Hi Star, I'll +1 this. I would like to see support for RBD snapshots as well, and maybe have a method to "backup" the snapshots to secondary storage. Right now for large volumes it can take an hour or more to finish the snapshot.
I have already discussed this with Wido, and was able to determine that even without using native RBD snapshots we could improve the copy time by saving the snaps as thin volumes instead of full raw files. Right now the snapshot code when using RBD specifically converts the volumes to a full raw file, when saving as a qcow2 image would use less space. When restoring a snapshot the code current specifically indicates the source image as being a raw file, but if we change the code to not indicate the source image type qemu-img should automatically detect it. We just need to see if that's the case with all of the supported versions of libvirt/qemu before submitting a pull request. Thank You, Logan Barfield Tranquil Hosting On Wed, May 27, 2015 at 9:18 PM, Star Guo <st...@ceph.me> wrote: > Hi everyone, > > Since I have test cloudstack 4.4.2 + kvm + rbd, deploy an instance is so > fast apart from the first deployment because copy template from secondary > storage (NFS) to primary storage (RBD). That is no problem. > However, when I do some volume operation, such as create snapshot, create > template, template deploy ect, it also take some time to finish because copy > data between primary storage and secondary storage. > So I think that if we support the same rbd as secondary storage, and use > ceph COW feature, it may reduce the time and just some seconds. (OpenStack > can make glance and cinder as the same rbd) > > Best Regards, > Star Guo >
