Openstack - (Manually) migrating (KVM) Nova compute virtual machines

13-06-2015 | Remy van Elst


Table of Contents


This guide shows you how to migrate KVM virtual machines with the Openstack Nova compute service, either manually or with the Openstack tooling.

Migrating compute instances is very usefull. It allows an administrator to free up a compute node for maintenance/updates. It also allows an administrator to better distribute resources between compute nodes.

Openstack provides a few different ways to migrate virtual machines from one compute node to another. Each option has different requirements and restrictions, for example:

  • You can't live-migrate without shared storage.
  • You can't live-migrate if you have a configdrive enabled.
  • You can't select the target host if you use the nova migrate (non-live) command.

Later on in this guide, I'll list the most common limitations per situation.

This article describes the most common migration scenario's including live and manual migration using native linux tools.

You can see all my Openstack related articles here. For example, how to build a High Available cluster with Ansible and Openstack.

If you like this article, consider sponsoring me by trying out a Digital Ocean VPS. With this link you'll get a $5 VPS for 2 months free (as in, you get $10 credit). (referral link)

It is tested on an Openstack cloud running Icehouse with KVM as the only hypervisor. If you run Xen or something else, the manual process is mostly the same, however might require specific adaptions.

Configure (live) migration

There are a few things you need to configure (live) migrations on an Openstack cloud. The Openstack documentation describes this very well, read that first.

If you have a cloud configured for live migration, read on.

Migration limitations

Openstack has two commands specific to virtual machine migration:

  • nova migrate $UUID
  • nova live-migration $UUID $COMPUTE-HOST

The nova migrate command shuts down an instance to move it to another hypervisor.

  • The instance is down for a period of time and sees this as a regular shutdown.
  • It is not possible to specify the compute host you want to migrate the instance to. (Read on to see how you can do that the dirty way).
  • This command does not require shared storage, the migrations can take a long time.
  • The Openstack cluster chooses the target hypervisor machine based on the free resources and availability.
  • The migrate command works with any type of instance.
  • The VM clock has no issues.

The nova live-migration command has almost no instance downtime.

  • The instance is suspended and does not see this as a shutdown.
  • The live-migration lets you specify the compute host you want to migrate to, however with some limitations.
    This requires shared storage, instances without a configdrive when block storage is used, or volume-backed instances.
  • The migration fails if there are not enough resources on the target hypervisor
  • The VM clock might be off.

Here are some examples when to use which option:

  • If it is important to choose the compute host or to have very little downtime you need to use the nova live-migration command.
  • If you don't want to choose the compute host, or you have a configdrive enabled, you need to use the nova migrate command.
  • If you need to specify the compute host and you have a configdrive enabled, you need to manually migrate the machine, or use a dirty trick to fool nova migrate.

All these options are described below in detail.

Hypervisor Capacity

Before you do a migration, check if the hypervisor host has enough free capacity for the VM you want to migrate:

nova host-describe compute-30

Example output:

+-------------+----------------------------------+-----+-----------+---------+
| HOST        | PROJECT                          | cpu | memory_mb | disk_gb |
+-------------+----------------------------------+-----+-----------+---------+
| compute-30 | (total)                          | 64  | 512880    | 5928    |
| compute-30 | (used_now)                       | 44  | 211104    | 892     |
| compute-30 | (used_max)                       | 44  | 315568    | 1392    |
| compute-30 | 4[...]0288                       | 1   | 512       | 20      |
| compute-30 | 4[...]0194                       | 20  | 4506      | 62      |
+-------------+----------------------------------+-----+-----------+---------+

In this table, the first row shows the total amount of resources available on the physical server. The second line shows the currently used resources. The third line shows the maximum used resources. The fourth line and below shows the resources available for each project.

If the VM flavor fits on this hypervisor, continue on with the manual migration. If not, free up some resources or choose another compute server.

If the hypervisor node lacks enough capacity, the migration will fail.

(Live) migration with nova live-migration

The live-migration command works with the following types of vm's/storage:

  • Shared storage: Both hypervisors have access to shared storage.
  • Block storage: No shared storage is required. Instances are backed by image-based root disks. Incompatible with read-only devices such as CD-ROMs and Configuration Drive (config_drive).
  • Volume storage: No shared storage is required. Instances are backed by iSCSI volumes rather than ephemeral disk.

The live-migration command requires the same CPU on both hypervisors. It is possible to set a generic CPU for the VM's, or a generic set of CPU features. This however does not work on versions lower than Kilo due to a bug where Nova compares the actual CPU instead of the virtual CPU. In my case, all the hypervisor machines are the same, lucky me. This is fixed in Kilo or later.

On versions older than Kilo, the Compute service does not use libvirt's live migration functionality by default, therefore guests are suspended before migration and might experience several minutes of downtime. This is because there is a risk that the migration process will never end. This can happen if the guest operating system uses blocks on the disk faster than they can be migrated.

To enable true live migration using libvirt's migrate functionality, see the Openstack documentation linked below.

Shared storage / Volume backed instances

A live-migration is very simple. Use the following command with an instance UUID and the name of the compute host:

nova live-migration $UUID $COMPUTE-HOST

If you have shared storage, or if the instance is volume backed, this will send the instances memory (RAM) content over to the destination host. The source hypervisor keeps track of which memory pages are modified on the source while the transfer is in progress. Once the initial bulk transfer is complete, pages changed in the meantime are transferred again. This is done repeatedly with (ideally) ever smaller increments.

As long as the differences can be transferred faster than the source VM dirties memory pages, at some point the source VM gets suspended. Final differences are sent to the target host and an identical machine started there. At the same time the virtual network infrastructure takes care of all traffic being directed to the new virtual machine. Once the replacement machine is running, the suspended source instance is deleted. Usually the actual handover takes place so quickly and seamlessly that all but very time sensitive applications ever notice anything.

You can check this by starting a ping to the VM you are live-migrating. It will stay online and when the VM is suspended and resumed on the target hypervisor, the ping responses will take a bit longer.

Block based storage (--block-migrate)

If you don't have shared storage and the VM is not backed by a volume as root disk (image based VM's) a live-migration requires an extra parameter:

nova live-migration ----block-migrate $UUID $COMPUTE-HOST

The process is almost exactly the same as described above. There is one extra step however. Before the memory contents is sent the disk content is copied over, without downtime. When the VM is suspended, both the memory contents and the disk contents (difference to the earlier copy) are sent over. The suspend action takes longer and might be noticable as downtime.

The --block-migrate option is incompatible with read only devices such as ISO CD/DVD drives and the Config Drive.

Migration with nova migrate

The nova migrate command shuts down an instance, copies over the disk to a hypervisor with enough free resources, starts it up there and removes it from the source hypervisor. The VM is shut down and will be down as long as the copying. With a migrate, the Openstack cluster chooses an compute-service enabled hypervisor with the most resources available. This works with any type of instance, with any type of backend storage.

A migrate is even simpler than a live-migration. Here's the syntax:

nova migrate $UUID

This is perfect for instances that are part of a clustered service, or when you have scheduled and communicated downtime for that specific VM. The downtime is dependent on the size of the disk and the speed of the (storage) network.

rsync over ssh is used to copy the actual disk, you can test the speed yourself with a few regular rsync tests, and combine that with the disksize to get an indication of the migration downtime.

Migrating to a speficic compute node, the dirty way

As seen above, we cannot migrate virtual machines to a specific compute node if the compute node does not have shared storage and the virtual machine has a configdrive enabled. You can force the Openstack cluster to choose a specific hypervisor by disabling the nova-compute service on all the other hypervisors. The VM's will keep running on there, only new virtual machines and migrations are not possible on those hypervisors.

If you have a lot of creating and removing of machines in your Openstack Cloud, this might be a bad idea. If you use (Anti) Affinity Groups, vm's created in there will also fail to start depending on the type of Affinity Group. See my article on Affinity Groups for more info on those.

Therefore, use this option with caution. If we have 5 compute nodes, compute-30 to compute-34 and we want to migrate the machine to compute-34, we need to disable the nova-compute service on all other hypervisors.

First check the state of the cluster:

nova service-list --binary nova-compute # or nova-conductor, nova-cert, nova-consoleauth, nova-scheduler

Example output:

+----+--------------+--------------+------+----------+-------+----------------------------+-----------------+
| Id | Binary       | Host         | Zone | Status   | State | Updated_at                 | Disabled Reason |
+----+--------------+--------------+------+----------+-------+----------------------------+-----------------+
| 7  | nova-compute | compute-30   | OS1  | enabled  | up    | 2015-06-13T17:04:27.000000 | -               |
| 8  | nova-compute | compute-31   | OS2  | enables  | up    | 2015-06-13T17:02:49.000000 | -               |
| 9  | nova-compute | compute-32   | OS2  | enabled  | up    | 2015-06-13T17:02:50.000000 | None            |
| 10 | nova-compute | compute-33   | OS2  | enabled  | up    | 2015-06-13T17:02:50.000000 | -               |
| 11 | nova-compute | compute-34   | OS1  | disabled | up    | 2015-06-13T17:02:49.000000 | Migrations Only |
+----+--------------+--------------+------+----------+-------+----------------------------+-----------------+

In this example we have 5 compute nodes, of which one is disabled with reason Migrations Only. In our case, before we started migrating we have enabled nova-compute on that hypervisor and disabled it on all the other hypervisors:

nova service-disable compute-30 nova-compute --reason 'migration to specific hypervisor the dirty way'
nova service-disable compute-31 nova-compute --reason 'migration to specific hypervisor the dirty way'
etc...

Now execute the nova migrate command. Since you've disabled all compute hypervisors except the target hypervisor, that one will be used as migration target.

All new virtual machines created during the migration will also be spawned on that specific hypervisor.

When the migration is finished, enable all the other compute nodes:

nova service-enable compute-30 nova-compute
nova service-enable compute-31 nova-compute
etc...

In our case, we would disable the compute-34 because it is for migrations only.

This is a bit dirty and might cause problems if you have monitoring on the cluster state or spawn a lot of machines all the time.

Manual migration to a specific compute node

As seen above, we cannot migrate virtual machines to a specific compute node if the compute node does not have shared storage and the virtual machine has a configdrive enabled. Since Openstack is just a bunch of wrappers around native Linux tools, we can manually migrate the machine and update the Nova database afterwards.

Do note that this part is specific to the storage you use. In this example we use local storage (or, a local folder on an NFS mount not shared with other compute nodes) and image-backed instances.

In my case, I needed to migrate an image-backed block storage instance to a non-shared storage node, but the instance had a configdrive enabled. Disabling the compute service everywhere is not an option, since the cluster was getting about a hundred new VM's every 5 minutes and that would overload the hypervisor node.

This example manually migrates a VM from compute-30 to compute-34. These nodes are in the same network and can access one another via SSH keys based on their hostname.

Shut down the VM first:

nova stop $VM_UUID

Also detach any volumes:

nova volume-detach $VM_UUID $VOLUME_UUID

Use the nova show command to see the specific hypervisor the VM is running on:

nova show UUID | grep hypervisor

Example output:

| OS-EXT-SRV-ATTR:hypervisor_hostname  | compute-30    |

Login to that hypervisor via SSH. Navigate to the folder where this instance is located, in our case, /var/lib/nova-compute/instances/$UUID.

The instance is booted from an image based root disk, named disk. qemu in our case diffs the root disk from the image the VM was created from. Therefore the new hypervisor also needs that backing image. Find out which file is the backing image:

cd /var/lib/nova-compute/instances/UUID/
qemu-img info disk # disk is the filename of the instance root disk

Example output:

  image: disk
  file format: qcow2
  virtual size: 32G (34359738368 bytes)
  disk size: 1.3G
  cluster_size: 65536
  backing file: /var/lib/nova-compute/instances/_base/d00[...]61
  Format specific information:
      compat: 1.1
      lazy refcounts: false 

The file /var/lib/nova-compute/instances/_base/d004f7f8d3f79a053fad5f9e54a4aed9e2864561 is the backing disk. Note that the long filename is not a UUID but a checksum of the specific image version. In my case it is a raw disk:

qemu-img info /var/lib/nova-compute/instances/_base/d00[...]61

Example output:

image: /var/lib/nova-compute/instances/_base/d00[...]61
file format: raw
virtual size: 8.0G (8589934592 bytes)
disk size: 344M

Check the target hypervisor for the existence of that image. If it is not there, copy that file to the target hypervisor first:

rsync -r --progress /var/lib/nova-compute/instances/_base/d00[...]61 -e ssh compute-34:/var/lib/nova-compute/instances/_base/d00[...]61

On the target hypervisor, set the correct permissions:

chown nova:nova /var/lib/nova-compute/instances/_base/d00[...]61

Copy the instance folder to the new hypervisor:

cd /var/lib/nova-compute/instances/
rsync -r --progress $VM_UUID -e ssh compute-34:/var/lib/nova-compute/instances/

Set the correct permissions on the folder on the target hypervisor:

chown nova:nova /var/lib/nova-compute/instances/$VM_UUID
chown nova:nova /var/lib/nova-compute/instances/$VM_UUID/disk.info 
chown nova:nova /var/lib/nova-compute/instances/libvirt.xml

chown libvirt:kvm /var/lib/nova-compute/instances/$VM_UUID/console.log 
chown libvirt:kvm /var/lib/nova-compute/instances/$VM_UUID/disk 
chown libvirt:kvm /var/lib/nova-compute/instances/$VM_UUID/disk.config

If you use other usernames and groups, change those in the command.

Log in to your database server. In my case that is a MySQL Galera cluster. Start up a MySQL command prompt in the nova database:

mysql nova

Execute the following command to update the nova database with the new hypervisor for this VM:

update instances set node='compute-34', host=node where uuid='$VM_UUID';

This was tested on an IceHouse database scheme, other versions might require other queries.

Use the nova show command to see if the new hypervisor is set. If so, start the VM:

nova start $VM_UUID

Attach any volumes that were detached earlier:

nova volume-attach $VM_UUID $VOLUME_UUID

Use the console to check if it all works:

nova get-vnc-console $VM_UUID novnc

Do note that you must check the free capacity yourself. The VM will work if there is not enough capacity, but you do run in to weird issues with the hypervisor like bad performance or killed processes (OOM's).

Conclusion

Openstack offers many ways to migrate machines from one compute node to another. Each way is applicable in certain scenario's, and if all else fails you can manually migrate machines using the underlying linux tools. This article gives you an overview of the most common migration ways and the scenario's when they are applicable. Happy migrating.

Further reading


Tags: cloud, cluster, compute, kvm, migrate, nova, openstack, qemu,