Newton (OSP10) has seen a variety of OpenvSwitch version supported. Initial release was with OvS2.5 and later it has been updated to OvS2.6, which is present for a long time of the support cycle. Recently, in the time-line with with queens (OSP13) release, OvS2.9 version is planned to be supported. In order to facilitate FFU (Fast Forward Upgrades) from newton to queens, bringing the support of OvS2.9 to newton (OSP10) will help in reduce the cluster downtime for the upgrade. This post details on the steps and major changes along with known issues.
Mapping OpenvSwitch version with TripleO
As you might be aware that there is no internal mapping for the incremental releases of newton (OSP10), it would be hard to map OvS version with TripleO. This table provides the OvS version agains the tripleo-heat-templates RPM version, which are released together.
|Date||OpenvSwitch Version||tripleo-heat-templates version|
Highlights of release
Newton release with OvS2.6 has the default vhost user mode as
dpdkvhostuser, in which OpenvSwitch acts as server and Qemu acts as client. But with this release along with OvS2.9, the default mode has been changed in neutron to
dpdkvhostuserclient, in which OpenvSwitch acts as client and Qemu acts as server. It provides the advantage of network reconnecting when OpenvSwitch is restarted (because of crash) without restart the guest VM.
Newton release with OvS2.6 is deployed by modifying the OpenvSwitch service file to make it to run with group as
qemu. With OvS2.9, this workaround of modifying OpenvSwitch service file is not required, as both OpenvSwitch and qemu will be configured to run with same group as
hugetlbfs. And the vhost sockets created by qemu on the directory
/var/lib/vhost_socketswill be accessible to OpenvSwitch running with this group.
Minor Update steps in Newton
Assuming that newton (OSP10) with OvS2.6 is deployed successfully as per the documentation and the undercloud is updated to the latest newton release, then here is the procedure to update the overcloud nodes with OVS-DPDK deployment:
Configure the vhostuser socket directory for the
dpdkvhostuserclientmode. The existing directory
/var/run/openvswitchdoes not have group write access needed for qemu to create the sockets. So, instead of exposing other files to the common group, the creation of vhostuser sockets has been moved to a new directory. Create an environment file
ovs-dpdk-update.yamlwith below contents:
parameter_defaults: NeutronVhostuserSocketDir: '/var/lib/vhost_sockets'
- Add a new environment file
environments/ovs-dpdk-permissions.yamlfor the deploy command to update the plan before starting the minor update. This environment file provides the new group
hugetlbfsvalue to the parameter
VhostuserSocketGroup, which will apply the group to the
/etc/libvirt/qemu.confto create qemu VMs with group as
hugetlbfs. And additionally create the new vhostuser socket directory with the specified permission.
Note: On a cluster with OVS-DPDK deployment, all the compute nodes will be configured with
hugetlbfsfor qemu to avoid confusion and allow easy adaptability of services in the existing nodes.
- Execute the plan update command with the above mentioned changes. Deploy
command with option
--update-plan-onlywill upload the latest templates after undercloud update to the swift container for the minor update to use. And also uploads the user provided additional environment files so that it can be added to the deployment. Once the plan is uploaded and rendered with Jinja2, the deploy command stops to allow the minor update command to continue.
openstack overcloud deploy --templates \ << include all the existing files >> \ -e /usr/share/openstack-tripleo-heat-templates/environments/ovs-dpdk-permissions.yaml \ -e ovs-dpdk-update.yaml \ --update-plan-only
- Add a new environment file
Configure the overcloud nodes have the correct newton (OSP10) repos configure to start the package update process.
Execute minor update which starts with the package update on all the overcloud nodes, followed by running the puppet step configuration.
yes "" | openstack overcloud update stack -i overcloud
Existing VMs and migration
Once the update is completed, it is required to restart the overcloud nodes to use the updated kernel and openvswitch. Restarting the overcloud nodes and starting the VM again will start the VM in the
dpdkvhostusermode only. In order to migrate to the
dpdkvhostuserclientmode, the VMs has to be either cold-migrated or re-created with a snapshot image of the existing VM.
Note: Live migration will fail with this forced mode changed. During the live migration with this change, nova will not re-create the virsh xml, but will try to use the same as existing mode (Qemu in client mode). But openvswitch and neutron will migrate the existing port to the new mode (OvS from server mode to client mode). This results in failure as both OvS and Qemu are in client mode.
The existing VM can be migrated in place before a reboot or cold migrated to another host (which has been restarted already), but both the migrations involve a downtime.
1. Migrate in-place with snapshot
The existing VMs with
dpdkvhostusermode can be migrated to client mode in- place without rebooting any of the nodes. As OvS2.6 has the support for
dpdkvhostuserclientmode, it is possible to re-create the instance in-place with OvS2.6 itself after minor update but before reboot. This method of migration works when the node is rebooted to OvS2.9 also. If in-place migration is mandatory, ensure the nova services are disabled on other nodes or other means of controlling the scheduler, so that it will be created on the same node.
# Stop the instance and create a snapshot image openstack server stop <instance> openstack server image create --name dpdk_vm1_migration_snapshot <instance> # Ensure the snapshot image is listed openstack image list # Delete the old instance so that the resources available to re-create the VM openstack server delete <instance> openstack server create --flavor m1.nano --nic net-id=<dpdk_net_id> \ --image dpdk_vm1_migration_snapshot --key-name <key_name> <new_instance>
2. Cold Migration
Cold migration of the VM also works to migrate the instance, but it has to be migrated to another node. So it is better to reboot the target node, so that kernel and OvS update takes effect before cold migrate.
openstack server migrate <instance> --block-migration --wait