Internet Engineering Task Force (IETF)                    A. Farrel, Ed.
Request for Comments: 7926                                      J. Drake
BCP: 206                                                Juniper Networks
Category: Best Current Practice                                 N. Bitar
ISSN: 2070-1721                                                    Nokia
                                                              G. Swallow
                                                     Cisco Systems, Inc.
                                                           D. Ceccarelli
                                                                Ericsson
                                                                X. Zhang
                                                                  Huawei
                                                               July 2016

      Problem Statement and Architecture for Information Exchange
           between Interconnected Traffic-Engineered Networks

Abstract

   In Traffic-Engineered (TE) systems, it is sometimes desirable to
   establish an end-to-end TE path with a set of constraints (such as
   bandwidth) across one or more networks from a source to a
   destination.  TE information is the data relating to nodes and TE
   links that is used in the process of selecting a TE path.  TE
   information is usually only available within a network.  We call such
   a zone of visibility of TE information a domain.  An example of a
   domain may be an IGP area or an Autonomous System.

   In order to determine the potential to establish a TE path through a
   series of connected networks, it is necessary to have available a
   certain amount of TE information about each network.  This need not
   be the full set of TE information available within each network but
   does need to express the potential of providing TE connectivity.
   This subset of TE information is called TE reachability information.

   This document sets out the problem statement for the exchange of TE
   information between interconnected TE networks in support of
   end-to-end TE path establishment and describes the best current
   practice architecture to meet this problem statement.  For reasons
   that are explained in this document, this work is limited to simple
   TE constraints and information that determine TE reachability.

Status of This Memo

   This memo documents an Internet Best Current Practice.

   This document is a product of the Internet Engineering Task Force
   (IETF).  It represents the consensus of the IETF community.  It has
   received public review and has been approved for publication by the
   Internet Engineering Steering Group (IESG).  Further information on
   BCPs is available in Section 2 of RFC 7841.

   Information about the current status of this document, any errata,
   and how to provide feedback on it may be obtained at
   http://www.rfc-editor.org/info/rfc7926.

Copyright Notice

   Copyright (c) 2016 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
   (http://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 Simplified BSD License text as described in Section 4.e of
   the Trust Legal Provisions and are provided without warranty as
   described in the Simplified BSD License.

Table of Contents

   1. Introduction ....................................................5
      1.1. Terminology ................................................6
           1.1.1. TE Paths and TE Connections .........................6
           1.1.2. TE Metrics and TE Attributes ........................6
           1.1.3. TE Reachability .....................................7
           1.1.4. Domain ..............................................7
           1.1.5. Server Network ......................................7
           1.1.6. Client Network ......................................7
           1.1.7. Aggregation .........................................7
           1.1.8. Abstraction .........................................8
           1.1.9. Abstract Link .......................................8
           1.1.10. Abstract Node or Virtual Node ......................8
           1.1.11. Abstraction Layer Network ..........................9
   2. Overview of Use Cases ...........................................9
      2.1. Peer Networks ..............................................9
      2.2. Client-Server Networks ....................................11
      2.3. Dual-Homing ...............................................15
      2.4. Requesting Connectivity ...................................15
           2.4.1. Discovering Server Network Information .............17
   3. Problem Statement ..............................................18
      3.1. Policy and Filters ........................................18
      3.2. Confidentiality ...........................................19
      3.3. Information Overload ......................................19
      3.4. Issues of Information Churn ...............................20
      3.5. Issues of Aggregation .....................................21
   4. Architecture ...................................................22
      4.1. TE Reachability ...........................................22
      4.2. Abstraction, Not Aggregation ..............................22
           4.2.1. Abstract Links .....................................23
           4.2.2. The Abstraction Layer Network ......................23
           4.2.3. Abstraction in Client-Server Networks ..............26
           4.2.4. Abstraction in Peer Networks .......................32
      4.3. Considerations for Dynamic Abstraction ....................34
      4.4. Requirements for Advertising Links and Nodes ..............35
      4.5. Addressing Considerations .................................36
   5. Building on Existing Protocols .................................36
      5.1. BGP-LS ....................................................37
      5.2. IGPs ......................................................37
      5.3. RSVP-TE ...................................................37
      5.4. Notes on a Solution .......................................37
   6. Application of the Architecture to Optical Domains and
      Networks .......................................................39
   7. Application of the Architecture to the User-Network Interface ..44
   8. Application of the Architecture to L3VPN Multi-AS Environments .46
   9. Scoping Future Work ............................................47
      9.1. Limiting Scope to Only Part of the Internet ...............47
      9.2. Working with "Related" Domains ............................47
      9.3. Not Finding Optimal Paths in All Situations ...............48
      9.4. Sanity and Scaling ........................................48
   10. Manageability Considerations ..................................48
      10.1. Managing the Abstraction Layer Network ...................49
      10.2. Managing Interactions of Abstraction Layer and
            Client Networks ..........................................49
      10.3. Managing Interactions of Abstraction Layer and
            Server Networks ..........................................50
   11. Security Considerations .......................................51
   12. Informative References ........................................52
   Appendix A. Existing Work .........................................58
      A.1. Per-Domain Path Computation ...............................58
      A.2. Crankback .................................................59
      A.3. Path Computation Element ..................................59
      A.4. GMPLS UNI and Overlay Networks ............................61
      A.5. Layer 1 VPN ...............................................62
      A.6. Policy and Link Advertisement .............................62
   Appendix B. Additional Features ...................................63
      B.1. Macro Shared Risk Link Groups .............................63
      B.2. Mutual Exclusivity ........................................64
   Acknowledgements ..................................................65
   Contributors ......................................................66
   Authors' Addresses ................................................67

1.  Introduction

   Traffic-Engineered (TE) systems such as MPLS-TE [RFC2702] and GMPLS
   [RFC3945] offer a way to establish paths through a network in a
   controlled way that reserves network resources on specified links.
   TE paths are computed by examining the Traffic Engineering Database
   (TED) and selecting a sequence of links and nodes that are capable of
   meeting the requirements of the path to be established.  The TED is
   constructed from information distributed by the Interior Gateway
   Protocol (IGP) running in the network -- for example, OSPF-TE
   [RFC3630] or ISIS-TE [RFC5305].

   It is sometimes desirable to establish an end-to-end TE path that
   crosses more than one network or administrative domain as described
   in [RFC4105] and [RFC4216].  In these cases, the availability of TE
   information is usually limited to within each network.  Such networks
   are often referred to as domains [RFC4726], and we adopt that
   definition in this document; viz.,

      For the purposes of this document, a domain is considered to be
      any collection of network elements within a common sphere of
      address management or path computational responsibility.  Examples
      of such domains include IGP areas and Autonomous Systems (ASes).

   In order to determine the potential to establish a TE path through a
   series of connected domains and to choose the appropriate domain
   connection points through which to route a path, it is necessary to
   have available a certain amount of TE information about each domain.
   This need not be the full set of TE information available within each
   domain but does need to express the potential of providing TE
   connectivity.  This subset of TE information is called TE
   reachability information.  The TE reachability information can be
   exchanged between domains based on the information gathered from the
   local routing protocol, filtered by configured policy, or statically
   configured.

   This document sets out the problem statement for the exchange of TE
   information between interconnected TE networks in support of end-to-
   end TE path establishment and describes the best current practice
   architecture to meet this problem statement.  The scope of this
   document is limited to the simple TE constraints and information
   (such as TE metrics, hop count, bandwidth, delay, shared risk)
   necessary to determine TE reachability: discussion of multiple
   additional constraints that might qualify the reachability can
   significantly complicate aggregation of information and the stability
   of the mechanism used to present potential connectivity, as is
   explained in the body of this document.

   Appendix A summarizes relevant existing work that is used to route TE
   paths across multiple domains.

1.1.  Terminology

   This section introduces some key terms that need to be understood to
   arrive at a common understanding of the problem space.  Some of the
   terms are defined in more detail in the sections that follow (in
   which case forward pointers are provided), and some terms are taken
   from definitions that already exist in other RFCs (in which case
   references are given, but no apology is made for repeating or
   summarizing the definitions here).

1.1.1.  TE Paths and TE Connections

   A TE connection is a Label Switched Path (LSP) through an MPLS-TE or
   GMPLS network that directs traffic along a particular path (the TE
   path) in order to provide a specific service such as bandwidth
   guarantee, separation of traffic, or resilience between a well-known
   pair of end points.

1.1.2.  TE Metrics and TE Attributes

   "TE metrics" and "TE attributes" are terms applied to parameters of
   links (and possibly nodes) in a network that is traversed by TE
   connections.  The TE metrics and TE attributes are used by path
   computation algorithms to select the TE paths that the TE connections
   traverse.  A TE metric is a quantifiable value (including measured
   characteristics) describing some property of a link or node that can
   be used as part of TE routing or planning, while a TE attribute is a
   wider term (i.e., including the concept of a TE metric) that refers
   to any property or characteristic of a link or node that can be used
   as part of TE routing or planning.  Thus, the delay introduced by
   transmission of a packet on a link is an example of a TE metric,
   while the geographic location of a router is an example of a more
   general attribute.

   Provisioning a TE connection through a network may result in dynamic
   changes to the TE metrics and TE attributes of the links and nodes in
   the network.

   These terms are also sometimes used to describe the end-to-end
   characteristics of a TE connection and can be derived according to a
   formula from the TE metrics and TE attributes of the links and nodes
   that the TE connection traverses.  Thus, for example, the end-to-end
   delay for a TE connection is usually considered to be the sum of the
   delay on each link that the connection traverses.

1.1.3.  TE Reachability

   In an IP network, reachability is the ability to deliver a packet to
   a specific address or prefix, i.e., the existence of an IP path to
   that address or prefix.  TE reachability is the ability to reach a
   specific address along a TE path.  More specifically, it is the
   ability to establish a TE connection in an MPLS-TE or GMPLS sense.
   Thus, we talk about TE reachability as the potential of providing TE
   connectivity.

   TE reachability may be unqualified (there is a TE path, but no
   information about available resources or other constraints is
   supplied); this is helpful especially in determining a path to a
   destination that lies in an unknown domain or that may be qualified
   by TE attributes and TE metrics such as hop count, available
   bandwidth, delay, and shared risk.

1.1.4.  Domain

   As defined in [RFC4726], a domain is any collection of network
   elements within a common sphere of address management or path
   computational responsibility.  Examples of such domains include IGP
   areas and ASes.

1.1.5.  Server Network

   A Server Network is a network that provides connectivity for another
   network (the Client Network) in a client-server relationship.  A
   Server Network is sometimes referred to as an underlay network.

1.1.6.  Client Network

   A Client Network is a network that uses the connectivity provided by
   a Server Network.  A Client Network is sometimes referred to as an
   overlay network.

1.1.7.  Aggregation

   The concept of aggregation is discussed in Section 3.5.  In
   aggregation, multiple network resources from a domain are represented
   outside the domain as a single entity.  Thus, multiple links and
   nodes forming a TE connection may be represented as a single link, or
   a collection of nodes and links (perhaps the whole domain) may be
   represented as a single node with its attachment links.

1.1.8.  Abstraction

   Section 4.2 introduces the concept of abstraction and distinguishes
   it from aggregation.  Abstraction may be viewed as "policy-based
   aggregation" where the policies are applied to overcome the issues
   with aggregation as identified in Section 3 of this document.

   Abstraction is the process of applying policy to the available TE
   information within a domain, to produce selective information that
   represents the potential ability to connect across the domain.  Thus,
   abstraction does not necessarily offer all possible connectivity
   options, but it presents a general view of potential connectivity
   according to the policies that determine how the domain's
   administrator wants to allow the domain resources to be used.

1.1.9.  Abstract Link

   An abstract link is the representation of the characteristics of a
   path between two nodes in a domain produced by abstraction.  The
   abstract link is advertised outside that domain as a TE link for use
   in signaling in other domains.  Thus, an abstract link represents the
   potential to connect between a pair of nodes.

   More details regarding abstract links are provided in Section 4.2.1.

1.1.10.  Abstract Node or Virtual Node

   An abstract node was defined in [RFC3209] as a group of nodes whose
   internal topology is opaque to an ingress node of the LSP.  More
   generally, an abstract node is the representation as a single node in
   a TE topology of some or all of the resources of one or more nodes
   and the links that connect them.  An abstract node may be advertised
   outside the domain as a TE node for use in path computation and
   signaling in other domains.

   The term "virtual node" has typically been applied to the aggregation
   of a domain (that is, a collection of nodes and links that operate as
   a single administrative entity for TE purposes) into a single entity
   that is treated as a node for the purposes of end-to-end traffic
   engineering.  Virtual nodes are often considered a way to present
   islands of single-vendor equipment in an optical network.

   Sections 3.5 and 4.2.2.1 provide more information about the uses and
   issues of abstract nodes and virtual nodes.

1.1.11.  Abstraction Layer Network

   The abstraction layer network is introduced in Section 4.2.2.  It may
   be seen as a brokerage-layer network between one or more server
   networks and one or more client networks.  The abstraction layer
   network is the collection of abstract links that provide potential
   connectivity across the server networks and on which path computation
   can be performed to determine edge-to-edge paths that provide
   connectivity as links in the client network.

   In the simplest case, the abstraction layer network is just a set of
   edge-to-edge connections (i.e., abstract links), but to make the use
   of server network resources more flexible, the abstract links might
   not all extend from edge to edge but might offer connectivity between
   server network nodes to form a more complex network.

2.  Overview of Use Cases

2.1.  Peer Networks

   The peer network use case can be most simply illustrated by the
   example in Figure 1.  A TE path is required between the source (Src)
   and destination (Dst), which are located in different domains.  There
   are two points of interconnection between the domains, and selecting
   the wrong point of interconnection can lead to a suboptimal path or
   even fail to make a path available.  Note that peer networks are
   assumed to have the same technology type -- that is, the same
   "switching capability", to use the term from GMPLS [RFC3945].

                    --------------      --------------
                   | Domain A     | x1 |     Domain Z |
                   |   -----      +----+       -----  |
                   |  | Src |     +----+      | Dst | |
                   |   -----      | x2 |       -----  |
                    --------------      --------------

                          Figure 1: Peer Networks

   For example, when Domain A attempts to select a path, it may
   determine that adequate bandwidth is available from Src through both
   interconnection points x1 and x2.  It may pick the path through x1
   for local policy reasons: perhaps the TE metric is smaller.  However,
   if there is no connectivity in Domain Z from x1 to Dst, the path
   cannot be established.  Techniques such as crankback may be used to
   alleviate this situation, but such techniques do not lead to rapid
   setup or guaranteed optimality.  Furthermore, RSVP signaling creates
   state in the network that is immediately removed by the crankback
   procedure.  Frequent events of this kind will impact scalability in a
   non-deterministic manner.  More details regarding crankback can be
   found in Appendix A.2.

   There are countless more complicated examples of the problem of peer
   networks.  Figure 2 shows the case where there is a simple mesh of
   domains.  Clearly, to find a TE path from Src to Dst, Domain A
   must not select a path leaving through interconnect x1, since
   Domain B has no connectivity to Domain Z.  Furthermore, in deciding
   whether to select interconnection x2 (through Domain C) or
   interconnection x3 through Domain D, Domain A must be sensitive to
   the TE connectivity available through each of Domains C and D,
   as well as the TE connectivity from each of interconnections x4 and
   x5 to Dst within Domain Z.  The problem may be further complicated
   when the source domain does not know in which domain the destination
   node is located, since the choice of a domain path clearly depends on
   the knowledge of the destination domain: this issue is obviously
   mitigated in IP networks by inter-domain routing [RFC4271].

   Of course, many network interconnection scenarios are going to be a
   combination of the situations expressed in these two examples.  There
   may be a mesh of domains, and the domains may have multiple points of
   interconnection.

                           --------------
                          |     Domain B |
                          |              |
                          |              |
                          /--------------
                         /
                        /x1
         --------------/                       --------------
        | Domain A     |                      |     Domain Z |
        |              |    --------------    |              |
        |  -----       | x2|     Domain C | x4|       -----  |
        | | Src |      +---+              +---+      | Dst | |
        |  -----       |   |              |   |       -----  |
        |              |    --------------    |              |
         --------------\                      /--------------
                        \x3                  /
                         \                  /
                          \                /x5
                           \--------------/
                           |     Domain D |
                           |              |
                           |              |
                            --------------

                     Figure 2: Peer Networks in a Mesh

2.2.  Client-Server Networks

   Two major classes of use case relate to the client-server
   relationship between networks.  These use cases have sometimes been
   referred to as overlay networks.  In both of these classes of
   use case, the client and server networks may have the same switching
   capability, or they may be built from nodes and links that have
   different technology types in the client and server networks.

   The first group of use cases, shown in Figure 3, occurs when domains
   belonging to one network are connected by a domain belonging to
   another network.  In this scenario, once connectivity is formed
   across the lower-layer network, the domains of the upper-layer
   network can be merged into a single domain by running IGP adjacencies
   and by treating the server-network-layer connectivity as links in the
   higher-layer network.  The TE relationship between the domains
   (higher and lower layers) in this case is reduced to determining what
   server network connectivity to establish, how to trigger it, how to
   route it in the server network, and what resources and capacity to
   assign within the server network layer.  As the demands in the
   higher-layer (client) network vary, the connectivity in the server
   network may need to be modified.  Section 2.4 explains in a little
   more detail how connectivity may be requested.

       ----------------                          ----------------
      | Client Network |                        | Client Network |
      |   Domain A     |                        |   Domain B     |
      |                |                        |                |
      |  -----         |                        |         -----  |
      | | Src |        |                        |        | Dst | |
      |  -----         |                        |         -----  |
      |                |                        |                |
       ----------------\                        /----------------
                        \x1                  x2/
                         \                    /
                          \                  /
                           \----------------/
                           | Server Network |
                           |     Domain     |
                           |                |
                            ----------------

                     Figure 3: Client-Server Networks

   The second class of use case relating to client-server networking is
   for Virtual Private Networks (VPNs).  In this case, as opposed to the
   former one, it is assumed that the client network has a different
   address space than that of the server network, where non-overlapping
   IP addresses between the client and the server networks cannot be
   guaranteed.  A simple example is shown in Figure 4.  The VPN sites
   comprise a set of domains that are interconnected over a core domain
   (i.e., the provider network) that is the server network in our model.

   Note that in the use cases shown in Figures 3 and 4 the client
   network domains may (and, in fact, probably do) operate as a single
   connected network.

          --------------                         --------------
         | Domain A     |                       |     Domain Z |
         | (VPN site)   |                       |   (VPN site) |
         |              |                       |              |
         |  -----       |                       |       -----  |
         | | Src |      |                       |      | Dst | |
         |  -----       |                       |       -----  |
         |              |                       |              |
          --------------\                       /--------------
                         \x1                 x2/
                          \                   /
                           \                 /
                            \---------------/
                            |  Core Domain  |
                            |               |
                            |               |
                            /---------------\
                           /                 \
                          /                   \
                         /x3                 x4\
          --------------/                       \--------------
         | Domain B     |                       |     Domain C |
         | (VPN site)   |                       |   (VPN site) |
         |              |                       |              |
         |              |                       |              |
          --------------                         --------------

                    Figure 4: A Virtual Private Network
   Both use cases in this section become "more interesting" when
   combined with the use case in Section 2.1 -- that is, when the
   connectivity between higher-layer domains or VPN sites is provided by
   a sequence or mesh of lower-layer domains.  Figure 5 shows how this
   might look in the case of a VPN.

        ------------                                   ------------
       | Domain A   |                                 |   Domain Z |
       | (VPN site) |                                 | (VPN site) |
       |  -----     |                                 |     -----  |
       | | Src |    |                                 |    | Dst | |
       |  -----     |                                 |     -----  |
       |            |                                 |            |
        ------------\                                 /------------
                     \x1                           x2/
                      \                             /
                       \                           /
                        \----------     ----------/
                        | Domain X |x5 | Domain Y |
                        | (core)   +---+ (core)   |
                        |          |   |          |
                        |          +---+          |
                        |          |x6 |          |
                        /----------     ----------\
                       /                           \
                      /                             \
                     /x3                           x4\
        ------------/                                 \------------
       | Domain B   |                                 |   Domain C |
       | (VPN site) |                                 | (VPN site) |
       |            |                                 |            |
        ------------                                   ------------

          Figure 5: A VPN Supported over Multiple Server Domains

2.3.  Dual-Homing

   A further complication may be added to the client-server relationship
   described in Section 2.2 by considering what happens when a client
   network domain is attached to more than one domain in the server
   network or has two points of attachment to a server network domain.
   Figure 6 shows an example of this for a VPN.

                               ------------
                              | Domain B   |
                              | (VPN site) |
       ------------           |  -----     |
      | Domain A   |          | | Src |    |
      | (VPN site) |          |  -----     |
      |            |          |            |
       ------------\           -+--------+-
                    \x1         |        |
                     \        x2|        |x3
                      \         |        |              ------------
                       \--------+-      -+--------     |   Domain C |
                       | Domain X | x8 | Domain Y | x4 | (VPN site) |
                       | (core)   +----+ (core)   +----+     -----  |
                       |          |    |          |    |    | Dst | |
                       |          +----+          +----+     -----  |
                       |          | x9 |          | x5 |            |
                       /----------      ----------\     ------------
                      /                            \
                     /                              \
                    /x6                            x7\
       ------------/                                  \------------
      | Domain D   |                                  |   Domain E |
      | (VPN site) |                                  | (VPN site) |
      |            |                                  |            |
       ------------                                    ------------

            Figure 6: Dual-Homing in a Virtual Private Network

2.4.  Requesting Connectivity

   The relationship between domains can be entirely under the control of
   management processes, dynamically triggered by the client network, or
   some hybrid of these cases.  In the management case, the server
   network may be asked to establish a set of LSPs to provide client
   network connectivity.  In the dynamic case, the client network may
   make a request to the server network exerting a range of controls
   over the paths selected in the server network.  This range extends
   from no control (i.e., a simple request for connectivity), through a
   set of constraints (latency, path protection, etc.), up to and
   including full control of the path and resources used in the server
   network (i.e., the use of explicit paths with label subobjects).

   There are various models by which a server network can be asked to
   set up the connections that support a service provided to the client
   network.  These requests may come from management systems, directly
   from the client network control plane, or through an intermediary
   broker such as the Virtual Network Topology Manager (VNTM) [RFC5623].

   The trigger that causes the request to the server network is also
   flexible.  It could be that the client network discovers a pressing
   need for server network resources (such as the desire to provision an
   end-to-end connection in the client network or severe congestion on a
   specific path), or it might be that a planning application has
   considered how best to optimize traffic in the client network or how
   to handle a predicted traffic demand.

   In all cases, the relationship between client and server networks is
   subject to policy so that server network resources are under the
   administrative control of the operator or the server network and are
   only used to support a client network in ways that the server network
   operator approves.

   As just noted, connectivity requests issued to a server network may
   include varying degrees of constraint upon the choice of path that
   the server network can implement.

   o  "Basic provisioning" is a simple request for connectivity.  The
      only constraints are the end points of the connection and the
      capacity (bandwidth) that the connection will support for the
      client network.  In the case of some server networks, even the
      bandwidth component of a basic provisioning request is superfluous
      because the server network has no facility to vary bandwidth and
      can offer connectivity only at a default capacity.

   o  "Basic provisioning with optimization" is a service request that
      indicates one or more metrics that the server network must
      optimize in its selection of a path.  Metrics may be hop count,
      path length, summed TE metric, jitter, delay, or any number of
      technology-specific constraints.

   o  "Basic provisioning with optimization and constraints" enhances
      the optimization process to apply absolute constraints to
      functions of the path metrics.  For example, a connection may be
      requested that optimizes for the shortest path but in any case
      requests that the end-to-end delay be less than a certain value.

      Equally, optimization may be expressed in terms of the impact on
      the network.  For example, a service may be requested in order to
      leave maximal flexibility to satisfy future service requests.

   o  "Fate diversity requests" ask the server network to provide a path
      that does not use any network resources (usually links and nodes)
      that share fate (i.e., can fail as the result of a single event)
      as the resources used by another connection.  This allows the
      client network to construct protection services over the server
      network -- for example, by establishing links that are known to be
      fate diverse.  The connections that have diverse paths need not
      share end points.

   o  "Provisioning with fate sharing" is the exact opposite of
      fate diversity.  In this case, two or more connections are
      requested to follow the same path in the server network.  This may
      be requested, for example, to create a bundled or aggregated link
      in the client network where each component of the client-layer
      composite link is required to have the same server network
      properties (metrics, delay, etc.) and the same failure
      characteristics.

   o  "Concurrent provisioning" enables the interrelated connection
      requests described in the previous two bullets to be enacted
      through a single, compound service request.

   o  "Service resilience" requests that the server network provide
      connectivity for which the server network takes responsibility to
      recover from faults.  The resilience may be achieved through the
      use of link-level protection, segment protection, end-to-end
      protection, or recovery mechanisms.

2.4.1.  Discovering Server Network Information

   Although the topology and resource availability information of a
   server network may be hidden from the client network, the service
   request interface may support features that report details about the
   services and potential services that the server network supports.

   o  Reporting of path details, service parameters, and issues such as
      path diversity of LSPs that support deployed services allows the
      client network to understand to what extent its requests were
      satisfied.  This is particularly important when the requests were
      made as "best effort".

   o  A server network may support requests of the form "If I were to
      ask you for this service, would you be able to provide it?" --
      that is, a service request that does everything except actually
      provision the service.

3.  Problem Statement

   The problem statement presented in this section is as much about the
   issues that may arise in any solution (and so have to be avoided) and
   the features that are desirable within a solution, as it is about the
   actual problem to be solved.

   The problem can be stated very simply and with reference to the use
   cases presented in the previous section.

      A mechanism is required that allows TE path computation in one
      domain to make informed choices about the TE capabilities and exit
      points from the domain when signaling an end-to-end TE path that
      will extend across multiple domains.

   Thus, the problem is one of information collection and presentation,
   not about signaling.  Indeed, the existing signaling mechanisms for
   TE LSP establishment are likely to prove adequate [RFC4726] with the
   possibility of minor extensions.  Similarly, TE information may
   currently be distributed in a domain by TE extensions to one of the
   two IGPs as described in OSPF-TE [RFC3630] and ISIS-TE [RFC5305], and
   TE information may be exported from a domain (for example,
   northbound) using link-state extensions to BGP [RFC7752].

   An interesting annex to the problem is how the path is made available
   for use.  For example, in the case of a client-server network, the
   path established in the server network needs to be made available as
   a TE link to provide connectivity in the client network.

3.1.  Policy and Filters

   A solution must be amenable to the application of policy and filters.
   That is, the operator of a domain that is sharing information with
   another domain must be able to apply controls to what information is
   shared.  Furthermore, the operator of a domain that has information
   shared with it must be able to apply policies and filters to the
   received information.

   Additionally, the path computation within a domain must be able to
   weight the information received from other domains according to local
   policy such that the resultant computed path meets the local
   operator's needs and policies rather than those of the operators of
   other domains.

3.2.  Confidentiality

   A feature of the policy described in Section 3.1 is that an operator
   of a domain may desire to keep confidential the details about its
   internal network topology and loading.  This information could be
   construed as commercially sensitive.

   Although it is possible that TE information exchange will take place
   only between parties that have significant trust, there are also use
   cases (such as the VPN supported over multiple server network domains
   described in Section 2.2) where information will be shared between
   domains that have a commercial relationship but a low level of trust.

   Thus, it must be possible for a domain to limit the shared
   information to only that which the computing domain needs to know,
   with the understanding that the less information that is made
   available the more likely it is that the result will be a less
   optimal path and/or more crankback events.

3.3.  Information Overload

   One reason that networks are partitioned into separate domains is to
   reduce the set of information that any one router has to handle.
   This also applies to the volume of information that routing protocols
   have to distribute.

   Over the years, routers have become more sophisticated, with greater
   processing capabilities and more storage; the control channels on
   which routing messages are exchanged have become higher capacity; and
   the routing protocols (and their implementations) have become more
   robust.  Thus, some of the arguments in favor of dividing a network
   into domains may have been reduced.  Conversely, however, the size of
   networks continues to grow dramatically with a consequent increase in
   the total amount of routing-related information available.
   Additionally, in this case, the problem space spans two or more
   networks.

   Any solution to the problems voiced in this document must be aware of
   the issues of information overload.  If the solution was to simply
   share all TE information between all domains in the network, the
   effect from the point of view of the information load would be to
   create one single flat network domain.  Thus, the solution must
   deliver enough information to make the computation practical (i.e.,
   to solve the problem) but not so much as to overload the receiving
   domain.  Furthermore, the solution cannot simply rely on the policies
   and filters described in Section 3.1 because such filters might not
   always be enabled.

3.4.  Issues of Information Churn

   As LSPs are set up and torn down, the available TE resources on links
   in the network change.  In order to reliably compute a TE path
   through a network, the computation point must have an up-to-date view
   of the available TE resources.  However, collecting this information
   may result in considerable load on the distribution protocol and
   churn in the stored information.  In order to deal with this problem
   even in a single domain, updates are sent at periodic intervals or
   whenever there is a significant change in resources, whichever
   happens first.

   Consider, for example, that a TE LSP may traverse ten links in a
   network.  When the LSP is set up or torn down, the resources
   available on each link will change, resulting in a new advertisement
   of the link's capabilities and capacity.  If the arrival rate of new
   LSPs is relatively fast, and the hold times relatively short, the
   network may be in a constant state of flux.  Note that the problem
   here is not limited to churn within a single domain, since the
   information shared between domains will also be changing.
   Furthermore, the information that one domain needs to share with
   another may change as the result of LSPs that are contained within or
   cross the first domain but that are of no direct relevance to the
   domain receiving the TE information.

   In packet networks, where the capacity of an LSP is often a small
   fraction of the resources available on any link, this issue is
   partially addressed by the advertising routers.  They can apply a
   threshold so that they do not bother to update the advertisement of
   available resources on a link if the change is less than a configured
   percentage of the total (or, alternatively, the remaining) resources.
   The updated information in that case will be disseminated based on an
   update interval rather than a resource change event.

   In non-packet networks, where link resources are physical switching
   resources (such as timeslots or wavelengths), the capacity of an LSP
   may more frequently be a significant percentage of the available link
   resources.  Furthermore, in some switching environments, it is
   necessary to achieve end-to-end resource continuity (such as using
   the same wavelength on the whole length of an LSP), so it is far more
   desirable to keep the TE information held at the computation points
   up to date.  Fortunately, non-packet networks tend to be quite a bit
   smaller than packet networks, the arrival rates of non-packet LSPs
   are much lower, and the hold times are considerably longer.  Thus,
   the information churn may be sustainable.

3.5.  Issues of Aggregation

   One possible solution to the issues raised in other subsections of
   this section is to aggregate the TE information shared between
   domains.  Two aggregation mechanisms are often considered:

   -  Virtual node model.  In this view, the domain is aggregated as if
      it was a single node (or router/switch).  Its links to other
      domains are presented as real TE links, but the model assumes that
      any LSP entering the virtual node through a link can be routed to
      leave the virtual node through any other link (although recent
      work on "limited cross-connect switches" may help with this
      problem [RFC7579]).

   -  Virtual link model.  In this model, the domain is reduced to a set
      of edge-to-edge TE links.  Thus, when computing a path for an LSP
      that crosses the domain, a computation point can see which domain
      entry points can be connected to which others, and with what TE
      attributes.

   Part of the nature of aggregation is that information is removed from
   the system.  This can cause inaccuracies and failed path computation.
   For example, in the virtual node model there might not actually be a
   TE path available between a pair of domain entry points, but the
   model lacks the sophistication to represent this "limited
   cross-connect capability" within the virtual node.  On the other
   hand, in the virtual link model it may prove very hard to aggregate
   multiple link characteristics: for example, there may be one path
   available with high bandwidth, and another with low delay, but this
   does not mean that the connectivity should be assumed or advertised
   as having both high bandwidth and low delay.

   The trick to this multidimensional problem, therefore, is to
   aggregate in a way that retains as much useful information as
   possible while removing the data that is not needed.  An important
   part of this trick is a clear understanding of what information is
   actually needed.

   It should also be noted in the context of Section 3.4 that changes in
   the information within a domain may have a bearing on what aggregated
   data is shared with another domain.  Thus, while the data shared is
   reduced, the aggregation algorithm (operating on the routers
   responsible for sharing information) may be heavily exercised.

4.  Architecture

4.1.  TE Reachability

   As described in Section 1.1, TE reachability is the ability to reach
   a specific address along a TE path.  The knowledge of TE reachability
   enables an end-to-end TE path to be computed.

   In a single network, TE reachability is derived from the Traffic
   Engineering Database (TED), which is the collection of all TE
   information about all TE links in the network.  The TED is usually
   built from the data exchanged by the IGP, although it can be
   supplemented by configuration and inventory details, especially in
   transport networks.

   In multi-network scenarios, TE reachability information can be
   described as "You can get from node X to node Y with the following TE
   attributes."  For transit cases, nodes X and Y will be edge nodes of
   the transit network, but it is also important to consider the
   information about the TE connectivity between an edge node and a
   specific destination node.  TE reachability may be qualified by TE
   attributes such as TE metrics, hop count, available bandwidth, delay,
   and shared risk.

   TE reachability information can be exchanged between networks so that
   nodes in one network can determine whether they can establish TE
   paths across or into another network.  Such exchanges are subject to
   a range of policies imposed by the advertiser (for security and
   administrative control) and by the receiver (for scalability and
   stability).

4.2.  Abstraction, Not Aggregation

   Aggregation is the process of synthesizing from available
   information.  Thus, the virtual node and virtual link models
   described in Section 3.5 rely on processing the information available
   within a network to produce the aggregate representations of links
   and nodes that are presented to the consumer.  As described in
   Section 3, dynamic aggregation is subject to a number of pitfalls.

   In order to distinguish the architecture described in this document
   from the previous work on aggregation, we use the term "abstraction"
   in this document.  The process of abstraction is one of applying
   policy to the available TE information within a domain, to produce
   selective information that represents the potential ability to
   connect across the domain.

   Abstraction does not offer all possible connectivity options (refer
   to Section 3.5) but does present a general view of potential
   connectivity.  Abstraction may have a dynamic element but is not
   intended to keep pace with the changes in TE attribute availability
   within the network.

   Thus, when relying on an abstraction to compute an end-to-end path,
   the process might not deliver a usable path.  That is, there is no
   actual guarantee that the abstractions are current or feasible.

   Although abstraction uses available TE information, it is subject to
   policy and management choices.  Thus, not all potential connectivity
   will be advertised to each client network.  The filters may depend on
   commercial relationships, the risk of disclosing confidential
   information, and concerns about what use is made of the connectivity
   that is offered.

4.2.1.  Abstract Links

   An abstract link is a measure of the potential to connect a pair of
   points with certain TE parameters.  That is, it is a path and its
   characteristics in the server network.  An abstract link represents
   the possibility of setting up an LSP, and LSPs may be set up over the
   abstract link.

   When looking at a network such as the network shown in Figure 7, the
   link from CN1 to CN4 may be an abstract link.  It is easy to
   advertise it as a link by abstracting the TE information in the
   server network, subject to policy.

   The path (i.e., the abstract link) represents the possibility of
   establishing an LSP from client network edge to client network edge
   across the server network.  There is not necessarily a one-to-one
   relationship between the abstract link and the LSP, because more than
   one LSP could be set up over the path.

   Since the client network nodes do not have visibility into the server
   network, they must rely on abstraction information delivered to them
   by the server network.  That is, the server network will report on
   the potential for connectivity.

4.2.2.  The Abstraction Layer Network

   Figure 7 introduces the abstraction layer network.  This construct
   separates the client network resources (nodes C1, C2, C3, and C4, and
   the corresponding links) and the server network resources (nodes CN1,
   CN2, CN3, and CN4, and the corresponding links).  Additionally, the
   architecture introduces an intermediary network layer called the
   abstraction layer.  The abstraction layer contains the client network
   edge nodes (C2 and C3), the server network edge nodes (CN1 and CN4),
   the client-server links (C2-CN1 and CN4-C3), and the abstract link
   (CN1-CN4).

   The client network is able to operate as normal.  Connectivity across
   the network can be either found or not found, based on links that
   appear in the client network TED.  If connectivity cannot be found,
   end-to-end LSPs cannot be set up.  This failure may be reported, but
   no dynamic action is taken by the client network.

   The server network also operates as normal.  LSPs across the server
   network between client network edges are set up in response to
   management commands or in response to signaling requests.

   The abstraction layer consists of the physical links between the two
   networks, and also the abstract links.  The abstract links are
   created by the server network according to local policy and represent
   the potential connectivity that could be created across the server
   network and that the server network is willing to make available for
   use by the client network.  Thus, in this example, the diameter of
   the abstraction layer network is only three hops, but an instance of
   an IGP could easily be run so that all nodes participating in the
   abstraction layer (and, in particular, the client network edge nodes)
   can see the TE connectivity in the layer.

    --    --                                  --    --
   |C1|--|C2|                                |C3|--|C4|   Client Network
    --   |  |                                |  |   --
         |  |                                |  |  . . . . . . . . . . .
         |  |                                |  |
         |  |                                |  |
         |  |    ---                  ---    |  |          Abstraction
         |  |---|CN1|================|CN4|---|  |         Layer Network
          --    |   |                |   |    --
                |   |                |   |   . . . . . . . . . . . . . .
                |   |                |   |
                |   |                |   |
                |   |   ---    ---   |   |                Server Network
                |   |--|CN2|--|CN3|--|   |
                 ---    ---    ---    ---

    Key
    --- Direct connection between two nodes
    === Abstract link

           Figure 7: Architecture for Abstraction Layer Network
   When the client network needs additional connectivity, it can make a
   request to the abstraction layer network.  For example, the operator
   of the client network may want to create a link from C2 to C3.  The
   abstraction layer can see the potential path C2-CN1-CN4-C3 and can
   set up an LSP C2-CN1-CN4-C3 across the server network and make the
   LSP available as a link in the client network.

   Sections 4.2.3 and 4.2.4 show how this model is used to satisfy the
   requirements for connectivity in client-server networks and in peer
   networks.

4.2.2.1.  Nodes in the Abstraction Layer Network

   Figure 7 shows a very simplified network diagram, and the reader
   would be forgiven for thinking that only client network edge nodes
   and server network edge nodes may appear in the abstraction layer
   network.  But this is not the case: other nodes from the server
   network may be present.  This allows the abstraction layer network to
   be more complex than a full mesh with access spokes.

   Thus, as shown in Figure 8, a transit node in the server network
   (here, the node is CN3) can be exposed as a node in the abstraction
   layer network with abstract links connecting it to other nodes in the
   abstraction layer network.  Of course, in the network shown in
   Figure 8, there is little if any value in exposing CN3, but if it had
   other abstract links to other nodes in the abstraction layer network
   and/or direct connections to client network nodes, then the resulting
   network would be richer.

    --    --                                     --    --     Client
   |C1|--|C2|                                   |C3|--|C4|    Network
    --   |  |                                   |  |   --
         |  |                                   |  |  . . . . . . . . .
         |  |                                   |  |
         |  |                                   |  |
         |  |   ---          ---          ---   |  |       Abstraction
         |  |--|CN1|========|CN3|========|CN5|--|  |      Layer Network
          --   |   |        |   |        |   |   --
               |   |        |   |        |   |  . . . . . . . . . . . .
               |   |        |   |        |   |
               |   |        |   |        |   |                 Server
               |   |   ---  |   |  ---   |   |                 Network
               |   |--|CN2|-|   |-|CN4|--|   |
                ---    ---   ---   ---    ---

         Figure 8: Abstraction Layer Network with Additional Node
   It should be noted that the nodes included in the abstraction layer
   network in this way are not "abstract nodes" in the sense of a
   virtual node described in Section 3.5.  Although it is the case that
   the policy point responsible for advertising server network resources
   into the abstraction layer network could choose to advertise abstract
   nodes in place of real physical nodes, it is believed that doing so
   would introduce significant complexity in terms of:

   -  Coordination between all of the external interfaces of the
      abstract node.

   -  Management of changes in the server network that lead to limited
      capabilities to reach (cross-connect) across the abstract node.
      There has been recent work on control-plane extensions to describe
      and operate devices (such as asymmetrical switches) that have
      limited cross-connect capabilities [RFC7579] [RFC7580].  These or
      similar extensions could be used to represent the same type of
      limitations, as they also apply in an abstract node.

4.2.3.  Abstraction in Client-Server Networks

   Figure 9 shows the basic architectural concepts for a client-server
   network.  The nodes in the client network are C1, C2, CE1, CE2, C3,
   and C4, where the client edge (CE) nodes are CE1 and CE2.  The core
   (server) network nodes are CN1, CN2, CN3, and CN4.  The interfaces
   CE1-CN1 and CE2-CN4 are the interfaces between the client and server
   networks.

   The technologies (switching capabilities) of the client and server
   networks may be the same or different.  If they are different, the
   client network traffic must be tunneled over a server network LSP.
   If they are the same, the client network LSP may be routed over the
   server network links, tunneled over a server network LSP, or
   constructed from the concatenation (stitching) of client network and
   server network LSP segments.

                      :                            :
      Client Network  :       Server Network       :  Client Network
                      :                            :
     --    --    ---                                  ---    --    --
    |C1|--|C2|--|CE1|................................|CE2|--|C3|--|C4|
     --    --   |   |    ---                  ---    |   |   --    --
                |   |===|CN1|================|CN4|===|   |
                |   |---|   |                |   |---|   |
                 ---    |   |   ---    ---   |   |    ---
                        |   |--|CN2|--|CN3|--|   |
                         ---    ---    ---    ---

     Key
     --- Direct connection between two nodes
     ... CE-to-CE LSP tunnel
     === Potential path across the server network (abstract link)

             Figure 9: Architecture for Client-Server Network

   The objective is to be able to support an end-to-end connection,
   C1-to-C4, in the client network.  This connection may support TE or
   normal IP forwarding.  To achieve this, CE1 is to be connected to CE2
   by a link in the client network.  This enables the client network to
   view itself as connected and to select an end-to-end path.

   As shown in the figure, three abstraction layer links are formed:
   CE1-CN1, CN1-CN2, and CN4-CE2.  A three-hop LSP is then established
   from CE1 to CE2 that can be presented as a link in the client
   network.

   The practicalities of how the CE1-CE2 LSP is carried across the
   server network LSP may depend on the switching and signaling options
   available in the server network.  The CE1-CE2 LSP may be tunneled
   down the server network LSP using the mechanisms of a hierarchical
   LSP [RFC4206], or the LSP segments CE1-CN1 and CN4-CE2 may be
   stitched to the server network LSP as described in [RFC5150].

   Section 4.2.2 has already introduced the concept of the abstraction
   layer network through an example of a simple layered network.  But it
   may be helpful to expand on the example using a slightly more complex
   network.

   Figure 10 shows a multi-layer network comprising client network nodes
   (labeled as Cn for n = 0 to 9) and server network nodes (labeled as
   Sn for n = 1 to 9).

                                              --     --
                                             |C3|---|C4|
                                             /--     --\
             --     --     --     --      --/           \--
            |C1|---|C2|---|S1|---|S2|----|S3|           |C5|
             --    /--     --\    --\     --\           /--
                  /           \--    \--     \--     --/    --
                 /            |S4|   |S5|----|S6|---|C6|---|C7|
                /             /--     --\    /--    /--     --
             --/    --     --/    --     \--/    --/
            |C8|---|C9|---|S7|---|S8|----|S9|---|C0|
             --     --     --     --      --     --

                 Figure 10: An Example Multi-Layer Network

   If the network in Figure 10 is operated as separate client and server
   networks, then the client network topology will appear as shown in
   Figure 11.  As can be clearly seen, the network is partitioned, and
   there is no way to set up an LSP from a node on the left-hand side
   (say C1) to a node on the right-hand side (say C7).

                                    --     --
                                   |C3|---|C4|
                                    --     --\
                    --     --                 \--
                   |C1|---|C2|                |C5|
                    --    /--                 /--
                         /                 --/    --
                        /                 |C6|---|C7|
                       /                  /--     --
                    --/    --          --/
                   |C8|---|C9|        |C0|
                    --     --          --

      Figure 11: Client Network Topology Showing Partitioned Network
   For reference, Figure 12 shows the corresponding server network
   topology.

                          --     --      --
                         |S1|---|S2|----|S3|
                          --\    --\     --\
                             \--    \--     \--
                             |S4|   |S5|----|S6|
                             /--     --\    /--
                          --/    --     \--/
                         |S7|---|S8|----|S9|
                          --     --      --

                    Figure 12: Server Network Topology

   Operating on the TED for the server network, a management entity or a
   software component may apply policy and consider what abstract links
   it might offer for use by the client network.  To do this, it
   obviously needs to be aware of the connections between the layers
   (there is no point in offering an abstract link S2-S8, since this
   could not be of any use in this example).

   In our example, after consideration of which LSPs could be set up in
   the server network, four abstract links are offered: S1-S3, S3-S6,
   S1-S9, and S7-S9.  These abstract links are shown as double lines on
   the resulting topology of the abstraction layer network in Figure 13.
   As can be seen, two of the links must share part of a path (S1-S9
   must share with either S1-S3 or S7-S9).  This could be achieved using
   distinct resources (for example, separate lambdas) where the paths
   are common, but it could also be done using resource sharing.

                                            --
                                           |C3|
                                           /--
                   --     --            --/
                  |C2|---|S1|==========|S3|
                   --     --\\          --\\
                             \\            \\
                              \\            \\--     --
                               \\            |S6|---|C6|
                                \\            --     --
                   --     --     \\--     --
                  |C9|---|S7|=====|S9|---|C0|
                   --     --       --     --

         Figure 13: Abstraction Layer Network with Abstract Links
   That would mean that when both paths S1-S3 and S7-S9 carry
   client-edge-to-client-edge LSPs, the resources on path S1-S9 are used
   and might be depleted to the point that the path is resource
   constrained and cannot be used.

   The separate IGP instance running in the abstraction layer network
   means that this topology is visible at the edge nodes (C2, C3, C6,
   C9, and C0) as well as at a Path Computation Element (PCE) if one is
   present.

   Now the client network is able to make requests to the abstraction
   layer network to provide connectivity.  In our example, it requests
   that C2 be connected to C3 and that C2 be connected to C0.  This
   results in several actions:

   1. The management component for the abstraction layer network asks
      its PCE to compute the paths necessary to make the connections.
      This yields C2-S1-S3-C3 and C2-S1-S9-C0.

   2. The management component for the abstraction layer network
      instructs C2 to start the signaling process for the new LSPs in
      the abstraction layer.

   3. C2 signals the LSPs for setup using the explicit routes
      C2-S1-S3-C3 and C2-S1-S9-C0.

   4. When the signaling messages reach S1 (in our example, both LSPs
      traverse S1), the server network may support them by a number of
      means, including establishing server network LSPs as tunnels,
      depending on the mismatch of technologies between the client and
      server networks.  For example, S1-S2-S3 and S1-S2-S5-S9 might be
      traversed via an LSP tunnel, using LSPs stitched together, or
      simply by routing the client network LSP through the server
      network.  If server network LSPs are needed, they can be signaled
      at this point.

   5. Once any server network LSPs that are needed have been
      established, S1 can continue to signal the client-edge-to-client-
      edge LSP across the abstraction layer, using the server network
      LSPs as either tunnels or stitching segments, or simply routing
      through the server network.

   6. Finally, once the client-edge-to-client-edge LSPs have been set
      up, the client network can be informed and can start to advertise
      the new TE links C2-C3 and C2-C0.  The resulting client network
      topology is shown in Figure 14.

                                      --   --
                                     |C3|-|C4|
                                     /--   --\
                                    /         \--
                          --     --/          |C5|
                         |C1|---|C2|          /--
                          --    /--\       --/    --
                               /    \     |C6|---|C7|
                              /      \    /--     --
                             /        \--/
                          --/    --   |C0|
                         |C8|---|C9|   --
                          --     --

         Figure 14: Connected Client Network with Additional Links

   7. Now the client network can compute an end-to-end path from C1
      to C7.

4.2.3.1.  A Server with Multiple Clients

   A single server network may support multiple client networks.  This
   is not an uncommon state of affairs -- for example, when the server
   network provides connectivity for multiple customers.

   In this case, the abstraction provided by the server network may vary
   considerably according to the policies and commercial relationships
   with each customer.  This variance would lead to a separate
   abstraction layer network maintained to support each client network.

   On the other hand, it may be that multiple client networks are
   subject to the same policies and the abstraction can be identical.
   In this case, a single abstraction layer network can support more
   than one client.

   The choices here are made as an operational issue by the server
   network.

4.2.3.2.  A Client with Multiple Servers

   A single client network may be supported by multiple server networks.
   The server networks may provide connectivity between different parts
   of the client network or may provide parallel (redundant)
   connectivity for the client network.

   In this case, the abstraction layer network should contain the
   abstract links from all server networks so that it can make suitable
   computations and create the correct TE links in the client network.

   That is, the relationship between the client network and the
   abstraction layer network should be one to one.

4.2.4.  Abstraction in Peer Networks

   Figure 15 shows the basic architectural concepts for connecting
   across peer networks.  Nodes from four networks are shown: A1 and A2
   come from one network; B1, B2, and B3 from another network; etc.  The
   interfaces between the networks (sometimes known as External Network
   Network Interfaces - ENNIs) are A2-B1, B3-C1, and C3-D1.

   The objective is to be able to support an end-to-end connection,
   A1-to-D2.  This connection is for TE connectivity.

   As shown in the figure, abstract links that span the transit networks
   are used to achieve the required connectivity.  These links form the
   key building blocks of the end-to-end connectivity.  An end-to-end
   LSP uses these links as part of its path.  If the stitching
   capabilities of the networks are homogeneous, then the end-to-end LSP
   may simply traverse the path defined by the abstract links across the
   various peer networks or may utilize stitching of LSP segments that
   each traverse a network along the path of an abstract link.  If the
   network switching technologies support or necessitate the use of LSP
   hierarchies, the end-to-end LSP may be tunneled across each network
   using hierarchical LSPs that each traverse a network along the path
   of an abstract link.

                 :                  :                  :
      Network A  :    Network B     :    Network C     :  Network D
                 :                  :                  :
       --    --     --    --    --     --    --    --     --    --
      |A1|--|A2|---|B1|--|B2|--|B3|---|C1|--|C2|--|C3|---|D1|--|D2|
       --    --    |  |   --   |  |   |  |   --   |  |    --    --
                   |  |========|  |   |  |========|  |
                    --          --     --          --

      Key
      --- Direct connection between two nodes
      === Abstract link across transit network

                    Figure 15: Architecture for Peering

   Peer networks exist in many situations in the Internet.  Packet
   networks may peer as IGP areas (levels) or as ASes.  Transport
   networks (such as optical networks) may peer to provide
   concatenations of optical paths through single-vendor environments
   (see Section 6).  Figure 16 shows a simple example of three peer
   networks (A, B, and C) each comprising a few nodes.

                 Network A    :     Network B      :   Network C
                              :                    :
           --     --      --  :  --     --     --  :  --     --
          |A1|---|A2|----|A3|---|B1|---|B2|---|B3|---|C1|---|C2|
           --     --\    /--  :  --    /--\    --  :  --     --
                     \--/     :       /    \       :
                     |A4|     :      /      \      :
                      --\     :     /        \     :
                   --    \--  :  --/          \--  :  --     --
                  |A5|---|A6|---|B4|----------|B6|---|C3|---|C4|
                   --     --  :  --            --  :  --     --
                              :                    :
                              :                    :

            Figure 16: A Network Comprising Three Peer Networks

   As discussed in Section 2, peered networks do not share visibility of
   their topologies or TE capabilities for scaling and confidentiality
   reasons.  That means, in our example, that computing a path from A1
   to C4 can be impossible without the aid of cooperating PCEs or some
   form of crankback.

   But it is possible to produce abstract links for reachability across
   transit peer networks and to create an abstraction layer network.
   That network can be enhanced with specific reachability information
   if a destination network is partitioned, as is the case with
   Network C in Figure 16.

   Suppose that Network B decides to offer three abstract links B1-B3,
   B4-B3, and B4-B6.  The abstraction layer network could then be
   constructed to look like the network in Figure 17.

                        --     --      --      --
                       |A3|---|B1|====|B3|----|C1|
                        --     --    //--      --
                                    //
                                   //
                                  //
                        --     --//     --     --
                       |A6|---|B4|=====|B6|---|C3|
                        --     --       --     --

     Figure 17: Abstraction Layer Network for the Peer Network Example

   Using a process similar to that described in Section 4.2.3, Network A
   can request connectivity to Network C, and abstract links can be
   advertised that connect the edges of the two networks and that can be
   used to carry LSPs that traverse both networks.  Furthermore, if
   Network C is partitioned, reachability information can be exchanged
   to allow Network A to select the correct abstract link, as shown in
   Figure 18.

                       Network A       :      Network C
                                       :
                 --     --      --     :     --       --
                |A1|---|A2|----|A3|=========|C1|.....|C2|
                 --     --\    /--     :     --       --
                           \--/        :
                           |A4|        :
                            --\        :
                         --    \--     :     --       --
                        |A5|---|A6|=========|C3|.....|C4|
                         --     --     :     --       --

      Figure 18: Tunnel Connections to Network C with TE Reachability

   Peer networking cases can be made far more complex by dual-homing
   between network peering nodes (for example, A3 might connect to B1
   and B4 in Figure 17) and by the networks themselves being arranged in
   a mesh (for example, A6 might connect to B4 and C1 in Figure 17).

   These additional complexities can be handled gracefully by the
   abstraction layer network model.

   Further examples of abstraction in peer networks can be found in
   Sections 6 and 8.

4.3.  Considerations for Dynamic Abstraction

   It is possible to consider a highly dynamic system where the server
   network adaptively suggests new abstract links into the abstraction
   layer, and where the abstraction layer proactively deploys new
   client-edge-to-client-edge LSPs to provide new links in the client
   network.  Such fluidity is, however, to be treated with caution.  In
   particular, in the case of client-server networks of differing
   technologies where hierarchical server network LSPs are used, this
   caution is needed for three reasons: there may be longer turn-up
   times for connections in some server networks; the server networks
   are likely to be sparsely connected; and expensive physical resources
   will only be deployed where there is believed to be a need for them.
   More significantly, the complex commercial, policy, and
   administrative relationships that may exist between client and server
   network operators mean that stability is more likely to be the
   desired operational practice.

   Thus, proposals for fully automated multi-layer networks based on
   this architecture may be regarded as forward-looking topics for
   research both in terms of network stability and with regard to
   economic impact.

   However, some elements of automation should not be discarded.  A
   server network may automatically apply policy to determine the best
   set of abstract links to offer and the most suitable way for the
   server network to support them.  And a client network may dynamically
   observe congestion, lack of connectivity, or predicted changes in
   traffic demand and may use this information to request additional
   links from the abstraction layer.  And, once policies have been
   configured, the whole system should be able to operate independently
   of operator control (which is not to say that the operator will not
   have the option of exerting control at every step in the process).

4.4.  Requirements for Advertising Links and Nodes

   The abstraction layer network is "just another network layer".  The
   links and nodes in the network need to be advertised along with their
   associated TE information (metrics, bandwidth, etc.) so that the
   topology is disseminated and so that routing decisions can be made.

   This requires a routing protocol running between the nodes in the
   abstraction layer network.  Note that this routing information
   exchange could be piggybacked on an existing routing protocol
   instance (subject to different switching capabilities applying to the
   links in the different networks, or to adequate address space
   separation) or use a new instance (or even a new protocol).  Clearly,
   the information exchanged is only information that has been created
   as part of the abstraction function according to policy.

   It should be noted that in many cases the abstract link represents
   the potential for connectivity across the server network but that
   no such connectivity exists.  In this case, we may ponder how the
   routing protocol in the abstraction layer will advertise topology
   information for, and over, a link that has no underlying
   connectivity.  In other words, there must be a communication channel
   between the abstraction layer nodes so that the routing protocol
   messages can flow.  The answer is that control-plane connectivity
   already exists in the server network and on the client-server edge
   links, and this can be used to carry the routing protocol messages
   for the abstraction layer network.  The same consideration applies to
   the advertisement, in the client network, of the potential
   connectivity that the abstraction layer network can provide, although
   it may be more normal to establish that connectivity before
   advertising a link in the client network.

4.5.  Addressing Considerations

   The network layers in this architecture should be able to operate
   with separate address spaces, and these may overlap without any
   technical issues.  That is, one address may mean one thing in the
   client network, yet the same address may have a different meaning in
   the abstraction layer network or the server network.  In other words,
   there is complete address separation between networks.

   However, this will require some care, both because human operators
   may well become confused, and because mapping between address spaces
   is needed at the interfaces between the network layers.  That mapping
   requires configuration so that, for example, when the server network
   announces an abstract link from A to B, the abstraction layer network
   must recognize that A and B are server network addresses and must map
   them to abstraction layer addresses (say P and Q) before including
   the link in its own topology.  And similarly, when the abstraction
   layer network informs the client network that a new link is available
   from S to T, it must map those addresses from its own address space
   to that of the client network.

   This form of address mapping will become particularly important in
   cases where one abstraction layer network is constructed from
   connectivity in multiple server networks, or where one abstraction
   layer network provides connectivity for multiple client networks.

5.  Building on Existing Protocols

   This section is non-normative and is not intended to prejudge a
   solutions framework or any applicability work.  It does, however,
   very briefly serve to note the existence of protocols that could be
   examined for applicability to serve in realizing the model described
   in this document.

   The general principle of protocol reuse is preferred over the
   invention of new protocols or additional protocol extensions, and it
   would be advantageous to make use of an existing protocol that is
   commonly implemented on network nodes and is currently deployed, or
   to use existing computational elements such as PCEs.  This has many
   benefits in network stability, time to deployment, and operator
   training.

   It is recognized, however, that existing protocols are unlikely to be
   immediately suitable to this problem space without some protocol
   extensions.  Extending protocols must be done with care and with
   consideration for the stability of existing deployments.  In extreme
   cases, a new protocol can be preferable to a messy hack of an
   existing protocol.

5.1.  BGP-LS

   BGP - Link State (BGP-LS) is a set of extensions to BGP, as described
   in [RFC7752].  Its purpose is to announce topology information from
   one network to a "northbound" consumer.  Application of BGP-LS to
   date has focused on a mechanism to build a TED for a PCE.  However,
   BGP's mechanisms would also serve well to advertise abstract links
   from a server network into the abstraction layer network or to
   advertise potential connectivity from the abstraction layer network
   to the client network.

5.2.  IGPs

   Both OSPF and IS-IS have been extended through a number of RFCs to
   advertise TE information.  Additionally, both protocols are capable
   of running in a multi-instance mode either as ships that pass in the
   night (i.e., completely separate instances using different address
   spaces) or as dual instances on the same address space.  This means
   that either OSPF or IS-IS could probably be used as the routing
   protocol in the abstraction layer network.

5.3.  RSVP-TE

   RSVP-TE signaling can be used to set up all TE LSPs demanded by this
   model, without the need for any protocol extensions.

   If necessary, LSP hierarchy [RFC4206] or LSP stitching [RFC5150] can
   be used to carry LSPs over the server network, again without needing
   any protocol extensions.

   Furthermore, the procedures in [RFC6107] allow the dynamic signaling
   of the purpose of any LSP that is established.  This means that when
   an LSP tunnel is set up, the two ends can coordinate into which
   routing protocol instance it should be advertised and can also agree
   on the addressing to be said to identify the link that will be
   created.

5.4.  Notes on a Solution

   This section is not intended to be prescriptive or dictate the
   protocol solutions that may be used to satisfy the architecture
   described in this document, but it does show how the existing
   protocols listed in the previous sections can be combined, with only
   minor modifications, to provide a solution.

   A server network can be operated using GMPLS routing and signaling
   protocols.  Using information gathered from the routing protocol, a
   TED can be constructed containing resource availability information
   and Shared Risk Link Group (SRLG) details.  A policy-based process
   can then determine which nodes and abstract links it wishes to
   advertise to form the abstraction layer network.

   The server network can now use BGP-LS to advertise a topology of
   links and nodes to form the abstraction layer network.  This
   information would most likely be advertised from a single point of
   control that made all of the abstraction decisions, but the function
   could be distributed to multiple server network edge nodes.  The
   information can be advertised by BGP-LS to multiple points within the
   abstraction layer (such as all client network edge nodes) or to a
   single controller.

   Multiple server networks may advertise information that is used to
   construct an abstraction layer network, and one server network may
   advertise different information in different instances of BGP-LS to
   form different abstraction layer networks.  Furthermore, in the case
   of one controller constructing multiple abstraction layer networks,
   BGP-LS uses the route target mechanism defined in [RFC4364] to
   distinguish the different applications (effectively abstraction layer
   network VPNs) of the exported information.

   Extensions may be made to BGP-LS to allow advertisement of Macro
   Shared Risk Link Groups (MSRLGs) (Appendix B.1) and the
   identification of mutually exclusive links (Appendix B.2), and to
   indicate whether the abstract link has been pre-established or not.
   Such extensions are valid options but do not form a core component of
   this architecture.

   The abstraction layer network may operate under central control or
   use a distributed control plane.  Since the links and nodes may be a
   mix of physical and abstract links, and since the nodes may have
   diverse cross-connect capabilities, it is most likely that a GMPLS
   routing protocol will be beneficial for collecting and correlating
   the routing information and for distributing updates.  No special
   additional features are needed beyond adding those extra parameters
   just described for BGP-LS, but it should be noted that the control
   plane of the abstraction layer network must run in an out-of-band
   control network because the data-bearing links might not yet have
   been established via connections in the server network.

   The abstraction layer network is also able to determine potential
   connectivity from client network edge to client network edge.  It
   will determine which client network links to create according to
   policy and subject to requests from the client network, and will take
   four steps:

   -  First, it will compute a path across the abstraction layer
      network.

   -  Then, if support of the abstract links requires the use of
      server network LSPs for tunneling or stitching and if those LSPs
      are not already established, it will ask the server layer to set
      them up.

   -  Then, it will signal the client-edge-to-client-edge LSP.

   -  Finally, the abstraction layer network will inform the client
      network of the existence of the new client network link.

   This last step can be achieved by either (1) coordination of the
   end points of the LSPs that span the abstraction layer (these points
   are client network edge nodes) using mechanisms such as those
   described in [RFC6107] or (2) using BGP-LS from a central controller.

   Once the client network edge nodes are aware of a new link, they will
   automatically advertise it using their routing protocol and it will
   become available for use by traffic in the client network.

   Sections 6, 7, and 8 discuss the applicability of this architecture
   to different network types and problem spaces, while Section 9 gives
   some advice about scoping future work.  Section 10 ("Manageability
   Considerations") is particularly relevant in the context of this
   section because it contains a discussion of the policies and
   mechanisms for indicating connectivity and link availability between
   network layers in this architecture.

6.  Application of the Architecture to Optical Domains and Networks

   Many optical networks are arranged as a set of small domains.  Each
   domain is a cluster of nodes, usually from the same equipment vendor
   and with the same properties.  The domain may be constructed as a
   mesh or a ring, or maybe as an interconnected set of rings.

   The network operator seeks to provide end-to-end connectivity across
   a network constructed from multiple domains, and so (of course) the
   domains are interconnected.  In a network under management control,
   such as through an Operations Support System (OSS), each domain is
   under the operational control of a Network Management System (NMS).

   In this way, an end-to-end path may be commissioned by the OSS
   instructing each NMS, and the NMSes setting up the path fragments
   across the domains.

   However, in a system that uses a control plane, there is a need for
   integration between the domains.

   Consider a simple domain, D1, as shown in Figure 19.  In this case,
   nodes A through F are arranged in a topological ring.  Suppose that
   there is a control plane in use in this domain and that OSPF is used
   as the TE routing protocol.

                            -----------------
                           |              D1 |
                           |      B---C      |
                           |     /     \     |
                           |    /       \    |
                           |   A         D   |
                           |    \       /    |
                           |     \     /     |
                           |      F---E      |
                           |                 |
                            -----------------

                    Figure 19: A Simple Optical Domain

   Now consider that the operator's network is built from a mesh of such
   domains, D1 through D7, as shown in Figure 20.  It is possible that
   these domains share a single, common instance of OSPF, in which case
   there is nothing further to say because that OSPF instance will
   distribute sufficient information to build a single TED spanning the
   whole network, and an end-to-end path can be computed.  A more likely
   scenario is that each domain is running its own OSPF instance.  In
   this case, each is able to handle the peculiarities (or, rather,
   advanced functions) of each vendor's equipment capabilities.

                  ------     ------     ------     ------
                 |      |   |      |   |      |   |      |
                 |  D1  |---|  D2  |---|  D3  |---|  D4  |
                 |      |   |      |   |      |   |      |
                  ------\    ------\    ------\    ------
                         \    |     \     |    \     |
                          \------    \------    \------
                          |      |   |      |   |      |
                          |  D5  |---|  D6  |---|  D7  |
                          |      |   |      |   |      |
                           ------     ------     ------

                Figure 20: A Mesh of Simple Optical Domains

   The question now is how to combine the multiple sets of information
   distributed by the different OSPF instances.  Three possible models
   suggest themselves, based on pre-existing routing practices.

   o  In the first model (the area-based model), each domain is treated
      as a separate OSPF area.  The end-to-end path will be specified to
      traverse multiple areas, and each area will be left to determine
      the path across the nodes in the area.  The feasibility of an
      end-to-end path (and, thus, the selection of the sequence of
      areas and their interconnections) can be derived using
      hierarchical PCEs.

      This approach, however, fits poorly with established use of the
      OSPF area: in this form of optical network, the interconnection
      points between domains are likely to be links, and the mesh of
      domains is far more interconnected and unstructured than we are
      used to seeing in the normal area-based routing paradigm.

      Furthermore, while hierarchical PCEs may be able to resolve this
      type of network, the effort involved may be considerable for more
      than a small collection of domains.

   o  Another approach (the AS-based model) treats each domain as a
      separate Autonomous System (AS).  The end-to-end path will be
      specified to traverse multiple ASes, and each AS will be left to
      determine the path across the nodes in the that AS.

      This model sits more comfortably with the established routing
      paradigm but causes a massive escalation of ASes in the global
      Internet.  It would, in practice, require that the operator use
      private AS numbers [RFC6996], of which there are plenty.

      Then, as suggested in the area-based model, hierarchical PCEs
      could be used to determine the feasibility of an end-to-end path
      and to derive the sequence of domains and the points of
      interconnection to use.  But just as in the area-based model, the
      scalability of this model using a hierarchical PCE must be
      questioned, given the sheer number of ASes and their
      interconnectivity.

      Furthermore, determining the mesh of domains (i.e., the inter-AS
      connections) conventionally requires the use of BGP as an
      inter-domain routing protocol.  However, not only is BGP not
      normally available on optical equipment, but this approach
      indicates that the TE properties of the inter-domain links would
      need to be distributed and updated using BGP -- something for
      which it is not well suited.

   o  The third approach (the Automatically Switched Optical Network
      (ASON) model) follows the architectural model set out by the ITU-T
      [G.8080] and uses the routing protocol extensions described in
      [RFC6827].  In this model, the concept of "levels" is introduced
      to OSPF.  Referring back to Figure 20, each OSPF instance running
      in a domain would be construed as a "lower-level" OSPF instance
      and would leak routes into a "higher-level" instance of the
      protocol that runs across the whole network.

      This approach handles the awkwardness of representing the domains
      as areas or ASes by simply considering them as domains running
      distinct instances of OSPF.  Routing advertisements flow "upward"
      from the domains to the high-level OSPF instance, giving it a full
      view of the whole network and allowing end-to-end paths to be
      computed.  Routing advertisements may also flow "downward" from
      the network-wide OSPF instance to any one domain so that it can
      see the connectivity of the whole network.

      Although architecturally satisfying, this model suffers from
      having to handle the different characteristics of different
      equipment vendors.  The advertisements coming from each low-level
      domain would be meaningless when distributed into the other
      domains, and the high-level domain would need to be kept
      up to date with the semantics of each new release of each vendor's
      equipment.  Additionally, the scaling issues associated with a
      well-meshed network of domains, each with many entry and exit
      points and each with network resources that are continually being
      updated, reduces to the same problem, as noted in the virtual link
      model.  Furthermore, in the event that the domains are under the
      control of different administrations, the domains would not want
      to distribute the details of their topologies and TE resources.

   Practically, this third model turns out to be very close to the
   methodology described in this document.  As noted in Section 6.1 of
   [RFC6827], there are policy rules that can be applied to define
   exactly what information is exported from or imported to a low-level
   OSPF instance.  [RFC6827] even notes that some forms of aggregation
   may be appropriate.  Thus, we can apply the following simplifications
   to the mechanisms defined in [RFC6827]:

   -  Zero information is imported to low-level domains.

   -  Low-level domains export only abstracted links as defined in this
      document and according to local abstraction policy, and with
      appropriate removal of vendor-specific information.

   -  There is no need to formally define routing levels within OSPF.

   -  Export of abstracted links from the domains to the network-wide
      routing instance (the abstraction routing layer) can take place
      through any mechanism, including BGP-LS or direct interaction
      between OSPF implementations.

   With these simplifications, it can be seen that the framework defined
   in this document can be constructed from the architecture discussed
   in [RFC6827], but without needing any of the protocol extensions
   defined in that document.  Thus, using the terminology and concepts
   already established, the problem may be solved as shown in Figure 21.
   The abstraction layer network is constructed from the inter-domain
   links, the domain border nodes, and the abstracted (cross-domain)
   links.

                                                       Abstraction Layer
      --             --    --             --    --             --
     |  |===========|  |--|  |===========|  |--|  |===========|  |
     |  |           |  |  |  |           |  |  |  |           |  |
   ..|  |...........|  |..|  |...........|  |..|  |...........|  |......
     |  |           |  |  |  |           |  |  |  |           |  |
     |  |  --   --  |  |  |  |  --   --  |  |  |  |  --   --  |  |
     |  |_|  |_|  |_|  |  |  |_|  |_|  |_|  |  |  |_|  |_|  |_|  |
     |  | |  | |  | |  |  |  | |  | |  | |  |  |  | |  | |  | |  |
      --   --   --   --    --   --   --   --    --   --   --   --
          Domain 1             Domain 2             Domain 3
     Key                                                   Optical Layer
       ...  Layer separation
       ---  Physical link
       ===  Abstract link

                Figure 21: The Optical Network Implemented
                   through the Abstraction Layer Network

7.  Application of the Architecture to the User-Network Interface

   The User-Network Interface (UNI) is an important architectural
   concept in many implementations and deployments of client-server
   networks, especially those where the client and server network have
   different technologies.  The UNI is described in [G.8080], and the
   GMPLS approach to the UNI is documented in [RFC4208].  Other
   GMPLS-related documents describe the application of GMPLS to specific
   UNI scenarios: for example, [RFC6005] describes how GMPLS can support
   a UNI that provides access to Ethernet services.

   Figure 1 of [RFC6005] is reproduced here as Figure 22.  It shows the
   Ethernet UNI reference model, and that figure can serve as an example
   for all similar UNIs.  In this case, the UNI is an interface between
   client network edge nodes and the server network.  It should be noted
   that neither the client network nor the server network need be an
   Ethernet switching network.

   There are three network layers in this model: the client network, the
   "Ethernet service network", and the server network.  The so-called
   Ethernet service network consists of links comprising the UNI links
   and the tunnels across the server network, and nodes comprising the
   client network edge nodes and various server network nodes.  That is,
   the Ethernet service network is equivalent to the abstraction layer
   network, with the UNI links being the physical links between the
   client and server networks, the client edge nodes taking the role of
   UNI Client-side (UNI-C) nodes, and the server edge nodes acting as
   the UNI Network-side (UNI-N) nodes.

        Client                                            Client
        Network       +----------+    +-----------+       Network
   -------------+     |          |    |           |     +-------------
         +----+ |     |  +-----+ |    |  +-----+  |     | +----+
   ------+    | |     |  |     | |    |  |     |  |     | |    +------
   ------+ EN +-+-----+--+ CN  +-+----+--+  CN +--+-----+-+ EN +------
         |    | |  +--+--|     +-+-+  |  |     +--+-----+-+    |
         +----+ |  |  |  +--+--+ | |  |  +--+--+  |     | +----+
                |  |  |     |    | |  |     |     |     |
   -------------+  |  |     |    | |  |     |     |     +-------------
                   |  |     |    | |  |     |     |
   -------------+  |  |     |    | |  |     |     |     +-------------
                |  |  |  +--+--+ | |  |  +--+--+  |     |
         +----+ |  |  |  |     | | +--+--+     |  |     | +----+
   ------+    +-+--+  |  | CN  +-+----+--+ CN  |  |     | |    +------
   ------+ EN +-+-----+--+     | |    |  |     +--+-----+-+ EN +------
         |    | |     |  +-----+ |    |  +-----+  |     | |    |
         +----+ |     |          |    |           |     | +----+
                |     +----------+    |-----------+     |
   -------------+           Server Networks             +-------------
        Client    UNI                               UNI   Client
        Network <----->                           <-----> Network
                          Scope of This Document

                        Legend:   EN  -  Client Network Edge Node
                                  CN  -  Server Network (Core) Node

                  Figure 22: Ethernet UNI Reference Model

   An issue that is often raised relates to how a dual-homed client
   network edge node (such as that shown at the bottom left-hand corner
   of Figure 22) can make determinations about how they connect across
   the UNI.  This can be particularly important when reachability across
   the server network is limited or when two diverse paths are desired
   (for example, to provide protection).  However, in the model
   described in this network, the edge node (the UNI-C node) is part of
   the abstraction layer network and can see sufficient topology
   information to make these decisions.  If the approach introduced in
   this document is used to model the UNI as described in this section,
   there is no need to enhance the signaling protocols at the GMPLS UNI
   nor to add routing exchanges at the UNI.

8.  Application of the Architecture to L3VPN Multi-AS Environments

   Serving Layer 3 VPNs (L3VPNs) across a multi-AS or multi-operator
   environment currently provides a significant planning challenge.
   Figure 6 shows the general case of the problem that needs to be
   solved.  This section shows how the abstraction layer network can
   address this problem.

   In the VPN architecture, the CE nodes are the client network edge
   nodes, and the PE nodes are the server network edge nodes.  The
   abstraction layer network is made up of the CE nodes, the CE-PE
   links, the PE nodes, and PE-PE tunnels that are the abstract links.

   In the multi-AS or multi-operator case, the abstraction layer network
   also includes the PEs (maybe Autonomous System Border Routers
   (ASBRs)) at the edges of the multiple server networks, and the PE-PE
   (maybe inter-AS) links.  This gives rise to the architecture shown in
   Figure 23.

   The policy for adding abstract links to the abstraction layer network
   will be driven substantially by the needs of the VPN.  Thus, when a
   new VPN site is added and the existing abstraction layer network
   cannot support the required connectivity, a new abstract link will be
   created out of the underlying network.

       ...........                                     .............
        VPN Site :                                     : VPN Site
        --   --  :                                     :  --   --
       |C1|-|CE| :                                     : |CE|-|C2|
        --  |  | :                                     : |  |  --
            |  | :                                     : |  |
            |  | :                                     : |  |
            |  | :                                     : |  |
            |  | :   --           --     --       --   : |  |
            |  |----|PE|=========|PE|---|PE|=====|PE|----|  |
             --  :  |  |         |  |   |  |     |  |  :  --
       ...........  |  |         |  |   |  |     |  |  ............
                    |  |         |  |   |  |     |  |
                    |  |         |  |   |  |     |  |
                    |  |         |  |   |  |     |  |
                    |  |  -   -  |  |   |  |  -  |  |
                    |  |-|P|-|P|-|  |   |  |-|P|-|  |
                     --   -   -   --     --   -   --

        Figure 23: The Abstraction Layer Network for a Multi-AS VPN
   It is important to note that each VPN instance can have a separate
   abstraction layer network.  This means that the server network
   resources can be partitioned and that traffic can be kept separate.

   This can be achieved even when VPN sites from different VPNs connect
   at the same PE.  Alternatively, multiple VPNs can share the same
   abstraction layer network if that is operationally preferable.

   Lastly, just as for the UNI discussed in Section 7, the issue of
   dual-homing of VPN sites is a function of the abstraction layer
   network and so is just a normal routing problem in that network.

9.  Scoping Future Work

   This section is provided to help guide the work on this problem.  The
   overarching view is that it is important to limit and focus the work
   on those things that are core and necessary to achieve the main
   function, and to not attempt to add unnecessary features or to
   over-complicate the architecture or the solution by attempting to
   address marginal use cases or corner cases.  This guidance is
   non-normative for this architecture description.

9.1.  Limiting Scope to Only Part of the Internet

   The scope of the use cases and problem statement in this document is
   limited to "some small set of interconnected domains."  In
   particular, it is not the objective of this work to turn the whole
   Internet into one large, interconnected TE network.

9.2.  Working with "Related" Domains

   Starting with this subsection, the intention of this work is to solve
   the TE interconnectivity for only "related" domains.  Such domains
   may be under common administrative operation (such as IGP areas
   within a single AS, or ASes belonging to a single operator) or may
   have a direct commercial arrangement for the sharing of TE
   information to provide specific services.  Thus, in both cases, there
   is a strong opportunity for the application of policy.

9.3.  Not Finding Optimal Paths in All Situations

   As has been well described in this document, abstraction necessarily
   involves compromises and removal of information.  That means that it
   is not possible to guarantee that an end-to-end path over
   interconnected TE domains follows the absolute optimal (by any
   measure of optimality) path.  This is taken as understood, and future
   work should not attempt to achieve such paths, which can only be
   found by a full examination of all network information across all
   connected networks.

9.4.  Sanity and Scaling

   All of the above points play into a final observation.  This work is
   intended to "bite off" a small problem for some relatively simple use
   cases as described in Section 2.  It is not intended that this work
   will be immediately (or even soon) extended to cover many large
   interconnected domains.  Obviously, the solution should, as far as
   possible, be designed to be extensible and scalable; however, it is
   also reasonable to make trade-offs in favor of utility and
   simplicity.

10.  Manageability Considerations

   Manageability should not be a significant additional burden.  Each
   layer in the network model can, and should, be managed independently.

   That is, each client network will run its own management systems and
   tools to manage the nodes and links in the client network: each
   client network link that uses an abstract link will still be
   available for management in the client network as any other link.

   Similarly, each server network will run its own management systems
   and tools to manage the nodes and links in that network just as
   normal.

   Three issues remain for consideration:

   -  How is the abstraction layer network managed?

   -  How is the interface between the client network and the
      abstraction layer network managed?

   -  How is the interface between the abstraction layer network and the
      server network managed?

10.1.  Managing the Abstraction Layer Network

   Management of the abstraction layer network differs from the client
   and server networks because not all of the links that are visible in
   the TED are real links.  That is, it is not possible to run
   Operations, Administration, and Maintenance (OAM) on the links that
   constitute the potential of a link.

   Other than that, however, the management of the abstraction layer
   network should be essentially the same.  Routing and signaling
   protocols can be run in the abstraction layer (using out-of-band
   channels for links that have not yet been established), and a
   centralized TED can be constructed and used to examine the
   availability and status of the links and nodes in the network.

   Note that different deployment models will place the "ownership" of
   the abstraction layer network differently.  In some cases, the
   abstraction layer network will be constructed by the operator of the
   server network and run by that operator as a service for one or more
   client networks.  In other cases, one or more server networks will
   present the potential of links to an abstraction layer network run by
   the operator of the client network.  And it is feasible that a
   business model could be built where a third-party operator manages
   the abstraction layer network, constructing it from the connectivity
   available in multiple server networks and facilitating connectivity
   for multiple client networks.

10.2.  Managing Interactions of Abstraction Layer and Client Networks

   The interaction between the client network and the abstraction layer
   network is a management task.  It might be automated (software
   driven), or it might require manual intervention.

   This is a two-way interaction:

   -  The client network can express the need for additional
      connectivity.  For example, the client network may try, and fail,
      to find a path across the client network and may request
      additional, specific connectivity (this is similar to the
      situation with the Virtual Network Topology Manager (VNTM)
      [RFC5623]).  Alternatively, a more proactive client network
      management system may monitor traffic demands (current and
      predicted), network usage, and network "hot spots" and may request
      changes in connectivity by both releasing unused links and
      requesting new links.

   -  The abstraction layer network can make links available to the
      client network or can withdraw them.  These actions can be in
      response to requests from the client network or can be driven by
      processes within the abstraction layer (perhaps reorganizing the
      use of server network resources).  In any case, the presentation
      of new links to the client network is heavily subject to policy,
      since this is both operationally key to the success of this
      architecture and the central plank of the commercial model
      described in this document.  Such policies belong to the operator
      of the abstraction layer network and are expected to be fully
      configurable.

      Once the abstraction layer network has decided to make a link
      available to the client network, it will install it at the link
      end points (which are nodes in the client network) such that it
      appears and can be advertised as a link in the client network.

   In all cases, it is important that the operators of both networks are
   able to track the requests and responses, and the operator of the
   client network should be able to see which links in that network are
   "real" physical links and which links are presented by the
   abstraction layer network.

10.3.  Managing Interactions of Abstraction Layer and Server Networks

   The interactions between the abstraction layer network and the server
   network are similar to those described in Section 10.2, but there is
   a difference in that the server network is more likely to offer up
   connectivity and the abstraction layer network is less likely to ask
   for it.

   That is, the server network will, according to policy that may
   include commercial relationships, offer the abstraction layer network
   a "set" of potential connectivity that the abstraction layer network
   can treat as links.  This server network policy will include:

   -  how much connectivity to offer

   -  what level of server network redundancy to include

   -  how to support the use of the abstract links
   This process of offering links from the server network may include a
   mechanism to indicate which links have been pre-established in the
   server network and can include other properties, such as:

   -  link-level protection [RFC4202]

   -  SRLGs and MSRLGs (see Appendix B.1)

   -  mutual exclusivity (see Appendix B.2)

   The abstraction layer network needs a mechanism to tell the server
   network which links it is using.  This mechanism could also include
   the ability to request additional connectivity from the server
   network, although it seems most likely that the server network will
   already have presented as much connectivity as it is physically
   capable of, subject to the constraints of policy.

   Finally, the server network will need to confirm the establishment of
   connectivity, withdraw links if they are no longer feasible, and
   report failures.

   Again, it is important that the operators of both networks are able
   to track the requests and responses, and the operator of the server
   network should be able to see which links are in use.

11.  Security Considerations

   Security of signaling and routing protocols is usually administered
   and achieved within the boundaries of a domain.  Thus, and for
   example, a domain with a GMPLS control plane [RFC3945] would apply
   the security mechanisms and considerations that are appropriate to
   GMPLS [RFC5920].  Furthermore, domain-based security relies strongly
   on ensuring that control-plane messages are not allowed to enter the
   domain from outside.

   In this context, additional security considerations arising from this
   document relate to the exchange of control-plane information between
   domains.  Messages are passed between domains using control-plane
   protocols operating between peers that have predictable relationships
   (for example, UNI-C to UNI-N, between BGP-LS speakers, or between
   peer domains).  Thus, the security that needs to be given additional
   attention for inter-domain TE concentrates on authentication of
   peers; assertion that messages have not been tampered with; and, to a
   lesser extent, protecting the content of the messages from
   inspection, since that might give away sensitive information about
   the networks.  The protocols described in Appendix A, which are
   likely to provide the foundation for solutions to this architecture,
   already include such protection and also can be run over protected
   transports such as IPsec [RFC6071], Transport Layer Security (TLS)
   [RFC5246], and the TCP Authentication Option (TCP-AO) [RFC5925].

   It is worth noting that the control plane of the abstraction layer
   network is likely to be out of band.  That is, control-plane messages
   will be exchanged over network links that are not the links to which
   they apply.  This models the facilities of GMPLS (but not of
   MPLS-TE), and the security mechanisms can be applied to the protocols
   operating in the out-of-band network.

12.  Informative References

   [G.8080]   International Telecommunication Union, "Architecture for
              the automatically switched optical network", ITU-T
              Recommendation G.8080/Y.1304, February 2012,
              <https://www.itu.int/rec/T-REC-G.8080-201202-I/en>.

   [GMPLS-ENNI]
              Bryskin, I., Ed., Doonan, W., Beeram, V., Ed., Drake, J.,
              Ed., Grammel, G., Paul, M., Kunze, R., Armbruster, F.,
              Margaria, C., Gonzalez de Dios, O., and D. Ceccarelli,
              "Generalized Multiprotocol Label Switching (GMPLS)
              External Network Network Interface (E-NNI): Virtual Link
              Enhancements for the Overlay Model", Work in Progress,
              draft-beeram-ccamp-gmpls-enni-03, September 2013.

   [RFC2702]  Awduche, D., Malcolm, J., Agogbua, J., O'Dell, M., and J.
              McManus, "Requirements for Traffic Engineering Over MPLS",
              RFC 2702, DOI 10.17487/RFC2702, September 1999,
              <http://www.rfc-editor.org/info/rfc2702>.

   [RFC3209]  Awduche, D., Berger, L., Gan, D., Li, T., Srinivasan, V.,
              and G. Swallow, "RSVP-TE: Extensions to RSVP for LSP
              Tunnels", RFC 3209, DOI 10.17487/RFC3209, December 2001,
              <http://www.rfc-editor.org/info/rfc3209>.

   [RFC3473]  Berger, L., Ed., "Generalized Multi-Protocol Label
              Switching (GMPLS) Signaling Resource ReserVation
              Protocol-Traffic Engineering (RSVP-TE) Extensions",
              RFC 3473, DOI 10.17487/RFC3473, January 2003,
              <http://www.rfc-editor.org/info/rfc3473>.

   [RFC3630]  Katz, D., Kompella, K., and D. Yeung, "Traffic Engineering
              (TE) Extensions to OSPF Version 2", RFC 3630,
              DOI 10.17487/RFC3630, September 2003,
              <http://www.rfc-editor.org/info/rfc3630>.

   [RFC3945]  Mannie, E., Ed., "Generalized Multi-Protocol Label
              Switching (GMPLS) Architecture", RFC 3945,
              DOI 10.17487/RFC3945, October 2004,
              <http://www.rfc-editor.org/info/rfc3945>.

   [RFC4105]  Le Roux, J.-L., Ed., Vasseur, J.-P., Ed., and J. Boyle,
              Ed., "Requirements for Inter-Area MPLS Traffic
              Engineering", RFC 4105, DOI 10.17487/RFC4105, June 2005,
              <http://www.rfc-editor.org/info/rfc4105>.

   [RFC4202]  Kompella, K., Ed., and Y. Rekhter, Ed., "Routing
              Extensions in Support of Generalized Multi-Protocol Label
              Switching (GMPLS)", RFC 4202, DOI 10.17487/RFC4202,
              October 2005, <http://www.rfc-editor.org/info/rfc4202>.

   [RFC4206]  Kompella, K. and Y. Rekhter, "Label Switched Paths (LSP)
              Hierarchy with Generalized Multi-Protocol Label Switching
              (GMPLS) Traffic Engineering (TE)", RFC 4206,
              DOI 10.17487/RFC4206, October 2005,
              <http://www.rfc-editor.org/info/rfc4206>.

   [RFC4208]  Swallow, G., Drake, J., Ishimatsu, H., and Y. Rekhter,
              "Generalized Multiprotocol Label Switching (GMPLS)
              User-Network Interface (UNI): Resource ReserVation
              Protocol-Traffic Engineering (RSVP-TE) Support for the
              Overlay Model", RFC 4208, DOI 10.17487/RFC4208,
              October 2005, <http://www.rfc-editor.org/info/rfc4208>.

   [RFC4216]  Zhang, R., Ed., and J.-P. Vasseur, Ed., "MPLS
              Inter-Autonomous System (AS) Traffic Engineering (TE)
              Requirements", RFC 4216, DOI 10.17487/RFC4216,
              November 2005, <http://www.rfc-editor.org/info/rfc4216>.

   [RFC4271]  Rekhter, Y., Ed., Li, T., Ed., and S. Hares, Ed., "A
              Border Gateway Protocol 4 (BGP-4)", RFC 4271,
              DOI 10.17487/RFC4271, January 2006,
              <http://www.rfc-editor.org/info/rfc4271>.

   [RFC4364]  Rosen, E. and Y. Rekhter, "BGP/MPLS IP Virtual Private
              Networks (VPNs)", RFC 4364, DOI 10.17487/RFC4364,
              February 2006, <http://www.rfc-editor.org/info/rfc4364>.

   [RFC4655]  Farrel, A., Vasseur, J.-P., and J. Ash, "A Path
              Computation Element (PCE)-Based Architecture", RFC 4655,
              DOI 10.17487/RFC4655, August 2006,
              <http://www.rfc-editor.org/info/rfc4655>.

   [RFC4726]  Farrel, A., Vasseur, J.-P., and A. Ayyangar, "A Framework
              for Inter-Domain Multiprotocol Label Switching Traffic
              Engineering", RFC 4726, DOI 10.17487/RFC4726,
              November 2006, <http://www.rfc-editor.org/info/rfc4726>.

   [RFC4847]  Takeda, T., Ed., "Framework and Requirements for Layer 1
              Virtual Private Networks", RFC 4847, DOI 10.17487/RFC4847,
              April 2007, <http://www.rfc-editor.org/info/rfc4847>.

   [RFC4874]  Lee, CY., Farrel, A., and S. De Cnodder, "Exclude Routes -
              Extension to Resource ReserVation Protocol-Traffic
              Engineering (RSVP-TE)", RFC 4874, DOI 10.17487/RFC4874,
              April 2007, <http://www.rfc-editor.org/info/rfc4874>.

   [RFC4920]  Farrel, A., Ed., Satyanarayana, A., Iwata, A., Fujita, N.,
              and G. Ash, "Crankback Signaling Extensions for MPLS and
              GMPLS RSVP-TE", RFC 4920, DOI 10.17487/RFC4920, July 2007,
              <http://www.rfc-editor.org/info/rfc4920>.

   [RFC5150]  Ayyangar, A., Kompella, K., Vasseur, JP., and A. Farrel,
              "Label Switched Path Stitching with Generalized
              Multiprotocol Label Switching Traffic Engineering
              (GMPLS TE)", RFC 5150, DOI 10.17487/RFC5150,
              February 2008, <http://www.rfc-editor.org/info/rfc5150>.

   [RFC5152]  Vasseur, JP., Ed., Ayyangar, A., Ed., and R. Zhang, "A
              Per-Domain Path Computation Method for Establishing
              Inter-Domain Traffic Engineering (TE) Label Switched Paths
              (LSPs)", RFC 5152, DOI 10.17487/RFC5152, February 2008,
              <http://www.rfc-editor.org/info/rfc5152>.

   [RFC5195]  Ould-Brahim, H., Fedyk, D., and Y. Rekhter, "BGP-Based
              Auto-Discovery for Layer-1 VPNs", RFC 5195,
              DOI 10.17487/RFC5195, June 2008,
              <http://www.rfc-editor.org/info/rfc5195>.

   [RFC5246]  Dierks, T. and E. Rescorla, "The Transport Layer Security
              (TLS) Protocol Version 1.2", RFC 5246,
              DOI 10.17487/RFC5246, August 2008,
              <http://www.rfc-editor.org/info/rfc5246>.

   [RFC5251]  Fedyk, D., Ed., Rekhter, Y., Ed., Papadimitriou, D.,
              Rabbat, R., and L. Berger, "Layer 1 VPN Basic Mode",
              RFC 5251, DOI 10.17487/RFC5251, July 2008,
              <http://www.rfc-editor.org/info/rfc5251>.

   [RFC5252]  Bryskin, I. and L. Berger, "OSPF-Based Layer 1 VPN
              Auto-Discovery", RFC 5252, DOI 10.17487/RFC5252,
              July 2008, <http://www.rfc-editor.org/info/rfc5252>.

   [RFC5305]  Li, T. and H. Smit, "IS-IS Extensions for Traffic
              Engineering", RFC 5305, DOI 10.17487/RFC5305,
              October 2008, <http://www.rfc-editor.org/info/rfc5305>.

   [RFC5440]  Vasseur, JP., Ed., and JL. Le Roux, Ed., "Path Computation
              Element (PCE) Communication Protocol (PCEP)", RFC 5440,
              DOI 10.17487/RFC5440, March 2009,
              <http://www.rfc-editor.org/info/rfc5440>.

   [RFC5441]  Vasseur, JP., Ed., Zhang, R., Bitar, N., and JL. Le Roux,
              "A Backward-Recursive PCE-Based Computation (BRPC)
              Procedure to Compute Shortest Constrained Inter-Domain
              Traffic Engineering Label Switched Paths", RFC 5441,
              DOI 10.17487/RFC5441, April 2009,
              <http://www.rfc-editor.org/info/rfc5441>.

   [RFC5523]  Berger, L., "OSPFv3-Based Layer 1 VPN Auto-Discovery",
              RFC 5523, DOI 10.17487/RFC5523, April 2009,
              <http://www.rfc-editor.org/info/rfc5523>.

   [RFC5553]  Farrel, A., Ed., Bradford, R., and JP. Vasseur, "Resource
              Reservation Protocol (RSVP) Extensions for Path Key
              Support", RFC 5553, DOI 10.17487/RFC5553, May 2009,
              <http://www.rfc-editor.org/info/rfc5553>.

   [RFC5623]  Oki, E., Takeda, T., Le Roux, JL., and A. Farrel,
              "Framework for PCE-Based Inter-Layer MPLS and GMPLS
              Traffic Engineering", RFC 5623, DOI 10.17487/RFC5623,
              September 2009, <http://www.rfc-editor.org/info/rfc5623>.

   [RFC5920]  Fang, L., Ed., "Security Framework for MPLS and GMPLS
              Networks", RFC 5920, DOI 10.17487/RFC5920, July 2010,
              <http://www.rfc-editor.org/info/rfc5920>.

   [RFC5925]  Touch, J., Mankin, A., and R. Bonica, "The TCP
              Authentication Option", RFC 5925, DOI 10.17487/RFC5925,
              June 2010, <http://www.rfc-editor.org/info/rfc5925>.

   [RFC6005]  Berger, L. and D. Fedyk, "Generalized MPLS (GMPLS) Support
              for Metro Ethernet Forum and G.8011 User Network Interface
              (UNI)", RFC 6005, DOI 10.17487/RFC6005, October 2010,
              <http://www.rfc-editor.org/info/rfc6005>.

   [RFC6071]  Frankel, S. and S. Krishnan, "IP Security (IPsec) and
              Internet Key Exchange (IKE) Document Roadmap", RFC 6071,
              DOI 10.17487/RFC6071, February 2011,
              <http://www.rfc-editor.org/info/rfc6071>.

   [RFC6107]  Shiomoto, K., Ed., and A. Farrel, Ed., "Procedures for
              Dynamically Signaled Hierarchical Label Switched Paths",
              RFC 6107, DOI 10.17487/RFC6107, February 2011,
              <http://www.rfc-editor.org/info/rfc6107>.

   [RFC6805]  King, D., Ed., and A. Farrel, Ed., "The Application of the
              Path Computation Element Architecture to the Determination
              of a Sequence of Domains in MPLS and GMPLS", RFC 6805,
              DOI 10.17487/RFC6805, November 2012,
              <http://www.rfc-editor.org/info/rfc6805>.

   [RFC6827]  Malis, A., Ed., Lindem, A., Ed., and D. Papadimitriou,
              Ed., "Automatically Switched Optical Network (ASON)
              Routing for OSPFv2 Protocols", RFC 6827,
              DOI 10.17487/RFC6827, January 2013,
              <http://www.rfc-editor.org/info/rfc6827>.

   [RFC6996]  Mitchell, J., "Autonomous System (AS) Reservation for
              Private Use", BCP 6, RFC 6996, DOI 10.17487/RFC6996,
              July 2013, <http://www.rfc-editor.org/info/rfc6996>.

   [RFC7399]  Farrel, A. and D. King, "Unanswered Questions in the Path
              Computation Element Architecture", RFC 7399,
              DOI 10.17487/RFC7399, October 2014,
              <http://www.rfc-editor.org/info/rfc7399>.

   [RFC7579]  Bernstein, G., Ed., Lee, Y., Ed., Li, D., Imajuku, W., and
              J. Han, "General Network Element Constraint Encoding for
              GMPLS-Controlled Networks", RFC 7579,
              DOI 10.17487/RFC7579, June 2015,
              <http://www.rfc-editor.org/info/rfc7579>.

   [RFC7580]  Zhang, F., Lee, Y., Han, J., Bernstein, G., and Y. Xu,
              "OSPF-TE Extensions for General Network Element
              Constraints", RFC 7580, DOI 10.17487/RFC7580, June 2015,
              <http://www.rfc-editor.org/info/rfc7580>.

   [RFC7752]  Gredler, H., Ed., Medved, J., Previdi, S., Farrel, A., and
              S. Ray, "North-Bound Distribution of Link-State and
              Traffic Engineering (TE) Information Using BGP", RFC 7752,
              DOI 10.17487/RFC7752, March 2016,
              <http://www.rfc-editor.org/info/rfc7752>.

   [RSVP-TE-EXCL]
              Ali, Z., Ed., Swallow, G., Ed., Zhang, F., Ed., and D.
              Beller, Ed., "Resource ReserVation Protocol-Traffic
              Engineering (RSVP-TE) Path Diversity using Exclude Route",
              Work in Progress, draft-ietf-teas-lsp-diversity-05,
              June 2016.

   [RSVP-TE-EXT]
              Zhang, F., Ed., Gonzalez de Dios, O., Ed., Hartley, M.,
              Ali, Z., and C. Margaria, "RSVP-TE Extensions for
              Collecting SRLG Information", Work in Progress,
              draft-ietf-teas-rsvp-te-srlg-collect-06, May 2016.

   [RSVP-TE-METRIC]
              Ali, Z., Swallow, G., Filsfils, C., Hartley, M., Kumaki,
              K., and R. Kunze, "Resource ReserVation Protocol-Traffic
              Engineering (RSVP-TE) extension for recording TE Metric of
              a Label Switched Path", Work in Progress,
              draft-ietf-teas-te-metric-recording-04, March 2016.

Appendix A.  Existing Work

   This appendix briefly summarizes relevant existing work that is used
   to route TE paths across multiple domains.  It is non-normative.

A.1.  Per-Domain Path Computation

   The mechanism for per-domain path establishment is described in
   [RFC5152], and its applicability is discussed in [RFC4726].  In
   summary, this mechanism assumes that each domain entry point is
   responsible for computing the path across the domain but that details
   regarding the path in the next domain are left to the next domain
   entry point.  The computation may be performed directly by the entry
   point or may be delegated to a computation server.

   This basic mode of operation can run into many of the issues
   described alongside the use cases in Section 2.  However, in practice
   it can be used effectively, with a little operational guidance.

   For example, RSVP-TE [RFC3209] includes the concept of a "loose hop"
   in the explicit path that is signaled.  This allows the original
   request for an LSP to list the domains or even domain entry points to
   include on the path.  Thus, in the example in Figure 1, the source
   can be told to use interconnection x2.  Then, the source computes the
   path from itself to x2 and initiates the signaling.  When the
   signaling message reaches Domain Z, the entry point to the domain
   computes the remaining path to the destination and continues the
   signaling.

   Another alternative suggested in [RFC5152] is to make TE routing
   attempt to follow inter-domain IP routing.  Thus, in the example
   shown in Figure 2, the source would examine the BGP routing
   information to determine the correct interconnection point for
   forwarding IP packets and would use that to compute and then signal a
   path for Domain A.  Each domain in turn would apply the same approach
   so that the path is progressively computed and signaled domain by
   domain.

   Although the per-domain approach has many issues and drawbacks in
   terms of achieving optimal (or, indeed, any) paths, it has been the
   mainstay of inter-domain LSP setup to date.

A.2.  Crankback

   Crankback addresses one of the main issues with per-domain path
   computation: What happens when an initial path is selected that
   cannot be completed toward the destination?  For example, what
   happens if, in Figure 2, the source attempts to route the path
   through interconnection x2 but Domain C does not have the right TE
   resources or connectivity to route the path further?

   Crankback for MPLS-TE and GMPLS networks is described in [RFC4920]
   and is based on a concept similar to the Acceptable Label Set
   mechanism described for GMPLS signaling in [RFC3473].  When a node
   (i.e., a domain entry point) is unable to compute a path further
   across the domain, it returns an error message in the signaling
   protocol that states where the blockage occurred (link identifier,
   node identifier, domain identifier, etc.) and gives some clues about
   what caused the blockage (bad choice of label, insufficient bandwidth
   available, etc.).  This information allows a previous computation
   point to select an alternative path, or to aggregate crankback
   information and return it upstream to a previous computation point.

   Crankback is a very powerful mechanism and can be used to find an
   end-to-end path in a multi-domain network if one exists.

   On the other hand, crankback can be quite resource-intensive, as
   signaling messages and path setup attempts may "wander around" in the
   network, attempting to find the correct path for a long time.  Since
   (1) RSVP-TE signaling ties up network resources for partially
   established LSPs, (2) network conditions may be in flux, and (3) most
   particularly, LSP setup within well-known time limits is highly
   desirable, crankback is not a popular mechanism.

   Furthermore, even if crankback can always find an end-to-end path, it
   does not guarantee that the optimal path will be found.  (Note that
   there have been some academic proposals to use signaling-like
   techniques to explore the whole network in order to find optimal
   paths, but these tend to place even greater burdens on network
   processing.)

A.3.  Path Computation Element

   The Path Computation Element (PCE) is introduced in [RFC4655].  It is
   an abstract functional entity that computes paths.  Thus, in the
   example of per-domain path computation (see Appendix A.1), both the
   source node and each domain entry point are PCEs.  On the other hand,
   the PCE can also be realized as a separate network element (a server)
   to which computation requests can be sent using the Path Computation
   Element Communication Protocol (PCEP) [RFC5440].

   Each PCE is responsible for computations within a domain and has
   visibility of the attributes within that domain.  This immediately
   enables per-domain path computation with the opportunity to offload
   complex, CPU-intensive, or memory-intensive computation functions
   from routers in the network.  But the use of PCEs in this way
   does not solve any of the problems articulated in Appendices A.1
   and A.2.

   Two significant mechanisms for cooperation between PCEs have been
   described.  These mechanisms are intended to specifically address the
   problems of computing optimal end-to-end paths in multi-domain
   environments.

   -  The Backward-Recursive PCE-Based Computation (BRPC) mechanism
      [RFC5441] involves cooperation between the set of PCEs along the
      inter-domain path.  Each one computes the possible paths from the
      domain entry point (or source node) to the domain exit point (or
      destination node) and shares the information with its upstream
      neighbor PCE, which is able to build a tree of possible paths
      rooted at the destination.  The PCE in the source domain can
      select the optimal path.

      BRPC is sometimes described as "crankback at computation time".
      It is capable of determining the optimal path in a multi-domain
      network but depends on knowing the domain that contains the
      destination node.  Furthermore, the mechanism can become quite
      complicated and can involve a lot of data in a mesh of
      interconnected domains.  Thus, BRPC is most often proposed for a
      simple mesh of domains and specifically for a path that will cross
      a known sequence of domains, but where there may be a choice of
      domain interconnections.  In this way, BRPC would only be applied
      to Figure 2 if a decision had been made (externally) to traverse
      Domain C rather than Domain D (notwithstanding that it could
      functionally be used to make that choice itself), but BRPC could
      be used very effectively to select between interconnections x1 and
      x2 in Figure 1.

   -  The Hierarchical PCE (H-PCE) [RFC6805] mechanism offers a parent
      PCE that is responsible for navigating a path across the domain
      mesh and for coordinating intra-domain computations by the child
      PCEs responsible for each domain.  This approach makes computing
      an end-to-end path across a mesh of domains far more tractable.
      However, it still leaves unanswered the issue of determining the
      location of the destination (i.e., discovering the destination
      domain) as described in Section 2.1.  Furthermore, it raises the
      question of who operates the parent PCE, especially in networks
      where the domains are under different administrative and
      commercial control.

   It should also be noted that [RFC5623] discusses how PCEs are used in
   a multi-layer network with coordination between PCEs operating at
   each network layer.  Further issues and considerations regarding the
   use of PCEs can be found in [RFC7399].

A.4.  GMPLS UNI and Overlay Networks

   [RFC4208] defines the GMPLS User-Network Interface (UNI) to present a
   routing boundary between an overlay (client) network and the server
   network, i.e., the client-server interface.  In the client network,
   the nodes connected directly to the server network are known as edge
   nodes, while the nodes in the server network are called core nodes.

   In the overlay model defined by [RFC4208], the core nodes act as a
   closed system and the edge nodes do not participate in the routing
   protocol instance that runs among the core nodes.  Thus, the UNI
   allows access to, and limited control of, the core nodes by edge
   nodes that are unaware of the topology of the core nodes.  This
   respects the operational and layer boundaries while scaling the
   network.

   [RFC4208] does not define any routing protocol extension for the
   interaction between core and edge nodes but allows for the exchange
   of reachability information between them.  In terms of a VPN, the
   client network can be considered as the customer network comprised of
   a number of disjoint sites, and the edge nodes match the VPN CE
   nodes.  Similarly, the provider network in the VPN model is
   equivalent to the server network.

   [RFC4208] is, therefore, a signaling-only solution that allows edge
   nodes to request connectivity across the server network and leaves
   the server network to select the paths for the LSPs as they traverse
   the core nodes (setting up hierarchical LSPs if necessitated by the
   technology).  This solution is supplemented by a number of signaling
   extensions, such as [RFC4874], [RFC5553], [RSVP-TE-EXCL],
   [RSVP-TE-EXT], and [RSVP-TE-METRIC], to give the edge node more
   control over the path within the server network and by allowing the
   edge nodes to supply additional constraints on the path used in the
   server network.  Nevertheless, in this UNI/overlay model, the edge
   node has limited information regarding precisely what LSPs could be
   set up across the server network and what TE services (diverse routes
   for end-to-end protection, end-to-end bandwidth, etc.) can be
   supported.

A.5.  Layer 1 VPN

   A Layer 1 VPN (L1VPN) is a service offered by a Layer 1 server
   network to provide Layer 1 connectivity (Time-Division Multiplexing
   (TDM), Lambda Switch Capable (LSC)) between two or more customer
   networks in an overlay service model [RFC4847].

   As in the UNI case, the customer edge has some control over the
   establishment and type of connectivity.  In the L1VPN context, three
   different service models have been defined, classified by the
   semantics of information exchanged over the customer interface: the
   management-based model, the signaling-based (a.k.a. basic) service
   model, and the signaling and routing (a.k.a. enhanced) service model.

   In the management-based model, all edge-to-edge connections are
   set up using configuration and management tools.  This is not a
   dynamic control-plane solution and need not concern us here.

   In the signaling-based (basic) service model [RFC5251], the CE-PE
   interface allows only for signaling message exchange, and the
   provider network does not export any routing information about the
   server network.  VPN membership is known a priori (presumably through
   configuration) or is discovered using a routing protocol [RFC5195]
   [RFC5252] [RFC5523], as is the relationship between CE nodes and
   ports on the PE.  This service model is much in line with GMPLS UNI
   as defined in [RFC4208].

   In the signaling and routing (enhanced) service model, there is an
   additional limited exchange of routing information over the CE-PE
   interface between the provider network and the customer network.  The
   enhanced model considers four different types of service models,
   namely the overlay extension, virtual node, virtual link, and per-VPN
   service models.  All of these represent particular cases of the TE
   information aggregation and representation.

A.6.  Policy and Link Advertisement

   Inter-domain networking relies on policy and management input to
   coordinate the allocation of resources under different administrative
   control.  [RFC5623] introduces a functional component called the VNTM
   for this purpose.

   An important companion to this function is determining how
   connectivity across the abstraction layer network is made available
   as a TE link in the client network.  Obviously, if the connectivity
   is established using management intervention, the consequent client
   network TE link can also be configured manually.  However, if
   connectivity from client edge to client edge is achieved using
   dynamic signaling, then there is need for the end points to exchange
   the link properties that they should advertise within the client
   network, and in the case of support for more than one client network,
   it will be necessary to indicate which client network or networks can
   use the link.  This capability it provided in [RFC6107].

Appendix B.  Additional Features

   This appendix describes additional features that may be desirable and
   that can be achieved within this architecture.  It is non-normative.

B.1.  Macro Shared Risk Link Groups

   Network links often share fate with one or more other links.  That
   is, a scenario that may cause a link to fail could cause one or more
   other links to fail.  This may occur, for example, if the links are
   supported by the same fiber bundle, or if some links are routed down
   the same duct or in a common piece of infrastructure such as a
   bridge.  A common way to identify the links that may share fate is to
   label them as belonging to a Shared Risk Link Group (SRLG) [RFC4202].

   TE links created from LSPs in lower layers may also share fate, and
   it can be hard for a client network to know about this problem
   because it does not know the topology of the server network or the
   path of the server network LSPs that are used to create the links in
   the client network.

   For example, looking at the example used in Section 4.2.3 and
   considering the two abstract links S1-S3 and S1-S9, there is no way
   for the client network to know whether links C2-C0 and C2-C3 share
   fate.  Clearly, if the client layer uses these links to provide a
   link-diverse end-to-end protection scheme, it needs to know that the
   links actually share a piece of network infrastructure (the server
   network link S1-S2).

   Per [RFC4202], an SRLG represents a shared physical network resource
   upon which the normal functioning of a link depends.  Multiple SRLGs
   can be identified and advertised for every TE link in a network.
   However, this can produce a scalability problem in a multi-layer
   network that equates to advertising in the client network the server
   network route of each TE link.

   Macro SRLGs (MSRLGs) address this scaling problem and are a form of
   abstraction performed at the same time that the abstract links are
   derived.  In this way, links that actually share resources in the
   server network are advertised as having the same MSRLG, rather than
   advertising each SRLG for each resource on each path in the server
   network.  This saving is possible because the abstract links are
   formulated on behalf of the server network by a central management
   agency that is aware of all of the link abstractions being offered.

   It may be noted that a less optimal alternative path for the abstract
   link S1-S9 exists in the server network (S1-S4-S7-S8-S9).  It would
   be possible for the client network request for C2-C0 connectivity to
   also ask that the path be maximally disjoint from path C2-C3.
   Although nothing can be done about the shared link C2-S1, the
   abstraction layer could make a request to use link S1-S9 in a way
   that is diverse from the use of link S1-S3, and this request could be
   honored if the server network policy allows it.

   Note that SRLGs and MSRLGs may be very hard to describe in the case
   of multiple server networks because the abstraction points will not
   know whether the resources in the various server layers share
   physical locations.

B.2.  Mutual Exclusivity

   As noted in the discussion of Figure 13, it is possible that some
   abstraction layer links cannot be used at the same time.  This arises
   when the potentiality of the links is indicated by the server
   network, but the use of the links would actually compete for server
   network resources.  Referring to Figure 13, this situation would
   arise when both link S1-S3 and link S7-S9 are used to carry LSPs: in
   that case, link S1-S9 could no longer be used.

   Such a situation need not be an issue when client-edge-to-client-edge
   LSPs are set up one by one, because the use of one abstraction layer
   link and the corresponding use of server network resources will cause
   the server network to withdraw the availability of the other
   abstraction layer links, and these will become unavailable for
   further abstraction layer path computations.

   Furthermore, in deployments where abstraction layer links are only
   presented as available after server network LSPs have been
   established to support them, the problem is unlikely to exist.

   However, when the server network is constrained but chooses to
   advertise the potential of multiple abstraction layer links even
   though they compete for resources, and when multiple client-edge-to-
   client-edge LSPs are computed simultaneously (perhaps to provide
   protection services), there may be contention for server network
   resources.  In the case where protected abstraction layer LSPs are
   being established, this situation would be avoided through the use of
   SRLGs and/or MSRLGs, since the two abstraction layer links that
   compete for server network resources must also fate-share across
   those resources.  But in the case where the multiple client-edge-to-
   client-edge LSPs do not care about fate sharing, it may be necessary
   to flag the mutually exclusive links in the abstraction layer TED so
   that path computation can avoid accidentally attempting to utilize
   two of a set of such links at the same time.

Acknowledgements

   Thanks to Igor Bryskin for useful discussions in the early stages of
   this work and to Gert Grammel for discussions on the extent of
   aggregation in abstract nodes and links.

   Thanks to Deborah Brungard, Dieter Beller, Dhruv Dhody, Vallinayakam
   Somasundaram, Hannes Gredler, Stewart Bryant, Brian Carpenter, and
   Hilarie Orman for review and input.

   Particular thanks to Vishnu Pavan Beeram for detailed discussions and
   white-board scribbling that made many of the ideas in this document
   come to life.

   Text in Section 4.2.3 is freely adapted from the work of Igor
   Bryskin, Wes Doonan, Vishnu Pavan Beeram, John Drake, Gert Grammel,
   Manuel Paul, Ruediger Kunze, Friedrich Armbruster, Cyril Margaria,
   Oscar Gonzalez de Dios, and Daniele Ceccarelli in [GMPLS-ENNI], for
   which the authors of this document express their thanks.

Contributors

   Gert Grammel
   Juniper Networks
   Email: ggrammel@juniper.net

   Vishnu Pavan Beeram
   Juniper Networks
   Email: vbeeram@juniper.net

   Oscar Gonzalez de Dios
   Email: ogondio@tid.es

   Fatai Zhang
   Email: zhangfatai@huawei.com

   Zafar Ali
   Email: zali@cisco.com

   Rajan Rao
   Email: rrao@infinera.com

   Sergio Belotti
   Email: sergio.belotti@alcatel-lucent.com

   Diego Caviglia
   Email: diego.caviglia@ericsson.com

   Jeff Tantsura
   Email: jeff.tantsura@ericsson.com

   Khuzema Pithewan
   Email: kpithewan@infinera.com

   Cyril Margaria
   Email: cyril.margaria@googlemail.com

   Victor Lopez
   Email: vlopez@tid.es

Authors' Addresses

   Adrian Farrel (editor)
   Juniper Networks

   Email: adrian@olddog.co.uk

   John Drake
   Juniper Networks

   Email: jdrake@juniper.net

   Nabil Bitar
   Nokia

   Email: nbitar40@gmail.com

   George Swallow
   Cisco Systems, Inc.
   1414 Massachusetts Ave.
   Boxborough, MA  01719

   Email: swallow@cisco.com

   Daniele Ceccarelli
   Ericsson
   Via A. Negrone 1/A
   Genova - Sestri Ponente
   Italy

   Email: daniele.ceccarelli@ericsson.com

   Xian Zhang
   Huawei Technologies

   Email: zhang.xian@huawei.com