Core Models

The XOS modeling framework provides a foundation for building CORD, but it is just a tool used to define a set of core models. It is these core models that provide a coherent interface for configuring, controlling, and applying policy to a CORD POD. This gives operators a way to specify and reason about the behavior of CORD, while allowing for a wide range of implementation choices for the underlying software components.

Services, Slices, and ServiceInstances

CORD's core starts with the Service model, which represents all functionality that can be on-boarded into CORD. This model is designed to meet two requirements. The first is to be implementation-agnostic, supporting both server-based implementations (e.g., legacy VNFs running in VMs and micro-services running in containers) and switch-based implementations (e.g., SDN control applications that install flow rules into white-box switches). The second is to be multi-tenant, supporting isolated and virtualized instances that can be created on behalf of both trusted and untrusted tenants.

To realize these two requirements, the Service model builds upon two other models—Slices and ServiceInstances—as shown in the figure below. Specifically, a Service is bound to one or more Slices, each of which represents a distributed resource container in which the Service runs. This resource container, in turn, consists of a scalable set of Instances (VMs or containers) and a set of Networks that interconnect those VMs and containers. Similarly, a Service is bound to one or more ServiceInstances, each of which represents the virtualized partition of the service allocated to some tenant.

Service

Slices model the compute and network resources used to implement a service. By creating and provisioning a Slice, the Service acquires the resources it needs to run the VNF image or the SDN control application that defines its behavior. Services are often bound to a single Slice, but multiple Slices are supported for Services implemented as a collection of micro-services that scales independently.

ServiceInstances model the virtualized/isolated partition of the service allocated to a particular tenant. It defines the context in which a tenant accesses and controls its virtualized instantiation of the Service. In practice, this means the ServiceInstance maintains tenant-specific state, whereas the Service maintains Service-wide state.

How ServiceInstances isolate tenants—and the extent of isolation (e.g., namespace isolation, failure isolation, performance isolation)—is an implementation choice. One option, as depicted by the dotted line in the figure shown above, is for each ServiceInstance to correspond to an underlying compute Instance. Because compute Instances provide isolation, the ServiceInstances are also isolated. But this is just one possible implementation. A second is that the ServiceInstance corresponds to a logically isolated partition of a horizontally scalable set of compute Instances. A third example is that each ServiceInstances corresponds to an isolated virtual network/channel implemented by some SDN control application. These three example implementations correspond to the vSG, vCDN, and vRouter Services in CORD, respectively.

One important takeaway is that ServiceInstances and compute Instances are not necessarily one-to-one: the former represents a virtualized instance of a service and the latter represents a virtualized instance of a compute resource. Only in certain limited cases is the first implemented by the latter.

M-CORD’s vSGW Service is a fourth example, one that is worth calling out because it does not fully utilize the degrees-of-freedom that the three models provide. vSGW is representative of many legacy VNFs in that it requires only one Slice that consists of a single VM (i.e., it does not necessarily leverage the Slice’s ability to scale across multiple compute Instances). And because the VNF was not designed to support multiple tenant contexts, there is no value in creating ServiceInstances (i.e., there is only Service-wide configuration). There is no harm in creating a ServiceInstance, representing the context in which all subscribers use the vSGW service, but doing so is not necessary since there is no need to control vSGW on a per-subscriber basis.

Service Graphs and Service Chains

Given a set of Services (and their corresponding Slices and ServiceInstances), CORD also defines two core models for interconnecting them: ServiceDependencies and ServiceInstanceLinks. The first defines a dependency of one Service on another, thereby forming a CORD-wide Service Graph. The second defines a dependency between a pair of ServiceInstances, thereby forming a per-subscriber Service Chain.

NOTE: Service Graphs and Service Chains are not explicit models in CORD, but rather, they are defined by a set of vertices (Services, ServiceInstances) and edges (ServiceDependency, ServiceInstanceLink).

The following figure illustrates an example service graph configured into CORD, along with an example collection of service chains. It does not show the related Slices.

Service Chain

This example is overly simplistic in three ways. One, the Service Graph is not necessarily linear. It generally forms an arbitrary mesh. Two, the Service Chains are not necessarily isomorphic to the Service Graph nor equivalent to each other. Each generally corresponds to a subscriber-specific path through the Service Graph. The path corresponding to one subscriber may be different from the path corresponding to another subscriber. Three, Service Chains are also not necessarily linear. In general, a Service Chain may include “forks” and “joins” that subscriber traffic might follow based on runtime decisions made on a packet-by-packet for flow-by-flow basis.

ServiceDependencies effectively define a template for how ServiceInterfaceLinks are implemented. For example, the ServiceDependency connecting some Service A (the subscriber_service) to some Service B (the provider_service) might indicate that they communicate in the data plane using one of the private networks associated with Service B. In general, this dependency is parameterized by a connect_method that defines how the two services are interconnected in the underlying network data plane. The design is general enough to interconnect two server-based services, two switch-based services, or a server-based and a switch-based service pair. This makes it possible to construct a service graph without regard to how the underlying services are implemented.

The ServiceInterfaceLink associated with a ServiceInstance of A and a the corresponding ServiceInstance of B would then record specific state about that data plane connection (e.g., what address each is known by).

Model Glossary

CORD's core models are defined by a set of xproto specifications. They are defined in their full detail in the source code (see core.xproto). The following summarizes these core models—along with the key relationships (bindings) among them—in words.

  • Service: Represents an elastically scalable, multi-tenant program, including the declarative state needed to instantiate, control, and scale functionality.

    • Bound to a set of Slices that contains the collection of virtualized resources (e.g., compute, network) in which the Service runs.

    • Bound to a set of ServiceInstances that record per-tenant context for a virtualized partition of the Service.

    In many CORD documents you will see mention of each service also having a "controller" which effectively corresponds to the Service model itself (i.e., its purpose is to generate a "control interface" for the service). There is no "Controller" model bound to a service. (Confusingly, CORD does include a Controller model, but it represents information about OpenStack. There is also a ServiceController construct in the TOSCA interface for CORD, which provides a means to load the Service model for a given service into CORD.)

  • ServiceDependency: Represents a dependency between a Subscriber service on a Provider service. The set of ServiceDependency and Service models defined in CORD collectively represent the edges and verticies of a Service Graph, but there is no explicit "ServiceGraph" model in CORD. The dependency between a pair of services is parameterized by the connect_method by which the service are interconnected in the data plane.Connect methods include:

    • None: The two services are not connected in the data plane.
    • Private: The two services are connected by a common private network.
    • Public: The two services are connected by a publicly routable network.
  • ServiceInstance: Represents an instance of a service instantiated on behalf of a particular tenant. This is a generalization of the idea of a Compute-as-a-Service spinning up individual "compute instances," or using another common example, the ServiceInstance corresponding to a Storage Service might be called a "Volume" or a "Bucket." Confusingly, there are also instances of a Service model that represent different services, but this is a consequence of standard modeling terminology, whereas ServiceInstance is a core model in CORD (and yes, there are instances of the ServiceInstance model).

  • ServiceInstanceLink: Represents a logical connection between ServiceInstances of two Services. A related model, ServiceInterface, types the ServiceInstanceLink between two ServiceInstances. A connected sequence of ServiceInstances and ServiceInstanceLinks form what is often called a Service Chain, but there is no explicit "ServiceChain" model in CORD.

  • Slice: Represents a distributed resource container that includes the compute and network resources that belong to (are used by) some Service.

    • Bound to a set of Instances that provide compute resources for the Slice.

    • Bound to a set of Networks that connect the slice's Instances to each other.

    • Bound to a default Flavor that represents a bundle of resources (e.g., disk, memory, and cores) allocated to an instance. Current flavors borrow from EC2.

    • Bound to a default Image that boots in each of the slice'sInstances. Each Image implies a virtualization layer (e.g., Docker, KVM).

  • Instance: Represents a single compute instance associated with a Slice and instantiated on some physical Node. Each Instance is of some isolation type: vm (implemented as a KVM virtual machine), container (implemented as a Docker container), or container_vm (implemented as a Docker container running inside a KVM virtual machine).

  • Network: Represents a virtual network associated with a Slice. The behavior of a given Networkis defined by a NetworkTemplate, which specifies a set of parameters, including visibility (set to public or private), access (set to direct or indirect), translation (set to noneor nat), and topology_kind (set to bigswitch, physical or custom). There is also a vtn_kind parameter (indicating the Network is manged by VTN), with possible settings: PRIVATE, PUBLIC, MANAGEMENT_LOCAL, MANAGEMENT_HOST, VSG, or ACCESS__AGENT.

  • Node: Represents a physical server that can be virtualized and host Instances.

    • Bound to the Site where the Node is physically located.
  • User: Represents an authenticated principal that is granted a set of privileges to invoke operations on a set of models, objects, and fields in the data model.

  • Privilege: Represents the right to perform a set of read, write, or grant operations on a set of models, objects, and fields.

  • Site: Represents a logical grouping of Nodes that are co-located at the same geographic location, which also typically corresponds to the nodes' location in the physical network. The typical use case involves one configuration of a CORD POD deployed at a single location, although the underlying core includes allows for multi-site deployments.

    • Bound to a set of Nodes located at the Site.