| Internet-Draft | 5G Transport Slices | January 2023 |
| Szarkowicz, et al. | Expires 17 July 2023 | [Page] |
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.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 17 July 2023.¶
Copyright (c) 2023 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
[I-D.ietf-teas-ietf-network-slices] introduces the 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 [I-D.ietf-teas-ietf-network-slices].¶
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 reuses many building blocks currently commonly used in communication service provider networks.¶
The reader may refer to [I-D.ietf-teas-ns-ip-mpls] for more advanced realization models.¶
A brief 5G overview is provided in Appendix B for readers convenience. The reader may refer to [RFC6459] and [TS-23.501] for more details about 3GPP network architectures.¶
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 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
The document uses the terms defined in [I-D.ietf-teas-ietf-network-slices].¶
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 adjacent domains. An ETN can be be a Provider Edge (PE) or a managed Customer Equipment (CE).¶
An extended list of abbreviations used in this document are listed in Appendix A.¶
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:¶
An 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.¶
[TS-28.530] 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.¶
An objective of TN Slicing is to isolate, guarantee, and 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 has two main flavors: Hard and Soft slicing. Hard slicing provides dedicated network capacity to slices, while soft slicing provides shared network capacity with guarantees for each slice.¶
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).¶
TN slice realization for 5G slices may combine elements of hard slicing in one part of the transport network, with elements of soft slicing in other parts of the transport network. Such as design is deployment-specific.¶
An optimized 5G network slicing architecture should integrate TN Slicing, 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 [I-D.ietf-teas-ietf-network-slices].¶
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.¶
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.¶
The datapath between two NFs may be decomposed into two segments based upon involved Orchestration domains:¶
TN Segment:¶
The realization of this segment is driven by the IETF Network Slice Controller (NSC) and the Transport Network Orchestrator (TNO). Generally, a TN Segment provides connectivity between two sites.¶
Local Segment:¶
This segment connects NFs within a given site or connects an NF to a TN. The realization of this segment is directly or indirectly driven by the 5G Orchestration without any involvement of the TNO. Generally, the Local Segment is a datapath local to a site. This 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 may 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) [I-D.ietf-teas-ietf-network-slices] may or may not be enforced on the ETN itself.¶
Figure 1 is a representation of the end-to-end datapath between NFs including Segments and ETNs (in practice PE or a managed CE), where applicable.¶
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).¶
Figure 3 provides samples to illustrate the different realizations of Local and TN Segments, as well the SDPs:¶
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 [RFC8969], we assume that this coordination is based upon IETF YANG data models and APIs (more details in further sections).¶
Figure 4 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 ([I-D.ietf-teas-ietf-network-slice-nbi-yang]) or a Attachement Circuit Service Interface ([I-D.boro-opsawg-teas-attachment-circuit]).¶
Note that the allocation of these resources (e.g., VLAN-ID or IP resources) 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.¶
There are multiple options to map a 5G network slice to IETF Network Slices:¶
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 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).¶
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 Figure 7.¶
At the time of writing, Section 6.2 of [NG.113] 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.¶
[I-D.ietf-teas-ietf-network-slices] 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 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 Section 4¶
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 Figure 7) 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.¶
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.¶
The 5G control plane relies on 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, IP addresses or DSCP, as documented in [I-D.gcdrb-teas-5g-network-slice-application].¶
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 domain and extended RAN/Core domains. 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.¶
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.¶
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 is simply an allocation method to allocate IPv6 addresses to NF loopbacks, without redefining IPv6 semantic. Different IPv6 address allocation schemes following this concept MAY be used, with one example allocation showed in Figure 11. 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 that provides a very easy way of identifying the packet as belonging to given S-NSSAI at any place in the transport domain. This might could be used, for example, to slectivelky enable per S-NSSAI monitoring, or any other per S-NSSAI handling, if required.¶
In the example, 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 S-NSSAI information. The 96-bit part of the address could be further divided, based for example on geographical location, or DC identification. 128 bits is wide enough to design an IPv6 addressing scheme, which is most suitable for particular 5G deployment.¶
Figure 12 shows an example 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 subnet, 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}.¶
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 [RFC4364]) for interconnecting multiple service domains:¶
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 Section 4.1.¶
In this option, service instances for different IETF Network Slice Services are instantiated outside the TN domain. These instances could be instantiated either on the compute, hosting mobile network functions (Figure 13, left hand side), or within the cloud infrastructure itself (e.g., on the top of the rack or leaf switch within cloud IP fabric (Figure 13, right hand side)). Between the TN domain and the (extended) RAN/CN domain, packets are 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 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 Figure 13. Therefore, for any slice-specific per hop behavior at the TN domain edge, the PE must be able to determine which label represents which slice. In the BGP control plane, when exchanging service prefixes between (extended) RAN/CN domains and TN domain, each slice might be represented by a unique BGP community, so tracking label assignment to the slice is possible. For example, in Figure 13, 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 is 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 Figure 13. 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.¶
for further study¶
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:¶
(1) 5G QoS¶
At this layer QoS treatment is indicated by the 5QI (5G QoS
indicator), as defined in {{TS-23.501}}. A 5QI is an ID that is
used as a reference to 5G QoS characteristics (e.g., scheduling
weights, admission thresholds, queue management thresholds, link
layer protocol configuration, etc.) in the RAN domain. Given the
fact 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, to allow differentiated treatment 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 in {{I-D.henry-tsvwg-diffserv-to-qci}}. Each slice
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'.
¶
(2) TN QoS¶
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 the 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 SDP), 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 TC. This layer of QoS will be referred as 'TN QoS Class', or 'TN QoS' for short, in this document.¶
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 (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 [O-RAN.WG9.XPSAAS]).¶
Therefore, from 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 (F2-U interface) might be 5QI-unaware, while at the same time the TN segment for 5G backhaul (N3 interface) might follow the 5QI-aware 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 Figure 14.¶
When the IP traffic is handed over at the SDP from the extended RAN or extended Core domains 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 Figure 15 for MPLS encapsulation, and in Figure 16 for SRv6 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 [RFC8754] provides more flexibility for TN CoS design, especially in combination with soft policing with in-profile/ out-profile, as discussed in Section 5.1.1.¶
Edge resources are controlled in a very 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.¶
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 (Section 7) is required to ensure proper isolation between slices in 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 attack, and equipment failure) and generates traffic volumes above the contracted rates.¶
The slice rates can be characterized with following parameters [I-D.ietf-teas-ietf-network-slice-nbi-yang]:¶
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 be dropped during any congestion event at any place in the TN domain.¶
2r3c (two-rate three-color) rate limiter¶
This was initially defined in [RFC2698], and its improved version in [RFC4115]. In essence, the traffic is assigned to one of the 3 categories:¶
An inbound 2c3r meter implemented with [RFC4115], compared to [RFC2698], 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 mode for 5QI-unaware mode, 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 Figure 17.¶
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:¶
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 [RFC7806].¶
Figure 18 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.¶
In the 5QI-aware model, potentially a large number of 5Q QoS Classes (the architecture scales to thousands of 5Q slices) is mapped (multiplexed) to up to 8 TN QoS Classes used in transport transit equipment, as outlined in Figure 19.¶
Given the fact 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 IETF Network Slices - can be 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.¶
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 AC link.¶
In 5QI-aware 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.¶
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¶
In the first option, the slice admission control is executed with traffic class granularity, as outlined in Figure 20. 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.¶
The second model combines the advantages of 5QI-unaware model (per slice admission control) with the per traffic class admission control, as outlined in Figure 20. 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 is allowed to use 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 Figure 21.¶
Figure 22 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.¶
Transit resource control is much simpler than Edge resource control. As outlined in Figure 19, 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 DHCP in outer header, as depicted in Figure 15 and Figure 16) 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 non-premium (best-effort) classes. Capacity planning and management, as described in Section 6, ensures that enough capacity is available to fulfill all approved slice requests.¶
A network operator might define various groups of tunnels, where each tunnel group is created with specific optimization criteria and constraints. This document refers to such tunnel groups as 'transport planes'. For example, transport plane A might represent tunnels optimized for latency, transport plane B for high capacity, transport plane C might represent tunnels using only the "upper half" of the transport network, and transport plane D might represent tunnels using only the "lower half" of the transport network. Figure 23 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. Section 6 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.¶
Similar to the QoS mapping models discussed in Section 4, 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).¶
As discussed in Section 4.1, 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 mode, the entire slice is mapped to a single transport plane, as depicted in Figure 24.¶
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 Classes (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.¶
In 5QI-aware mode, 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 class, as depicted in Figure 25.¶
This section describes the information conveyed by the SMO to the transport controller with respect to slice bandwidth requirements.¶
Figure 26 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.¶
Let us consider 5G Slice X that uses some of the network functions in the three DCs. If the slice has latency requirements, the SMO will have taken those into account when deciding which network functions in which DC would participate in the slice. As a result of that placement decision, the three DCs shown are involved in 5G Slice X, rather than other DCs. In order to make this determination, the SMO needs information from the NSC about the 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.¶
Figure 27 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 Section 4, 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.¶
[I-D.ietf-teas-ietf-network-slice-nbi-yang] 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 [I-D.ietf-opsawg-sap].¶
Depending on the bandwidth model used in the network (Section 6.1), 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 [RFC8299] or the IETF YANG Data Model for L2VPN Service Delivery [RFC8466] could be used instead of [I-D.ietf-teas-ietf-network-slice-nbi-yang], 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 Section 5. The SMO can use [I-D.ietf-teas-ietf-network-slice-nbi-yang] to request low-latency transport for a given slice if required. However, [RFC8299] or [RFC8466] 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 used in [RFC9182] and [RFC9291].¶
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 [I-D.ietf-teas-rfc3272bis].¶
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 [I-D.ietf-lsr-flex-algo], 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 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 Section 5. 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 Figure 26, 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.¶
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 Section 5. 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.¶
This document does not make any IANA request.¶
IETF Network Slices considerations are discussed in Section 6 of [I-D.ietf-teas-ietf-network-slices].¶
TBC.¶
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¶
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 [TS-23.501].¶
[TS-23.501] 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. Figure 28 outlines an example 5GS architecture with a subset of possible network functions and network interfaces.¶
Similar to previous versions of 3GPP mobile networks [RFC6459], a 5G mobile network is split into the following four major domains (Figure 29):¶
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.¶
The 5G Core Network (5GC) is made up of a set of Network Functions (NFs) which fall into two main categories (Figure 30):¶
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 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 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).¶
Figure 31 depicts a disaggregated RAN with NFs and interfaces.¶
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) ([TR-GSTR-TN5G]):¶
Figure 32 illustrates the different segments of the Transport Network with the relevant Network Functions.¶
It is worth mentioning that a given part of the transport network can carry several 5G transport segments concurrently, as outlined in Figure 33. 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).¶
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.¶
To be added later¶