Network Working Group H.W. Braun
Request for Comments: 1093 Merit
 February 1989
 The NSFNET Routing Architecture
Status of this Memo
 This document describes the routing architecture for the NSFNET
 centered around the new NSFNET Backbone, with specific emphasis on
 the interface between the backbone and its attached networks.
 Distribution of this memo is unlimited.
Introduction
 This document describes the routing architecture for the NSFNET
 centered around the new NSFNET Backbone, with specific emphasis on
 the interface between the backbone and its attached networks. It
 reflects and augments thoughts described in [1], discussions during
 the Internet Engineering Task Force meeting at the San Diego
 Supercomputing Center in March 1988, discussions on mailing lists,
 especially on a backbone/regional network working group mailing list,
 and a final discussion held at the IBM T.J. Watson Research Center in
 Yorktown, NY, on the 21st of March 1988. The Yorktown meeting was
 attended by Hans-Werner Braun (Merit), Scott Brim (Cornell
 University), Mark Fedor (NYSERNet), Jeff Honig (Cornell University),
 and Jacob Rekhter (IBM). Thanks also to: Milo Medin (NASA), John Moy
 (Proteon) and Greg Satz (Cisco) for discussing this document by email
 and/or phone.
 Understanding of [1] is highly recommended prior to reading this
 document.
1. Routing Overview
 The new NSFNET backbone forms the core of the overall NSFNET, which
 connects to regional networks (or regional backbones) as well as to
 peer networks (other backbones like the NASA Science Network or the
 ARPANET). The NSFNET core uses a SPF based internal routing
 protocol, adapted from the IS-IS protocol submitted by ANSI for
 standardization to the ISO. The ANSI IS-IS protocol is based upon
 work done at Digital Equipment Corporation. Its adaptation to the
 Internet environment requires additional definitions, most notably to
 the addressing structure, which will be described in a later
 document. This adaptation was largely done by Jacob Rekhter of IBM
 Research in Yorktown, NY. The RCP/PSP routing architecture was
 largely implemented by Rick Boivie and his colleagues at IBM TCS in
 Milford, CT. The adaptation of EGP to the NSS routing code and the
 new requirements was done jointly by Jeff Honig (who spent about a
 week to work on this at IBM Research) and Jacob Rekhter. Jeff is
 integrating the changes done for the new EGP requirements into the
 "gated" distributions.
 The IGP derives routing tables from Internet address information.
 This information is flooded throughout the NSFNET core, and the
 individual NSS nodes create or update their routing information after
 running the SPF algorithm over the flooded information. A detailed
 description of the NSFNET backbone IGP will be documented in a future
 document.
 The routing interface between the NSFNET core and regional backbones
 as well as peer networks utilizes the Exterior Gateway Protocol
 (EGP). The EGP/IGP consistency and integrity at the interface points
 is ensured by filtering mechanisms according to individual nodal
 routing policy data bases [1]. EGP is selected as the routing
 interface of choice between the NSFNET backbone and its regional
 attachments due to its widespread implementation as well its ability
 to utilize autonomous system designators and to allow for effective
 firewalls between systems. In the longer run the hope is to replace
 the EGP interface with a new inter Autonomous System protocol. Such a
 new protocol should also allow to move the filtering of network
 numbers or Autonomous Network number groups to the regional gateways
 in order for the regional gateways to decide as to what routing
 information they wish to receive.
 A general model is to ensure consistent routing information between
 peer networks. This means that, e.g., the NSFNET core will have the
 same sets of Internet network numbers in its routing tables as are
 present in the ARPANET core. However, the redistribution of this
 routing information is tightly controlled and based on Autonomous
 System numbers. For example, ARPANET routes with the ARPANET
 Autonomous System number will not be redistributed into regional or
 other peer networks. If an NSFNET internal path exists to such a
 network known to the ARPANET it may be redistributed into regional
 networks, subject to further policy verification. Generally it may be
 necessary to have different trust models for peer and subordinate
 networks, while giving a greater level of trust to peer networks.
 The described use of EGP, which is further elaborated on in [1]
 requires bidirectional translation of network information between the
 IGP in use and EGP.
2. Conclusions reached during the discussions
 The following conclusions were reached during the meeting and in
 subsequent discussions:
 No DDN-only routes (ARPANET/MILNET) shall be announced into the
 regional backbones. This is a specific case of the ability to
 suppress information from specific Autonomous Systems, as
 described later.
 Regional backbones are required to use an unique Autonomous System
 number. Announcements from non-sanctioned autonomous systems,
 relative to a particular site, will not be believed and will
 instead trigger an alarm to the Network Operations Center.
 Regional backbone attachments must not require routes to local
 subnets. This means that the locally attached network needs to
 use a flat space, without subnet bits, at least from the NSS point
 of view. The reason for this is that the EGP information
 exchanged between the regional gateway and the NSS cannot include
 subnet information. Therefore the NSS has no knowledge of remote
 subnets. The safest way to get around this limitation is to use a
 non-subnetted network (like a separate Class-C network) at the
 interface between a regional backbone and the NSFNET backbone.
 The other way is to use Proxy-ARP while having just the NSS think
 that the network is not subnetted. In the latter case care must be
 taken so that the E-PSP uses the proper local IP broadcast
 address.
 Routing information received by the NSS from regional gateways
 will be verified on both network number and autonomous system
 number.
 Metric reconstitution is done on a per-network basis. The NSS
 will construct the fixed metric it will use for a given network
 number from its internal data base. Network metrics given to the
 NSS via EGP will be ignored. The metrics used are a result of an
 ordered list of preferred paths as supplied by the regional
 backbones and the attached campuses. This metric is of relevance
 only to the NSFNET core itself. The mechanisms are further
 explained in [1].
 Global metric reconstitution by Autonomous System numbers is
 necessary in specific cases, such as peer networks. An example is
 that ARPANET routes will be reconstituted to a global metric, as
 determined by the NSS.
 EGP announcements into regional networks will use a fixed metric.
 The metric used shall be "128." The 128-metric is somewhat
 arbitrarily chosen to be high enough so that a regional backbone
 will get a metric high enough from the NSFNET Core AS to allow a
 comparison against other (most likely internal) routes. "128" is
 also consistent with [2].
 Peer network routes (e.g., ARPANET routes) are propagated through
 the NSS structure.
 No DEFAULT routing information is distributed within the NSFNET
 backbone, as the NSFNET core has the combined routing knowledge of
 the attached regional and peer networks.
 We do not expect the requirement for damping of routing update
 frequencies, at least initially. The frequency of net up/down
 changes combined with the available bandwidth and CPU capacity do
 not let the frequency of SPF floodings appear as being a major
 problem. Simple metric changes as heard by a NSS via EGP will not
 trigger updates.
 An allowed list of Source Autonomous System information will be
 used to convert from the IGP to EGP, on a Destination Autonomous
 System number basis, to allow for specific exclusion of definable
 remote Autonomous System information.
 EGP must only announce networks for which the preferred path is
 via the IGP. This means in particular that the EGP peer will
 never announce via EGP what it learned via EGP on the same
 interface, not even if the information was received from a third
 EGP peer. This will avoid the back-distribution of information
 learned via that same interface. The EGP peers of regional
 gateways must only announce information belonging to their own
 Autonomous System.
 EGP will be used in interior mode only.
 The regional backbones are responsible for generating DEFAULT
 routing information at their option. One possibility is to
 generate an IGP default on a peer base as long as the NSS EGP
 connection is working. The EGP information will not include a
 special indication for DEFAULT.
 It is highly desirable to have direct peer-peer connections, to
 ease the implementation of a consistent routing data base.
 A single Autonomous System number may not be used with two E-PSPs
 at the same time as long as the two E-PSP's belong to the same
 NSS. Otherwise the same Autonomous System number can be used from
 multiple points of attachment to the backbone and therefore can
 talk to more than one E-PSP. However, this may result in
 suboptimal routing unless multiple announcements are properly
 engineered according to [1].
 The administrator of the regional networks should be warned that
 improper routing implementations within the region may create
 suboptimal regional routing by using this restriction if no care
 is taken in that:
 Only networks belonging to their own Autonomous System get
 preferred over NSFNET backbone paths; this may extend to a
 larger virtual Autonomous System if backdoor paths are
 effectively implemented.
 IGP implementations should not echo back routing information
 heard via the same path.
 If two regional networks decide to implement a backdoor
 connection between themselves, then the backdoor must have a
 firewall in so that information about their own Autonomous
 System cannot flow in from the other Autonomous System. That
 is, a regional network must not allow information about
 networks that are interior to its Autonomous System to enter
 via exterior routes. Likewise, if a regional network is
 connected to the NSFNET via two NSS connections, the NSS cannot
 send back information about the Autonomous System into the
 Autonomous System where it originated. The end effect is that
 partitions within an Autonomous System will not be healed by
 using the NSS system. In addition, if three or more regionals
 connect to each other via multiple back-door paths, it is
 imperative that all back-door paths have firewalls that ensure
 that the above restrictions are imposed. These actions are
 necessary to prevent routing loops that involve the NSS system.
 Furthermore routing information should only be accepted from
 another regional backbone via backdoor paths for networks which
 are positively desired to be reached via this same backdoor
 path.
3. EGP requirements for attached gateways
 The following EGP requirements are necessary for attached gateways;
 they may require changes in existing vendor products:
 IGP to EGP routing exchanges need to be bidirectional. This
 feature should be selectable by the gateway administrator, and by
 default be configured OFF.
 The metric used when translating from EGP to IGP should be
 configurable.
 It must be possible for IGP information to override EGP
 information, so that the internal paths are preferred over
 external paths. Overriding EGP information on an absolute basis,
 where an external path would never be used as long as there is an
 internal one, is acceptable.
 The ability to do route filtering in the regional gateways on a
 per net basis is highly desirable to allow the regional gateways
 to do a further selection as to what routes they would want to
 redistribute into their network.
 The existence of an EGP connection should optionally lead to the
 generation of a DEFAULT announcement for propagation via the IGP.
 The DEFAULT metric should be independently configurable.
 EGP routes with a metric of "128" should be acceptable. In most
 cases the regional backbone should ignore the EGP metric.
 The regional gateways must only announce networks known to their
 own Autonomous System. At the very least they must not
 redistribute routing information via EGP for routes previously
 learned via EGP.
 It would be beneficial if the regional IGPs would tag routes as
 being EGP derived.
 If the EGP peer (e.g., a NSS) terminates the EGP exchange the
 previously learned routes should expire in a timely fashion.
4. References
 [1] Rekhter, J., "EGP and Policy Based Routing in the New NSFNET
 Backbone", T.J. Watson Research Center, IBM Corporation, March
 1988. Also as RFC 1092, February 1989.
 [2] Mills, D., "Autonomous Confederations", RFC 975, M/A-COM
 Linkabit, February 1986.
 [3] Mills, D., "Exterior Gateway Formal Specification", RFC 904,
 M/A-COM Linkabit, April 1984.
 [4] "Exterior Gateway Protocol, Version 3, Revisions and Extensions,"
 Working Notes of the IETF WG on EGP, Marianne L. Gardner and
 Mike Karels, February 1988.
 [5] "Management and Operation of the NSFNET Backbone Network,"
 proposal to the National Science Foundation, Merit Computer
 Network, August 1987.
5. Appendix
 The following are extensions implemented for the "gated" EGP
 implementation, as designed by Jeff Honig of the Cornell University
 Theory Center. These extensions are still in the design stage and
 may be changed over time. They are included here as an
 implementation example.
 Changes to egpneighbor clause:
 egpneighbor <address> metricin <metric>
 egpmetricout <egpmetric>
 ASin <as>
 ASout <as>
 nogendefault
 acceptdefault
 defaultout <egpmetric>
 validate
 metricin <metric>
 If specified, the metric of all nets received from this
 neighbor are set to <metric>.
 egpmetricout <egpmetric>
 If specified, the metric of all nets sent to this neighbor,
 except default, are set to <egpmetric>.
 ASin <as>
 If specified, EGP packets received from this neighbor must
 specify this AS number of an EGP error packet is generated.
 The AS number is only checked at neighbor acquisition time.
 ASout <as>
 If specified, this AS number is used on all EGP packets sent
 to thiqs neighbor
 nogendefault
 If specified, this neighbor is not considered when
 generating a gateway default.
 acceptdefault
 If specified, the default will be accepted from this
 neighbor, otherwise it will be explicitly ignored.
 defaultout <egpmetric>
 If specified, the internally generated default is send to
 this neighbor in EGP updates. Default learned from other
 gateways is not propogated.
 validate
 If specifed, all nets learned from this EGP neighbor must
 have a corresponding 'validAS' clause or they will be
 ignored.
 Addition of a validAS clause:
 validAS <net> AS <as> metric <metric>
 This clause specifies which AS a network may be learned from and
 what internal metric to use when the net is learned. The
 specifies the 'validate' option. Note that more than one may be
 learned from more than one AS.
 Addition of sendAS and donotsendAS clauses:
 These clauses control the announcement of exterior (currently only
 EGP) routes. Normally, exterior routes are not considered for
 announcement. When the 'sendAS' or 'donotsendAS' clauses are
 used, the announce/donotannounce, egpnetsreachable and other
 restrictions still apply. The 'sendAS' and 'donotsendAS' clauses
 are mutually exclusive by autonomous system.
 sendAS <as0> ASlist <as1> <as2> ...
 This clause specifies that only nets learned from as1, as2, ...
 may be propogated to as0.
 donotsendAS <as0> ASlist <as1> <as2> ...
 This clause specifies that nets learned from as1, as2, ... may
 not be propogated to <as0>, all other nets are propogated.
 An example of a "/etc/gated.conf" file could include the following:
 #
 RIP supplier
 #
 autonomousystem (regional AS)
 #
 egpneighbor (NSS address) ASin (NSS AS) nogendefault
 metricin (metric)
 #
 sendAS (NSS AS) ASlist (regional AS)
 #
 Where:
 Regional AS Is the AS number of the regional network
 NSS address Is the IP address of the local NSS
 NSS AS Is the AS number the NSFNET backbone
 Metric Is the gated internal (time delay) metric that
 EGP learned routes should have. This is the
 metric used on output after conversion to a RIP
 metric. Some values are:
 HELLO RIP
 100 1
 148 2
 219 3
 325 4
 481 5
Author's Address:
 Hans-Werner Braun
 University of Michigan
 Computing Center
 1075 Beal Avenue
 Ann Arbor, MI 48109
 Phone: (313) 763-4897
 Email: HWB@MCR.UMICH.EDU