Dell EMC ScaleIO as ephemeral storage backend

Add support for ScaleIO as an ephemeral storage backend.

Problem description

Nova does not support Dell EMC ScaleIO - another type of software-defined storage [1] - as an ephemeral storage backend. ScaleIO is already supported by Cinder (and os-brick) for volume hosting. It is reasonable (and asked by customers) to let Nova use ScaleIO for ephemeral disks.

The required set of supported features should be the same as the one implemented for Ceph, including live migration, except features related to optimized interaction with Glance, like clone or direct snapshot, because Glance does not support ScaleIO yet. The target virtual driver should be libvirt.

Main ScaleIO traits which have to be considered in a special way during development are:

  • ScaleIO devices must be explicitly connected to a compute host before usage, and disconnected after that.

  • ScaleIO can provide volumes which sizes are multiples of 8 Gb only.

Use Cases

  • A cloud operator selects ScaleIO as the ephemeral backend for certain compute nodes. He/she ensures that no instance is running on the nodes. He/she ensures that SDC (ScaleIO Data Client) is properly installed on the nodes, and makes appropriate changes to nova-compute configuration files to setup them to use ScaleIO cluster. After restart nova-compute services on the nodes new instances will use ScaleIO for all ephemeral disks.

There is no other use case specific for ScaleIO in addition to existing ones to manage instances by end user and/or cloud operator. Most valuable existing use cases are: create/delete, live migrate, resize/revert resize, create snapshot, host evacuation.

Proposed change

  • A new images type, a new image backend class, and a new image model class will be introduced and will be used throughout all nova.virt.libvirt package as it is done for rbd and lvm image types.

  • Required operations to connect/disconnect instance’s disks will be called in several appropriate places of nova.virt.libvirt.driver. These calls will be wrapped by if clause to perform them for the new backend explicitly, as it is done for rbd and other backends.

  • ScaleIO image backend will fail to create disks with sizes differ from multiples of 8 Gb. The only exception is rescue disk, which is created with no size specified. In this case the backend will approximate the rescue image size up to 8 Gb multiple.

  • Config drive will not be hosted on ScaleIO, but will be placed in instance local directory, as it happens by default. With iso9660 format this will not require much space (vfat format is supported by Nova for legacy only).


Instead of to implement the separate ephemeral storage backend for ScaleIO, it is possible to develop Cinder ephemeral storage backend, which obviously may serve any kind of physical storage supported by Cinder (including ScaleIO). Although the previous attempt [2] to get such backend was stopped, recently this idea was discussed in the community again. However there is even no new spec for this task at the moment. Also the implementation seems to be much more complex, probably needing additional refactoring, because the backend is intended particularly to replace Ceph backend, which implementation (explicit and implicit) is very distributed throughout whole libvirt driver. But Ceph backend is used widely, so it is important for the task to be completed to support in Cinder backend each feature of Ceph backend. Because that it looks that Cinder backend cannot be implemented pretty soon, but requires 2-3 release cycles. On the other hand we have ScaleIO backend which is already implemented for previous OpenStack releases (since Juno), and also available for various deployment tools [3]. Another question in this theme is if ScaleIO backend brings much complexity to a seamless transition system from volume based ephemeral backends to Cinder backend. Since there are more than one such backends (LVM, Ceph), and Cinder is able to acquire unmanaged volumes, the transition system probably will have much common code with small backend-related parts. So it will not be a big extra work to add support for ScaleIO there as well.

Instead of failing to run instances on flavors with wrong disk sizes it is possible:

  1. Approximate disk size to the closest multiple of 8 Gb. Write a recommendation for cloud operators into Nova documentation to use multiples of 8 Gb for disk sizes.

    Instances will get disks bigger than it is requested. This makes usage of statistics gathered by Nova, Ceilometer, etc. nonsense. Also snapshots size will be bigger than it is expected.

  2. The same as previous but also fix instance properties (like root_gb).

    Nova is not designed to have flavor and instance properties different at the moment.

  3. Create a partition with requested size on ScaleIO volume with approximated size and pass the partition to libvirt as the device.

    This solution can work with thin provisioning type only. Additional investigations are required to proof that this idea can work, including resize feature.

Data model impact


REST API impact


Security impact


Notifications impact


Other end user impact


Performance Impact


Other deployer impact

ScaleIO cluster must be installed and configured. A protection domain and a storage pool to store ephemeral disks must be created in the cluster. ScaleIO SDC component must be installed on affected compute nodes. These SDCs must be registered with the cluster.

Nova’s compute config file must contain these options to let ScaleIO image backend work:

images_type = sio

rest_server_ip = <ScaleIO Gateway IP>
rest_server_username = <ScaleIO Gateway user>
rest_server_password = <ScaleIO Gateway user password>
default_protection_domain_name = <ScaleIO protection domain>
default_storage_pool_name = <ScaleIO pool name>

Optional parameters:

default_provisioning_type = <thick (default) or thin>
verify_server_certificate = <True or False (default)>
server_certificate_path = <Path to the certificate>
default_sdcguid = <ScaleIO SDC guid of compute host>
rest_server_port = <ScaleIO Gateway port, 443 by default>

To get more flexibility, extra_spec of flavors can contain:




ScaleIO protection domain


ScaleIO pool name


thick or thin

If a key is not set in a flavor, default value is used from config file.

Each flavor intended to be used with ScaleIO image backend must have disk sizes in multiples of 8 Gb. A flavor with zero root disk size may be used to create volume backed instances only, because that it is not recommended to create such flavor. At the same time zero ephemeral/swap size is processed by usual way.

If an OpenStack deployment has compute nodes, which use any other image backend, and there is necessary to have ScaleIO-incompatible flavors there, it is possible to get this with scheduler filtering using flavor extra specs and host aggregates.

Developer impact




Primary assignee:

Feodor Tersin <>

Work Items

  • Implement all proposed changes at once.


A new library siolib [4] is used to communicate with ScaleIO via REST API.


  • Unit tests.

  • Will be tested with a new thirdparty ScaleIO CI, using the same or similar infrastructure as ScaleIO Cinder CI does.

No new Tempest test will be introduced.

Documentation Impact

Documentation about usage of ScaleIO for ephemeral storage backend will be added.

There is a list of potentially touched doc pages with required changes.

Document page


Operations Guide - Architecture Compute Nodes - Instance Storage Solutions

Mention Shared Block Storage as a ComputeNode Storage, and mention ScaleIO there

High Availability Guide - Configuring Storage for high availability - Storage back end

Mention ScaleIO there as an option for ephemeral storage

Administrator Guide - Compute - System administration - Advanced configuration

Add a section which explains how to configure Compute to use ScaleIO, what to tune in flavors, how to isolate ScaleIO hosts from other flavors

Configuration Reference - Compute service - Overview of nova.conf

Mention scaleio section

Configuration Reference - Compute service - The full set of available options

Describe full scaleio section

Configuration Reference - Compute service - Hypervisors - KVM - Configure Compute backing storage

Add sio to the list of available images_type option values




Release Name