]> Juniper Networks

Wien Austria kszarkowicz@juniper.net

Juniper Networks

Rennes France rroberts@juniper.net

Juniper Networks

London United Kingdom jlucek@juniper.net

Juniper Networks

Sunnyvale United States of America jdrake@juniper.net

Orange

France mohamed.boucadair@orange.com

Telefonica

Ronda de la Comunicacion, s/n Madrid Spain luismiguel.contrerasmurillo@telefonica.com http://lmcontreras.com/

Ribbon Communications

Tel Aviv Israel ivan.bykov@rbbn.com

Ciena

Ottawa Canada rrokui@ciena.com

Verizon

Dallas, TX United States of America luay.jalil@verizon.com

XL Axiata

Jakarta Indonesia benyds@xl.co.id

Rakuten

Bangalore India amit.dhamija@rakuten.com

British Telecom

London United Kingdom mojdeh.amani@bt.com

Routing TEAS L3VPN L2VPN Slice Service 5G slicing is a feature that was introduced by the 3rd Generation Partnership Project (3GPP) in mobile networks. This feature covers slicing requirements for all mobile domains, including the Radio Access Network (RAN), Core Network (CN), and Transport Network (TN). This document describes a basic IETF Network Slice realization model in IP/MPLS networks with a focus on the Transport Network fulfilling 5G slicing connectivity requirements. This realization model reuses many building blocks currently commonly used in service provider networks.

Introduction defines a framework for network slicing in the context of networks built using IETF technologies. The IETF network slicing framework introduces the concept of a Network Resource Partition (NRP), which is simply a collection of resources identified in the underlay network. There could be multiple realizations of high-level IETF Network Slice and NRP concepts, where each realization might be optimized for the different network slicing use cases that are listed in . This document describes a basic - using only single NRP - IETF Network Slice realization model in IP/MPLS networks, with a focus on fulfilling 5G slicing connectivity requirements. This IETF Network Slice realization model leverages many building blocks currently commonly used in service provider networks. The reader may refer to for more advanced realization models. A brief 5G overview is provided in for readers' convenience. The reader may refer to and for more details about 3GPP network architectures.

Conventions and Definitions The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here. The document uses the terms defined in . This document makes use of the following terms:

Service Management and Orchestration (SMO):

O-RAN management/orchestration entity

Edge Transport Node (ETN):

Node, under the transport domain orchestration, that stitches the transport domain to an adjacent domain (e.g., enterprise network, data center, peer provider network). An ETN can be be a Provider Edge (PE) or a managed Customer Equipment (CE).

An extended list of abbreviations used in this document is provided in .

5G Network Slicing Integration in Transport Networks

5G Network Slicing versus Transport Network Slicing Network slicing has a different meaning in the 3GPP mobile and transport worlds. Hence, for the sake of precision and whithout seeking to be exhaustive, this section provides a brief description of the objectives of 5G Network Slicing and Transport Network Slicing:

• 5G Network Slicing: The objective of 5G Network Slicing is to provide dedicated resources of the whole 5G infrastructure to some users/customers, applications, or Public Land Mobile Networks (PLMNs) (e.g., RAN sharing). These resources are from the Transport Network (TN), RAN, and Core Network Functions and the underlying infrastructure. defines 5G Network Slicing by introducing the concept of Network Slice Subnet (NSS) to represent slices within each of these domains: RAN, CN, and TN (i.e., RAN NSS, CN NSS, and TN NSS). As per 3GPP specifications, an NSS can be shared or dedicated to a single slice.
• TN Slicing: The objective of TN Slicing is to isolate, guarantee, or prioritize Transport Network resources for slices. Examples of such resources are: buffers, link capacity, or even Routing Information Base (RIB) and Forwarding Information Base (FIB). TN Slicing provides various degrees of sharing of resources between slices. For example, the network capacity can be shared by all slices, usually with a guaranteed minimum per slice, or each individual slice can be allocated dedicated network capacity. Parts of a given network may use the former, while others use the latter. For example, shared TN resources could be provided in the backhaul, and dedicated TN resources could be provided in the midhaul. There are different options to implement TN slices based upon tools, such as VRFs (Virtual Routing and Forwarding instances) for logical separation, QoS (Quality of Service), or TE (Traffic Engineering). A 5G network slicing architecture should integrate TN Slicing for an optimal control of SLAs, however, it is possible to implement 5G Network Slicing without TN Slicing, as explained in the next section. TN Slicing is implemented using IETF technologies, therefore, inline with . In this document, the term "IETF Network Slice" (IETF NS, or INS in short) is used to describe the slice in the Transport Network domain of the overall 5G architecture, composed from RAN, TN, and CN domains.

NF-to-NF Datapath vs. Transport Network The 3GPP specifications loosely define the Transport Network and its integration in RAN and CN domains: the role of the Transport Network is to interconnect Network Functions (NFs). In other words, it is the end-to-end datapath between two NFs. In practice, this end-to-end datapath results often from a non-uniform architecture made up of several segments managed by the same or distinct organizations. This document defines the Transport Network with a service provider scope. That is, the TN extends up to the PE or the CE if it is also managed by the TN Orchestration. Additionally, we assume that the Transport Network is IP, MPLS, or SRv6 capable.

Segmentation of the NF-to-NF Datapath The datapath between NFs may be decomposed into two type of segments based on Orchestration domains:

• TN Segment: The TN Segment provides connectivity between two sites that host NFs. The realization of this segment is driven by the IETF Network Slice Controller (NSC) and the Transport Network Orchestrator (TNO).
• Local Segment: The Local segment either connects two NFs within a given site or connects a NF to the TN. In the first case, the realization of the segment is driven by the 5G Orchestration without any involvement of the IETF NSC or TNO. In the second case, the realization of this segment partially relies on the IETF NSC/TNO for the configuration of the TN-side of the segment (e.g., the configuration of the attachment circuit on a PE interface). Generally, the Local Segment is a datapath local to a site with a potential extension to reach the TN. A site can be (but not limited to): a Data Center (DC), a Point of Presence (PoP), a Central Office (CO), or a virtualized infrastructure in a Public Cloud.

Note that more complex scenarios can be considered (for example, adding an extra segmentation of TN or Local Segments). Additionally, sites can be of different types (such as Edge, Data Center, or Public Cloud), each with specific network design, hardware dependencies, management interface, and diverse networking technologies (e.g., MPLS, SRv6, VXLAN, or L2VPN vs. L3VPN). The objective of this section is to clarify the scope of the Transport Network rather than to cover random technology or design combination. The realization of IETF Network Slices (i.e., connectivity with performance commitments) applies to the TN Segments. We consider Local Segments as an extension of the connectivity of the RAN/CN domain without slice-specific performances requirements by assuming that the local infrastructure is overprovisioned and implements traditional QoS/Scheduling logic. Also, since the TN domain can extend either to the CE or to the PE, we introduce the term Edge Transport Node (ETN) to denote this boundary. The ETN is, therefore, a Transport node that stitches Local Segments and TN Segments. Note that depending on the design, the placement of the Service Demarcation Point (SDP) may or may not be enforced on the ETN itself. is a representation of the end-to-end datapath between NFs including Segments and ETNs (in practice PE or a managed CE), where applicable.

Segmentation of the NF-NF Datapath

NFs may also be placed in the same site and interconnected via a Local Segment. In such a case, there is no TN Segment (i.e., no Transport Network Node is present in the datapath).

NF-to-NF Datapath within the Same Site

provides samples to illustrate the different realizations of Local and TN Segments, as well the SDPs:

• Layer 2 vs. Layer 3 Local Segment: The Local Segment can interconnect the NF and the ETN thanks to a unique VLAN/LAN with no intermediate routing hop (the simplest example is an NF directly connected to a PE): A1, A2, A3, and A4. Alternatively, the NF interfaces may be attached in a different VLAN/LAN than the ETN interface assuming some additional local routing capabilities between the ETN and the NF (e.g., CE, IP Fabric): B1, B2, B3, and B4.
• ETN: It can be either a PE (A3, A4, B3, and B4) or a CE if it is managed by the TN Orchestration (A1, A2, B1, and B2).
• SDP: It can be located in many places as per Section b 4.2 of : A1/B1 for case (1), A2/B2 for case (2), A3/B3 for case (3), and A4/B4 for case (4).
• Redundancy/Scale-out: No example of redundancy/multihoming/scale- out is provided for the sake of simplicification. Nonetheless, each node/NF can be represented by multiple instances (e.g., multiple containers in a cloud architecture).

Examples of various combinations of Local Segments, ETN, and SDP

Orchestration of Local Segment Terminations at ETNs The interconnection between a 5G site and the Transport Network is made up of shared networking resources. More precisely, the Local Segment terminates to an interface of the ETN, which must be configured with consistent dataplane network information (e.g., VLAN- ID and IP addresses/subnets). Hence, the realization of this interconnection requires a coordination between the Service Management and Orchestration (SMO) and the Transport Orchestration (IETF NSC). In this document, and aligned with , we assume that this coordination is based upon standard YANG data models and APIs (more details in further sections). is a basic example of a Layer 3 CE-PE link realization with shared network resources, such as VLAN-ID and IP prefixes, which must be passed between Orchestrators via the Network Slice Service Interface () or an Attachement Circuit Service Interface ().

An Example of Data Exchange

Note that the allocation of these resources (e.g., VLAN-IDs or IP prefixes) can be either managed by the SMO or the Transport Network. In other words, the initial SMO request for the creation of a new IETF Network Slice on a given 5G site may or may not include all network resources. In the latter case, this information is exchanged in a second step.

5G Slice to IETF Network Slice Mapping There are multiple options to map a 5G network slice to IETF Network Slices:

• 1 to N: A single 5G Network Slice can map to multiple IETF Network Slices (1 to N). One example of such a case is the separation of the 5G Control Plane and User Plane: this use case is represented in where a slice (EMBB) is deployed with a separation of User Plane and Control Plane at the TN.
• N to 1: Multiple 5G Network Slices may rely upon the same IETF Network Slice (i.e., in semantic, two RAN/CN NSSes uses a shared TN NSS). In such a case, the Service Level Agreement (SLA) differentiation of slices would be entirely controlled at 5G Control Plane, for example, with appropriate placement strategies: this use case is represented in , where a User Plane Function (UPF) for the URLLC slice is instantiated at the edge cloud close to the gNB CU-UP User Plane for better latency/jitter control, while the 5G Control Plane and the UPF for EMBB slice are instantiated in the regional cloud.
• N to M: The 5G to IETF Network Slice mapping combines both approaches with a mix of shared and dedicated associations.

1 (5G Slice) to N (IETF Network Slice) Mapping

N (5G Slice) to 1 (IETF Network Slice) Mapping

Note that the actual realization of the mapping depends on several factors, such as the actual business cases, the NF vendor capabilities, the NF vendor reference designs, as well as service provider or even legal requirements. Specifically, the actual mapping is a design choice of service operators that may be a function of, e.g., the number of instantiated slices, requested services, or local engineering capabilities and guidelines. However, operators should carefully consider means to ease slice migration strategies (e.g., move from 1-to-1 mapping to N-to-1).

First 5G Slice versus Subsequent Slices A 5G Network Slice is fully functional with both 5G Control Plane and User Plane capabilities (i.e., dedicated NF functions or contexts). In this regard, the creation of the "first slice" is subject to a specific logic since it must deploy both CP and UP. This is not the case for the deployment of subsequent slices because they can share the same CP of the first slice, while instantiating dedicated UP. An example of an incremental deployment is depicted in . At the time of writing (2023), Section 6.2 of specifies that the eMBB slice (SST=1 and no SD) should be supported globally. This 5G slice would be the first slice in any 5G deployment. Note that the actual realization of the mapping depends on several factors such as the actual business cases, the NF vendor capabilities, the NF vendor reference designs, as well as service providers or even legal requirements.

First and Subsequent Slice Deployment

High-Level Overview of the Realization Model introduces the concept of the Network Resource Partition (NRP), which is defined as a collection of resources identified in the underlay network. In the basic realization model described in this document, a single NRP is used with the following characteristics:

• L2VPN/L3VPN service instances for logical separation: This realization model of transport for 5G slices assumes Layer-3 delivery for midhaul and backhaul transport connections, and a Layer 2 or Layer 3 (eCPRI supports both) delivery model for fronthaul connections. L2VPN/L3VPN service instances might be used as basic form of logical slice separation. Further, using service instances results in additional outer header (as packets are encapsulated/decapsulated at the nodes performing PE functions) providing clean discrimantion between 5G QoS and TN QoS, as explained in .
• Fine-grained resource control at the ETN: This is sometimes called 'admission control' or 'traffic conditioning'. The main purpose is the enforcement of the bandwidth contract for the slice right at the edge of the transport domain where the traffic is handed-off between the transport domain and the 5G domains (i.e., RAN/Core). The toolset used here is granular ingress policing (rate limiting) to enforce contracted bandwidths per slice and, potentially, per traffic class within the slice. Out-of-contract traffic might be immediately dropped, or marked as high drop probability traffic, which is more likely to be dropped somewhere at the transit if congestion occurs. In the egress direction at the edge of the transport domain, hierarchical schedulers/shapers can be deployed, providing guaranteed rates per slice, as well as guarantees per traffic class within the slice. In the managed CE use cases (use cases A1, A2, B1, and B2 depicted in ) edge admission control could be distributed between CE and PE, where one part of the edge admission control is implemented on CE, and another part of the edge admission control is implemented on PE.
• Coarse resource control at the TN transit (non-attachment circuits) links of the transport domain, using a single NRP, spanning the entire TN domain. Transit nodes do not maintain any state of individual slices. Instead, only a flat (non-hierarchical) QoS model is used on transit links with up to 8 traffic classes. At the transport domain edge, traffic-flows from multiple slice services are mapped to the limited number of traffic classes used on transit links.
• Capacity planning/management for efficient usage of TN edge and TN transit resources: The role of capacity management is to ensure the transport capacity can be utilized without causing any bottlenecks. The toolset used here can range from careful network planning, to ensure more less equal traffic distribution (i.e., equal cost load balancing), to advanced traffic engineering techniques, with or without bandwidth reservations, to force more consistent load distribution even in non-ECMP friendly network topologies.

Resource Allocation in with single NRP Slicing Model

The 5G control plane relies upon S-NSSAI (Single Network Slice Selection Assistance Information: 32-bit slice identifier) for slice identification. The S-NSSAI is not visible to the transport domain, so instead 5G functions can expose the 5G slices to the transport domain by mapping to explicit L2/L3 identifiers such as VLAN-ID, IP addresses, or Differentiated Services Code Point (DSCP) as documented in .

VLAN Hand-off In this option, the IETF Network Slice, fulfilling connectivity requirements between NFs of some 5G slice, is represented at the SDP by a VLAN, or double VLANs (commonly known as QinQ). Each VLAN can represent a distinct logical interface on the attachment circuits, hence it provides the possibility to place these logical interfaces in distinct L2 or L3 service instances and implement separation between slices via service instances. Since the 5G interfaces are IP based interfaces (the only exception could be the F2 fronthaul- interface, where eCPRI with Ethernet encapsulation is used), this VLAN is typically not transported across the TN domain. Typically, it has only local significance at a particular SDP. For simplification it is recommended to rely on a same VLAN identifier for all ACs, when possible. However, SDPs for a same slice at different locations may also use different VLAN values. Therefore, a VLAN to IETF Network Slice mapping table MUST be maintained for each AC, and the VLAN allocation MUST be coordinated between TN orchestration and local segment orchestration. Thus, while VLAN hand-off is simple from the NF point of view, it adds complexity due to the requirement of maintaining mapping tables for each SDP.

5G Slice with VLAN Hand-off

IP Hand-off In this option, the slices in the transport domain are instantiated by IP tunnels (for example, IPsec, GTP-U tunnel) established between NFs. The transport for a single 5G slice is constructed with multiple such tunnels, since a typical 5G slice contains many NFs - especially DUs and CUs. If a shared NF (i.e., an NF that serves multiple slices, for example a shared DU) is deployed, multiple tunnels from shared NF are established, each tunnel representing a single slice. As opposed to the VLAN hand-off case, there is no logical interface representing slice on the PE, hence all slices are handled within single service instance. On the other hand, similarly to the VLAN hand-off case, a mapping table tracking IP to IETF Network Slice mapping is required.

5G Slice with IP Hand-off

The mapping table can be simplified if, e.g., IPv6 addressing is used to address NFs. An IPv6 address is a 128-bit long field, while the S-NSSAI is a 32-bit field: Slice/Service Type (SST): 8 bits, Slice Differentiator (SD): 24 bits. 32 bits, out of 128 bits of the IPv6 address, may be used to encode the S-NSSAI, which makes an IP to Slice mapping table unnecessary. This mapping is simply a local allocation method to allocate IPv6 addresses to NF loopbacks, without redefining IPv6 semantic. Different IPv6 address allocation schemes following this mapping approach may be used, with one example allocation showed in . Note that this addressing scheme is local to a node; intermediary nodes are not required to associate any additional semantic with IPv6 address. One benefit of embedding the S-NSSAI in the IPv6 address is providing a very easy way of identifying the packet as belonging to given S-NSSAI at any place in the transport domain. This might be used, for example, to selectively enable per S-NSSAI monitoring, or any other per S-NSSAI handling, if required.

An Example of S-NSSAI embedded into IPv6

In the example shown in , the most significant 96 bits of the IPv6 address are unique to NF, but do not carry any slice-specific information, while the least significant 32 bits are used to embed the S-NSSAI information. The 96-bit part of the address may be further divided based, for example, on the geographical location or the DC identification. shows an example of slicing deployment, where S-NSSAI is embedded into IPv6 addresses used by NFs. NF-A has a set of tunnel termination points, with unique per-slice IP addresses allocated from the 2001:db8::a:0:0/96 prefix, while NF-B uses set of tunnel termination points with per-slice IP addresses allocated from 2001:db8::b:0:0/96. This example shows two slices: eMBB (SST=1) and MIoT (SST=3). Therefore, for eMBB the tunnel IP addresses are auto- derived (without the need for a mapping table) as {2001:db8::a:100:0, 2001:db8::b:100:0}, while for MIoT (SST=3) tunnel uses {2001:db8::a:300:0, 2001:db8::b:300:0}.

Deployment example with S-NSSAI embedded into IPv6

MPLS Label Hand-off In this option, the service instances representing different slices are created directly on the NF, or within the cloud infrastructure hosting the NF, and attached to the TN domain. Therefore, the packet is MPLS encapsulated outside the TN domain with native MPLS encapsulation, or MPLSoUDP encapsulation, depending on the capability of the NF or cloud infrastructure, with the service label depicting the slice. There are three major methods (based upon Section 10 of ) for interconnecting multiple service domains:

• Option 10A (): VRF-to-VRF connections.
• Option 10B (): redistribution of labeled VPN routes with next-hop change at domain boundaries.
• Option 10C (): redistribution of labeled VPN routes without next-hop change + redistribution of labeled transport routes with next-hop change at domain boundaries.

Option 10A In this option, MPLS is not used in VRF-to-VRF hand-offs, since services are terminated at the boundary of each domain, and VLAN hand-off is in place between the domains. Thus, this option is the same as VLAN hand-off, described in .

Option 10B In this option, L3VPN service instances for different IETF Network Slice Services are instantiated outside the TN domain. These L3VPN service instances could be instantiated either on the compute, hosting mobile network functions (, left hand side), or within the cloud infrastructure itself (e.g., on the top of the rack or leaf switch within cloud IP fabric (, right hand side)). On the local segment connected to ETN packets are already MPLS encapsulated (or MPLSoUDP encapsulated, if cloud or compute infrastructure doesn't support native MPLS encapsulation). Therefore, the PE uses neither a VLAN nor an IP address for slice identification at the SDP, but instead uses the MPLS label.

MPLS Hand-off: Option B

MPLS labels are allocated dynamically, especially in Option 10B deployments, where at the domain boundaries service prefixes are reflected with next-hop self, and new label is dynamically allocated, as visible in (e.g., labels A, A' and A" for the first depicted slice). Therefore, for any slice-specific per hop behavior at the TN domain edge, the ETN must be able to determine which label represents which slice. In the BGP control plane, when exchanging service prefixes over local segment, each slice might be represented by a unique BGP community, so tracking label assignment to the slice is possible. For example, in , for the slice identified with COM=1, ETN advertises a dynamically allocated label A". Since, based on the community, the label to slice association is known, ETN can use this dynamically allocated label A" to identify incoming packets as belonging to slice 1, and execute appropriate edge per hop behavior. It is worth noting that slice identification in the BGP control plane might be at the prefix granularity. In extreme case, each prefix can have different community representing a different slice. Depending on the business requirements, each slice could be represented by a different service instance, as outlined in . In that case, the route target extended community might be used as slice differentiator. In another deployment, all prefixes (representing different slices) might be handled by single 'mobile' service instance, and some other BGP attribute (e.g., a standard community) might be used for slice differentiation. Or there could be a deployment that groups multiple slices together into a single service instance, resulting in a handful of service instances. In any case, fine-grained per-hop behavior at the edge of TN domain is possible.

Option 10C for further study

QoS Mapping Models The resources are managed via various QoS policies deployed in the network. QoS mapping models to support 5G slicing connectivity implemented over packet switched transport uses two layers of QoS that are discussed in the following subsections.

5G QoS Layer QoS treatment is indicated in the 5G QoS layer by the 5QI (5G QoS indicator), as defined in . A 5QI is an identifier (ID) that is used as a reference to 5G QoS characteristics (e.g., scheduling weights, admission thresholds, queue management thresholds, and link layer protocol configuration) in the RAN domain. Given that 5QI applies to the RAN domain, it is not visible to the TN domain. Therefore, if 5QI-aware treatment is desired in the TN domain as well, 5G network functions might set DSCP with a value representing 5QI so that differentiated treatment can implemented in TN domain as well. Based on these DSCP values, at SDP of each TN segment used to construct transport for given 5G slice, very granular QoS enforcement might be implemented. The mapping between 5QI and DSCP is out of scope for this document. Mapping recommendations are documented, e.g., in . Each slice service might have flows with multiple 5QIs, thus there could be many different 5QIs being deployed. 5QIs (or, more precisely, corresponding DSCP values) are visible to the TN domain at SDP (i.e., at the edge of the TN domain). In this document, this layer of QoS will be referred as '5G QoS Class' ('5G QoS' in short), or '5G DSCP'.

TN QoS Layer Control of the TN resources on transit links, as well as traffic scheduling/prioritization on transit links, is based on a flat (non-hierarchical) QoS model in this IETF Network Slice realization. That is, IETF Network Slices are assigned dedicated resources (e.g., QoS queues) at the edge of the TN domain (at SDPs), while all IETF Network Slices are sharing resources (sharing QoS queues) on the transit links of the TN domain. Typical router hardware can support up to 8 traffic queues per port, therefore the architecture assumes 8 traffic queues per port support in general. At this layer, QoS treatment is indicated by QoS indicator specific to the encapsulation used in the TN domain, and it could be DSCP or MPLS Traffic Class (TC). This layer of QoS will be referred as 'TN QoS Class', or 'TN QoS' for short, in this document.

QoS Realization Models While 5QI might be exposed to the TN domain, via the DSCP value (corresponding to specific 5QI value) set in the IP packet generated by NFs, some 5G deployments might use 5QI in the RAN domain only, without requesting per 5QI differentiated treatment from the TN domain. This can be due to an NF limitation (e.g., no capability to set DSCP), or it might simply depend on the overall slicing deployment model. The O-RAN Alliance, for example, defines a phased approach to the slicing, with initial phases utilizing only per slice, but not per 5QI, differentiated treatment in the TN domain (Annex F of ). Therefore, from a QoS perspective, the 5G slicing connectivity realization architecture defines two high-level realization models for slicing in the transport domain: a 5QI-unaware model and a 5QI- aware model. Both slicing models in the transport domain could be used concurrently within the same 5G slice. For example, the TN segment for 5G midhaul (F1-U interface) might be 5QI-aware, while at the same time the TN segment for 5G backhaul (N3 interface) might follow the 5QI-unaware model. These models are further elaborated in the following two subsections.

5QI-unaware Model In 5QI-unaware mode, the DSCP values in the packets received from NF at SDP are ignored. In the TN domain, there is no QoS differentiation at the 5G QoS Class level. The entire IETF Network Slice is mapped to single TN QoS Class, and, therefore, to a single QoS queue on the routers in the TN domain. With a small number of deployed 5G slices (for example only two 5G slices: eMBB and MIoT), it is possible to dedicate a separate QoS queue for each slice on transit routers. However, with introduction of private/enterprises slices, as the number of 5G slices (and thus corresponding IETF Network Slices) increases, a single QoS queue on transit links serves multiple slices with similar characteristics. QoS enforcement on transit links is fully coarse (single NRP, sharing resources among all IETF Network Slices), as displayed in .

Slice to TN QoS Mapping (5QI-unaware Model)

When the IP traffic is handed over at the SDP from the local segment to the TN domain, the PE encapsulates the traffic into MPLS (if MPLS transport is used in the TN domain), or IPv6 - optionally with some additional headers (if SRv6 transport is used in the TN domain), and sends out the packets on the TN transit link. The original IP header retains the DCSP marking (which is ignored in 5QI-unaware mode), while the new header (MPLS or IPv6) carries QoS marking (MPLS Traffic Class bits for MPLS encapsulation, or DSCP for SRv6/IPv6 encapsulation) related to TN CoS. Based on TN QoS Class marking, per hop behavior for all IETF Network Slices is executed on TN links. TN domain transit routers do not evaluate the original IP header for QoS-related decisions. This model is outlined in for MPLS encapsulation, and in for SRv6 encapsulation.

QoS with MPLS Encapsulation

QoS with IPv6 Encapsulation

From the QoS perspective, both options are similar. However, there is one difference between the two options. The MPLS TC is only 3 bits (8 possible combinations), while DSCP is 6 bits (64 possible combinations). Hence, SRv6 provides more flexibility for TN CoS design, especially in combination with soft policing with in-profile/ out-profile traffic, as discussed in . Edge resources are controlled in a granular, fine-grained manner, with dedicated resource allocation for each IETF Network Slice. The resource control/enforcement happens at each SDP in two directions: inbound and outbound.

Inbound Edge Resource Control The main aspect of inbound edge resource control is per-slice traffic capacity enforcement. This kind of enforcement is often called 'admission control' or 'traffic conditioning'. The goal of this inbound enforcement is to ensure that the traffic above the contracted rate is dropped or deprioritized, depending on the business rules, right at the edge of TN domain. This, combined with appropriate network capacity planning/management () is required to ensure proper isolation between slices in a scalable manner. As a result, traffic of one slice has no influence on the traffic of other slices, even if the slice is misbehaving (e.g., DDoS attacks or node/link failures) and generates traffic volumes above the contracted rates. The slice rates can be characterized with following parameters :

• CIR: Committed Information Rate (i.e., guaranteed bandwidth)
• PIR: Peak Information Rate (i.e., maximum bandwidth)

These parameters define the traffic characteristics of the slice and are part of SLO parameter set provided by the SMO to IETF NSC. Based on these parameters the inbound policy can be implemented using one of following options:

• 1r2c (single-rate two-color) rate limiter This is the most basic rate limiter, which meters at the SDP a traffic stream of given slice and marks its packets as in-contract (below contracted CIR) or out-of-contract (above contracted CIR). In-contract packets are accepted and forwarded. Out-of contract packets are either dropped right at the SDP (hard rate limiting), or remarked (with different MPLS TC or DSCP TN markings) to signify 'this packet should be dropped in the first place, if there is a congestion' (soft rate limiting), depending on the business policy of the operator. In the second case, while packets above CIR are forwarded at the SDP, they are subject to being dropped during any congestion event at any place in the TN domain.
• 2r3c (two-rate three-color) rate limiter This was initially defined in , and its improved version in . In essence, the traffic is assigned to one of the these three categories:
  • Green, for traffic under CIR
  • Yellow, for traffic between CIR and PIR
  • Red, for traffic above PIR
  An inbound 2r3c meter implemented with , compared to , is more 'customer friendly' as it doesn't impose outbound peak-rate shaping requirements on customer edge (CE) devices. 2r3c meters in general give greater flexibility for edge enforcement regarding accepting the traffic (green), de- prioritizing and potentially dropping the traffic during congestion (yellow), or hard dropping the traffic (red).

Inbound edge enforcement model for 5QI-unaware model, where all packets belonging to the slice are treated the same way in the TN domain (no 5Q QoS Class differentiation in the TN domain) is outlined in .

Ingress Slice Admission Control (5QI-unware Model)

Outbound Edge Resource Control While inbound slice admission control at the transport edge is mandatory in the model, outbound edge resource control might not be required in all use cases. Use cases that specifically call for outbound edge resource control are:

• Slices use both CIR and PIR parameters, and transport edge links (attachment circuits) are dimensioned to fulfil the aggregate of slice CIRs. If at any given time, some slices send the traffic above CIR, congestion in outbound direction on the transport edge link might happen. Therefore, fine-grained resource control to guarantee at least CIR for each slice is required.
• Any-to-Any (A2A) connectivity constructs are deployed, again resulting in potential congestion in outbound direction on the transport edge links, even if only slice CIR parameters are used. This again requires fine-grained resource control per slice in outbound direction at transport edge links.

As opposed to inbound edge resource control, typically implemented with rate-limiters/policers, outbound resource control is typically implemented with a weighted/priority queuing, potentially combined with optional shapers (per slice). A detailed analysis of different queuing mechanisms is out of scope for this document, but is provided in . outlines the outbound edge resource control model at the transport network layer for 5QI-unaware slices. Each slice is assigned a single egress queue. The sum of slice CIRs, used as the weight in weighted queueing model, MUST NOT exceed the physical capacity of the attachment circuit. Slice requests above this limit MUST be rejected by the NSC, unless an already established slice with lower priority, if such exists, is preempted.

Ingress Slice Admission control (5QI-unaware Model)

5QI-aware Model In the 5QI-aware model, potentially a large number of 5G QoS Classes, represented via DSCP set by NFs (the architecture scales to thousands of 5G slices) is mapped (multiplexed) to up to 8 TN QoS Classes used in transport transit equipment, as outlined in .

Slice 5Q QoS to TN QoS Mapping (5QI-aware Model)

Given that in large scale deployments (large number of 5G slices), the number of potential 5G QoS Classes is much higher than the number of TN QoS Classes, multiple 5G QoS Classes with similar characteristics - potentially from different slices - would be grouped with common operator-defined TN logic and mapped to a same TN QoS Class when transported in the TN domain. That is, common per hop behavior (PHB) is executed on transit TN routers for all packets grouped together. An example of this approach is outlined in . Note that the numbers indicated in (S-NSSAI, 5QI, DSCP, queue, etc.) are provided for illustration purposes only and shoudl not be considered as deployment guidance.

Example of 3GPP QoS Mapped to TN QoS

In current SDO progress of 3GPP (Rel.17) and O-RAN the mapping of 5QI to DSCP is not expected in per-slice fashion, where 5QI to DSCP mapping may vary from 3GPP slice to 3GPP slice, hence the mapping of 5G QoS DSCP values to TN QoS Classes may be rather common. Like in 5QI-unaware model, the original IP header retains the DCSP marking corresponding to 5QI (5G QoS Class), while the new header (MPLS or IPv6) carries QoS marking related to TN QoS Class. Based on TN QoS Class marking, per hop behavior for all aggregated 5G QoS Classes from all IETF Network Slices is executed on TN links. TN domain transit routers do not evaluate original IP header for QoS related decisions. The original DSCP marking retained in the original IP header is used at the PE for fine-grained per slice and per 5G QoS Class inbound/outbound enforcement on the AC. In 5QI-aware model, compared to 5QI-unware model, edge resources are controlled in an even more granular, fine-grained manner, with dedicated resource allocation for each IETF Network Slice and dedicated resource allocation for number of traffic classes (most commonly up 4 or 8 traffic classes, depending on the HW capability of the equipment) within each IETF Network Slice.

Inbound Edge Resource Control Compared to the 5QI-unware model, admission control (traffic conditioning) in the 5QI-aware model is more granular, as it enforces not only per slice capacity constraints, but may as well enforce the constraints per 5G QoS Class within each slice. 5G slice using multiple 5QIs can potentially specify rates in one of the following ways:

• Rates per traffic class (CIR or CIR+PIR), no rate per slice (sum of rates per class gives the rate per slice).
• Rate per slice (CIR or CIR+PIR), and rates per prioritized (premium) traffic classes (CIR only). Best effort traffic class uses the bandwidth (within slice CIR/PIR) not consumed by prioritized classes.

In the first option, the slice admission control is executed with traffic class granularity, as outlined in . In this model, if a premium class doesn't consume all available class capacity, it cannot be reused by non-premium (i.e., Best Effort) class.

Ingress Slice Admission Control (5QI-aware Model)

The second model combines the advantages of 5QI-unaware model (per slice admission control) with the per traffic class admission control, as outlined in . Ingress admission control is at class granularity for premium classes (CIR only). Non-premium class (i.e., Best Effort) has no separate class admission control policy, but it is allowed to use the entire slice capacity, which is available at any given moment. I.e., slice capacity, which is not consumed by premium classes. It is a hierarchical model, as depicted in .

Ingress Slice Admission Control (5QI-aware) - Hierarchical

Outbound Edge Resource Control outlines the outbound edge resource control model at the transport network layer for 5QI-aware slices. Each slice is assigned multiple egress queues. The sum of queue weights (equal to 5Q QoS CIRs within the slice) CIRs MUST NOT exceed the CIR of the slice itself. And, similarly to the 5QI-aware model, the sum of slice CIRs MUST NOT exceed the physical capacity of the attachment circuit.

Egress Slice Admission Control (5QI-aware)

Transit Resource Control Transit resource control is much simpler than Edge resource control. As outlined in , at the edge, 5Q QoS Class marking (represented by DSCP related to 5QI set by mobile network functions in the packets handed off to the TN) is mapped to the TN QoS Class. Based in TN QoS Class, when the packet is encapsulated with outer header (MPLS or IPv6), TN QoS Class marking (MPLS TC or IPv6 DSCP in outer header, as depicted in and ) is set in the outer header. PHB on transit is based exclusively on that TN QoS Class marking, i.e., original 5G QoS Class DSCP is not taken into consideration on transit. Transit resource control does not use any inbound interface policy, but only outbound interface policy, which is based on priority queue combined with weighted or deficit queuing model, without any shaper. The main purpose of transit resource control is to ensure that during network congestion events, for example caused by network failures and temporary rerouting, premium classes are prioritized, and any drops only occur in traffic that was de-prioritized by ingress admission control or in non-premium (best-effort) classes. Capacity planning and management, as described in , ensures that enough capacity is available to fulfill all approved slice requests.

Transport Planes Mapping Models A network operator might define various tunnel groups, where each tunnel group is created with specific optimization criteria and constraints. This document refers to such tunnel groups as 'transport planes'. For example, a transport plane "A" might represent tunnels optimized for latency, and transport plane "B" represent tunnels optimized for high capacity. depicts an example of a simple network with two transport planes. These transport planes might be realized via various IP/MPLS techniques, for example Flex-Algo or RSVP/SR traffic engineering tunnels with or without PCE, and with or without bandwidth reservations. discusses in detail different bandwidth models that can be deployed in the transport network. However, discussion about how to realize or orchestrate transport planes is out of scope for this document.

Transport Planes

Note that there could be multiple tunnels within a single transport plane between any pair of PEs. For readibility, shows only single tunnel per transport plane for [ingress PE, egress PE] pair. Similar to the QoS mapping models discussed in , for mapping to transport planes at the ingress PE, both 5QI-unaware and 5QI-aware modes are defined. In essence, entire slices can be mapped to transport planes without 5G QoS consideration (5QI-unaware mode), or flows with different 5G QoS Classes, even if they are from the same slice, might be mapped to different transport planes (5QI-aware mode).

5QI-unaware Model As discussed in , in the 5QI-unware model, the TN domain doesn't take into account 5G QoS during execution of per-hop behavior. The entire slice is mapped to single TN QoS Class, therefore the entire slice is subject to the same per-hop behavior. Similarly, in 5QI-unaware transport plane mapping model, the entire slice is mapped to a single transport plane, as depicted in .

Slice to Transport Plane Mapping (5QI-unaware Model)

It is worth noting that there is no strict correlation between TN QoS Classes and Transport Planes. The TN domain can be operated with e.g., 8 TN QoS Classes (representing 8 hardware queues in the routers), and 2 Transport Planes (e.g., latency optimized transport plane using link latency metrics for path calculation, and transport plane following IGP metrics). TN QoS Class determines the per-hop behavior when the packets are transiting through the TN domain, while Transport Plane determines the path, optimized or constrained based on operator's business criteria, that the packets use to transit through the TN domain.

5QI-aware Model In 5QI-aware model, the traffic can be mapped to transport planes at the granularity of 5G QoS Class. Given that the potential number of transport planes is limited, packets from multiple 5G QoS Classes with similar characteristics are mapped to a common transport plane, as depicted in .

Slice to Transport Plane mapping (5QI-aware Model)

Capacity Planning/Management This section describes the information conveyed by the SMO to the transport controller with respect to slice bandwidth requirements. shows three DCs that contain instances of network functions. Also shown are PEs that have links to the DCs. The PEs belong to the transport network. Other details of the transport network, such as P-routers and transit links are not shown. Also details of the DC infrastructure such as switches and routers are not shown. The SMO is aware of the existence of the network functions and their locations. However, it is not aware of the details of the transport network. The transport controller has the opposite view - it is aware of the transport infrastructure and the links between the PEs and the DCs, but is not aware of the individual network functions.

An Example of Multi-DC Architecture

Let us consider 5G Slice "X" that uses some of the network functions in the three DCs. If this slice has latency requirements, the SMO will have taken those into account when deciding which NF instances in which DC is to be invoked for this slice. As a result of such a placement decision, the three DCs shown are involved in 5G Slice "X", rather than other DCs. For its decision-making, the SMO needs information from the NSC about the observed latency between DCs. Preferably, the NSC would present the topology in an abstracted form, consisting of point-to-point abstracted links between pairs of DCs and associated latency and optionally delay variation and link loss values. It would be valuable to have a mechanism for the SMO to inform the NSC which DC-pairs are of interest for these metrics - there may be of order thousands of DCs, but the SMO will only be interested in these metrics for a small fraction of all the possible DC-pairs, i.e. those in the same region of the network. The mechanism for conveying the information will be discussed in a future version of this document. shows the matrix of bandwidth demands for 5G slice "X". Within the slice, multiple network function instances might be sending traffic from DCi to DCj. However, the SMO sums the associated demands into one value. For example, NF1A and NF1B in DC1 might be sending traffic to multiple NFs in DC2, but this is expressed as one value in the traffic matrix: the total bandwidth required for 5G Slice X from DC1 to DC2 (8 units). Each row in the right-most column in the traffic matrix shows the total amount of traffic going from a given DC into the transport network, regardless of the destination DC. Note that this number can be less than the sum of DC-to-DC demands in the same row, on the basis that not all the network functions are likely to be sending at their maximum rate simultaneously. For example, the total traffic from DC1 for Slice X is 11 units, which is less than the sum of the DC-to-DC demands in the same row (13 units). Note, as described in , a slice may have per-QoS class bandwidth requirements, and may have CIR and PIR limits. This is not included in the example, but the same principles apply in such cases.

Inter-DC Traffic Demand Matrix

can be used to convey all of the information in the traffic matrix to the IETF NSC. The IETF NSC applies policers corresponding to the last column in the traffic matrix to the appropriate PE routers, in order to enforce the bandwidth contract. For example, it applies a policer of 11 units to PE1A and PE1B that face DC1, as this is the total bandwidth that DC1 sends into the transport network corresponding to Slice X. Also, the controller may apply shapers in the direction from the TN to the DC, if otherwise there is the possibility of a link in the DC being oversubscribed. Note that a peer NF endpoint of an AC can be identified using 'peer-sap-id' as defined in . Depending on the bandwidth model used in the network (), the other values in the matrix, i.e., the DC-to-DC demands, may not be directly applied to the transport network. Even so, the information may be useful to the IETF NSC for capacity planning and failure simulation purposes. If, on the other hand, the DC-to-DC demand information is not used by the IETF NSC, the IETF YANG Data Model for L3VPN Service Delivery or the IETF YANG Data Model for L2VPN Service Delivery could be used instead of , as they support conveying the bandwidth information in the right-most column of the traffic matrix. The transport network may be implemented in such a way that it has various types of paths, for example low-latency traffic might be mapped onto a different transport path to other traffic (for example a particular flex-algo or a particular set of TE LSPs), as discussed in . The SMO can use to request low-latency transport for a given slice if required. However, or do not support requesting a particular transport-type, e.g., low-latency. One option is to augment these models to convey this information. This can be achieved by reusing the 'underlay- transport' construct defined in and .

Bandwidth Models This section describes three bandwidth management schemes that could be employed in the transport network. Many variations are possible, but each example describes the salient points of the corresponding scheme. Schemes 2 and 3 use TE; other variations on TE are possible as described in .

Scheme 1: Shortest Path Forwarding (SPF) Shortest path forwarding is used according to the IGP metric. Given that some slices are likely to have latency SLOs, the IGP metric on each link can be set to be in proportion to the latency of the link. In this way, all traffic follows the minimum latency path between endpoints. In Scheme 1, although the operator provides bandwidth guarantees to the slice customers, there is no explicit end-to-end underpinning of the bandwidth SLO, in the form of bandwidth reservations across the transport network. Rather, the expected performance is achieved via capacity planning, based on traffic growth trends and anticipated future demands, in order to ensure that network links are not over- subscribed. This scheme is analogous to that used in many existing business VPN deployments, in that bandwidth guarantees are provided to the customers but are not explicitly underpinned end to end across the transport network. A variation on the scheme is that Flex-Algo, defined in , is used, for example one Flex-Algo could use latency-based metrics and another Flex-Algo could use the IGP metric. There would be a many-to-one mapping of slices to Flex- Algos. While Scheme 1 is technically feasible, it is vulnerable to unexpected changes in traffic patterns and/or network element failures resulting in congestion. This is because, unlike Schemes 2 and 3 that employ TE, traffic cannot be diverted from the shortest path.

Scheme 2: TE LSPs with Fixed Bandwidth Reservations Scheme 2 uses RSVP-TE or SR-TE LSPs with fixed bandwidth reservations. By "fixed", we mean a value that stays constant over time, unless the SMO communicates a change in slice bandwidth requirements, due to the creation or modification of a slice. Note that the "reservations" would be in the mind of the transport controller - it is not necessary (or indeed possible for SR-TE) to reserve bandwidth at the network layer. The bandwidth requirement acts as a constraint whenever the controller (re)computes the path of an LSP. There could be a single mesh of LSPs between endpoints that carry all of the traffic types, or there could be a small handful of meshes, for example one mesh for low-latency traffic that follows the minimum latency path and another mesh for the other traffic that follows the minimum IGP metric path, as described in . There would be a many-to-one mapping of slices to LSPs. The bandwidth requirement from DCi to DCj is the sum of the DCi-DCj demands of the individual slices. For example, if only Slice X and Slice Y are present, then the bandwidth requirement from DC1 to DC2 is 12 units (8 units for Slice X and 4 units for Slice Y). When the SMO requests a new slice, the transport controller, in its mind, increments the bandwidth requirement according to the requirements of the new slice. For example, in , suppose a new slice is instantiated that needs 0.8 Gbps from DC1 to DC2. The transport controller would increase its notion of the bandwidth requirement from DC1 to DC2 from 12 Gbps to 12.8 Gbps to accommodate the additional expected traffic. In the example, each DC has two PEs facing it for reasons of resilience. The transport controller needs to determine how to map the DC1 to DC2 bandwidth requirement to bandwidth reservations of TE LSPs from DC1 to DC2. For example, if the routing configuration is arranged such that in the absence of any network failure, traffic from DC1 to DC2 always enters PE1A and goes to PE2A, the controller reserves 12.8 Gbps of bandwidth on the LSP from PE1A to PE2A. If, on the other hand, the routing configuration is arranged such that in the absence of any network failure, traffic from DC1 to DC2 always enters PE1A and is load-balanced across PE2A and PE2B, the controller reserves 6.4 Gbps of bandwidth on the LSP from PE1A to PE2A and 6.4 Gbps of bandwidth on the LSP from PE1A to PE2B. It might be tricky for the transport controller to be aware of all conditions that change the way traffic lands on the various PEs, and therefore know that it needs to change bandwidth reservations of LSPs accordingly. For example, there might be an internal failure within DC1 that causes traffic from DC1 to land on PE1B, rather than PE1A. The transport controller may not be aware of the failure and therefore may not know that it now needs to apply bandwidth reservations to LSPs from PE1B to PE2A/PE2B.

Scheme 3: TE LSPs without Bandwidth Reservation Like Scheme 2, Scheme 3 uses RSVP-TE or SR-TE LSPs. There could be a single mesh of LSPs between endpoints that carry all of the traffic types, or there could be a small handful of meshes, for example one mesh for low-latency traffic that follows the minimum latency path and another mesh for the other traffic that follows the minimum IGP metric path, as described in . There would be a many-to-one mapping of slices to LSPs. The difference between Scheme 2 and Scheme 3 is that Scheme 3 does not have fixed bandwidth reservations for the LSPs. Instead, actual measured data-plane traffic volumes are used to influence the placement of TE LSPs. One way of achieving this is to use distributed RSVP-TE with auto-bandwidth. Alternatively, the transport controller can use telemetry-driven automatic congestion avoidance. In this approach, when the actual traffic volume in the data plane on given link exceeds a threshold, the controller, knowing how much actual data plane traffic is currently travelling along each RSVP or SR-TE LSP, can tune the paths of one or more LSPs using the link such that they avoid that link. It would be undesirable to move a minimum-latency LSP rather than another type of LSP in order to ease the congestion, as the new path will typically have a higher latency, if the minimum-latency LSP is currently following the minimum-latency path. This can be avoided by designing the algorithms described in the previous paragraph such that they avoid moving minimum-latency LSPs unless there is no alternative.

IANA Considerations This document does not make any IANA request.

Security Considerations IETF Network Slices considerations are discussed in Section 6 of . TBC.

References Normative References Old Dog Consulting Juniper Networks Ciena NTT Futurewei Telefonica Microsoft Inc. This document describes network slicing in the context of networks built from IETF technologies. It defines the term "IETF Network Slice" and establishes the general principles of network slicing in the IETF context. The document discusses the general framework for requesting and operating IETF Network Slices, the characteristics of an IETF Network Slice, the necessary system components and interfaces, and how abstract requests can be mapped to more specific technologies. The document also discusses related considerations with monitoring and security. This document also provides definitions of related terms to enable consistent usage in other IETF documents that describe or use aspects of IETF Network Slices. In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements. RFC 2119 specifies common key words that may be used in protocol specifications. This document aims to reduce the ambiguity by clarifying that only UPPERCASE usage of the key words have the defined special meanings. This document describes a method by which a Service Provider may use an IP backbone to provide IP Virtual Private Networks (VPNs) for its customers. This method uses a "peer model", in which the customers' edge routers (CE routers) send their routes to the Service Provider's edge routers (PE routers); there is no "overlay" visible to the customer's routing algorithm, and CE routers at different sites do not peer with each other. Data packets are tunneled through the backbone, so that the core routers do not need to know the VPN routes. [STANDARDS-TRACK] This document defines a Two Rate Three Color Marker (trTCM), which can be used as a component in a Diffserv traffic conditioner. This memo provides information for the Internet community. This document describes a two-rate, three-color marker that has been in use for data services including Frame Relay services. This marker can be used for metering per-flow traffic in the emerging IP and L2 VPN services. The marker defined here is different from previously defined markers in the handling of the in-profile traffic. Furthermore, this marker doesn't impose peak-rate shaping requirements on customer edge (CE) devices. This memo provides information for the Internet community. This document defines a YANG data model that can be used to configure a Layer 2 provider-provisioned VPN service. It is up to a management system to take this as an input and generate specific configuration models to configure the different network elements to deliver the service. How this configuration of network elements is done is out of scope for this document. The YANG data model defined in this document includes support for point-to-point Virtual Private Wire Services (VPWSs) and multipoint Virtual Private LAN Services (VPLSs) that use Pseudowires signaled using the Label Distribution Protocol (LDP) and the Border Gateway Protocol (BGP) as described in RFCs 4761 and 6624. The YANG data model defined in this document conforms to the Network Management Datastore Architecture defined in RFC 8342. As a complement to the Layer 3 Virtual Private Network Service Model (L3SM), which is used for communication between customers and service providers, this document defines an L3VPN Network Model (L3NM) that can be used for the provisioning of Layer 3 Virtual Private Network (L3VPN) services within a service provider network. The model provides a network-centric view of L3VPN services. The L3NM is meant to be used by a network controller to derive the configuration information that will be sent to relevant network devices. The model can also facilitate communication between a service orchestrator and a network controller/orchestrator. This document defines an L2VPN Network Model (L2NM) that can be used to manage the provisioning of Layer 2 Virtual Private Network (L2VPN) services within a network (e.g., a service provider network). The L2NM complements the L2VPN Service Model (L2SM) by providing a network-centric view of the service that is internal to a service provider. The L2NM is particularly meant to be used by a network controller to derive the configuration information that will be sent to relevant network devices. Also, this document defines a YANG module to manage Ethernet segments and the initial versions of two IANA-maintained modules that include a set of identities of BGP Layer 2 encapsulation types and pseudowire types. Informative References ITU-T 3GPP 3GPP O-RAN Alliance GSMA Cisco Systems Inc. Juniper Networks Huawei Technologies Comcast Ericsson Ericsson ZTE Corporation ZTE Corporation IBM Corporation Telefonica Ciena Verizon Realizing network slices may require the Service Provider to have the ability to partition a physical network into multiple logical networks of varying sizes, structures, and functions so that each slice can be dedicated to specific services or customers. Multiple network slices can be realized on the same network while ensuring slice elasticity in terms of network resource allocation. This document describes a scalable solution to realize network slicing in IP/MPLS networks by supporting multiple services on top of a single physical network by relying on compliant domains and nodes to provide forwarding treatment (scheduling, drop policy, resource usage) on to packets that carry identifiers that indicate the slicing service that is to be applied to the packets. The use of cellular broadband for accessing the Internet and other data services via smartphones, tablets, and notebook/netbook computers has increased rapidly as a result of high-speed packet data networks such as HSPA, HSPA+, and now Long-Term Evolution (LTE) being deployed. Operators that have deployed networks based on 3rd Generation Partnership Project (3GPP) network architectures are facing IPv4 address shortages at the Internet registries and are feeling pressure to migrate to IPv6. This document describes the support for IPv6 in 3GPP network architectures. This document is not an Internet Standards Track specification; it is published for informational purposes. Data models provide a programmatic approach to represent services and networks. Concretely, they can be used to derive configuration information for network and service components, and state information that will be monitored and tracked. Data models can be used during the service and network management life cycle (e.g., service instantiation, service provisioning, service optimization, service monitoring, service diagnosing, and service assurance). Data models are also instrumental in the automation of network management, and they can provide closed-loop control for adaptive and deterministic service creation, delivery, and maintenance. This document describes a framework for service and network management automation that takes advantage of YANG modeling technologies. This framework is drawn from a network operator perspective irrespective of the origin of a data model; thus, it can accommodate YANG modules that are developed outside the IETF. Huawei Technologies Huawei Technologies Ciena Cisco Systems, Inc China Mobile Cisco Systems, Inc This document defines a YANG model for the IETF Network Slice service. The model can be used by an IETF Network Slice customer to manage IETF Network Slices. Orange Juniper Telefonica Nokia Huawei Technologies This document specifies a YANG service data model for Attachment Circuits (ACs). This model can be used for the provisioning of ACs prior or during service provisioning (e.g., Network Slice Service). The document specifies also a module that updates other service and network modules with the required information to bind specific services to ACs that are created using the AC service model. Also, the document specifies a set of reusable groupings. Whether a service model reuses structures defined in the AC models or simply include an AC reference is a design choice of these service models. Relying upon the AC service model to manage ACs over which a service is delivered has the merit to decorrelate the management of a service vs. upgrade the AC components to reflect recent AC technologies or features. Huawei Technologies Telefonica Ciena Huawei Technologies Ribbon Communications Network Slicing is one of the core features in 5G, which provides different network service as independent logical networks. To provide 5G network slices service, an end-to-end network slice needs to consists of 3 major types of network segments: Radio Access Network (RAN), Mobile Core Network (CN) and Transport Network (TN). This document describes the application of IETF network slice in providing 5G end-to-end network slices, including the network slice identification mapping, network slice parameter mapping and 5G IETF Network Slice NBI. Cisco Cisco Telefonica As communication devices become more hybrid, smart devices include more media-rich communication applications, and the boundaries between telecommunication and other applications becomes less clear. Simultaneously, as the end-devices become more mobile, application traffic transits more often between enterprise networks, the Internet, and cellular telecommunication networks, sometimes using simultaneously more than one path and network type. In this context, it is crucial that quality of service be aligned between these different environments. However, this is not always the case by default, and cellular communication networks use a different QoS nomenclature from the Internet and enterprise networks. This document specifies a set of 3rd Generation Partnership Project (3GPP) Quality of Service (QoS) Class Identifiers (QCI) and 5G QoS Identifiers (5QI) to Differentiated Services Code Point (DSCP) mappings, to reconcile the marking recommendations offered by the 3GPP with the recommendations offered by the IETF, so as to maintain a consistent QoS treatment between cellular networks and the Internet. This mapping can be used by enterprises or implementers expecting traffic to flow through both types of network, and wishing to align the QoS treatment applied to one network under their control with the QoS treatment applied to the other network. Segment Routing can be applied to the IPv6 data plane using a new type of Routing Extension Header called the Segment Routing Header (SRH). This document describes the SRH and how it is used by nodes that are Segment Routing (SR) capable. This note discusses queuing and marking/dropping algorithms. While these algorithms may be implemented in a coupled manner, this note argues that specifications, measurements, and comparisons should decouple the different algorithms and their contributions to system behavior. Orange Telefonica Nokia Huawei Nokia This document defines a YANG data model for representing an abstract view of the provider network topology that contains the points from which its services can be attached (e.g., basic connectivity, VPN, network slices). Also, the model can be used to retrieve the points where the services are actually being delivered to customers (including peer networks). This document augments the 'ietf-network' data model by adding the concept of Service Attachment Points (SAPs). The SAPs are the network reference points to which network services, such as Layer 3 Virtual Private Network (L3VPN) or Layer 2 Virtual Private Network (L2VPN), can be attached. One or multiple services can be bound to the same SAP. Both User-Network Interface (UNI) and Network-to- Network Interface (NNI) are supported in the SAP data model. This document defines a YANG data model that can be used for communication between customers and network operators and to deliver a Layer 3 provider-provisioned VPN service. This document is limited to BGP PE-based VPNs as described in RFCs 4026, 4110, and 4364. This model is intended to be instantiated at the management system to deliver the overall service. It is not a configuration model to be used directly on network elements. This model provides an abstracted view of the Layer 3 IP VPN service configuration components. It will be up to the management system to take this model as input and use specific configuration models to configure the different network elements to deliver the service. How the configuration of network elements is done is out of scope for this document. This document obsoletes RFC 8049; it replaces the unimplementable module in that RFC with a new module with the same name that is not backward compatible. The changes are a series of small fixes to the YANG module and some clarifications to the text. Old Dog Consulting This document describes the principles of traffic engineering (TE) in the Internet. The document is intended to promote better understanding of the issues surrounding traffic engineering in IP networks and the networks that support IP networking, and to provide a common basis for the development of traffic engineering capabilities for the Internet. The principles, architectures, and methodologies for performance evaluation and performance optimization of operational networks are also discussed. This work was first published as RFC 3272 in May 2002. This document obsoletes RFC 3272 by making a complete update to bring the text in line with best current practices for Internet traffic engineering and to include references to the latest relevant work in the IETF. Cisco Systems, Inc. Juniper Networks, Inc. Cisco Systems, Inc. Cisco Systems, Inc Edward Jones IGP protocols historically compute the best paths over the network based on the IGP metric assigned to the links. Many network deployments use RSVP-TE or Segment Routing - Traffic Engineering (SR-TE) to steer traffic over a path that is computed using different metrics or constraints than the shortest IGP path. This document specifies a solution that allows IGPs themselves to compute constraint-based paths over the network. This document also specifies a way of using Segment Routing (SR) Prefix-SIDs and SRv6 locators to steer packets along the constraint-based paths.

Acronyms and Abbreviations 3GPP: 3rd Generation Partnership Project 5GC: 5G Core 5QI: 5G QoS Indicator A2A: Any-to-Any AC: Attachment Circuit AMF: Access and Mobility Management Function AUSF: Authentication Server Function BBU: Baseband Unit BH: Backhaul BS: Base Station CE: Customer Edge CIR: Committed Information Rate CN: Core Network CoS: Class of Service CP: Control Plane CSP: Communication Service Provider CU: Centralized Unit CU-CP: Centralized Unit Control Plane CU-UP: Centralized Unit User Plane DC: Data Center DDoS: Distributed Denial of Services DN: Data Network DSCP: Differentiated Services Code Point DU: Distributed Unit eCPRI: enhanced Common Public Radio Interface FH: Fronthaul FIB: Forwarding Information Base GPRS: Generic Packet Radio Service gNB: gNodeB GTP: GPRS Tunneling Protocol GTP-U: GPRS Tunneling Protocol User plane HW: Hardware ID: Identifier IGP: Interior Gateway Protocol IP: Internet Protocol L2VPN: Layer 2 Virtual Private Network L3VPN: Layer 3 Virtual Private Network LSP: Label Switched Path MH: Midhaul MIoT: Massive Internet of Things MPLS: Multiprotocol Label Switching NF: Network Function NR: New Radio NRF: Network Function Repository NRP: Network Resource Partition NSC: Network Slice Controller NSS: Network Slice Subnet PE: Provider Edge PIR: Peak Information Rate PLMN: Public Land Mobile Network PSTN: Public Switched Telephony Network QoS: Quality of Service RAN: Radio Access Network RF: Radio Frequency RIB: Routing Information Base RSVP: Resource Reservation Protocol RU: Radio Unit SD: Slice Differentiator SDP: Service Demarcation Point SLA: Service Level Agreement SLO: Service Level Objective SMF: Session Management Function SMO: Service Management and Orchestration S-NSSAI: Single Network Slice Selection Assistance Information SST: Slice/Service Type SR: Segment Routing SRv6: Segment Routing version 6 TC: Traffic Class TE: Traffic Engineering TN: Transport Network TS: Technical Specification UDM: Unified Data Management UE: User Equipment UP: User Plane UPF: User Plane Function URLLC: Ultra Reliable Low Latency Communication VLAN: Virtual Local Area Network VNF: Virtual Network Function VPN: Virtual Private Network VRF: Virtual Routing and Forwarding VXLAN: Virtual Extensible Local Area Network

An Overview of 5G Networking This section provides a brief introduction to 5G mobile networking with a perspective on the Transport Network. This section does not intend to replace or define 3GPP architecure, it just provides a brief overview for readers that do not have a mobile background. For more comprehensive information, refer to .

Key Building Blocks defines the Network Functions (UPF, AMF, etc.) that compose the 5G System (5GS) Architecture together with related interfaces (e.g., N1, N2). This architecture has native Control and User Plane separation, and the Control Plane leverages a service- based architecture. outlines an example 5GS architecture with a subset of possible network functions and network interfaces.

5GS Architecture and Service-based Interfaces

Similar to previous versions of 3GPP mobile networks , a 5G mobile network is split into the following four major domains ():

• UE, MS, MN, and Mobile: The terms UE (User Equipment), MS (Mobile Station), MN (Mobile Node), and mobile refer to the devices that are hosts with the ability to obtain Internet connectivity via a 3GPP network. An MS is comprised of the Terminal Equipment (TE) and a Mobile Terminal (MT). The terms UE, MS, MN, and mobile are used interchangeably within this document.
• Radio Access Network (RAN): Provides wireless connectivity to the UE devices via radio. It is made up of the Antenna that transmits and receives signals to the UE and the Base Station that digitizes the signal and converts the RF data stream to IP packets.
• Core Network (CN): Controls the CP of the RAN and provides connectivity to the Data Network (e.g., the Internet or a private VPN). The Core Network hosts dozens of services such as authentication, phone registry, charging, access to PSTN and handover.
• Transport Network (TN): Provides connectivity between sites where 5G Network Functions are located. The TN may connect sites from the RAN to the Core Network, as well as sites within the RAN or within the CN. This connectivity is achieved using IP.

Building Blocks of 5G Architecture (A High-Level Representation)

Core Network (CN) The 5G Core Network (5GC) is made up of a set of NFs which fall into two main categories ():

• 5GC User Plane: The User Plane Function (UPF) is the interconnect point between the mobile infrastructure and the Data Network (DN). It interfaces with the RAN via the N3 interface by encapsulating/ decapsulating the User Plane Traffic in GTP Tunnels (aka GTP-U or Mobile User Plane).
• 5GC Control Plane: The 5G Control Plane is made up of a comprehensive set of Network Functions. An exhaustive list and description of these entities is out of the scope of this document. The following NFs and interfaces are worth mentioning, since their connectivity may rely on the Transport Network:
  • the AMF (Access and Mobility Function) connects with the RAN control plane over the N2 interface
  • the SMF controls the 5GC UPF via the N4 interface

5G Core Network (CN)

Radio Access Network (RAN) The RAN connects cellular wireless devices to a mobile Core Network. The RAN is made up of three components, which form the Radio Base Station:

• The Baseband Unit (BBU) provides the interface between the Core Network and the Radio Network. It connects to the Radio Unit and is responsible for the baseband signal processing to packet.
• The Radio Unit (RU) is located close to the Antenna and controlled by the BBU. It converts the Baseband signal received from the BBU to a Radio frequency signal.
• The Antenna converts the electric signal received from the RU to radio waves

The 5G RAN Base Station is called a gNodeB (gNB). It connects to the Core Network via the N3 (User Plane) and N2 (Control Plane) interfaces. The 5G RAN architecture supports RAN disaggregation in various ways. Notably, the BBU can be split into a DU (Distributed Unit) for digital signal processing and a CU (Centralized Unit) for RAN Layer 3 processing. Furthermore, the CU can be itself split into Control Plane (CU-CP) and User Plane (CU-UP). depicts a disaggregated RAN with NFs and interfaces.

RAN Disaggregation

Transport Network (TN) The 5G transport architecture defines three main segments for the Transport Network, which are commonly referred to as Fronthaul (FH), Midhaul (MH), and Backhaul (BH) :

• Fronthaul happens before the BBU processing. In 5G, this interface is based on eCPRI (Enhanced CPRI) with native Ethernet or IP encapsulation.
• Midhaul is optional: this segment is introduced in the BBU split presented in Appendix B.3, where Midhaul network refers to the DU- CU interconnection (i.e., F1 interface). At this level, all traffic is encapsulated in IP (signaling and user plane).
• Backhaul happens after BBU processing. Therefore, it maps to the interconnection between the RAN and the Core Network. All traffic is also encapsulated in IP.

illustrates the different segments of the Transport Network with the relevant Network Functions.

5G Transport Segments

It is worth mentioning that a given part of the transport network can carry several 5G transport segments concurrently, as outlined in . This is because different types of 5G network functions might be placed in the same location (e.g., the UPF from one slice might be placed in the same location as the CU-UP from another slice).

Concurrent 5G Transport Segments

Acknowledgments The authors would like to thank Adrian Farrel, Joel Halpern and Tarek Saad for their reviews of this document and for providing valuable feedback on it.

Contributors To be added later