Network Working Group                                         J. Uttaro
Internet Draft                                                     AT&T
Intended status: Standards Track                    V. Van den Schrieck
May 25,
Nov 26, 2012                                     Individual Contributor
Expires: Nov 25, 2012 May 26, 2013                                       P. Francois
                                                         IMDEA Networks
                                                            R. Fragassi
                                                             A. Simpson
                                                           P. Mohapatra
                                                          Cisco Systems

        Best Practices for Advertisement of Multiple Paths in IBGP

Status of this Memo

   This Internet-Draft is submitted in full conformance with the
   provisions of BCP 78 and BCP 79.

   Internet-Drafts are working documents of the Internet Engineering
   Task Force (IETF), its areas, and its working groups.  Note that
   other groups may also distribute working documents as Internet-

   Internet-Drafts are draft documents valid for a maximum of six months
   and may be updated, replaced, or obsoleted by other documents at any
   time.  It is inappropriate to use Internet-Drafts as reference
   material or to cite them other than as "work in progress."

   The list of current Internet-Drafts can be accessed at

   The list of Internet-Draft Shadow Directories can be accessed at

   This Internet-Draft will expire on Nov 25, 2012. May 26, 2013.

Copyright Notice

   Copyright (c) 2012 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
   ( 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.


   Add-Paths is a BGP enhancement that allows a BGP router to advertise
   multiple distinct paths for the same prefix/NLRI. This provides a
   number of potential benefits, including reduced routing churn, faster
   convergence and better loadsharing.

   This document provides recommendations to implementers of Add-Paths
   so that network operators have the tools needed to address their
   specific applications and to manage the scalability impact of Add-
   Paths. A router implementing Add-Paths may learn many paths for a
   prefix and must decide which of these to advertise to peers. This
   document analyses different algorithms for making this selection and
   provides recommendations based on the target application.

Table of Contents

   1. Introduction...................................................4
   2. Terminology....................................................4
   3. Add-Paths Applications.........................................5
      3.1. Fast Connectivity Restoration.............................5
      3.2. Load Balancing............................................7
      3.3. Churn Reduction...........................................7
      3.4. Suppression of MED-Related Persistent Route Oscillation...7
   4. Implementation Guidelines......................................8
      4.1. Capability Negotiation....................................8
      4.2. Receiving Multiple Paths..................................9
      4.3. Advertising Multiple Paths................................9
         4.3.1. Path Selection Modes................................11
   Advertise All Paths............................11
   Advertise N Paths..............................12
   Advertise All AS-Wide Best Paths...............12
   Advertise ALL AS-Wide Best and Next-Best Paths
            (Double AS Wide)........................................13
         4.3.2. Derived Modes from Bounding the Number of Advertised
   5. Deployment Considerations.....................................14
      5.1. Introducing Add-Paths into an Existing Network...........14
      5.2. Scalability Considerations...............................17
      5.3. Routing Consistency Considerations.......................17
      5.4. Consistency between Advertised Paths and Forwarding Paths18
      5.5. Routing Churn............................................19
   6. Security Considerations.......................................19
   7. IANA Considerations...........................................19
   8. Conclusions...................................................19
   9. References....................................................19
      9.1. Normative References.....................................19
      9.2. Informative References...................................19
   10. Acknowledgments..............................................20
   Appendix A. Other Path Selection Modes...........................21
      A.1. Advertise Neighbor-AS Group Best Path....................21
      A.2. Best LocPref/Second LocPref..............................21
      A.3. Advertise Paths at decisive step -1......................22

1. Introduction

   The BGP Add-Paths capability enhances current BGP implementations by
   allowing a BGP router to exchange with its BGP peers more than one
   path for the same destination/NLRI. The base BGP standard [RFC 4271]
   does not provide for such a capability. If a BGP router learns
   multiple paths for the same NLRI (from multiple peers), it selects
   only one as its best path and advertises the best path to its peers.
   The primary goal of Add-Paths is to increase the visibility of paths
   within an iBGP system.  This has the effect of improving robustness
   in case of failure, reducing the number of BGP messages exchanged
   during such an event, and offering the potential for faster re-
   convergence. Through careful selection of the paths to be advertised,
   Add-Paths can also prevent routing oscillations.

   The purpose of this document is to provide the necessary
   recommendations to the implementers of Add-Paths so that network
   operators have the tools needed to address their specific
   applications and to manage the scalability impact of Add-Paths while
   maintaining routing consistency.  A router implementing Add-Paths may
   learn many paths for a prefix and must decide which of these to
   advertise to peers. This document analyses different algorithms for
   making this selection and provides recommendations based on the
   target application.

2. Terminology

   In this document the following terms are used:

   Add-Paths peer: refers a peer with which the local system has agreed
   to receive and/or send NLRI with path identifiers

   Primary path: A path toward a prefix that is considered a best path
   by the BGP decision process [RFC 4271] and actively used for
   forwarding traffic to that prefix. A router may have multiple primary
   paths for a prefix if it implements multipath.

   Diverse path: A BGP path associated with a different BGP next-hop and
   BGP router than some other set of paths. The BGP router associated
   with a path is inferred from the ORIGINATOR_ID attribute or, if there
   is none, the BGP Identifier of the peer that advertised the path.

   Backup path: A diverse path with respect to the primary paths toward
   a prefix. The backup path can be used to forward traffic to the
   destination if the primary paths fail.

   Optimal backup path: The backup path that will be selected as the new
   best path for a prefix when all primary paths are removed/withdrawn.

   AS-Wide preferred paths: All paths that are considered as best when
   applying rules of the BGP decision process up to the IGP tie-break.

   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   document are to be interpreted as described in [RFC-2119].

3. Add-Paths Applications

   [draft-pmohapat] presents the applications that would benefit from
   multiple paths advertisement in iBGP.  They are summarized in the
   following subsections.

   3.1. Fast Connectivity Restoration

   With the dissemination of backup paths, fast connectivity restoration
   and convergence can be achieved.  If a router has a backup path, it
   can directly select that path as best upon failure of the primary
   path.  This minimizes packet loss in the dataplane.  Sending multiple
   paths in iBGP allows routers to receive backup paths when path
   visibility is not sufficient with classical BGP.  This is especially
   useful when Route Reflection is used.

   Consider a network such as the one depicted in Figure 1 and suppose
   that none of the routers support Add-Paths. AS1 receives from AS3 2
   paths (A and B) to a particular destination XYZ. Suppose path A is
   preferred over path B due to path A having a lower MED (multi-exit

   AS1 uses a route reflector RR1 to reduce the scale of its IBGP mesh.
   If the routers in AS1 are not configured for best-external then RR1
   knows about only path A during steady state because router B
   suppresses/withdraws its advertisement of path (B) to RR1. If the
   routers in AS1 do support best-external then RR1 may have both paths
   in its Adj-RIB-IN, but regardless of the best-external configuration
   RR1 can only advertise its best path A to its peers, including router

              ========        =====================
              =  +---+        +---+           +---+
              =  |RTR|________|RTR|           |RTR|
              =  | E |        | A |           | C |
              =  +---+Path A->+---+    AS1    +---+
              =      =        =    \         /    =
              =      =        =     \       /     =
              =      =        =      \     /      =
              =      =        =       \   /       =
              = AS3  =        =       +---+       =
              =      =        =       |RR |       =
              =      =        =       | 1 |       =
              =      =        =       +---+       =
              =      =        =       /   \       =
              =      =        =      /     \      =
              =      =        =     /       \     =
              =      =        =    /         \    =
              =  +---+Path B->+---+           +---+
              =  |RTR|  ______|RTR|           |RTR|
              =  | F |        | B |           | D |
              =  +---+        +---+           +---+
              ========        =====================

                        Figure 1: Example Topology

   Under these circumstances consider the steps required to restore
   traffic from router D to destination XYZ when the link between Router
   A and Router E fails. (Assume that router A set next-hop to self when
   advertising path A and that router B is not configured for best-

   1. Router A sends a BGP UPDATE message withdrawing its advertisement
      of path (A).

   2. RR1 receives the withdrawal, and propagates it to its other client
      peers, routers B, C and D.

   3. When router B receives the withdrawal of path (A) it reruns its
      decision process and selects path (B) as its new best path. Router
      B advertises path (B) to RR1.

   4. RR1 reruns its decision process and selects path (B) as its new
      best path. RR1 advertises path (B) to client peers A, C and D.

   5. Router D reruns its decisions process, determines path (B) to be
      the best path, and updates its forwarding table. After this step
      traffic from router D to destination XYZ is restored (the traffic
      path has changed from A to B).

   With the use of Add-Paths, the convergence time for the above path
   failure example can be reduced considerably. The main reason for the
   improvement is that Add-Paths allows router D to be aware of more
   than one path to destination XYZ prior to the failure of the best
   path (A). In steady-state (with no failures) router B decides, as
   before, that path (A) is its best path but because of its Add-Paths
   (or best-external) configuration it also advertises path (B) to RR1.
   Using Add-Paths RR1 can advertise both learned paths to its IBGP
   peers, including router D. Now consider again the scenario where the
   link between Router A and Router E fails. In this case, with Add-
   Paths, fewer steps are required to achieve re-convergence:

   1. Router A sends a BGP UPDATE message withdrawing its advertisement
      of path (A).

   2. RR1 receives the withdrawal, and propagates it to its other client
      peers, routers B, C and D.

   3. Router D receives the withdrawal, reruns the decision process and
      updates the forwarding entry for destination XYZ.

   3.2. Load Balancing

   Increased path diversity allows routers to install several paths in
   their forwarding tables in order to load balance traffic across those

   3.3. Churn Reduction

   When Add-Paths is used in an AS, the availability of additional
   backup paths means failures can be recovered locally with much less
   path exploration in iBGP and therefore less updates disseminated in
   eBGP.  When the preferred backup path is the post-convergence path,
   churn is minimized.

   3.4. Suppression of MED-Related Persistent Route Oscillation

   As described in [oscillation], Add-Paths is a valuable tool in
   helping to stop persistent route oscillations caused by comparison of
   paths based on MED in topologies where route reflectors or the
   confederation structure hide some paths. With the appropriate path
   selection algorithm Add-Paths stops these route oscillations because
   the same set of paths are consistently advertised by the route
   reflector or the confederation border router and the routers
   receiving this set of paths make stable routing decisions about the
   best path.

4. Implementation Guidelines

   This section discusses recommendations for the implementation of Add-
   Paths. The following topics are addressed:

     .  Considerations related to Add-Paths capability negotiation

     .  Receiving BGP routes from Add-Paths peers

     .  Advertising BGP routes to Add-Paths peers. This section
        discusses various path selection algorithms, which are the
        procedures available to an Add-Paths speaker for deciding which
        set of paths to advertise to an Add-Paths peer for particular

   4.1. Capability Negotiation

       +---+           +---+
       | A |  <-BGP->  | B |
       +---+           +---+

                       Figure 2: BGP Peering Example

   In Figure 2, in order for a router A to receive multiple paths per
   NLRI from peer B, for a particular address family (AFI=x, SAFI=y),
   the BGP capabilities advertisements during session setup must
   indicate that peer B wants to send multiple paths for AFI=x, SAFI=y
   and that router A is willing to receive multiple paths for AFI=x,
   SAFI=y. Similarly, in order for router A to send multiple paths per
   NLRI to peer B, for a particular address family (AFI=x, SAFI=y), the
   BGP capabilities advertisements must indicate that router A wants to
   send multiple paths for AFI=x, SAFI=y and peer B is willing to
   receive multiple paths for AFI=x, SAFI=y. Refer to [Add-Paths] for
   details of the Add-Paths capabilities advertisement.

   The capabilities of the local router MUST be configurable per peer
   and per address family, and SHOULD support the ability to configure
   send-only operation or receive-only operation. The default mode of
   operation shall be to both send and receive.

   4.2. Receiving Multiple Paths

   Currently, per standard BGP behavior, if a BGP router receives an
   advertisement of an NLRI and path from a specific peer and that peer
   subsequently advertises the same NLRI with different path information
   (e.g. a different NEXT_HOP and/or different path attributes) the new
   path effectively overwrites the existing path.

   When Add-Paths has been negotiated with the peer, the newly
   advertised path should be stored in the RIB-IN along with all of the
   paths previously advertised (and not withdrawn) by the peer.

   When an Add-Paths speaker has negotiated to receive multiple paths
   for (AFIx, SAFIy) from a peer all advertisements and withdrawals of
   NLRI within that address family from that peer MUST include a path
   identifier, as described in [Add-Paths]. The path identifiers have no
   significance to the receiving peer. If the combination of NLRI and
   path identifier in an advertisement from a peer is unique (does not
   match an existing route in the RIB-IN from that peer) then the route
   is added to the RIB-IN. If the combination of NLRI and path
   identifier in a received advertisement is the same as an existing
   route in the RIB-IN from the peer then the new route replaces the
   existing one. If the combination of NLRI and path identifier in a
   received withdrawal matches an existing route in the RIB-IN from the
   peer then that route shall be removed from the RIB-IN.

   A BGP UPDATE message from an Add-Paths peer may advertise and
   withdraw more than one NLRI belonging to one or more address
   families. In this case Add-Paths may be supported for some of the
   address families and not others. In this situation the receiving BGP
   router should not expect that all of the path identifiers in the
   UPDATE message will be the same.

   4.3. Advertising Multiple Paths

   [Add-Paths] specifies how to encode the advertisement of multiple
   paths towards the same NLRI over an iBGP session, but provides no
   details about which set of multiple paths should be advertised.  In
   this section, four path selection algorithms are described and
   compared with each other. These 4 algorithms are considered to be the
   most useful across the widest range of deployment scenarios. The list
   of possible path selection algorithms is much larger and for the
   interested reader Appendix A provides information about other path
   selection modes that were considered in historical versions of this

   In comparing any two path selection algorithms the following factors
   should be taken into account:

   Control Plane Load: When a router receives multiples paths for a
   prefix from an iBGP client it has to store more paths in its Adj-Rib-

   Control Plane Stress: Coping with multiple iBGP paths has two
   implications on the computation that a router has to handle. First,
   it has to compute the paths to send to its peers, i.e. more than the
   best path.  Second, it also has to handle the potential churn related
   to the exchange of those multiple paths.

   MED/IGP oscillations: BGP sometimes suffers from routing oscillations
   when the physical topology differs from the logical topology, or when
   the MED attribute is used.  This is due to the limited path
   visibility when a single path is advertised and Route Reflection is
   used.  Increasing the path visibility by advertising multiple paths
   can help solve this issue.

   Path optimality: When a single path is advertised, border routers do
   not always receive the optimal path. As an example, Route Reflectors
   typically send a single path chosen based on their own IGP tie-break
   (although modifications to this are proposed in [BGP-ORR]).
   Increasing path visibility would also help routers to learn the path
   that is best suited for them w.r.t. the IGP tie-break.

   Backup path optimality: Multiple paths advertisement gives routers
   the opportunity to have a backup path.  However, some backup paths
   are better than others.  Indeed, when a link failure occurs, if a
   router already knows its post-convergence path, the BGP re-
   convergence is straightforward and traffic is less impacted by the
   transient use of non-best forwarding paths.

   Convergence time: Advertising multiple paths in iBGP has an impact on
   the convergence time of the BGP system.  More paths need to be
   exchanged, but on the other hand, the routing information is
   propagated faster. With an increased path visibility, there is less
   path exploration during the convergence.  Also, with the availability
   of backup paths, convergence time in case of failure is also reduced.

   Target application: Depending on the application type, the number of
   paths to advertise for a prefix will vary. For example, for fast
   connectivity restoration, it may be sufficient to advertise only 2
   paths to a peer so that it will have the best path and the optimal
   backup path. For load balancing purposes, it may be desirable to
   advertise more paths, but inclusion of the optimal backup path in the
   set may be less critical. For route oscillation elimination, it is
   required to advertise all group-best paths for a prefix.

4.3.1. Path Selection Modes

   The following subsections describe the 4 main path selection modes
   considered in this draft. Each mode is considered either MANDATORY or
   OPTIONAL. A MANDATORY mode MUST be supported by any implementation
   that claims compliance with this document. An OPTIONAL made may be
   supported by some but not all implementations.

   The path selection mode and any parameters applicable to the mode
   MUST be configurable per AFI/SAFI and per peer and SHOULD be
   configurable per prefix. To illustrate the value of this flexibility,
   consider a prefix P that belongs to an address family F requiring
   path IDs to be included with every NLRI (e.g. due to the Add-Paths
   capability negotiation with the peer). If P is one of a number of
   prefixes that would not benefit from the advertisement of multiple
   paths then it is perfectly valid to send only the best path. Advertise All Paths

   A simple rule for advertising multiple paths in iBGP is to advertise
   to iBGP peers all received paths minus those blocked by export
   filters or applicable split horizon rules.  This solution is easy to
   implement, but the counterpart is that all those paths need to be
   stored by all routers that receive them, which can be quite
   expensive.  If a path to a prefix P is advertised to N border
   routers, with a Full Mesh of iBGP sessions, all routers have N paths
   in their Adj-RIB-Ins.  If Route Reflection is used and each client is
   connected to 2 Route Reflectors, it may learn up to 2*N paths.

   This solution gives a perfect path visibility to all routers, thus
   limiting churn and losses of connectivity in case of failure. Indeed,
   this allows routers to select their optimal primary path, and to
   switch on their optimal backup path in case of failure.

   However, as more paths are exchanged, the number of BGP messages
   disseminated during the initial iBGP convergence can be high, and
   convergence may be slower.

   Routing oscillations are prevented with this rule, because a router
   won't need to withdraw a previously advertised path when its best
   path changes.

   This path selection mode is OPTIONAL. Advertise N Paths

   Another solution is for a router to advertise a maximum of N paths to
   iBGP peers.  Here, the computational cost is the selection of the N
   paths. Indeed, there must be a ranking of the paths in order to
   advertise the most interesting ones.  A way for a router to select N
   paths is to run N times its decision process. At each iteration of
   the process only those paths not selected during a previous iteration
   and those with a different NEXT_HOP and BGP Identifier (or Originator
   ID) combination from previously-selected paths are eligible for
   consideration. In this mode the paths actually advertised to a peer
   are the eligible paths (up to N) minus those blocked by export
   filters or applicable split horizon rules. The memory cost is
   bounded: a router receives a maximum of N paths for each prefix from
   each peer. With N equal to 2, all routers know at least two paths and
   can provide local recovery in case of failure.  If multipath routing
   is to be deployed in the AS, N can be increased to provide more
   alternate paths to the routers.

   Path optimality and backup path optimality are not guaranteed, i.e.
   it is possible that the optimal path of a router (w.r.t. IGP tie-
   break) is not contained in the set of paths advertised by its Route
   Reflector. However, as the number of paths that it receives is higher
   than without Add-Paths, it is possible that the chosen nexthop is
   closer to the router in terms of IGP cost than the nexthop that would
   have been chosen without Add-Paths.

   This solution helps to reduce routing oscillations, but not in all
   cases.  Indeed, path visibility is still constrained by the maximum
   number of paths, and configurations with routing oscillations still

   This path selection mode is MANDATORY. The default value of N MUST be
   2.  The value of N MUST be configurable and MAY be upper bounded by
   an implementation.

   The default value of 2 ensures the availability of a backup path (if
   2 or more paths have been received) while maintaining minimum impact
   to memory and churn.  If Add-N with N equal to 2 is insufficient to
   meet another objective (e.g. loadsharing or MED/IGP oscillation)
   there is always a large enough value of N that can selected, if N is
   configurable, to meet that objective. Advertise All AS-Wide Best Paths

   Another choice is to consider the set of paths with the same AS-wide
   preference [Basu-ibgp-osc], i.e. the paths that all routers would
   select based on the rules of the decision process that are not
   router-dependent (i.e.  Local-preference, ASPath length and MED
   rules).  Thus, for a given router, those paths only differ by the IGP
   cost to the nexthop or by the tie-breaking rules. The paths actually
   advertised to a peer are the set of AS-wide best paths minus those
   blocked by export filters or applicable split horizon rules.

   The computational cost is reduced, as a router only has to send the
   paths remaining before applying the IGP tie-breaking rule.  However,
   it is difficult to predict how many paths will be stored, as it
   depends on the number of eBGP sessions on which this prefix is
   advertised with the best AS-wide preference.

   With this rule, the routing system is optimal: all routers can choose
   their best path (or best paths if multipath is used) based on their
   router-specific preferences, i.e. the IGP cost to the nexthop. Hot
   potato routing is respected.  Also, MED oscillations are prevented,
   because the path visibility among the AS-wide preferred paths is

   The existence of a backup path is not guaranteed. If only one path
   with the AS-wide best attributes exists, there is no backup path
   disseminated.  However, if such a path exists, it is optimal as it
   has the same AS-wide preference as the primary

   This path selection mode is OPTIONAL. Advertise ALL AS-Wide Best and Next-Best Paths (Double
          AS Wide)

   This variant of "Advertise All AS Wide Best Paths" trades-off the
   number of paths being propagated within the iBGP system for post-
   convergence alternate paths availability and routing stability. A BGP
   speaker running this mode will select, as candidates for
   advertisement, its AS Wide Best paths, plus all the AS Wide Best
   paths obtained when removing the first ones from consideration. The
   paths actually advertised to a peer are the double-AS_wide candidate
   paths minus those blocked by export filters or applicable split
   horizon rules.

   Under this mode, a BGP speaker knows multiple AS-Wide best paths or
   the AS-Wide best path and all the second AS-Wide best paths, so that
   routing optimality and backup path availability are ensured. Note
   that the post-convergence paths will be known by each BGP node in an
   AS supporting this mode.

   The computation complexity of this mode is relatively low as it
   requires to run the usual BGP Decision Process up to and including
   the MED rule. The set of paths remaining after that step form the AS-
   Wide best paths.  Next, a best path selection algorithm is run up to
   and including the MED rule, based on the paths that are not in the
   set of AS-Wide best paths.

   The number of paths for a prefix p, known by a given router of the
   AS, is the number of AS-Wide best and second AS-Wide best paths found
   at the Borders of the AS.

   MED Oscillations are avoided by this mode, both for the primary and
   alternate paths being picked under this mode.

   This path selection mode is OPTIONAL.

4.3.2. Derived Modes from Bounding the Number of Advertised Paths

   For some of the modes discussed in section 4.3.1 the number of paths
   selected by the algorithm (M) is not predictable in advance, and
   depends on factors such as network topology. For such modes,
   implementations MAY support the ability to limit the number of
   advertised paths to some value N that is less than M.

   It must be noted that the resulting derivative mode may no longer
   meet the properties stated in section 4.3.1 (which assumes N=M). This
   is particularly true for the MED oscillation avoidance property. The
   use of such bounds thus needs to be considered carefully in
   deployments where MED oscillation avoidance is a key goal of
   deploying Add-path. If fast recovery is the main objective then it is
   reasonable and sufficient to set N to 2.  If the main goal is
   improved load-balancing then limiting N to number of ECMP paths
   supported by the forwarding planes of the receiving routers is also a
   reasonable practice.

5. Deployment Considerations

   This section proposes a potential strategy for introducing Add-Paths
   into an existing network and discusses considerations related to
   scalability, routing consistency and routing churn.

   5.1. Introducing Add-Paths into an Existing Network

   There are many possible ways that Add-Paths can be introduced into an
   existing deployed network. It is not a practical goal for this
   document to list all of these options and discuss the pros and cons
   for each one. It is however valuable to consider an example migration
   strategy that may be relatively common among layer 3 service
   providers that currently use route reflectors for scaling. This
   example migration strategy is attractive for several reasons:

     1. It involves incremental steps that allow the impact of Add-
        Paths to be carefully evaluated before proceeding to the next

     2. It recognizes the fact that many routers will require at least
        a software upgrade to support Add-Paths, and it will not be
        practical to upgrade all of these routers all at once.

     3. It reduces convergence time (in stages) with a relatively
        moderate increase in router memory and CPU demands.

   The example migration strategy assumes a starting point of a deployed
   network with one or more RR clusters. None of the routers in the
   network support Add-Paths without an upgrade, but some do support
   best-external. Two of the clusters in this network are shown in
   Figure 3. In cluster 2, PE1, PE2, RRy and RRz are configured for
   best-external. This makes RRy and RRz aware of all external paths
   received by PEs in cluster 2 and ensures that RRy and RRz can
   advertise a path to the RRs in cluster 1 if it happens that the best
   overall route is learned from cluster 1. It doesn't however allow
   other clusters to be aware of more than one path per prefix learned
   by cluster 2.

       ==========               ==================
       =        =               =                =
       =    +---+               +---+      +---+ =
       =    |RR |---------------|RR |  <-BE|   | =
       =    |a  |               |y  |------|PE1| =
       =    |   |               |   |      |   | =
       =    +---+               +---+      +---+ =
       =      | =               = |  \    /      =
       =      | =               = |   \  /       =
       =      | =               = |    \/        =
       =      | =               = |    /\        =
       =      | =               = |   /  \       =
       =      | =               = |  /    \      =
       =    +---+               +---+      +---+ =
       =    |RR |---------------|RR |------|   | =
       =    |b  |               |z  |  <-BE|PE2| =
       =    |   |               |   |      |   | =
       =    +---+               +---+      +---+ =
       =        =               =                =
       ==========               ==================
      RR Cluster 1                 RR Cluster 2

                   Figure 3: RR Cluster Before Add-Paths

   The following sequence of steps occurs in the example migration

   1. The route reflectors are upgraded in each cluster, one by one, to
     support Add-Paths. This allows the intra- and (eventually) inter-
     cluster RR-to-RR sessions to start using Add-Paths. All RRs are
     configured to use the Add-N, N=2 path selection algorithm. The
     effect of this step is to slightly reduce convergence time when
     the best and second-best paths for a prefix are learned by a
     single cluster (such as cluster 2 in Figure 3).

   2. The clients are upgraded in each cluster, one by one, to support
     Add-Paths. On the RRs Add-Paths is configured to use the Add-N,
     N=2 path selection algorithm towards upgraded client peers. At
     this step clients are configured in the receive-only Add-Paths
     mode.  This means that best-external continues to operate as
     before in the client-to-RR direction. The effect of this step is
     to ensure that all clients have two paths per prefix for ECMP or
     fast failover, assuming at least 2 paths are available.

   3. The clients are re-configured to use Add-Paths in the transmit
     direction towards their RR peers. This causes Add-Paths to replace
     the best-external behavior. The effect of this step is to free up
     CPU and memory resources related to the storage of paths that are
     third best or worse. If a cluster such as the one in Figure 3 had
     50 clients, and 10 of these learned an external route for the same
     prefix, then the RRs in that cluster would need to store up to 12
     paths for that prefix. This would be true even if the 2 best
     overall paths came from another cluster. Contrast this with the
     use of Add-Paths in the client-to-RR direction. For the same case
     the route reflectors need only store the 2 paths learned from non-
     client peers.

   5.2. Scalability Considerations

   In terms of scalability, we note that advertising multiple paths per
   prefix requires more memory and state than the current behavior of
   advertising the best path only. A BGP speaker that does not implement
   Add-Paths maintains send state information in its prefix data
   structure per neighbor as a way to determine that the prefix has been
   advertised to the neighbor. With Add-Paths, this information has to
   be replicated on a per path basis that needs to be advertised.
   Mathematically, if "send state" size per prefix is 's' bytes, number
   of neighbors is 'n', and number of paths being advertised is 'p',
   then the current memory requirement for BGP "send state" = n * s
   bytes; with Add-Paths, it becomes n * s * p bytes. In practice, this
   value may be reduced with implementation optimizations similar to
   attribute sharing.  Receiving multiple paths per prefix also requires
   more memory and state since each path is a separate entry in the Adj-

   5.3. Routing Consistency Considerations

   As discussed in previous sections Add-Paths can help routers select
   more optimal paths and it can help deal with certain route
   oscillation conditions arising from incomplete knowledge of the
   available paths.  But depending on the path selection algorithm and
   how it is used Add-Paths is not immune to its own cases of routing
   inconsistencies. If the BGP routers within an AS do not make
   consistent routing decisions about how to reach a particular
   destination, route oscillations may occur and these route
   oscillations may result in traffic loss.

   Optimizing an Add-Paths deployment for scalability may run counter to
   routing consistency goals, and in these circumstances operators have
   to decide the correct tradeoff for their particular deployment. For
   example the Advertise All Paths mode, if applied to many prefixes, is
   far from ideal from a scalability perspective but it does guarantee
   routing consistency and correctness. A path selection mode that
   allows better control over scalability is the Advertise N paths mode,
   but this is susceptible to routing inconsistency. First, if the N
   paths do not include the best path from each neighbor AS group then
   route oscillation cannot be precluded. Second, if the advertising
   router (e.g. an RR) advertises N paths to peer_n and M paths to
   peer_m, and N < M, care must be exercised to ensure that all paths
   advertised to peer_n are included in the paths advertised to peer_m.
   This can be assured as long as the advertising router has strictly
   ordered all of its paths.

   5.4. Consistency between Advertised Paths and Forwarding Paths

   When using Add-Paths, routers may advertise paths that they have not
   selected as best, and that they are thus not using for traffic
   forwarding.  This is generally not an issue if encapsulation is used
   in the AS as described in [RFC4364] and all forwarding decisions,
   including by the tunnel egress router, are based on label information
   - i.e. if only the ingress router performs an IP FIB lookup.  In this
   situation the dataplane path followed by the packets is the one
   intended by the ingress router, and corresponds to the control plane
   path it selected.

   On the other hand, if Add-Paths is used in a network without
   encapsulation, some scenarios can result in forwarding deflection or
   loops.  Such forwarding anomalies already occur without Add-Paths,
   when the routers on the forwarding path do not have a synchronized
   view of the best path.  They will deflect the traffic to their own
   local view of the best path, and, when multiple deflections occur,
   forwarding loops can occur.  With Add-Paths, the issue can be
   exacerbated due to routers advertising non-best paths. As discussed
   above, encapsulation can help with this issue, but only to the extent
   that it allows downstream routers to forward without an IP FIB

   A first example of such issue is when the Local-Pref of non-primary
   paths received over iBGP sessions is modified.  The ingress router
   may thus select as best a path non-preferred by the egress, and the
   egress router will thus deflect the traffic.

   Another example is when the best path is selected based on tie-
   breaking rule.  When the ingress and the egress base their path
   selection on the router-id of the neighbor that advertised the path
   to them, the result may be different for each of them.  This specific
   issue is described and solved in [draft-pmohapat].

   5.5. Routing Churn

   As noted in section 3.3 using Add-Paths between IBGP peers can help
   to reduce routing churn with EBGP peers. This benefit does however
   come at the cost of potentially increased churn between the IBGP Add-
   Paths peers. In a non Add-Paths deployment a change in the preference
   order of non-best paths requires no updates to be sent to peers. But
   when a router has Add-Paths peers changes in non-best path preference
   may no longer be invisible and increased route churn may be
   observable. Choosing the right path selection mode and parameters -
   for example not setting N unnecessarily large in the Add-N mode, is
   important to minimizing this additional churn.

6. Security Considerations


7. IANA Considerations


8. Conclusions


9. References

   9.1. Normative References

   [RFC2119]        Bradner, S., "Key words for use in RFCs to Indicate
                    Requirement Levels", BCP 14, RFC 2119, March 1997.

   9.2. Informative References

   [Add-Paths]      Walton, D., Retana, A., Chen E., Scudder J.,
                    "Advertisement of Multiple Paths in BGP", February
                    6, 2010. draft-
                    ietf-idr-add-paths-07, June 17, 2012.

   [draft-pmohapat] Mohapatra, P., Fernando, R., Filsfils, C., and R.
                    Raszuk, "Fast Connectivity Restoration Using BGP
                    Add-path", draft-pmohapat-idr-fast-conn-restore-
                    00.txt (work in progress), September 2008.
                    02.txt, Oct 3, 2011.

   [oscillation]    Walton, D., Retana, A., Chen, E., Scudder, J., "BGP
                    Persistent Route Oscillation Solutions", draft-
                    walton-bgp-route-oscillation-stop-03.txt, May 10,
                    walton-bgp-route-oscillation-stop-06.txt, June 14,

   [Basu-ibgp-osc]  Basu, A., Ong, C., Rasala, A., Sheperd, B., and G.
                    Wilfong, "Route oscillations in iBGP with Route
                    Reflection", Sigcomm 2002.

   [BGP-ORR]        Raszuk, R., Cassar, C., Aman, E., Decraene, B., "BGP
                    Optimal Route Reflection", draft-raszuk-bgp-optimal-
                    route-reflection-01, March 11, 2011.

   [RFC4271]        Rekhter, Y., Li, T., Hares, S., "A Border Gateway
                    Protocol 4 (BGP-4), January 2006.

10. Acknowledgments

   This document was prepared using

Appendix A.                 Other Path Selection Modes

A.1. Advertise Neighbor-AS Group Best Path

   [walton-osc] proposes that a router groups its paths based on the
   neighbor AS from which it was learned, and to advertise the best path
   in each of those groups.

   The control plane stress induced by this solution is the computation
   of the per-neighbor path group, and the application of the decision
   process to each of them.  The Control-Plane load is bounded by the
   number of neighboring ASes advertising a prefix, which cannot be
   known a-priori.

   Path optimality and backup path optimality are not guaranteed, as the
   paths advertised are not all the AS-wide preferred paths. Backup path
   availability is not guaranteed.  Indeed, if only one AS advertises
   this prefix, even on multiple eBGP sessions, only one of the paths
   may be selected and advertised.

A.2. Best LocPref/Second LocPref

   This selection method consists in grouping the paths by Local
   Preference.  A router sends to its peers all paths with the highest
   Local Preference.  If there is only a single path with the highest
   Local Preference, it also sends all paths with the second best Local

   This method ensures that all routers know all paths with the best
   local preference.  As local preference are often related to the type
   of peering of the peer the path comes from, this ensures that in case
   of failure, routers have a backup path of equivalent quality.  This
   prevents for example that a router switches temporarily on a peer
   path while an alternate path from a customer is available but hidden
   at the border of the AS.  Such a situation could result in a
   temporary withdrawal of the prefix on some eBGP sessions when the
   router selects the path via the peer.

   The advertisement of the Second Local Preference occurs when there is
   no alternate path with the same quality as the best path.  This way,
   fast convergence is still ensured.  Backup path is optimal, as it has
   the second AS-Wide preference, which becomes the AS-wide best
   preference upon failure of the primary one.

   Sending all the paths with a given Local Preference also has a
   positive impact on routing optimality. Indeed, this allows border
   routers to have an increased path visibility and to choose their best
   path based on their own criteria.

   The computational cost of this solution is reduced when there are
   several paths with the best local preference.  In this case, it is
   sufficient to stop the decision process after the first rule to have
   the set of paths to be advertised.  When it is necessary to advertise
   the paths with second local-preference, the additional cost is to
   apply a second time the first rule of the decision process, which is
   still reasonable.  The memory cost depends on the number of paths
   with the best local preference.

A.3. Advertise Paths at decisive step -1

   When the goal is to provide fast recovery by advertising candidate
   post-reconvergence paths, one can choose to stop the decision process
   just before the step where only one path remains.  If the decision
   process comes to IGP tie-break, all remaining paths are advertised.
   This way, routers advertise as many paths as possible with a quality
   as similar as possible.

   This path selection is an intermediary solution between the two
   preceding ones.  Here, instead of stopping the decision process at
   the local preference step or the IGP step, we stop it before the rule
   that removes the best potential backup paths.  This way, we minimize
   the number of paths to advertise while guaranteeing the presence of a
   backup path.  Primary and backup path optimality is ensured, as all
   paths with the same AS-wide preference as the best paths are included
   in the set of paths advertised.

Authors' Addresses

   Jim Uttaro
   200 S. Laurel Avenue
   Middletown, NJ 07748 USA

   Virginie Van den Schrieck

   Pierre Francois
   IMDEA Networks
   Avenida del Mar Mediterraneo
   Leganes  28919

   Roberto Fragassi
   600 Mountain Avenue
   Murray Hill, New Jersey

   Adam Simpson
   600 March Road
   Ottawa, Ontario K2K 2E6

   Pradosh Mohapatra
   Cisco Systems
   170 W. Tasman Drive
   San Jose, CA 95134 USA