Jump to: navigation, search

OpenDaylight Controller:MD-SAL:MD-SAL Document Review:Architecture


The technologies in the design of the OpenDaylight controller software architecture lie at the intersection of three recent technology trends in software engineering and networking: Software Defined Networking (SDN) innovations, model driven software engineering (MDSE) and model driven network management. Because SDN is driven (in part) by the demands for increased network programmability (enabling the networking agility component of the movement toward increased virtualization, abstraction and programmatic control), the controller functions as both “middleware” in solution delivery and “foundry” in application development. The OpenDaylight controller software architecture defines patterns of application creation and interaction (APIs and RPC governance), underlying application services (e.g. message routing, formatting and data storage) and has ultimately lead to new development tools for an environment enabled for open source collaboration.

The following sections are organized as follows. Section Background provides a brief background of these key software and networking technologies/approaches used in the model-driven controller architecture/design. Section Requirements describes the context and requirements for the OpenDaylight controller architecture/design. Finally, Section Architecture describes the architecture of the OpenDaylight controller.


Software Defined Networking

Software-defined networking emerged from work done at UC Berkeley and Stanford University around 2008. The original SDN network architecture proposed decoupling of network control from packet forwarding and a direct programmability of the network control function [1]. Network intelligence is (logically) centralized in software-based SDN controllers, which maintain a global view of the network. As a result, the network appears to the applications and policy engines as a single, logical switch. This architecture has been further evolved in the Open Networking Foundation (ONF), an organization that also drives the evolution of the OpenFlow protocol standard. OpenFlow allows the network control function to remotely program the packet forwarding function.

Over the years, multiple controllers were developed for this architecture: Beacon, Floodlight, NOX, POX, Ryu and more recently ONOS. Each controller implemented OpenFlow as the sole southbound protocol towards the packet forwarding function, and provided access to its control plane functions through northbound REST APIs. These controller platforms supported mostly network control plane applications, with the “killer app” being network virtualization.

While decoupling of control and forwarding functions is a valid and useful SDN use case, SDN can be defined more broadly than by the OpenFlow protocol abstractions. OpenFlow only supports retrieval and programming of data related to the network forwarding function. However, the network is a source of a vast amount of other useful data that can be utilized by SDN applications. Moreover, network data can be retrieved and networks can be programmed through a variety of protocols and APIs at different levels of information abstraction, such as BGP, NETCONF, Netflow/IP-Fix, SNMP, or vendor-specific CLIs. Basically, SDN defines a generic feedback/control/policy loop between applications and the network: applications get data/information from the network, process the data and make policy/control decisions that are driven back into the network, as shown in the following figure:


SDN use cases now encompass, in addition to Data Center specific orchestration applications, a variety of diverse SDN applications that harvest network information and use network APIs to program the network. Examples are, in addition to virtualization (VTN or Open DOVE), applications such as multi-layer path optimization, network analytics, Group Policy configuration, Service Function Chaining or Security.

Model-Driven Software Engineering

OpenDaylight leverages Model Driven Software Engineering (MDSE) defined by the Object Management Group (OMG). MDSE describes a framework based on consistent relationships between (different) models, standardized mappings and patterns that enable model generation and, by extension, code/API generation from models. This generalization can overlay any specific modeling language. Although OMG focus their MDA solution on UML, YANG has emerged as the data modeling language for the networking domain.

The models created are portable (cross platform) and can be leveraged to describe business policies, platform operation or component specific capability (vertically) as well as (in the case of SDN) manifest at the service interfaces provided by a network control entity (horizontally and vertically within an application service framework).

MDSE enables run-time mediation and inter-model mapping, for example, if the models are of two different domain types, such as J2EE and UML

Model-Driven Network Management/Programmability

The model-driven approach is being increasingly used in the networking domain to describe the functionality of network devices [4], services [15], policies [18], [19], and network APIs [16], [17]. The protocols of choice are NETCONF and RESTCONF; the modeling language of choice is YANG.

NETCONF is an IETF network management protocol that defines configuration and operational conceptual data stores and a set of Create, Retrieve, Update, Delete (CRUD) operations that can be used to access these data stores. In addition to the data stores CRUD operations, NETCONF also supports simple Remote Procedure Call (RPC) and Notification operations. NETCONF operations are realized on top of a simple Call (RPC) layer. NETCONF uses an XML- based data encoding for the configuration and operational data, as well as for its protocol messages.

RESTCONF is a REST-like protocol that provides a programmatic interface over HTTP for accessing data defined in YANG, using the data stores defined in NETCONF. Configuration data and state data are exposed as resources that can be retrieved with the HTTP GET method. Resources representing configuration data can be modified with the HTTP DELETE, PATCH, POST, and PUT methods. Data is encoded in either XML or JSON.

YANG was originally developed to model configuration and state data in network devices, but it can also be used to describe other network constructs, such as services, policies, protocols, or subscribers. YANG is tree- structured rather than object-oriented; data is structured into a tree and it can contain complex types, such as lists and unions. In addition to data definitions, YANG supports constructs to model Remote Procedure Calls (RPCs) and Notifications, which make it suitable for use as an Interface Description Language (IDL) in a model-driven system.


The SDN controller should be both a platform for deploying SDN applications and provide (or be associated with) an SDN application development environment.

An SDN controller platform should be built to meet the following key requirements:

  • Flexibility: The controller must be able to accommodate a variety of diverse applications; at the same time, controller applications should use a common framework and programming model and provide consistent APIs to their clients. This is important for troubleshooting, system integration, and for combining applications into higher-level orchestrated workflows.
  • Scale the development process: There should be no common controller infrastructure subsystem where a plugin would have to add code. The architecture must allow for plugins to be developed independently of each other and of the controller infrastructure, and it must support relatively short system integration times. Currently, there are 18 active OpenDaylight projects. Fifteen more projects are in the proposal stage. OpenDaylight projects are largely autonomous, and are being developed by independent teams, with little coordination between them.
  • Run-time Extensibility: The controller must be able to load new protocol and service/application plugins at run-time. The controller’s infrastructure should adapt itself to data schemas (models) that are either ingested from dynamically loaded plugins or discovered from devices. Run-time extensibility allows the controller to adapt to network changes (new devices and/or new features) and avoids the lengthy release cycle typical for legacy EMS/NMS systems, where each new feature in a network device results in a manual change of the device model in the NMS/EMS.
  • Performance & scale: A controller must be able to perform well for a variety of different loads/applications in a diverse set of environments; however, performance should not be achieved at the expense of modularity. The controller architecture should allow for horizontal scaling in clustered/cloud environments.

To support development of SDN applications, an SDN controller should also provide (or be associated with) an application development environment that should meet the following key requirements:

  • Code generation from models should be used to enforce standard API contracts and to generate boilerplate code performing repetitive and error-prone tasks, such as parameter range checking.
    • Code generation should produce functionally equivalent APIs for different language bindings.
  • Use a domain-specific modeling language to describe internal and external system behavior; this fosters co- operation between developers and network domain experts and facilitates system integration.
    • A domain-specific modeling language and code generation tools should enable rapid evolution APIs and protocols (Agility).
    • Domain-specific language/technologies/tools used in the controller must be usable for modeling of generic network constructs, such as services, service chains, subscriber managements and policies.
  • Modeling tools for the controller should be aligned with modeling tools for devices. Then, a common tool chain can be used for both, and device models can be re-used in the controller, creating a zero-touch path between the device and a controller application/plugin that uses its models.
  • The tool chain should support code generation for model-to-model adaptations for services and devices.


History and Overview

OpenDaylight was originally inspired by Beacon, which introduced the use of Open Service Gateway Interface (OSGi) that is key for modularity and run-time loading of components (plugins) into the controller. The initial “bootstrap” contribution to ODL introduced a major innovation: the Service Adaptation Layer (SAL) that separated southbound (SB) protocol plugins and northbound (NB) service/application plugins. The architecture comprised three layers: SB protocol plugins, the SAL, and the NB Application/Service Functions., as shown in the following figure.


The SB (protocol) plugins interface with network devices. The SAL adapted the SB plugin functions to higher-level Application/Service functions, which provide the controller’s NB APIs to applications. Examples of Application/Service functions are the Topology Exporter, the Inventory Manager and the OpenFlow Statistics Manager. The layered architecture allows the controller to support multiple southbound protocols (via SB protocol plugins) and to provide a uniform set of services and APIs to applications through a common set of NB APIs.

In the original SAL (referred to as the “API-Driven SAL”, AD-SAL), developers had to define the SAL APIs used by the plugins and code the adaptation functionality between the northbound and southbound plugins. It quickly became apparent that hand-coding the SAL APIs and adaptations to support new plugin functionality would not scale in an initiative of ODP’s magnitude.

Evolution to the Model-Driven Service Adaptation Layer

To address the requirements for the controller listed in Section Requirements , a new model-driven architecture was proposed and implemented for the Service Adaptation Layer (referred to as the “Model-Driven SAL”, MD-SAL). The architecture is built around concepts, protocols and modeling language described in detail in Section Background.

Controller plugins can be either data/service Providers or data/service Consumers. A Provider provides data/services through its APIs. A Consumer consumes services/data provided by one or more Providers. For example, the OpenFlow protocol plugin provides services to add, modify or delete flows from switches connected to the plugin. One of the consumers of the OpenFlow plugin’s services is the Forwarding Rules Manager that provides higher-level flow programming services to controller clients.

The OpenDaylight development environment includes tooling that generates this code (codecs and Java APIs). The tooling preserves YANG data type hierarchies, retains data tree hierarchy (providing normal Java compile-time type safety) and data addressing hierarchies. A plugin’s APIs are resolved when the plugin is loaded into the controller. The SAL does not contain any plugin-specific code or APIs and is therefore a generic plumbing that can adapt itself to any plugins or services/applications loaded into the controller.

From the infrastructure’s point of view, there is no difference between a protocol plugin and an application/service plugin. All plugin life-cycles are the same, each plugin is an OSGi bundle that contains models defining the plugin’s APIs. Since all plugins are the same to the controller infrastructure, the architecture can be drawn as shown in the following figure, which also shows how in a cluster of multiple controller instances an MD- SAL instance within a JVM container will be connected to the cluster’s message bus and data store.


OpenDaylight YANG Models

The Hydrogen production code contains approximately 110 YANG models; 3 models were defined in IETF RFCs, 8 in IETF drafts, the rest have been defined for specifically for the controller and its applications. Approximately 10 models describe IETF protocol PDUs, 27 models describe various aspects of the OpenFlow protocol and 35 models define the controller’s internal wiring and configuration. Approximately 15 additional models were used for prototyping and proof-of- concept work.

TODO: Helium statistics

Model-Driven SAL (MD-SAL) Detailed Description

TODO: Add a summary of the three things that MD-SAL is: RPC, Notification, Data Store, all described in yang.

Common Concepts

The common concepts used in the MD-SAL design are as follows. An RPC is a one-to-one call triggered by a Consumer, which may be processed by a Provider either local or remote. A Notification is an event, which a Consumer may be interested in to receive, and which is triggered / originated in a Provider. The Data Store is a conceptual data tree, which is described by YANG schemas. A Path is a unique locator of a leaf or sub-tree in the conceptual data tree. Finally, a Mount is a logically-nested MD-SAL instance, which may be using a separate set of YANG models; it supports its own RPCs and Notifications and it allows for reusing device models and a context in network-wide contexts without having to redefine the device models in the controller (“Mount”, as its name suggests, is basically a logical mount of a remote conceptual data store).

MD-SAL Functionality and Data Representations

The MD-SAL provides a variety of functions required for adaptation between Providers and Consumers. First, it routes RPC calls between Consumers and Providers (RPC Call Router). Second, it provides a subscription-based mechanism for delivery of Notifications from Publishers to Subscribers (Notification Broker). Third, it routes data reads from Consumers to a particular data store and coordinates data changes between Providers (Data Broker). Finally, it creates and manages Mounts (Mount Manager).

The implementation of the above SAL functions requires the use of two data representations and two sets of SAL Plugin APIs. The Binding-Independent data format/APIs is a Data Object Model (DOM) representation of YANG trees. This format is suitable for generic components, such as the data store, the NETCONF Connector, RESTCONF, which can derive behavior from a YANG model itself. The Binding- Aware data format/APIs is a specific YANG to Java language binding, which specifies how Java Data Transfer Objects (DTOs) and APIs are generated from YANG model. The API definition for these DTOs, interfaces for invoking / implementing RPCs, interfaces containing Notification callbacks are generated at compile time. Codecs to translate between the Java DTOs and DOM representation are generated on demand at run time. Note that the functionality and performance requirements for both data representations are the same.

MD-SAL Design

The data handling functionality is separated into two distinct brokers: a binding-independent DOM Broker that interprets YANG models at runtime and is the core component of the MD-SAL runtime, and a Binding-Aware Broker that exposes Java APIs for plugins using binding-aware representation of data (Java DTOs. These brokers, along with their supporting components are shown in the following figure:


The DOM Broker uses YANG data APIs to describe data and Instance Identifiers specific to YANG to describe paths to data in the system. Data structures in the Binding-Aware Broker that are visible to applications are generated from YANG models in YANG tools. The DOM Broker relies on presence of YANG schemas, which are interpreted at runtime for functionality-specific purposes, such as RPC routing, data store organization, and validation of paths.

The Binding-Aware Broker relies on Java APIs, which are generated from YANG models, and on common properties of Java DTOs, which are enforced by code generation. Therefore data transfer optimizations (zero-copy) are possible when a data Consumer and a data Provider are both Binding-Aware.

The Binding-Aware Broker connects to the DOM Broker through the BA-BI Connector, so that Binding-Aware Consumer/Provider applications/plugins can communicate with their respective binding-independent counterparts. The BA-BI Connector, together with the Mapping Service, the Schema Service, the Codec Registry and the Codec Generator implement dynamic late binding: the codecs that translate YANG data representations between a binding-independent (DOM) format and DTOs, which are specific to Java bindings, are auto-generated on demand.

The physical Data Store is pluggable – MD-SAL provides an SPI through which different data store implementations can be plugged in.

The Mount concept and the support for APIs generated from models allow for applications talking to NETCONF devices to be compiled directly against device models – there is no need for controller-level models that represent devices. Device models are loaded into the controller from a NETCONF device when the controller connects to the device, and apps can work directly with them.