Quality of Service (QoS): Guaranteed Minimum Bandwidth

Most Networking Quality of Service (QoS) features are implemented solely by OpenStack Neutron and they are already documented in the QoS configuration chapter of the Networking Guide. Some more complex QoS features necessarily involve the scheduling of a cloud server, therefore their implementation is shared between OpenStack Nova, Neutron and Placement. As of the OpenStack Stein release the Guaranteed Minimum Bandwidth feature is like the latter.

This Networking Guide chapter does not aim to replace Nova or Placement documentation in any way, but it still hopes to give an overall OpenStack-level guide to understanding and configuring a deployment to use the Guaranteed Minimum Bandwidth feature.

A guarantee of minimum available bandwidth can be enforced on two levels:

  • Scheduling a server on a compute host where the bandwidth is available. To be more precise: scheduling one or more ports of a server on a compute host’s physical network interfaces where the bandwidth is available.

  • Queueing network packets on a physical network interface to provide the guaranteed bandwidth.

In short the enforcement has two levels:

  • (server) placement and

  • data plane.

Since the data plane enforcement is already documented in the QoS chapter, here we only document the placement-level enforcement.

Limitations

  • A pre-created port with a minimum-bandwidth rule must be passed when booting a server (openstack server create). Passing a network with a minimum-bandwidth rule at boot is not supported because of technical reasons (in this case the port is created too late for Neutron to affect scheduling).

  • Bandwidth guarantees for ports can only be requested on networks backed by a physical network (physnet).

  • In Stein there is no support for networks with multiple physnets. However some simpler multi-segment networks are still supported:

    • Networks with multiple segments all having the same physnet name.

    • Networks with only one physnet segment (the other segments being tunneled segments).

  • If you mix ports with and without bandwidth guarantees on the same physical interface then the ports without a guarantee may starve. Therefore mixing them is not recommended. Instead it is recommended to separate them by Nova host aggregates.

  • Changing the guarantee of a QoS policy (adding/deleting a minimum_bandwidth rule, or changing the min_kbps field of a minimum_bandwidth rule) is only possible while the policy is not in effect. That is ports of the QoS policy are not yet used by Nova. Requests to change guarantees of in-use policies are rejected.

  • Changing the QoS policy of the port with new minimum_bandwidth rules changes placement allocations from Victoria release. If the VM was booted with port without QoS policy and minimum_bandwidth rules the port update succeeds but placement allocations will not change. The same is true if the port has no binding:profile, thus no placement allocation record exists for it. But if the VM was booted with a port with QoS policy and minimum_bandwidth rules the update is possible and the allocations are changed in placement as well.

Note

As it is possible to update a port to remove the QoS policy, updating it back to have QoS policy with minimum_bandwidth rule will not result in placement allocation record, only the dataplane enforcement will happen.

Note

updating the minimum_bandwidth rule of a QoS policy that is attached to a port which is bound to a VM is still not possible.

  • The first data-plane-only Guaranteed Minimum Bandwidth implementation (for SR-IOV egress traffic) was released in the Newton release of Neutron. Because of the known lack of placement-level enforcement it was marked as “best effort” (5th bullet point). Since placement-level enforcement was not implemented bandwidth may have become overallocated and the system level resource inventory may have become inconsistent. Therefore for users of the data-plane-only implementation a migration/healing process is mandatory (see section On Healing of Allocations) to bring the system level resource inventory to a consistent state. Further operations that would reintroduce inconsistency (e.g. migrating a server with minimum_bandwidth QoS rule, but no resource allocation in Placement) are rejected now in a backward-incompatible way.

  • The Guaranteed Minimum Bandwidth feature is not complete in the Stein release. Not all Nova server lifecycle operations can be executed on a server with bandwidth guarantees. Since Stein (Nova API microversion 2.72+) you can boot and delete a server with a guarantee and detach a port with a guarantee. Since Train you can also migrate and resize a server with a guarantee. Support for further server move operations (for example evacuate, live-migrate and unshelve after shelve-offload) is to be implemented later. For the definitive documentation please refer to the Port with Resource Request chapter of the OpenStack Compute API Guide.

  • If an SR-IOV physical function is configured for use by the neutron-openvswitch-agent, and the same physical function’s virtual functions are configured for use by the neutron-sriov-agent then the available bandwidth must be statically split between the corresponding resource providers by administrative choice. For example a 10 Gbps SR-IOV capable physical NIC could be treated as two independent NICs - a 5 Gbps NIC (technically the physical function of the NIC) added to an Open vSwitch bridge, and another 5 Gbps NIC whose virtual functions can be handed out to servers by neutron-sriov-agent.

Placement pre-requisites

Placement must support microversion 1.29. This was first released in Rocky.

Nova pre-requisites

Nova must support microversion 2.72. This was first released in Stein.

Not all Nova virt drivers are supported, please refer to the Virt Driver Support section of the Nova Admin Guide.

Neutron pre-requisites

Neutron must support the following API extensions:

  • agent-resources-synced

  • port-resource-request

  • qos-bw-minimum-ingress

These were all first released in Stein.

Supported drivers and agents

In release Stein the following agent-based ML2 mechanism drivers are supported:

  • Open vSwitch (openvswitch) vnic_types: normal, direct

  • SR-IOV (sriovnicswitch) vnic_types: direct, macvtap

neutron-server config

The placement service plugin synchronizes the agents’ resource provider information from neutron-server to Placement.

Since neutron-server talks to Placement you need to configure how neutron-server should find Placement and authenticate to it.

/etc/neutron/neutron.conf (on controller nodes):

[DEFAULT]
service_plugins = placement,...
auth_strategy = keystone

[placement]
auth_type = password
auth_url = https://controller/identity
password = secret
project_domain_name = Default
project_name = service
user_domain_name = Default
username = placement

If a vnic_type is supported by default by multiple ML2 mechanism drivers (e.g. vnic_type=direct by both openvswitch and sriovnicswitch) and multiple agents’ resources are also meant to be tracked by Placement, then the admin must decide which driver to take ports of that vnic_type by prohibiting the vnic_type for the unwanted drivers. Use ovs_driver.vnic_type_prohibit_list in this case. Valid values are all the supported_vnic_types of the respective mechanism drivers.

/etc/neutron/plugins/ml2/ml2_conf.ini (on controller nodes):

[ovs_driver]
vnic_type_prohibit_list = direct

[sriov_driver]
#vnic_type_prohibit_list = direct

neutron-openvswitch-agent config

Set the agent configuration as the authentic source of the resources available. Set it on a per-bridge basis by ovs.resource_provider_bandwidths. The format is: bridge:egress:ingress,... You may set only one direction and omit the other.

Note

egress / ingress is meant from the perspective of a cloud server. That is egress = cloud server upload, ingress = download.

Egress and ingress available bandwidth values are in kilobit/sec (kbps).

If desired, resource provider inventory fields can be tweaked on a per-agent basis by setting ovs.resource_provider_inventory_defaults. Valid values are all the optional parameters of the update resource provider inventory call.

/etc/neutron/plugins/ml2/ovs_agent.ini (on compute and network nodes):

[ovs]
bridge_mappings = physnet0:br-physnet0,...
resource_provider_bandwidths = br-physnet0:10000000:10000000,...
#resource_provider_inventory_defaults = step_size:1000,...

neutron-sriov-agent config

The configuration of neutron-sriov-agent is analog to that of neutron-openvswitch-agent. However look out for:

  • The different .ini section names as you can see below.

  • That neutron-sriov-agent allows a physnet to be backed by multiple physical devices.

  • Of course refer to SR-IOV physical functions instead of bridges in sriov_nic.resource_provider_bandwidths.

/etc/neutron/plugins/ml2/sriov_agent.ini (on compute nodes):

[sriov_nic]
physical_device_mappings = physnet0:ens5,physnet0:ens6,...
resource_provider_bandwidths = ens5:40000000:40000000,ens6:40000000:40000000,...
#resource_provider_inventory_defaults = step_size:1000,...

Propagation of resource information

The flow of information is different for available and used resources.

The authentic source of available resources is neutron agent configuration - where the resources actually exist, as described in the agent configuration sections above. This information is propagated in the following chain: neutron-l2-agent -> neutron-server -> Placement.

From neutron agent to server the information is included in the configurations field of the agent heartbeat message sent on the message queue periodically.

# as admin
$ openstack network agent list --agent-type open-vswitch --host devstack0
+--------------------------------------+--------------------+-----------+-------------------+-------+-------+---------------------------+
| ID                                   | Agent Type         | Host      | Availability Zone | Alive | State | Binary                    |
+--------------------------------------+--------------------+-----------+-------------------+-------+-------+---------------------------+
| 5e57b85f-b017-419a-8745-9c406e149f9e | Open vSwitch agent | devstack0 | None              | :-)   | UP    | neutron-openvswitch-agent |
+--------------------------------------+--------------------+-----------+-------------------+-------+-------+---------------------------+

# output shortened and pretty printed
# note: 'configurations' on the wire, but 'configuration' in the cli
$ openstack network agent show -f value -c configuration 5e57b85f-b017-419a-8745-9c406e149f9e
{'bridge_mappings': {'physnet0': 'br-physnet0'},
 'resource_provider_bandwidths': {'br-physnet0': {'egress': 10000000,
                                                  'ingress': 10000000}},
 'resource_provider_inventory_defaults': {'allocation_ratio': 1.0,
                                          'min_unit': 1,
                                          'reserved': 0,
                                          'step_size': 1},
 ...
}

Re-reading the resource related subset of configuration on SIGHUP is not implemented. The agent must be restarted to pick up and send changed configuration.

Neutron-server propagates the information further to Placement for the resources of each agent via Placement’s HTTP REST API. To avoid overloading Placement this synchronization generally does not happen on every received heartbeat message. Instead the re-synchronization of the resources of one agent is triggered by:

  • The creation of a network agent record (as queried by openstack network agent list). Please note that deleting an agent record and letting the next heartbeat to re-create it can be used to trigger synchronization without restarting an agent.

  • The restart of that agent (technically start_flag being present in the heartbeat message).

Both of these can be used by an admin to force a re-sync if needed.

The success of a synchronization attempt from neutron-server to Placement is persisted into the relevant agent’s resources_synced attribute. For example:

# as admin
$ openstack network agent show -f value -c resources_synced 5e57b85f-b017-419a-8745-9c406e149f9e
True

resources_synced may take the value True, False and None:

  • None: No sync was attempted (normal for agents not reporting Placement-backed resources).

  • True: The last sync attempt was completely successful.

  • False: The last sync attempt was partially or utterly unsuccessful.

In case resources_synced is not True for an agent, neutron-server does try to re-sync on receiving every heartbeat message from that agent. Therefore it should be able to recover from transient errors of Neutron-Placement communication (e.g. Placement being started later than Neutron).

It is important to note that the restart of neutron-server does not trigger any kind of re-sync to Placement (to avoid an update storm).

As mentioned before, the information flow for resources requested and (if proper) allocated is different. It involves a conversation between Nova, Neutron and Placement.

  1. Neutron exposes a port’s resource needs in terms of resource classes and traits as the admin-only resource_request attribute of that port.

  2. Nova reads this and incorporates it as a numbered request group into the cloud servers overall allocation candidate request to Placement.

  3. Nova selects (schedules) and allocates one candidate returned by Placement.

  4. Nova informs Neutron when binding the port of which physical network interface resource provider had been selected for the port’s resource request in the binding:profile.allocation sub-attribute of that port.

For details please see slides 13-15 of a (pre-release) demo that was presented on the Berlin Summit in November 2018.

Sample usage

Physnets and QoS policies (together with their rules) are usually pre-created by a cloud admin:

# as admin

$ openstack network create net0 \
    --provider-network-type vlan \
    --provider-physical-network physnet0 \
    --provider-segment 100

$ openstack subnet create subnet0 \
    --network net0 \
    --subnet-range 10.0.4.0/24

$ openstack network qos policy create policy0

$ openstack network qos rule create policy0 \
    --type minimum-bandwidth \
    --min-kbps 1000000 \
    --egress

$ openstack network qos rule create policy0 \
    --type minimum-bandwidth \
    --min-kbps 1000000 \
    --ingress

Then a normal user can use the pre-created policy to create ports and boot servers with those ports:

# as an unprivileged user

# an ordinary soft-switched port: ``--vnic-type normal`` is the default
$ openstack port create port-normal-qos \
    --network net0 \
    --qos-policy policy0

# alternatively an SR-IOV port, unused in this example
$ openstack port create port-direct-qos \
    --network net0 \
    --vnic-type direct \
    --qos-policy policy0

$ openstack server create server0 \
    --flavor cirros256 \
    --image cirros-0.5.1-x86_64-disk \
    --port port-normal-qos

On Healing of Allocations

Since Placement carries a global view of a cloud deployment’s resources (what is available, what is used) it may in some conditions get out of sync with reality.

One important case is when the data-plane-only Minimum Guaranteed Bandwidth feature was used before Stein (first released in Newton). Since before Stein guarantees were not enforced during server placement the available resources may have become overallocated without notice. In this case Placement’s view and the reality of resource usage should be made consistent during/after an upgrade to Stein.

Another case stems from OpenStack not having distributed transactions to allocate resources provided by multiple OpenStack components (here Nova and Neutron). There are known race conditions in which Placement’s view may get out of sync with reality. The design knowingly minimizes the race condition windows, but there are known problems:

  • If a QoS policy is modified after Nova read a port’s resource_request but before the port is bound its state before the modification will be applied.

  • If a bound port with a resource allocation is deleted. The port’s allocation is leaked. https://bugs.launchpad.net/nova/+bug/1820588

Note

Deleting a bound port has no known use case. Please consider detaching the interface first by openstack server remove port instead.

Incorrect allocations may be fixed by:

  • Moving the server, which will delete the wrong allocation and create the correct allocation as soon as move operations are implemented (not in Stein unfortunately). Moving servers fixes local overallocations.

  • The need for an upgrade-helper allocation healing tool is being tracked in bug 1819923.

  • Manually, by using openstack resource provider allocation set /delete.

Debugging

  • Are all components running at least the Stein release?

  • Is the placement service plugin enabled in neutron-server?

  • Is resource_provider_bandwidths configured for the relevant neutron agent?

  • Is resource_provider_bandwidths aligned with bridge_mappings or physical_device_mappings?

  • Was the agent restarted since changing the configuration file?

  • Is resource_provider_bandwidths reaching neutron-server?

# as admin
$ openstack network agent show ... | grep configurations

Please find an example in section Propagation of resource information.

  • Did neutron-server successfully sync to Placement?

# as admin
$ openstack network agent show ... | grep resources_synced

Please find an example in section Propagation of resource information.

  • Is the resource provider tree correct? Is the root a compute host? One level below the agents? Two levels below the physical network interfaces?

$ openstack --os-placement-api-version 1.17 resource provider list
+--------------------------------------+------------------------------------------+------------+--------------------------------------+--------------------------------------+
| uuid                                 | name                                     | generation | root_provider_uuid                   | parent_provider_uuid                 |
+--------------------------------------+------------------------------------------+------------+--------------------------------------+--------------------------------------+
| 3b36d91e-bf60-460f-b1f8-3322dee5cdfd | devstack0                                |          2 | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd | None                                 |
| 4a8a819d-61f9-5822-8c5c-3e9c7cb942d6 | devstack0:NIC Switch agent               |          0 | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd |
| 1c7e83f0-108d-5c35-ada7-7ebebbe43aad | devstack0:NIC Switch agent:ens5          |          2 | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd | 4a8a819d-61f9-5822-8c5c-3e9c7cb942d6 |
| 89ca1421-5117-5348-acab-6d0e2054239c | devstack0:Open vSwitch agent             |          0 | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd |
| f9c9ce07-679d-5d72-ac5f-31720811629a | devstack0:Open vSwitch agent:br-physnet0 |          2 | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd | 89ca1421-5117-5348-acab-6d0e2054239c |
+--------------------------------------+------------------------------------------+------------+--------------------------------------+--------------------------------------+
  • Does Placement have the expected traits?

# as admin
$ openstack --os-placement-api-version 1.17 trait list | awk '/CUSTOM_/ { print $2 }' | sort
CUSTOM_PHYSNET_PHYSNET0
CUSTOM_VNIC_TYPE_DIRECT
CUSTOM_VNIC_TYPE_DIRECT_PHYSICAL
CUSTOM_VNIC_TYPE_MACVTAP
CUSTOM_VNIC_TYPE_NORMAL
  • Do the physical network interface resource providers have the proper trait associations and inventories?

# as admin
$ openstack --os-placement-api-version 1.17 resource provider trait list RP-UUID
$ openstack --os-placement-api-version 1.17 resource provider inventory list RP-UUID
  • Does the QoS policy have a minimum-bandwidth rule?

  • Does the port have the proper policy?

  • Does the port have a resource_request?

# as admin
$ openstack port show port-normal-qos | grep resource_request
  • Was the server booted with a port (as opposed to a network)?

  • Did nova allocate resources for the server in Placement?

# as admin
$ openstack --os-placement-api-version 1.17 resource provider allocation show SERVER-UUID
  • Does the allocation have a part on the expected physical network interface resource provider?

# as admin
$ openstack --os-placement-api-version 1.17 resource provider show --allocations RP-UUID
  • Did placement manage to produce an allocation candidate list to nova during scheduling?

  • Did nova manage to schedule the server?

  • Did nova tell neutron which physical network interface resource provider was allocated to satisfy the bandwidth request?

# as admin
$ openstack port show port-normal-qos | grep binding.profile.*allocation
  • Did neutron manage to bind the port?