Storage: Copy-on-write cloning for RBD-backed disks

Currently RBD-backed ephemeral disks are created by downloading an image from Glance to a local file, then uploading that file into RBD. Even if the file is cached, uploading may take a long time, since ‘rbd import’ is synchronous and slow. If the image is already stored in RBD by Glance, there’s no need for any local copies - it can be cloned to a new image for a new disk without copying the data at all.

Problem description

The primary use case that benefits from this change is launching an instance from a Glance image where Ceph RBD backend is enabled for both Glance and Nova, and Glance images are stored in RBD in RAW format.

Following problems are addressed:

  • Disk space on compute nodes is wasted by caching an additional copy of the image on each compute node that runs instances from that image.

  • Disk space in Ceph is wasted by uploading a full copy of an image instead of creating a copy-on-write clone.

  • Network capacity is wasted by downloading the image from RBD to a compute node the first time that node launches an instance from that image, and by uploading the image to RBD every time a new instance is launched from the same image.

  • Increased time required to launch an instance reduces elasticity of the cloud environment and increases the number of in-flight operations that have to be maintained by Nova.

Proposed change

Extract RBD specific utility code into a new file, align its structure and provided functionality in line with similar code in Cinder. This includes the volume cleanup code that should be converted from rbd CLI to using the RBD library.

Add utility functions to support cloning, including checks whether image exists and whether it can be cloned.

Add direct_fetch() method to nova.virt.libvirt.imagebackend, make its implementation in the Rbd subclass try to clone the image when possible. Following criteria are used to determine that the image can be cloned:

  • Image location uses the rbd:// schema and contains a valid reference to an RBD snapshot;

  • Image location references the same Ceph cluster as Nova configuration;

  • Image disk format is ‘raw’;

  • RBD snapshot referenced by image location is accessible by Nova.

Extend fetch_to_raw() in nova.virt.images to try direct_fetch() when a new optional backend parameter is passed. Make the libvirt driver pass the backend parameter.

Instead of calling disk.get_disk_size() directly from verify_base_size(), which assumes the disk is stored locally, add a new method that is overridden by the Rbd subclass to get the disk size.


An alternative implementation based on the image-multiple-location blueprint ( was tried in Icehouse. It was ultimately reverted, which can be attributed to a sum of multiple reasons:

This design has a significantly smaller footprint, and is mostly isolated to the RBD image backend in the libvirt driver.

Data model impact


REST API impact


Security impact


Notifications impact


Other end user impact

When Ceph RBD backend is enabled for Glance and Nova, there will be a noticeable difference in time and resource consumption when launching instances from Glance images in RAW and non-RAW formats.

Performance Impact

In the primary use case defined in the Problem description section above, there will be a significant performance improvement.

In other use cases, libvirt driver will introduce one more API call to Glance to retrieve a list of image locations when RBD backend is enabled. The performance impact of that call is insignificant compared to the time and resources it takes to fetch a full image from Glance.

Other deployer impact


Developer impact




Primary assignee:


Other contributors:


Work Items

Current implementation (see References) consists of following changes:

  • Move libvirt RBD utilities to a new file

  • Use library instead of CLI to cleanup RBD volumes

  • Enable cloning for rbd-backed ephemeral disks




This is a non-functional change with no impact on the test cases that need to be covered.

There is work currently going on to get all of tempest running against an environment using Ceph in the OpenStack CI environment. The first step is ceph support for devstack, which you can see here:

There’s also a test devstack patch with forces ceph to be enabled, which results in all of the devstack jobs being run with ceph enabled. You can find that here:

There are some tests failing (14 and 15 the first couple of runs). However, that also means that the vast majority of tests that cover this code (anything that spawns an instance) are passing. So, we at least have a way to run these tests on demand against master. Once the devstack patch merges, we will enable a job that can run against patches in all projects (perhaps experimental to start with).

Fuel CI also includes a suite of tests for OpenStack deployments with Ceph:

Documentation Impact



Mailing list discussions:

Current implementation:,n,z