ceph is your backend, check out this post about disk_format and image caching, maybe that helps.
Are the instances created with horizon volume based instances or do they have ephemeral disks?