Network Working Group                                         J. Rekhter
Request for Comments: 1092                  T. J. Watson Research Center
                                                           February 1989

EGP and Policy Based Routing in the New NSFNET Backbone

Status of this Memo

   This memo discusses implementation decisions for routing issues in
   the NSFNET, especially in the NSFNET Backbone.  Of special concern is
   the restriction of routing information to advertize the best route as
   established by a policy decision.  Distribution of this memo is
   unlimited.

Introduction

   The NSFNET backbone routes packets between the Regionals Networks to
   which it is connected, (i.e., the packets arriving at a backbone
   entry node are routed to an exit node).  How they travel through the
   network is determined by two components:

     the NSFNET backbone routing protocol/algorithm, and

     additional information about the externally connected networks.

   This paper is concerned with how reachability information between the
   external networks and the NSFNET backbone is exchanged so that
   packets can be routed to the correct destination by using a
   reasonable path.

EGP as reachability protocol

   The EGP (Exterior Gateway Protocol) routing method will be used to
   exchange reachability information between the NSFNET backbone and the
   regional networks.

   There are several problems with using EGP as a reachability protocol
   for routing in a meshed environment.  Some EGP components require
   further definitions for the NSFNET backbone - regional network
   interactions.  It should be noted that the use of EGP is only viewed
   as an interim measure until better inter autonomous system protocols
   are defined and widely deployed for gateways used by regional
   networks.

   The following is a list of some EGP problems and issues:

      The EGP model assumes an engineered spanning tree topology,

      however, the NSFNET (due to the presence of backdoor routes) does
      not fit into this model.  In the NSFNET the same network may be
      advertized as reachable by more than one regional network.
      Besides the fact that the overall NSFNET does not fit into a
      spanning tree model there are serious concerns with the concept
      of the "core" (central to the EGP) and its obvious deficiencies.

      While EGP is going to isolate intra-Regional routing from the
      intra-NSFNET-Backbone routing, it does not address the issue of
      false information which may be supplied by regional networks.
      EGP by itself does not protect a particular network from unwanted
      and unsolicited representation by some regional network.  As an
      example, if network N1 is reachable through regional network R1
      as well as through regional network R2, EGP has no provisions to
      specify one of these paths as a primary and one as a secondary,
      since there is not generally accepted interpretation of EGP
      metrics today.  Also, there is nothing in EGP which can prevent one
      or more regional networks from advertizing other networks (in
      particular, networks which belong to other regional networks) as
      reachable with zero distance.  This could result in the creation
      of a "black hole" or at least in suboptimal IP routing.

      EGP by itself has no provisions to guarantee that routes through
      the NSFNET Backbone will be preferred over routes through the
      backdoor routers or vice versa.

Policy Based Routing

   Looking at the problems listed above the appearance of the new
   factors like autonomy and mutual trust becomes obvious.  While trying
   to achieve the routing functionality required for the new NSFNET
   backbone we should realize that one of our primary concerns has to be
   the accommodation of those new factors.

   This means that some kind of a rudimentary Policy Based Routing
   method becomes imperative.  We would like to emphasize, however, that
   we are not talking about complete Policy Based Routing, but that we
   are rather concerned about supporting a minimum subset of a policy
   functionality to be an initial solution to the above mentioned
   problems.  This requires support and cooperation between the
   management of each of the networks connected to the NSFNET backbone.

   We need to support the ability of a particular network N, which
   belongs to one of the regional networks, to establish a bilateral
   agreement with one or more regional networks of the type "network N
   can be reached via one or more regional networks (RN1, RN2, ...
   RNx)".  This allows each network to select one or more
   representatives at the regional network level.  Once this agreement

   is established the information will be available to:

     The network which initiated the agreement.

     The management of the regional network(s) with whom this
     agreement has been established.

     The NSFNET backbone Network Operation Center where it will be
     entered into the Routing Policy Data Base which will be available
     through the NSFNET information services.

   Supporting multiple routes to the NSFNET core requires the guarantee
   that for a certain network N, no regional network other than the
   one(s) selected by N, will advertize N as reachable, which
   necessitates that the NSFNET core will ignore unauthorized
   advertisements for network N.

EGP and Rudimentary Policy Based Routing

   Each network which belongs to the NSFNET will select a specific
   regional network as its primary representative to the NSFNET core by
   bilateral agreement with the management of same regional network as
   well as the NSFNET backbone management.  The same network can
   furthermore select an arbitrary number of other regional networks as
   their secondary, tertiary, etc., representative by establishing
   bilateral agreements with the management of the corresponding
   regional networks as well as the NSFNET backbone management.

   Reachability information supplied by each regional network will be
   distributed to all other NSS nodes of the NSFNET Backbone.  We would
   like to emphasize that we are not going to flood EGP packets
   internally within the backbone, but to rather use the learned
   information for the interior gateway protocol, which uses the ANSI
   IS-IS protocol.

   The implementation allows for a defined regional network to advertize
   a particular leaf network in the EGP NR packets with a distance of
   zero.  Secondary representatives may advertize the same network with
   distance one or higher.  If the path through the primary regional
   representative is available all secondary paths will be ignored.  If
   the path through the primary regional representative goes down (which
   will be discovered via the EGP NR information), the next path with
   the lowest available EGP metric will be used.

   We will also be able to detect and report unsolicited
   representations.  This will be done by examining (on a periodic
   basis) all reachability information obtained via EGP.  The result
   will be compared against the Routing Policy Data Base which will hold

   information about all bilateral agreements between networks and their
   regional representatives.  Any mismatch will cause an alarm to the
   Network Operations Center.  For example, network N established a
   bilateral agreement with the regional network R1 electing it as its
   primary representative. The EGP NR record received from the regional
   network R5 advertizes the network N as reachable with distance zero.
   By comparing the Routing Policy Data Base entry for the network N
   with the EGP NR record a mismatch will be detected and an alarm is
   forwarded to the Network Operation Center.

   Since the whole scheme is based on a combination of the network
   number and the autonomous system number, to allow for further
   verification, it is also important to insure the correctness of the
   autonomous system numbers as advertized by the regionals networks to
   the NSFNET core.

   The autonomous system number validation for each regional network
   will be performed at the NSS which connects the particular leaf
   network to the NSFNET backbone.  All discrepancies wil be reported to
   the Network Operations Center.

   The NSFNET backbone will be considered as a separate Autonomous
   System with its own autonomous system number.

Backbone versus Backdoor Routes

   There are instances where regional networks prefer paths through some
   backdoor route over paths through the NSFNET backbone.  Therefore,
   the reachability information advertized by the NSFNET core to the
   regional networks (via EGP NR records) will always use a fixed metric
   of 128 for all routes.  This may aid to encourage traffic to flow
   through backdoors, if desired and available.

   The regional networks can use a variety of techniques to determine
   how they route traffic for any particular network at their own
   option.

What do we expect from the Regional Networks

   Each regional network should get its own Autonomous System number.

   The connection between regional networks to NSFNET backbone will be
   done via EGP.  It is the responsibility of the regional backbone to
   provide an EGP functionality via the attachment to the E-PSP
   dedicated to the regional network.

   The EGP functionality may require a translation of network numbers in
   and out of the regional network.  In any case, the NSFNET backbone

   expects individual network numbers of the leaf networks of the
   regional network, as long as they should be advertised, and will
   announce individual networks known to the NSFNET core to the regional
   network.

   The EGP support should includes the ability to configure EGP metrics
   from some statically definable configuration table.  If the EGP
   metrics cannot be defined or if they are not fixed the metric
   determination will be done by the NSFNET backbone routers, as taken
   from their databases, themselves.  In that case, it is the
   responsibility of the regional network to provide the NSFNET backbone
   management with the metric data to allow for proper use of metrics.

   We also expect each regional network to handle all bilateral
   agreements with its leaf networks regarding Policy Based Routing and
   supply a copy of those agreements to the NSFNET backbone management.

Acknowledgements

   I would like to express my thanks to Barry Appelman (T.J. Watson
   Research Center, IBM Corp.) and Hans-Werner Braun (Merit) for their
   contributions to this document.

Author's Address

   Jacob Rekhter
   T.J. Watson Research Center
   IBM Corporation
   P.O. Box 218
   Yorktown Heights, NY 10598

   Phone: (914) 945-3896

   Email: YAKOV@IBM.COM