Quality of Service

Quality of Service advanced service is designed as a service plugin. The service is decoupled from the rest of Neutron code on multiple levels (see below).

QoS extends core resources (ports, networks) without using mixins inherited from plugins but through an ml2 extension driver.

Details about the DB models, API extension, and use cases can be found here: qos spec .

Service side design

  • neutron.extensions.qos: base extension + API controller definition. Note that rules are subattributes of policies and hence embedded into their URIs.
  • neutron.services.qos.qos_plugin: QoSPlugin, service plugin that implements ‘qos’ extension, receiving and handling API calls to create/modify policies and rules.
  • neutron.services.qos.notification_drivers.manager: the manager that passes object notifications down to every enabled notification driver.
  • neutron.services.qos.notification_drivers.qos_base: the interface class for pluggable notification drivers that are used to update backends about new {create, update, delete} events on any rule or policy change.
  • neutron.services.qos.notification_drivers.message_queue: MQ-based reference notification driver which updates agents via messaging bus, using RPC callbacks.
  • neutron.core_extensions.base: Contains an interface class to implement core resource (port/network) extensions. Core resource extensions are then easily integrated into interested plugins. We may need to have a core resource extension manager that would utilize those extensions, to avoid plugin modifications for every new core resource extension.
  • neutron.core_extensions.qos: Contains QoS core resource extension that conforms to the interface described above.
  • neutron.plugins.ml2.extensions.qos: Contains ml2 extension driver that handles core resource updates by reusing the core_extensions.qos module mentioned above. In the future, we would like to see a plugin-agnostic core resource extension manager that could be integrated into other plugins with ease.

Supported QoS rule types

Any plugin or Ml2 mechanism driver can claim support for some QoS rule types by providing a plugin/driver class property called ‘supported_qos_rule_types’ that should return a list of strings that correspond to QoS rule types (for the list of all rule types, see: neutron.extensions.qos.VALID_RULE_TYPES).

In the most simple case, the property can be represented by a simple Python list defined on the class.

For Ml2 plugin, the list of supported QoS rule types is defined as a common subset of rules supported by all active mechanism drivers.

Note: the list of supported rule types reported by core plugin is not enforced when accessing QoS rule resources. This is mostly because then we would not be able to create any rules while at least one ml2 driver in gate lacks support for QoS (at the moment of writing, linuxbridge is such a driver).

Database models

QoS design defines the following two conceptual resources to apply QoS rules for a port or a network:

  • QoS policy
  • QoS rule (type specific)

Each QoS policy contains zero or more QoS rules. A policy is then applied to a network or a port, making all rules of the policy applied to the corresponding Neutron resource.

When applied through a network association, policy rules could apply or not to neutron internal ports (like router, dhcp, load balancer, etc..). The QosRule base object provides a default should_apply_to_port method which could be overridden. In the future we may want to have a flag in QoSNetworkPolicyBinding or QosRule to enforce such type of application (for example when limiting all the ingress of routers devices on an external network automatically).

From database point of view, following objects are defined in schema:

  • QosPolicy: directly maps to the conceptual policy resource.
  • QosNetworkPolicyBinding, QosPortPolicyBinding: defines attachment between a Neutron resource and a QoS policy.
  • QosBandwidthLimitRule: defines the only rule type available at the moment.

All database models are defined under:

  • neutron.db.qos.models

QoS versioned objects

There is a long history of passing database dictionaries directly into business logic of Neutron. This path is not the one we wanted to take for QoS effort, so we’ve also introduced a new objects middleware to encapsulate the database logic from the rest of the Neutron code that works with QoS resources. For this, we’ve adopted oslo.versionedobjects library and introduced a new NeutronObject class that is a base for all other objects that will belong to the middle layer. There is an expectation that Neutron will evolve into using objects for all resources it handles, though that part was obviously out of scope for the QoS effort.

Every NeutronObject supports the following operations:

  • get_by_id: returns specific object that is represented by the id passed as an argument.
  • get_objects: returns all objects of the type, potentially with a filter applied.
  • create/update/delete: usual persistence operations.

Base object class is defined in:

  • neutron.objects.base

For QoS, new neutron objects were implemented:

  • QosPolicy: directly maps to the conceptual policy resource, as defined above.
  • QosBandwidthLimitRule: class that represents the only rule type supported by initial QoS design.

Those are defined in:

  • neutron.objects.qos.policy
  • neutron.objects.qos.rule

For QosPolicy neutron object, the following public methods were implemented:

  • get_network_policy/get_port_policy: returns a policy object that is attached to the corresponding Neutron resource.
  • attach_network/attach_port: attach a policy to the corresponding Neutron resource.
  • detach_network/detach_port: detach a policy from the corresponding Neutron resource.

In addition to the fields that belong to QoS policy database object itself, synthetic fields were added to the object that represent lists of rules that belong to the policy. To get a list of all rules for a specific policy, a consumer of the object can just access the corresponding attribute via:

  • policy.rules

Implementation is done in a way that will allow adding a new rule list field with little or no modifications in the policy object itself. This is achieved by smart introspection of existing available rule object definitions and automatic definition of those fields on the policy class.

Note that rules are loaded in a non lazy way, meaning they are all fetched from the database on policy fetch.

For Qos<type>Rule objects, an extendable approach was taken to allow easy addition of objects for new rule types. To accomodate this, fields common to all types are put into a base class called QosRule that is then inherited into type-specific rule implementations that, ideally, only define additional fields and some other minor things.

Note that the QosRule base class is not registered with oslo.versionedobjects registry, because it’s not expected that ‘generic’ rules should be instantiated (and to suggest just that, the base rule class is marked as ABC).

QoS objects rely on some primitive database API functions that are added in:

  • neutron.db.api: those can be reused to fetch other models that do not have corresponding versioned objects yet, if needed.
  • neutron.db.qos.api: contains database functions that are specific to QoS models.

RPC communication

Details on RPC communication implemented in reference backend driver are discussed in a separate page.

One thing that should be mentioned here explicitly is that RPC callback endpoints communicate using real versioned objects (as defined by serialization for oslo.versionedobjects library), not vague json dictionaries. Meaning, oslo.versionedobjects are on the wire and not just used internally inside a component.

One more thing to note is that though RPC interface relies on versioned objects, it does not yet rely on versioning features the oslo.versionedobjects library provides. This is because Liberty is the first release where we start using the RPC interface, so we have no way to get different versions in a cluster. That said, the versioning strategy for QoS is thought through and described in the separate page.

There is expectation that after RPC callbacks are introduced in Neutron, we will be able to migrate propagation from server to agents for other resources (f.e. security groups) to the new mechanism. This will need to wait until those resources get proper NeutronObject implementations.

The flow of updates is as follows:

  • if a port that is bound to the agent is attached to a QoS policy, then ML2 plugin detects the change by relying on ML2 QoS extension driver, and notifies the agent about a port change. The agent proceeds with the notification by calling to get_device_details() and getting the new port dict that contains a new qos_policy_id. Each device details dict is passed into l2 agent extension manager that passes it down into every enabled extension, including QoS. QoS extension sees that there is a new unknown QoS policy for a port, so it uses ResourcesPullRpcApi to fetch the current state of the policy (with all the rules included) from the server. After that, the QoS extension applies the rules by calling into QoS driver that corresponds to the agent.
  • on existing QoS policy update (it includes any policy or its rules change), server pushes the new policy object state through ResourcesPushRpcApi interface. The interface fans out the serialized (dehydrated) object to any agent that is listening for QoS policy updates. If an agent have seen the policy before (it is attached to one of the ports it maintains), then it goes with applying the updates to the port. Otherwise, the agent silently ignores the update.

Agent side design

To ease code reusability between agents and to avoid the need to patch an agent for each new core resource extension, pluggable L2 agent extensions were introduced. They can be especially interesting to third parties that don’t want to maintain their code in Neutron tree.

Extensions are meant to receive handle_port events, and do whatever they need with them.

  • neutron.agent.l2.agent_extension: This module defines an abstract extension interface.
  • neutron.agent.l2.extensions.manager: This module contains a manager that allows to register multiple extensions, and passes handle_port events down to all enabled extensions.
  • neutron.agent.l2.extensions.qos defines QoS L2 agent extension. It receives handle_port and delete_port events and passes them down into QoS agent backend driver (see below). The file also defines the QosAgentDriver interface. Note: each backend implements its own driver. The driver handles low level interaction with the underlying networking technology, while the QoS extension handles operations that are common to all agents.

Agent backends

At the moment, QoS is supported by Open vSwitch and SR-IOV ml2 drivers.

Each agent backend defines a QoS driver that implements the QosAgentDriver interface:

  • Open vSwitch (QosOVSAgentDriver);
  • SR-IOV (QosSRIOVAgentDriver).

Open vSwitch

Open vSwitch implementation relies on the new ovs_lib OVSBridge functions:

  • get_egress_bw_limit_for_port
  • create_egress_bw_limit_for_port
  • delete_egress_bw_limit_for_port

An egress bandwidth limit is effectively configured on the port by setting the port Interface parameters ingress_policing_rate and ingress_policing_burst.

That approach is less flexible than linux-htb, Queues and OvS QoS profiles, which we may explore in the future, but which will need to be used in combination with openflow rules.

SR-IOV

SR-IOV bandwidth limit implementation relies on the new pci_lib function:

  • set_vf_max_rate

As the name of the function suggests, the limit is applied on a Virtual Function (VF).

ip link interface has the following limitation for bandwidth limit: it uses Mbps as units of bandwidth measurement, not kbps, and does not support float numbers. So in case the limit is set to something less than 1000 kbps, it’s set to 1 Mbps only. If the limit is set to something that does not divide to 1000 kbps chunks, then the effective limit is rounded to the nearest integer Mbps value.

Configuration

To enable the service, the following steps should be followed:

On server side:

  • enable qos service in service_plugins;
  • set the needed notification_drivers in [qos] section (message_queue is the default);
  • for ml2, add ‘qos’ to extension_drivers in [ml2] section.

On agent side (OVS):

  • add ‘qos’ to extensions in [agent] section.

Testing strategy

All the code added or extended as part of the effort got reasonable unit test coverage.

Neutron objects

Base unit test classes to validate neutron objects were implemented in a way that allows code reuse when introducing a new object type.

There are two test classes that are utilized for that:

  • BaseObjectIfaceTestCase: class to validate basic object operations (mostly CRUD) with database layer isolated.
  • BaseDbObjectTestCase: class to validate the same operations with models in place and database layer unmocked.

Every new object implemented on top of one of those classes is expected to either inherit existing test cases as is, or reimplement it, if it makes sense in terms of how those objects are implemented. Specific test classes can obviously extend the set of test cases as they see needed (f.e. you need to define new test cases for those additional methods that you may add to your object implementations on top of base semantics common to all neutron objects).

Functional tests

Additions to ovs_lib to set bandwidth limits on ports are covered in:

  • neutron.tests.functional.agent.test_ovs_lib

API tests

API tests for basic CRUD operations for ports, networks, policies, and rules were added in:

  • neutron.tests.api.test_qos