BGP Color-Aware Routing Problem Statement

This document explores the scope, use-cases and requirements for a BGP based routing solution to establish end-to-end intent-aware paths across a multi-domain service provider network environment. The targeted design outcome is to define the technology and protocol extensions that may be required in a manner that addresses the widest application. The problem that the document initially focuses on is the BGP-based delivery of an intent across several transport domains. To do this, it describes existing intent-aware routing solutions that are deployed and then extends the solution scope and architecture to BGP. The problem space is then widened to include any intent (including NFV chains and their location), any dataplane and the application of the intent-based routing to the Service/VPN routes. All of this is detailed in the rest of the document.

Color-Aware Routing (CAR) establishes routed paths that satisfy specific intent in a network. This section describes the basic concepts that define CAR and the protocols that currently support it. The figure below is used as reference.

Intent in routing may be any combination of the following behaviors: Topology path selection (e.g. minimize metric, avoid resource) NFV service insertion (e.g. service chain steering) Per-hop behavior (e.g. QoS for 5G slice) An intent-aware routed path may be within a single network domain or across multiple domains.

Color is a 32-bit numerical value that is associated with an intent, as defined in [I-D.ietf-spring-segment-routing-policy]

An Egress PE E2 colors a BGP service (e.g., VPN) route V/v to indicate the particular intent that E2 requests for the traffic bound to V/v. The color (C) is encoded as a BGP Color Extended community .

(E2, C) is a color-aware route to E2 which satisfies the intent associated with color C. Multiple technologies already provide color-aware paths in solutions that are widely deployed. SR Policy [I-D.ietf-spring-segment-routing-policy] IGP Flex-Algo [I-D.ietf-lsr-flex-algo] In the context of large-scale SR-MPLS networks, SR Policy is applicable to both intra-domain and inter-domain deployments; whereas IGP Flex-Algo is better suited to intra-domain scenarios.

An ingress PE E1 automatically steers V-destined packets onto a Color-Aware path bound to (E2, C). If several such paths exist, a preference scheme is used to select the best path: E.g. IGP Flex-Algo first, then SR Policy.

If E1 and E2 are in different domains, E1 may request an SR-PCE in its domain for a path to (E2, C). The SR-PCE (or a set of them) computes the end-to-end path and installs it at E1 as an SR Policy. The end-to-end color-aware path may seamlessly cross multiple domains.

An operator with an existing Seamless-MPLS/BGP-LU inter-domain deployment may prefer a BGP based extension as a more incremental approach There may be an expectation that BGP would support a larger scale Trust boundaries in an inter-domain deployment leads to a preference for a BGP peering based solution

BGP Color-Aware Routing (CAR) is a new BGP solution which signals intent-aware routes to reach a given destination (e.g., E2). (E2, C) represents a BGP hop-by-hop distributed route that builds an inter-domain color-aware path to E2 for color C.

As seen above, multiple technologies exist that provide color-aware routing in a network. A BGP based solution must be compliant with the existing principles that apply to them: Service routes MUST be colored using BGP Color Extended-Community to request intent V/v via E, colored with C Colored service routes MUST be automatically steered on an appropriate color-aware path V/v via E with C is steered via (E, C) (E, C) provided by any color-aware technology or protocol Color-aware routes MAY resolve recursively via other color-aware routes (E, C) via N recursively resolves via (N, C) Here is a brief example that illustrates these principles.

In the figure above, all the nodes are part of an inter-domain network under a single authority and with a consistent color-to-intent mapping: Color C1 is mapped to "low delay" Flex-Algo FA1 is mapped to "low delay" and hence to C1 Color C2 is mapped to "low delay and avoid resource R" Flex-Algo FA2 is mapped to "low delay and avoid resource R" and hence to C2 E1 receives two BGP colored service routes from E2: V/v with BGP Color Extended community C1 W/w with BGP Color Extended community C2 E1 has the following inter-domain color-aware paths: (E2, C1) provided by BGP CAR which recursively resolves via intra-domain color-aware paths: (N1, C1) provided by IGP FA1 in Domain1 (N2, C1) provided by SR Policy bound to color C1 in Domain2 (E2, C2) provided by SR Policy E1 automatically steers the received colored service routes as follows: V/v via (E2, C1) provided by BGP CAR W/w via (E2, C2) provided by SR Policy The example illustrates the benefits provided by leveraging the architectural principles: Seamless co-existence of multiple color-aware technologies, e.g., BGP CAR and SR Policy V/v is steered on BGP CAR color-aware path W/w is steered on SR Policy color-aware path Seamless and complementary interworking between different color-aware technologies V/v is steered on a BGP CAR color-aware path that is itself resolved within domain 2 onto an SR Policy bound to the color of V/v

A color domain represents a collection of one or more network (IGP/BGP) domains with a single, consistent color-to-intent mapping Color re-mapping may happen at color domain boundaries

Ingress PE E1 steers packets destined for a service (VPN) route V/v via BGP Color-Aware Route R/r to E2 Per-Destination Steering: Incoming packets on E1 match BGP service route V/v to be steered based on the destination IP address of the packets. Per-Flow Steering: Incoming packets on E1 match BGP service route V/v to be steered based on the combination of the destination IP address and additional elements in the packet header (i.e., IP flow). Such a packet lookup may recurse on a forwarding array where some of the entries are BGP color-aware routes to E2. A given flow is mapped to a specific entry in this array i.e. via a specific BGP color-aware route to E2.

The BGP CAR solution must support the following intents bound to a color: Minimization of a cost metric vs a latency metric Minimization of different metric types, static and dynamic Exclusion/Inclusion of SRLG and/or Link Affinity and/or minimum MTU/number of hops Bandwidth management In the inter-domain context, exclusion/inclusion of entire domains, and border routers Inclusion of one or several virtual network function chains Located in a regional domain and/or core domain, in a DC Localization of the virtual network function chains Some functions may be desired in the regional DC or vice versa Per-Destination and Per-Flow steering

The BGP CAR route may be a transport route or a service route (in this document, we use the term VPN instead of service for simplicity).

Transport Intent Intent-aware routing between PEs connected across multiple transit domains Set up BGP based end-to-end paths stitching intent-aware intra-domain segments The network diagram below illustrates the reference network topology used in this section for Transport CAR:

The following network design assumptions apply to the reference topology above, as an example: Independent ISIS/OSPF SR instance in each domain. eBGP peering link between ASBRs (121-211, 121-212, 122-211, 122-212, 231-321, 231-322, 232-321 and 232-322). Peering links have equal cost metric. Peering links have delay configured or measured as shown by "D". D=50 for cross peering links. VPN service is running from PE31 to PE11 via service RRs (S-RRn in figure). The following sections illustrate a few examples of intent use-cases applicable to transport routes.

In the reference topology of Each domain has Algo 0 and Flex Algo 128 Algo 0 is for minimum cost metric(cost optimized). Flex Algo 128 definition is for minimum delay (low latency). Cost Optimized Color C1 - Minimum cost intent. (Here, a BGP CAR route with Color C1 is being used, instead of BGP-LU.) On PE11, VPN routes colored with C1 are steered via (C1, PE31) BGP CAR route BGP CAR for C1 sets up path(s) between PEs for end-to-end minimum cost. (2) These paths traverse over intra-domain Algo 0 in each domain and account for the peering link cost between ASBRs. Example: PE11 learns (C1, PE31) CAR route via several equal paths: One such path is through FA0 to node 121, links 121-211, FA0 to 231, link 231-321, FA0 to PE31 Another such path is through FA0 to node 122, link 122-212, FA0 to 232, link 232-322, FA0 to PE31. Minimize latency Color C2 - Minimum latency intent. On PE11, VPN routes colored with C2 are steered via (C2, PE31) BGP CAR route. BGP CAR for C2 advertises paths between PEs for minimum end-to-end delay. (2) These paths traverse over intra-domain Flex Algo 128 in each domain and account for peering link delay between ASBRs. (3) Example: PE11 learns (C2, PE31) BGP CAR route and best path is through FA128 to node 122, link 122-212, FA128 to 232, link 232-322, FA128 to PE31.

Color C3 - Intent to Minimize cost metric and avoid purple links In the reference topology of Each domain has Flex Algo 129 and some links have purple affinity. Flex Algo 129 definition is set to minimum cost metric and avoid purple links (within domain). Peering cross links are colored purple by policy. On PE11, VPN routes colored with C3 are steered via (C3, PE31) BGP CAR route. BGP CAR for C3 sets up paths between PEs for minimum end-to-end cost and avoiding purple link affinity. These paths traverse over intra domain Flex Algo 129 in each domain and accounts for peering link cost between ASBR and avoiding purple links. Example: PE11 learns (C3, PE31) BGP CAR route via 2 paths. First path is through FA 129 to node 121, link 121-211, FA129 to 231, link 231-321, FA129 to PE31. Second path is through FA129 to node 122, link 122-212, FA129 to 232, link 232-322, FA129 to PE31.

Color C4 - Avoid sending selected traffic via Domain 3 VPN routes advertised from PEs with Color C4 BGP CAR for Color C4 should only set up paths between PE11 and PE41 that exclude Domain 3

Color intent C5 - Routing via min-cost paths C6 - Routing via a local NFV service chain situated at E11 C7 - Routing via a centrally located NFV service chain situated at E21 Forwarding of packets from PE11 towards PE31: (C5, PE31) mapped packets are sent via nodes 121, 231 to PE31 (C6, PE31) mapped packets are sent to E11 and then post-service chain, via 121, 231 to PE31 (C7, PE31) mapped packets are sent via 121 to E21 and then post-service chain, via 231 to PE31

VPN (Service layer) intent Extend the signaling of intent awareness end-to-end: CE site to CE site across provider networks Provide ability for a CE to select paths through specific PEs for a given intent Example-1: Certain intent in transport not available via specific PEs Example-2: Certain CE-PE connection does not support specific intent Example-3: Site access via certain CE does not support specific intent. For instance, link connecting a specific CE to a DC hosting loss-sensitive service may have better quality than a link from another CE Provide ability for a CE to send traffic indicating a specific intent (via suitable encapsulation) to the PE for optimal steering. Intent aware routing support for multiple service (VPN) interworking models Beyond options such as iBGP or Inter-AS Option C that inherently extend from PE to PE Inter-AS Option A Inter-AS Option B GW based interworking(L3VPN, EVPN) Interworking with existing L3VPN deployments, both PEs and CEs The network diagram below illustrates the reference network topology used in this section for VPN CAR.

The following network design assumptions apply to the reference topology above, as an example: Independent ISIS/OSPF SR instance in each AS. eBGP peering link between VPN ASBRs 121-211, 121-212, 122-211, 122-212. VPN service is running between PEs via service RRs in each AS to local ASBRs. Between ASBRs, its Option-B i.e. next hop self for VPN SAFI. CE1 is dual homed to PE11 and PE12. Similarly, CE2 is dual homed to PE21 and PE22. Peering links have equal cost metric Peering links have delay configured or measured as shown by "D". CE2 advertises prefix V/24 to CE1. It is advertised as RD:V/24 between PEs, including color-awareness The following sections illustrate a few examples of intent use-cases applicable to VPN (service) routes.

In the reference topology of Each AS has Flex Algo 0 and 128. Flex Algo 0 is for minimum cost metric(cost optimized). Flex Algo 128 definition is for minimum delay (low latency). Cost Optimized Color C1 - Minimum cost intent. On CE1, flows requiring cost optimized paths to V/24 are steered over (C1, V/24) route. BGP CAR for C1 sets up paths between CEs for minimum end-to-end cost. This advertisement needs BGP CAR between PE-CE for V/24 prefix and color C1 awareness. It also needs BGP VPN CAR between PEs and ASBRs for RD:V/24 prefix and color C1 awareness (C1, RD:V/24). Paths traverse over PE-CE links, intra-domain Flex Algo 0 in each AS and peering links between ASBRs, minimizing cost for VPN. Example: CE1 learns (C1, V/24) CAR route through several equal cost paths: One path is through link CE1-PE11, FA0 to 121, link 121-211, FA0 to PE21 and link PE21-CE2. Another such path is through CE1-PE12, FA0 to node 122, link 122-212, FA0 to PE22, link PE22-CE2. Minimize latency Color C2 - Minimum latency intent On CE1, flows requiring low latency paths to prefix V/24 are steered over (C2, V/24) CAR route. BGP CAR for C2 sets up paths between CEs for minimum end-to-end delay. This advertisement needs BGP CAR between PE-CE for V/24 prefix and color C2 awareness. It also needs BGP VPN CAR between PEs and ASBR for RD:V/24 prefix and color C2 awareness (C2, RD:V/24). Paths traverse over intra-domain Flex Algo 128 in each AS and accounts for inter ASBR link delays and PE-CE link delays for the VPN. Example: CE1 learns (C2, V/24) CAR best route through link CE1-PE12, FA128 to 122, link 122-212, FA128 to PE22 and link PE22-CE2.

Color C3 - Intent to Minimize cost metric and avoid purple links In the reference topology of Each AS has Flex Algo 129 and some links have purple affinity. Flex Algo 129 definition is set to minimum cost metric and avoid purple links (within AS). ASBR cross links are colored purple by policy. Bottom PE-CE links are colored purple as well by policy On CE1, flows requiring minimum cost path avoiding purple links to V/24 are steered over (C3, V/24) BGP CAR route. BGP CAR for C3 setup paths between CEs for minimum end-to-end cost and avoiding purple link affinity. This advertisement needs BGP CAR between PE-CE for V/24 prefix and color C3 awareness It also needs BGP VPN CAR between PEs and ASBRs for RD:V/24 prefix and color C3 awareness (C3, RD:V/24). The path avoids purple PE-CE links, traverses over intra-domain Flex Algo 129 in each AS and avoids purple links between VPN ASBRs. Example: CE1 learns (C3, V/24) CAR route through link CE1-PE11, FA129 to 121, link 121-211, FA129 to PE21 and link PE21-CE2.

Color intent C1 - Routing via NFV service chain comprising of [S1, S2] attached to E11 and E12 C2 - Routing via NFV service [S3] attached to E21 CE1, CE2, CE3 are sites of VPN1. Prefix V1/24 colored with C1 from CE2, and advertised as RD:V1/24 with C1 by PE31 to PE11 via S-RR Prefix V2/24 colored with C2 from CE3, and advertised as RD:V2/24 with C2 by PE32 to PE11 via SS-RR From PE11: [V1/24, C1] mapped packets are sent via S1, S2 and then routed to PE31, CE2 [V2/24, C2] mapped packets are sent via S3 and then routed to PE32, CE3

The figure below shows a reference large-scale multi-domain network topology for targeted deployments. E1 and E2 are PEs; the other nodes are border routers between domains in different tiers of the network. A VPN route is advertised via service RRs (S-RR) between an egress PE (E2) and an ingress PE (E1). BGP must provide reachability from E1 to E2 based on various intent.

The solution must support the following : Co-existence, compatibility and interworking with currently deployed SR-PCE based multi-domain color-aware solution Support different multi-domain deployment designs Multiple IGP domains within a single AS (Seamless MPLS) Inter-connect at node level (ABR) Multiple BGP AS domains Inter-connect via peering links (ASBR) Support end-to-end path crossing transport domains with different technologies and encapsulations LDP-MPLS RSVP-TE-MPLS SR-MPLS SRv6 IPv4/IPv6 Support interworking between domains with different encapsulations (e.g, SR-MPLS and SRv6) Support multiple transport encapsulations within a domain for co-existence and migration Provide a BGP-based control-plane solution for the use-case illustrated in [RFC8604] together with deployment design guidelines for the leverage of anycast and binding SIDs.

Support for massive scaled transport network Number of Remote PE's: >= 300k Number of Colors C: >= 5 Scalable MPLS dataplane solution With one label per (C, Remote PE), the 1M MPLS dataplane does not work. A notion of hierarchy or segment list is required. E.g. the SR-PCE builds the end-to-end path as a list of segments such that no single node needs to support a data-plane scaling in the order of (Remote PE * C) The solution is thus not a direct extension of BGP-LU Additionally, PE and transit nodes (ABRs) may be devices with limited forwarding table space Devices may have constraints on packet processing (e.g., label operations, number of labels pushed) and performance Ability to abstract the topology from remote domains - for scale, stability and faster convergence Abstracting PE and/or ABR related state and network events Support for an Emulated-PULL model for the BGP signaling The SR-PCE solution natively supports a PULL model: when PE1 installs a VPN route V/v via (C, PE2), PE1 requests its serving SR-PCE to compute the SR Policy to (C, PE2). I.e. PE1 does not learn unneeded SR policies. BGP Signaling is natively a PUSH model. Emulated-PULL refers to the ability for a BGP CAR node PE1 to "subscribe" to (C, PE2) route such that only the related paths are signaled to PE1. The subscription and related filtering solution must apply to any BGP CAR node Transport CAR routes Ability for a node (PE/ABR/RR) to signal interest for routes of specific colors. PEs only learn routes that they need - remote VPN endpoints (PEs/ASBRs) or transit nodes (ABRs, ASBRs). ABRs also only learn and propagate routes they need locally in domain Service/VPN CAR routes Ability for a node (PE) to signal interest for a specific (Egress PE, Color) transport route CEs learn routes that they need - interested colors PEs learn routes that they need - interested VPNs, colors Automation of the subscription/filter route Similar to the SR-PCE solution, when an ingress PE1 installs VPN V/v via (C, PE2), PE1 originates its subscription/filter route for (C, PE2). Efficient propagation and processing of subscription/filter routes. Ability to perform aggregation and suppression of subscription/filter routes at nodes in the route propagation path to reduce explosion and churn in propagation of the filter routes themselves. The solution may be optional for networks that do not have the large scaling requirements

It is useful to analyze the multiple scaling requirements and specifically the data plane constraints in the context of a few common reference designs and use-cases. A couple of example scenarios are listed below for reference. Seamless-MPLS design, with IGP Flex-Algo in each domain

Inter-AS Option C VPN design, with IGP Flex-Algo in each domain, and eBGP peering between domains

The BGP CAR solution should provide high network availability for typical deployment topologies, with minimum loss of connectivity in different network failure scenarios. The network failure scenarios, applicable technologies and design options described in should be used as a reference. In the Seamless-MPLS reference topology in previous section: Failure of intra-domain links should limit loss of connectivity (LoC) to < 50ms. E.g., PE11 to a P node (not shown), 121 to a P node in Domain1 or Domain2) Failure of an intra-domain node (P node in any domain) should limit LoC to < 50ms Failure of an ABR node (e.g., 121, 231) should limit LoC to < 1sec Failure of a remote PE node (e.g., PE3) should limit LoC to < 1sec In the Inter-AS Option C VPN reference topology in previous section: Failure of intra-domain links should limit LoC to < 50ms. E.g., PE11 to a P node (not shown), 121 to a P node in Domain1 or Domain2) Failure of an intra-domain node (P node in any domain) should limit LoC to < 50ms Failure of an ASBR node (e.g., 121, 211) should limit LoC to < 1sec Failure of a remote PE node (e.g., PE3) should limit LoC to < 1sec Failure of an external link (e.g., 121-211) should limit LoC to < 1sec The solution should explore and describe additional techniques and design options that are applicable to further improve handling of the failure cases listed above.

Support signaling and distribution of different Color-Aware routes to reach a participating node, e.g., a PE. Intent should be indicated by the notion of a Color as defined in SR Policy Architecture. Signal different instances of a prefix distinguished by color Signal intent associated with a given route Support for a flexible NLRI definition to accommodate both efficiency of processing (e.g., packing) and future extensibility Avoid limitations associated with existing SAFI NLRI definitions. For example, 24-bit label. Support for validation of paths Reachability of next-hop in control plane Availability and programming of encapsulation in data plane Validation of intent Next-hop resolution for Color-Aware route Flexibility to use different intra-domain and inter-domain mechanisms - IGP-FA, SR-TE, RSVP-TE, IGP, BGP-LU etc. Recursive resolution over BGP Color-Aware routes Ability to carry end-to-end cumulative metric for a given color Support setting up an end-to-end Color-Aware path using a different/less preferred or best-effort paths in domains where a particular intent is not available Separation of transport and VPN service semantics. Allow for different route distribution planes for service vs transport routes. Support signaling of different transport encapsulations Support for signaling multiple encapsulations for co-existence and migration Generation of BGP Color-Aware routes sourced from IGP-FA, SR-TE policies and BGP-LU from a domain Support signaling across domains with different color mappings for a given intent.

Multicast service intent

Many people contributed to this document. The authors would especially like to thank Jim Uttaro for his guidance on the work and feedback on many aspects of the problem statement. We would also like to thank Daniel Voyer, Luay Jalil and Robert Raszuk for their review and valuable suggestions. We also express our appreciation to Bruno Decreane, Keyur Patel, Jim Guichard, Alex Bogdanov, Dirk Steinberg, Hannes Gredler and Xiaohu Hu for discussions on several topics that have helped provide input to the document. We also thank Huaimo Chen for his valuable review comments. The authors would like to thank Stephane Litkowski for his detailed review and for making valuable suggestions to improve the quality of the document. We would also like to thank Kamran Raza and Kris Michelson for their review and comments on the document and to Simon Spraggs, Jose Liste and Jiri Chaloupka for their early inputs on the problem statement.