RFC 1322






Network Working Group                                          D. Estrin
Request for Comments:  1322                                          USC
                                                              Y. Rekhter
                                                                     IBM
                                                                 S. Hotz
                                                                     USC
                                                                May 1992


               A Unified Approach to Inter-Domain Routing

Status of this Memo



   This memo provides information for the Internet community.  It does
   not specify an Internet standard.  Distribution of this memo is
   unlimited.

Abstract



   This memo is an informational RFC which outlines one potential
   approach for inter-domain routing in future global internets.  The
   focus is on scalability to very large networks and functionality, as
   well as scalability, to support routing in an environment of
   heterogeneous services, requirements, and route selection criteria.

   Note: The work of D. Estrin and S. Hotz was supported by the National
   Science Foundation under contract number NCR-9011279, with matching
   funds from GTE Laboratories.  The work of Y. Rekhter was supported by
   the Defense Advanced Research Projects Agency, under contract
   DABT63-91-C-0019.  Views and conclusions expressed in this paper are
   not necessarily those of the Defense Advanced Research Projects
   Agency and National Science Foundation.

1.0 Motivation



   The global internet can be modeled as a collection of hosts
   interconnected via transmission and switching facilities.  Control
   over the collection of hosts and the transmission and switching
   facilities that compose the networking resources of the global
   internet is not homogeneous, but is distributed among multiple
   administrative authorities.  Resources under control of a single
   administration form a domain.  In order to support each domain's
   autonomy and heterogeneity, routing consists of two distinct
   components: intra-domain (interior) routing, and inter-domain
   (exterior) routing.  Intra-domain routing provides support for data
   communication between hosts where data traverses transmission and
   switching facilities within a single domain.  Inter-domain routing
   provides support for data communication between hosts where data



Estrin, Rekhter & Hotz                                          [Page 1]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   traverses transmission and switching facilities spanning multiple
   domains.  The entities that forward packets across domain boundaries
   are called border routers (BRs).  The entities responsible for
   exchanging inter-domain routing information are called route servers
   (RSs).  RSs and BRs may be colocated.

   As the global internet grows, both in size and in the diversity of
   routing requirements, providing inter-domain routing that can
   accommodate both of these factors becomes more and more crucial.  The
   number and diversity of routing requirements is increasing due to:
   (a) transit restrictions imposed by source, destination, and transit
   networks, (b) different types of services offered and required, and
   (c) the presence of multiple carriers with different charging
   schemes.  The combinatorial explosion of mixing and matching these
   different criteria weighs heavily on the mechanisms provided by
   conventional hop-by-hop routing architectures ([ISIS10589, OSPF,
   Hedrick88, EGP]).

   Current work on inter-domain routing within the Internet community
   has diverged in two directions: one is best represented by the Border
   Gateway Protocol (BGP)/Inter-Domain Routeing Protocol (IDRP)
   architectures ([BGP91, Honig90, IDRP91]), and another is best
   represented by the Inter-Domain Policy Routing (IDPR) architecture
   ([IDPR90, Clark90]).  In this paper we suggest that the two
   architectures are quite complementary and should not be considered
   mutually exclusive.

   We expect that over the next 5 to 10 years, the types of services
   available will continue to evolve and that specialized facilities
   will be employed to provide new services.  While the number and
   variety of routes provided by hop-by-hop routing architectures with
   type of service (TOS) support (i.e., multiple, tagged routes) may be
   sufficient for a large percentage of traffic, it is important that
   mechanisms be in place to support efficient routing of specialized
   traffic types via special routes.  Examples of special routes are:
   (1) a route that travels through one or more transit domains that
   discriminate according to the source domain, (2) a route that travels
   through transit domains that support a service that is not widely or
   regularly used.  We refer to all other routes as generic.

   Our desire to support special routes efficiently led us to
   investigate the dynamic installation of routes ([Breslau-Estrin91,
   Clark90, IDPR90]).  In a previous paper ([Breslau-Estrin91]), we
   evaluated the algorithmic design choices for inter-domain policy
   routing with specific attention to accommodating source-specific and
   other "special" routes.  The conclusion was that special routes are
   best supported with source-routing and extended link-state
   algorithms; we refer to this approach as source-demand routing



Estrin, Rekhter & Hotz                                          [Page 2]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   [Footnote:  The Inter-Domain Policy Routing (IDPR) architecture uses
   these techniques.].  However, a source-demand routing architecture,
   used as the only means of inter-domain routing, has scaling problems
   because it does not lend itself to general hierarchical clustering
   and aggregation of routing and forwarding information.  For example,
   even if a particular route from an intermediate transit domain X, to
   a destination domain Y is shared by 1,000 source-domains, IDPR
   requires that state for each of the 1,000 routes be setup and
   maintained in the transit border routers between X and Y.  In
   contrast, an alternative approach to inter-domain routing, based on
   hop-by-hop routing and a distributed route-computation algorithm
   (described later), provides extensive support for aggregation and
   abstraction of reachability, topology, and forwarding information.
   The Border Gateway Protocol (BGP) and Inter-Domain Routeing Protocol
   (IDRP) use these techniques ([BGP91, IDRP91]).  While the BGP/IDRP
   architecture is capable of accommodating very large numbers of
   datagram networks, it does not provide support for specialized
   routing requirements as flexibly and efficiently as IDPR-style
   routing.

1.1 Overview of the Unified Architecture



   We want to support special routes and we want to exploit aggregation
   when a special route is not needed.  Therefore, our scalable inter-
   domain routing architecture consists of two major components:
   source-demand routing (SDR), and node routing (NR).  The NR component
   computes and installs routes that are shared by a significant number
   of sources.  These generic routes are commonly used and warrant wide
   propagation, consequently, aggregation of routing information is
   critical.  The SDR component computes and installs specialized routes
   that are not shared by enough sources to justify computation by NR
   [Footnote: Routes that are only needed sporadically (i.e., the demand
   for them is not continuous or otherwise predictable) are also
   candidates for SDR.].  The potentially large number of different
   specialized routes, combined with their sparse utilization, make them
   too costly to support with the NR mechanism.

   A useful analogy to this approach is the manufacturing of consumer
   products.  When predictable patterns of demand exist, firms produce
   objects and sell them as "off the shelf" consumer goods.  In our
   architecture NR provides off-the-shelf routes.  If demand is not
   predictable, then firms accept special orders and produce what is
   demanded at the time it is needed.  In addition, if a part is so
   specialized that only a single or small number of consumers need it,
   the  consumer may repeatedly special order the part, even if it is
   needed in a predictable manner, because the consumer does not
   represent a big enough market for the producer to bother managing the
   item as part of its regular production.  SDR provides such special



Estrin, Rekhter & Hotz                                          [Page 3]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   order, on-demand routes.

   By combining NR and SDR routing we propose to support inter-domain
   routing in internets of practically-unlimited size, while at the same
   time providing efficient support for specialized routing
   requirements.

   The development of this architecture does assume that routing
   requirements will be diverse and that special routes will be needed.
   On the other hand, the architecture does not depend on assumptions
   about the particular types of routes demanded or on the distribution
   of that demand.  Routing will adapt naturally over time to changing
   traffic patterns and new services by shifting computation and
   installation of particular types of routes between the two components
   of the hybrid architecture [Footnote: Before continuing with our
   explanation of this architecture, we wish to state up front that
   supporting highly specialized routes for all source-destination pairs
   in an internet, or even anything close to that number, is not
   feasible in any routing architecture that we can foresee.  In other
   words, we do not believe that any foreseeable routing architecture
   can support unconstrained proliferation of user requirements and
   network services.  At the same time, this is not necessarily a
   problem.  The capabilities of the architecture may in fact exceed the
   requirements of the users.  Moreover, some of the requirements that
   we regard as infeasible from the inter-domain routing point of view,
   may be supported by means completely outside of routing.
   Nevertheless, the caveat is stated here to preempt unrealistic
   expectations.].

   While the packet forwarding functions of the NR and SDR components
   have little or no coupling with each other, the connectivity
   information exchange mechanism of the SDR component relies on
   services provided by the NR component.

1.2 Outline



   The remainder of this report is organized as follows.  Section 2
   outlines the requirements and priorities that guide the design of the
   NR and SDR components.  Sections 3 and 4 describe the NR and SDR
   design choices, respectively, in light of these requirements.
   Section 5 describes protocol support for the unified architecture and
   briefly discusses transition issues.  We conclude with a brief
   summary.








Estrin, Rekhter & Hotz                                          [Page 4]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


2.0 Architectural Requirements and Priorities



   In order to justify our design choices for a scalable inter-domain
   routing architecture, we must articulate our evaluation criteria and
   priorities.  This section defines complexity, abstraction, policy,
   and type of service requirements.

2.1 Complexity



   Inter-domain routing complexity must be evaluated on the basis of the
   following performance metrics: (1) storage overhead, (2)
   computational overhead, and (3) message overhead.  This evaluation is
   essential to determining the scalability of any architecture.

2.1.1 Storage Overhead



   The storage overhead of an entity that participates in inter-domain
   routing comes from two sources: Routing Information Base (RIB), and
   Forwarding Information Base (FIB) overhead.  The RIB contains the
   routing information that entities exchange via the inter-domain
   routing protocol; the RIB is the input to the route computation.  The
   FIB contains the information that the entities use to forward the
   inter-domain traffic; the FIB is the output of the route computation.
   For an acceptable level of storage overhead, the amount of
   information in both FIBs and RIBs should grow significantly slower
   than linearly (e.g., close to logarithmically) with the total number
   of domains in an internet.  To satisfy this requirement with respect
   to the RIB, the architecture must provide mechanisms for either
   aggregation and abstraction of routing and forwarding information, or
   retrieval of a subset of this information on demand.  To satisfy this
   requirement with respect to the FIB, the architecture must provide
   mechanisms for either aggregation of the forwarding information (for
   the NR computed routes), or dynamic installation/tear down of this
   information (for the SDR computed routes).

   Besides being an intrinsically important evaluation metric, storage
   overhead has a direct impact on computational and bandwidth
   complexity.  Unless the computational complexity is fixed (and
   independent of the total number of domains), the storage overhead has
   direct impact on the computational complexity of the architecture
   since the routing information is used as an input to route
   computation. Moreover, unless the architecture employs incremental
   updates, where only changes to the routing information are
   propagated, the storage overhead has direct impact on the bandwidth
   overhead of the architecture since the exchange of routing
   information constitutes most of the bandwidth overhead.





Estrin, Rekhter & Hotz                                          [Page 5]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


2.1.2 Computational Overhead



   The NR component will rely primarily on precomputation of routes.  If
   inter-domain routing is ubiquitous, then the precomputed routes
   include all reachable destinations.  Even if policy constraints make
   fully ubiquitous routing impossible, the precomputed routes are
   likely to cover a very large percentage of all reachable
   destinations.  Therefore the complexity of this computation must be
   as small as possible.  Specifically, it is highly desirable that the
   architecture would employ some form of partial computation, where
   changes in topology would require less than complete recomputation.
   Even if complete recomputation is necessary, its complexity should be
   less than linear with the total number of domains.

   The SDR component will use on-demand computation and caching.
   Therefore the complexity of this computation can be somewhat higher.
   Another reason for relaxed complexity requirements for SDR is that
   SDR is expected to compute routes to a smaller number of destinations
   than is NR (although SDR route computation may be invoked more
   frequently).

   Under no circumstances is computational complexity allowed to become
   exponential (for either the NR or SDR component).

2.1.3 Bandwidth Overhead



   The bandwidth consumed by routing information distribution should be
   limited.  However, the possible use of data compression techniques
   and the increasing speed of network links make this less important
   than route computation and storage overhead.  Bandwidth overhead may
   be further contained by using incremental (rather than complete)
   exchange of routing information.

   While storage and bandwidth overhead may be interrelated, if
   incremental updates are used then bandwidth overhead is negligible in
   the steady state (no changes in topology), and is independent of the
   storage overhead.  In other words, use of incremental updates
   constrains the bandwidth overhead to the dynamics of the internet.
   Therefore, improvements in stability of the physical links, combined
   with techniques to dampen the effect of topological instabilities,
   will make the bandwidth overhead even less important.

2.2 Aggregation



   Aggregation and abstraction of routing and forwarding information
   provides a very powerful mechanism for satisfying storage,
   computational, and bandwidth constraints.  The ability to aggregate,
   and subsequently abstract, routing and forwarding information is



Estrin, Rekhter & Hotz                                          [Page 6]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   essential to the scaling of the architecture [Footnote: While we can
   not prove that there are no other ways to achieve scaling, we are not
   aware of any mechanism other than clustering that allows information
   aggregation/abstraction.  Therefore, the rest of the paper assumes
   that clustering is used for information aggregation/abstraction.].
   This is especially true with respect to the NR component, since the
   NR component must be capable of providing routes to all or almost all
   reachable destinations.

   At the same time, since preserving each domain's independence and
   autonomy is one of the crucial requirements of inter-domain routing,
   the architecture must strive for the maximum flexibility of its
   aggregation scheme, i.e., impose as few preconditions, and as little
   global coordination, as possible on the participating domains.

   The Routing Information Base (RIB) carries three types of
   information: (1) topology (i.e., the interconnections between domains
   or groups of domains), (2) network layer reachability, and (3)
   transit constraint.  Aggregation of routing information should
   provide a reduction of all three components.  Aggregation of
   forwarding information will follow from reachability information
   aggregation.

   Clustering (by forming routing domain confederations) serves the
   following aggregation functions: (1) to hide parts of the actual
   physical topology, thus abstracting topological information, (2) to
   combine a set of reachable destination entities into a single entity
   and reduce storage overhead, and (3) to express transit constraints
   in terms of clusters, rather than individual domains.

   As argued in [Breslau-Estrin91], the architecture must allow
   confederations to be formed and changed without extensive
   configuration and coordination; in particular, forming a
   confederation should not require global coordination (such as that
   required in ECMA ([ECMA89]).  In addition, aggregation should not
   require explicit designation of the relative placement of each domain
   relative to another; i.e., domains or confederations of domains
   should not be required to agree on a partial ordering (i.e., who is
   above whom, etc.).












Estrin, Rekhter & Hotz                                          [Page 7]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   The architecture should allow different domains to use different
   methods of aggregation and abstraction.  For example, a research
   collaborator at IBM might route to USC as a domain-level entity in
   order to take advantage of some special TOS connectivity to, or even
   through, USC.  Whereas, someone else at Digital Equipment Corporation
   might see information at the level of the California Educational
   Institutions Confederation, and know only that USC is a member.
   Alternatively, USC might see part of the internal structure within
   the IBM Confederation (at the domain's level), whereas UCLA may route
   based on the confederation of IBM domains as a whole.

   Support for confederations should be flexible.  Specifically, the
   architecture should allow confederations to overlap without being
   nested, i.e., a single domain, or a group of domains may be part of
   more than one confederation.  For example, USC may be part of the
   California Educational Institutions Confederation and part of the US
   R&D Institutions Confederation; one is not a subset of the other.
   Another example: T.J.  Watson Research Center might be part of
   NYSERNET Confederation and part of IBM-R&D-US Confederation.  While
   the above examples describe cases where overlap consists of a single
   domain, there may be other cases where multiple domains overlap.  As
   an example consider the set of domains that form the IBM
   Confederation, and another set of domains that form the DEC
   Confederation.  Within IBM there is a domain IBM-Research, and
   similarly within DEC there is a domain DEC-Research.  Both of these
   domains could be involved in some collaborative effort, and thus have
   established direct links.  The architecture should allow restricted
   use of these direct links, so that other domains within the IBM
   Confederation would not be able to use it to talk to other domains
   within the DEC Confederation.  A similar example exists when a
   multinational corporation forms a confederation, and the individual
   branches within each country also belong to their respective country
   confederations.  The corporation may need to protect itself from
   being used as an inter-country transit domain (due to internal, or
   international, policy).  All of the above examples illustrate a
   situation where confederations overlap, and it is necessary to
   control the traffic traversing the overlapping resources.

   While flexible aggregation should be accommodated in any inter-domain
   architecture, the extent to which this feature is exploited will have
   direct a effect on the scalability associated with aggregation.  At
   the same time, the exploitation of this feature depends on the way
   addresses are assigned.  Specifically, scaling associated with
   forwarding information depends heavily on the assumption that there
   will be general correspondence between the hierarchy of address
   registration authorities, and the way routing domains and routing
   domain confederations are organized (see Section 2.6).




Estrin, Rekhter & Hotz                                          [Page 8]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


2.3 Routing Policies



   Routing policies that the architecture must support may be broadly
   classified into transit policies and route selection policies
   [Breslau-Estrin 91, Estrin89].

   Restrictions imposed via transit policies may be based on a variety
   of factors.  The architecture should be able to support restrictions
   based on the source, destination, type of services (TOS), user class
   identification (UCI), charging, and path [Estrin89 , Little89].  The
   architecture must allow expression of transit policies on all routes,
   both NR and SDR.  Even if NR routes are widely used and have fewer
   source or destination restrictions, NR routes may have some transit
   qualifiers (e.g., TOS, charging, or user-class restriction).  If the
   most widely-usable route to a destination has policy qualifiers, it
   should be advertiseable by NR and the transit constraints should be
   explicit.

   Route selection policies enable each domain to select a particular
   route among multiple routes to the same destination.  To maintain
   maximum autonomy and independence between domains, the architecture
   must support heterogeneous route selection policies, where each
   domain can establish its own criteria for selecting routes.  This
   argument was made earlier with respect to source route selection
   ([IDPR90, Clark90, Breslau-Estrin91]).  In addition, each
   intermediate transit domain must have the flexibility to apply its
   own selection criteria to the routes made available to it by its
   neighbors.  This is really just a corollary to the above; i.e., if we
   allow route selection policy to be expressed for NR routes, we can
   not assume all domains will apply the same policy.  The support for
   dissimilar route selection policies is one of the key factors that
   distinguish inter-domain routing from intra-domain routing ([ECMA89,
   Estrin89]).  However, it is a non-goal of the architecture to support
   all possible route selection policies.  For more on unsupported route
   selection policies see Section 2.3.2 below.

2.3.1 Routing Information Hiding



   The architecture should not require all domains within an internet to
   reveal their connectivity and transit constraints to each other.
   Domains should be able to enforce their transit policies while
   limiting the advertisement of their policy and connectivity
   information as much as possible; such advertisement will be driven by
   a "need to know" criteria.  Moreover, advertising a transit policy to
   domains that can not use this policy will increase the amount of
   routing information that must be stored, processed, and propagated.
   Not only may it be impractical for a router to maintain full inter-
   domain topology and policy information, it may not be permitted to



Estrin, Rekhter & Hotz                                          [Page 9]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   obtain this information.

2.3.2 Policies Not Supported



   In this and previous papers we have argued that a global inter-domain
   routing architecture should support a wide range of policies.  In
   this section we identify several types of policy that we explicitly
   do not propose to support.  In general our reasoning is pragmatic; we
   think such policy types are either very expensive in terms of
   overhead, or may lead to routing instabilities.

   1. Route selection policies contingent on external behavior.
      The route selection process may be modeled by a function that
      assigns a non-negative integer to a route, denoting the degree
      of preference.  Route selection applies this function to all
      feasible routes to a given destination, and selects the route
      with the highest value.  To provide a stable environment, the
      preference function should not use as an input the status and
      attributes of other routes (either to the same or to a
      different destination).

   2. Transit policies contingent on external behavior.
      To provide a stable environment, the domain's transit policies
      can not be automatically affected by any information external
      to the domain.  Specifically, these policies can not be modified,
      automatically, in response to information about other domains'
      transit policies, or routes selected by local or other domains.
      Similarly, transit policies can not be automatically modified
      in response to information about performance characteristics of
      either local or external domains.

   3. Policies contingent on external state (e.g., load).
      It is a non-goal of the architecture to support load-sensitive
      routing for generic routes.  However, it is possible that some
      types of service may employ load information to select among
      alternate SDR routes.

   4. Very large number of simultaneous SDR routes.
      It is a non-goal of the architecture to support a very large
      number of simultaneous SDR routes through any single router.
      Specifically, the FIB storage overhead associated with SDR
      routes must be comparable with that of NR routes, and should
      always be bound by the complexity requirements outlined earlier
      [Foonote: As discussed earlier, theoretically the state overhead
      could grow O(N^2) with the number of domains in an internet.
      However, there is no evidence or intuition to suggest that
      this will be a limiting factor on the wide utilization of SDR,
      provided that NR is available to handle generic routes.].



Estrin, Rekhter & Hotz                                         [Page 10]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


2.4 Support for TOS Routing



   Throughout this document we refer to support for type of service
   (TOS) routing.  There is a great deal of research and development
   activity currently underway to explore network architectures and
   protocols for high-bandwidth, multimedia traffic.  Some of this
   traffic is delay sensitive, while some requires high throughput.  It
   is unrealistic to assume that a single communication fabric will be
   deployed homogeneously across the internet (including all
   metropolitan, regional, and backbone networks) that will support all
   types of traffic uniformly.  To support diverse traffic requirements
   in a heterogeneous environment, various resource management
   mechanisms will be used in different parts of the global internet
   (e.g., resource reservation of various kinds) [ST2-90, Zhang91].

   In this context, it is the job of routing protocols to locate routes
   that can potentially support the particular TOS requested.  It is
   explicitly not the job of the general routing protocols to locate
   routes that are guaranteed to have resources available at the
   particular time of the route request.  In other words, it is not
   practical to assume that instantaneous resource availability will be
   known at all remote points in the global internet.  Rather, once a
   TOS route has been identified, an application requiring particular
   service guarantees will attempt to use the route (e.g., using an
   explicit setup message if so required by the underlying networks).
   In Section 4 we describe additional services that may be provided to
   support more adaptive route selection for special TOS [Footnote:
   Adaptive route selection implies adaptability only during the route
   selection process.  Once a route is selected, it is not going to be a
   subject to any adaptations, except when it becomes infeasible.].

2.5 Commonality between Routing Components



   While it is acceptable for the NR and SDR components to be
   dissimilar, we do recognize that such a solution is less desirable --
   all other things being equal.  In theory, there are advantages in
   having the NR and SDR components use similar algorithms and
   mechanisms.  Code and databases could be shared and the architecture
   would be more manageable and comprehensible.  On the other hand,
   there may be some benefit (e.g., robustness) if the two parts of the
   architecture are heterogeneous, and not completely inter-dependent.
   In Section 5 we list several areas in which opportunities for
   increased commonality (unification) will be exploited.

2.6 Interaction with Addressing



   The proposed architecture should be applicable to various addressing
   schemes.  There are no specific assumptions about a particular



Estrin, Rekhter & Hotz                                         [Page 11]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   address structure, except that this structure should facilitate
   information aggregation, without forcing rigid hierarchical routing.

   Beyond this requirement, most of the proposals for extending the IP
   address space, for example, can be used in conjunction with our
   architecture.  But our architecture itself does not provide (or
   impose) a particular solution to the addressing problem.












































Estrin, Rekhter & Hotz                                         [Page 12]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


3.0 Design Choices for Node Routing (NR)



   This section addresses the design choices made for the NR component
   in light of the above architectural requirements and priorities.  All
   of our discussion of NR assumes hop-by-hop routing.  Source routing
   is subject to different constraints and is used for the complementary
   SDR component.

3.1 Overview of NR



   The NR component is designed and optimized for an environment where a
   large percentage of packets will travel over routes that can be
   shared by multiple sources and that have predictable traffic
   patterns.  The efficiency of the NR component improves when a number
   of source domains share a particular route from some intermediate
   point to a destination.  Moreover, NR is best suited to provide
   routing for inter-domain data traffic that is either steady enough to
   justify the existence of a route, or predictable, so that a route may
   be installed when needed (based on the prediction, rather than on the
   actual traffic).  Such routes lend themselves to distributed route
   computation and installation procedures.

   Routes that are installed in routers, and information that was used
   by the routers to compute these routes, reflect the known operational
   state of the routing facilities (as well as the policy constraints)
   at the time of route computation.  Route computation is driven by
   changes in either operational status of routing facilities or policy
   constraints.  The NR component supports route computation that is
   dynamically adaptable to both changes in topology and policy.  The NR
   component allows time-dependent selection or deletion of routes.
   However, time dependency must be predictable (e.g., advertising a
   certain route only after business hours) and routes should be used
   widely enough to warrant inclusion in NR.

   The proposed architecture assumes that most of the inter-domain
   conversations will be forwarded via routes computed and installed by
   the NR component.

   Moreover, the exchange of routing information necessary for the SDR
   component depends on facilities provided by the NR component; i.e.,
   NR policies must allow SDR reachability information to travel.
   Therefore, the architecture requires that all domains in an internet
   implement and participate in NR.  Since scalability (with respect to
   the size of the internet) is one of the fundamental requirements for
   the NR component, it must provide multiple mechanisms with various
   degrees of sophistication for information aggregation and
   abstraction.




Estrin, Rekhter & Hotz                                         [Page 13]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   The potential reduction of routing and forwarding information depends
   very heavily on the way addresses are assigned and how domains and
   their confederations are structured.  "If there is no correspondence
   between the address registration hierarchy and the organisation of
   routeing domains, then ... each and every routeing domain in the OSIE
   would need a table entry potentially at every boundary IS of every
   other routeing domain" ([Oran89]).  Indeed, scaling in the NR
   component is almost entirely predicated on the assumption that there
   will be general correspondence between the hierarchy of address
   assignment authorities and the way routing domains are organised, so
   that the efficient and frequent aggregation of routing and forwarding
   information will be possible.  Therefore, given the nature of inter-
   domain routing in general, and the NR component in particular,
   scalability of the architecture depends very heavily on the
   flexibility of the scheme for information aggregation and
   abstraction, and on the preconditions that such a scheme imposes.
   Moreover, given a flexible architecture, the operational efficiency
   (scalability) of the global internet, or portions thereof, will
   depend on tradeoffs made between flexibility and aggregation.

   While the NR component is optimized to satisfy the common case
   routing requirements for an extremely large population of users, this
   does not imply that routes produced by the NR component would not be
   used for different types of service (TOS).  To the contrary, if a TOS
   becomes sufficiently widely used (i.e., by multiple domains and
   predictably over time), then it may warrant being computed by the NR
   component.

   To summarize, the NR component is best suited to provide routes that
   are used by more than a single domain.  That is, the efficiency of
   the NR component improves when a number of source domains share a
   particular route from some intermediate point to a destination.
   Moreover, NR is best suited to provide routing for inter-domain data
   traffic that is either steady enough to justify the existence of a
   route, or predictable, so that a route may be installed when needed,
   (based on the prediction, rather than on the actual traffic).

3.2 Routing Algorithm Choices for NR



   Given that a NR component based on hop-by-hop routing is needed to
   provide scalable, efficient inter-domain routing, the remainder of
   this section considers the fundamental design choices for the NR
   routing algorithm.

   Typically the debate surrounding routing algorithms focuses on link
   state and distance vector protocols.  However, simple distance vector
   protocols (i.e., Routing Information Protocol [Hedrick88]), do not
   scale because of convergence problems.  Improved distance vector



Estrin, Rekhter & Hotz                                         [Page 14]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   protocols, such as those discussed in [Jaffee82, Zaumen91, Shin87],
   have been developed to address this issue using synchronization
   mechanisms or additional path information.  In the case of inter-
   domain routing, having additional path information is essential to
   supporting policy.  Therefore, the algorithms we consider for NR are
   link state and one we call path vector (PV).  Whereas the
   characteristics of link state algorithms are generally understood
   (for example, [Zaumen 91]), we must digress from our evaluation
   discussion to describe briefly the newer concept of the PV algorithm
   [Footnote: We assume that some form of SPF algorithm will be used to
   compute paths over the topology database when LS algorithms are used
   [Dijkstra59, OSPF]].

3.2.1 Path Vector Protocol Overview



   The Border Gateway Protocol (BGP) (see [BGP91]) and the Inter Domain
   Routing Protocol (IDRP) (see [IDRP91]) are examples of path vector
   (PV) protocols [Footnote: BGP is an inter-autonomous system routing
   protocol for TCP/IP internets.  IDRP is an OSI inter-domain routing
   protocol that is being progressed toward standardization within ISO.
   Since in terms of functionality BGP represents a proper subset of
   IDRP, for the rest of the paper we will only consider IDRP.].

   The routing algorithm employed by PV bears a certain resemblance to
   the traditional Bellman-Ford routing algorithm in the sense that each
   border router advertises the destinations it can reach to its
   neighboring BRs.  However,the PV routing algorithm augments the
   advertisement of reachable destinations with information that
   describes various properties of the paths to these destinations.

   This information is expressed in terms of path attributes.  To
   emphasize the tight coupling between the reachable destinations and
   properties of the paths to these destinations, PV defines a route as
   a pairing between a destination and the attributes of the path to
   that destination.  Thus the name, path-vector protocol, where a BR
   receives from its neighboring BR a vector that contains paths to a
   set of destinations [Footnote: The term "path-vector protocol" bears
   an intentional similarity to the term "distance-vector protocol",
   where a BR receives from each of its neighbors a vector that contains
   distances to a set of destinations.].  The path, expressed in terms
   of the domains (or confederations) traversed so far, is carried in a
   special path attribute which records the sequence of routing domains
   through which the reachability information has passed.  Suppression
   of routing loops is implemented via this special path attribute, in
   contrast to LS and distance vector which use a globally-defined
   monotonically-increasing metric for route selection [Shin87].

   Because PV does not require all routing domains to have homogeneous



Estrin, Rekhter & Hotz                                         [Page 15]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   criteria (policies) for route selection, route selection policies
   used by one routing domain are not necessarily known to other routing
   domains.  To maintain the maximum degree of autonomy and independence
   between routing domains, each domain which participates in PV may
   have its own view of what constitutes an optimal route.  This view is
   based solely on local route selection policies and the information
   carried in the path attributes of a route.  PV standardizes only the
   results of the route selection procedure, while allowing the
   selection policies that affect the route selection to be non-standard
   [Footnote: This succinct observation is attributed to Ross Callon
   (Digital Equipment Corporation).].

3.3 Complexity



   Given the above description of PV algorithms, we can compare them to
   LS algorithms in terms of the three complexity parameters defined
   earlier.

3.3.1 Storage Overhead



   Without any aggregation of routing information, and ignoring storage
   overhead associated with transit constraints, it is possible to show
   that under some rather general assumptions the average case RIB
   storage overhead of the PV scheme for a single TOS ranges between
   O(N) and O(Nlog(N)), where N is the total number of routing domains
   ([Rekhter91]).  The LS RIB, with no aggregation of routing
   information, no transit constraints, a single homogeneous route
   selection policy across all the domains, and a single TOS, requires a
   complete domain-level topology map whose size is O(N).

   Supporting heterogeneous route selection and transit policies with
   hop-by-hop forwarding and LS requires each domain to know all other
   domains route selection and transit policies.  This may significantly
   increase the amount of routing information that must be stored by
   each domain.  If the number of policies advertised grows with the
   number of domains, then the storage could become unsupportable.  In
   contrast, support for heterogeneous route selection policies has no
   effect on the storage complexity of the PV scheme (aside from simply
   storing the local policy information).  The presence of transit
   constraints in PV results in a restricted distribution of routing
   information, thus further reducing storage overhead.  In contrast,
   with LS no such reduction is possible since each domain must know
   every other domain's transit policies.  Finally, some of the transit
   constraints (e.g., path sensitive constraints) can be expressed in a
   more concise form in PV (see aggregation discussion below).

   The ability to further restrict storage overhead is facilitated by
   the PV routing algorithm, where route computation precedes routing



Estrin, Rekhter & Hotz                                         [Page 16]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   information dissemination, and only routing information associated
   with the routes selected by a domain is distributed to adjacent
   domains.  In contrast, route selection with LS is decoupled from the
   distribution of routing information, and has no effect on such
   distribution.

   While theoretically routing information aggregation can be used to
   reduce storage complexity in both LS and PV, only aggregation of
   topological information would yield the same gain for both.
   Aggregating transit constraints with LS may result in either reduced
   connectivity or less information reduction, as compared with PV.
   Aggregating heterogeneous route selection policies in LS is highly
   problematic, at best.  In PV, route selection policies are not
   distributed, thus making aggregation of route selection policies a
   non-issue [Footnote: Although a domain's selection policies are not
   explicitly distributed, they have an impact on the routes available
   to other domains.  A route that may be preferred by a particular
   domain, and not prohibited by transit restrictions, may still be
   unavailable due to the selection policies of some intermediate
   domain.  The ability to compute and install alternative routes that
   may be lost using hop-by-hop routing (either LS of PV) is the
   critical functionality provided by SDR.].

   Support for multiple TOSs has the same impact on storage overhead for
   both LS and for PV.

   Potentially the LS FIB may be smaller if routes are computed at each
   node on demand.  However, the gain of such a scheme depends heavily
   on the traffic patterns, memory size, and caching strategy.  If there
   is not a high degree of locality, severely degraded performance can
   result due to excessive overall computation time and excessive
   computation delay when forwarding packets to a new destination.  If
   demand driven route computation is not used for LS, then the size of
   the FIB would be the same for both LS and PV.

















Estrin, Rekhter & Hotz                                         [Page 17]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


3.3.2 Route Computation Complexity



   Even if all domains employ exactly the same route selection policy,
   computational complexity of PV is smaller than that of LS.  The PV
   computation consists of evaluating a newly arrived route and
   comparing it with the existing one [Footnote: Some additional checks
   are required when an update is received to insure that a routing loop
   has not been created.].  Whereas, conventional LS computation
   requires execution of an SPF algorithm such as Dijkstra's.

   With PV, topology changes only result in the recomputation of routes
   affected by these changes, which is more efficient than complete
   recomputation.  However, because of the inclusion of full path
   information with each distance vector, the effect of a topology
   change may propagate farther than in traditional distance vector
   algorithms.  On the other hand, the number of affected domains will
   still be smaller with PV than with conventional LS hop-by-hop
   routing.  With PV, only those domains whose routes are affected by
   the changes have to recompute, while with conventional LS hop-by-hop
   routing all domains must recompute.  While it is also possible to
   employ partial recomputation with LS (i.e., when topology changes,
   only the affected routes are recomputed), "studies suggest that with
   a very small number of link changes (perhaps 2) the expected
   computational complexity of the incremental update exceeds the
   complete recalculation" ([ANSI87-150R]).  However these checks are
   much simpler than executing a full SPF algorithm.

   Support for heterogeneous route selection policies has serious
   implications for the computational complexity.  The major problem
   with supporting heterogeneous route selection policies with LS is the
   computational complexity of the route selection itself.
   Specifically, we are not aware of any algorithm with less than
   exponential time complexity that would be capable of computing routes
   to all destinations, with LS hop-by-hop routing and heterogeneous
   route selection policies.  In contrast, PV allows each domain to make
   its route selection autonomously, based only on local policies.
   Therefore support for dissimilar route selection policies has no
   negative implications for the complexity of route computation in PV.
   Similarly, providing support for path-sensitive transit policies in
   LS implies exponential computation, while in PV such support has no
   impact on the complexity of route computation.

   In summary, because NR will rely primarily on precomputation of
   routes, aggregation is essential to the long-term scalability of the
   architecture.  To the extent aggregation is facilitated with PV, so
   is reduced computational complexity.  While similar arguments may be
   made for LS, the aggregation capabilities that can be achieved with
   LS are more problematic because of LS' reliance on consistent



Estrin, Rekhter & Hotz                                         [Page 18]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   topology maps at each node.  It is also not clear what additional
   computational complexity will be associated with aggregation of
   transit constraints and heterogeneous route selection policies in LS.

3.3.3 Bandwidth Overhead



   PV routing updates include fully-expanded information.  A complete
   route for each supported TOS is advertised.  In LS, TOS only
   contributes a factor increase per link advertised, which is much less
   than the number of complete routes.  Although TOSs may be encoded
   more efficiently with LS than with PV, link state information is
   flooded to all domains, while with PV, routing updates are
   distributed only to the domains that actually use them.  Therefore,
   it is difficult to make a general statement about which scheme
   imposes more bandwidth overhead, all other factors being equal.

   Moreover, this is perhaps really not an important difference, since
   we are more concerned with the number of messages than with the
   number of bits (because of compression and greater link bandwidth, as
   well as the increased physical stability of links).

3.4 Aggregation



   Forming confederations of domains, for the purpose of consistent,
   hop-by-hop, LS route computation, requires that domains within a
   confederation have consistent policies.  In addition, LS
   confederation requires that any lower level entity be a member of
   only one higher level entity.  In general, no intra-confederation
   information can be made visible outside of a confederation, or else
   routing loops may occur as a result of using an inconsistent map of
   the network at different domains.  Therefore, the use of
   confederations with hop-by-hop LS is limited because each domain (or
   confederation) can only be a part of one higher level confederation
   and only export policies consistent with that confederation (see
   examples in Section 2.2).  These restrictions are likely to impact
   the scaling capabilities of the architecture quite severely.

   In comparison, PV can accommodate different confederation definitions
   because looping is avoided by the use of full path information.
   Consistent network maps are not needed at each route server, since
   route computation precedes routing information dissemination.
   Consequently, PV confederations can be nested, disjoint, or
   overlapping.  A domain (or confederation) can export different
   policies or TOS as part of different confederations, thus providing
   the extreme flexibility that is crucial for the overall scaling and
   extensibility of the architecture.

   In summary, aggregation is essential to achieve acceptable complexity



Estrin, Rekhter & Hotz                                         [Page 19]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   bounds, and flexibility is essential to achieve acceptable
   aggregation across the global, decentralized internet.  PV's
   strongest advantage stems from its flexibility.

3.5 Policy



   The need to allow expression of transit policy constraints on any
   route (i.e., NR routes as well as SDR routes), by itself, can be
   supported by either LS or PV.  However, the associated complexities
   of supporting transit policy constraints are noticeably higher for LS
   than for PV.  This is due to the need to flood all transit policies
   with LS, where with PV transit policies are controlled via restricted
   distribution of routing information.  The latter always imposes less
   overhead than flooding.

   While all of the transit constraints that can be supported with LS
   can be supported with PV, the reverse is not true.  In other words,
   there are certain transit constraints (e.g., path-sensitive transit
   constraints) that are easily supported with PV, and are prohibitively
   expensive (in terms of complexity) to support in LS.  For example, it
   is not clear how NR with LS could support something like ECMA-style
   policies that are based on hierarchical relations between domains,
   while support for such policies is straightforward with PV.

   As indicated above, support for heterogeneous route selection
   policies, in view of its computational and storage complexity, is
   impractical with LS hop-by-hop routing.  In contrast, PV can
   accommodate heterogeneous route selection with little additional
   overhead.

3.6 Information Hiding



   PV has a clear advantage with respect to selective information
   hiding.  LS with hop-by-hop routing hinges on the ability of all
   domains to have exactly the same information; this clearly
   contradicts the notion of selective information hiding.  That is, the
   requirement to perform selective information hiding is unsatisfiable
   with LS hop-by-hop routing.

3.7 Commonality between NR and SDR Components



   In [Breslau-Estrin91] we argued for the use of LS in conjunction with
   SDR.  Therefore there is some preference for using LS with NR.
   However, there are several reasons why NR and SDR would not use
   exactly the same routing information, even if they did use the same
   algorithm.  Moreover, there are several opportunities for unifying
   the management (distribution and storage) of routing and forwarding
   information, even if dissimilar algorithms are used.



Estrin, Rekhter & Hotz                                         [Page 20]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   In considering the differences between NR and SDR we must address
   several areas:

     1. Routing information and distribution protocol: LS for SDR is
        quite different from the LS in NR.  For example, SDR LS need
        not aggregate domains; to the contrary SDR LS  requires detailed
        information to generate special routes.

        In addition, consistency requirements (essential for NR) are
        unnecessary for the SDR component.  Therefore LS information for
        the SDR component can be retrieved on-demand, while the NR
        component must use flooding of topology information.

     2. Route computation algorithm: It is not clear whether route
        computation algorithm(s)  can be shared between the SDR and NR
        components, given the difficulty of supporting  heterogeneous
        route selection policies in NR.

     3. Forwarding information: The use of dissimilar route computation
        algorithms does not preclude common handling of packet
        forwarding.  Even if LS were used for NR, the requirement would
        be the same, i.e., that the forwarding agent can determine
        whether to use a NR  precomputed route or an SDR installed route
        to forward a particular data packet.

   In conclusion, using similar algorithms and mechanisms for SDR and NR
   components would have benefits.  However, these benefits do not
   dominate the other factors as discussed before.























Estrin, Rekhter & Hotz                                         [Page 21]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


3.8 Summary



   Given the performance complexity issues associated with global
   routing, aggregation of routing information is essential; at the same
   time we have argued that such aggregation must be flexible.  Given
   the difficulties of supporting LS hop-by-hop routing in the presence
   of (a) flexible aggregation, (b) heterogeneous route selection
   policies, and (c) incomplete or inconsistent routing information, we
   see no alternative but to employ PV for the NR component of our
   architecture.

   Based on the above tradeoffs, our NR component employs a PV
   architecture, where route computation and installation is done in a
   distributed fashion by the routers participating in the NR component
   [Footnote: Packet forwarding along routes produced by the NR
   component can be accomplished by either source routing or hop-by-hop
   routing.  The latter is the primary choice because it reduces the
   amount of state in routers (if route setup is employed), or avoids
   encoding an explicit source route in network layer packets.  However,
   the architecture does not preclude the use of source routing (or
   route setup) along the routes computed, but not installed, by the NR
   component.].

   The distributed algorithm combines some of the features of link state
   with those of distance vector algorithms; in addition to next hop
   information, the NR component maintains path attributes for each
   route (e.g., the list of domains or routing domain confederations
   that the routing information has traversed so far).  The path
   attributes that are carried along with a route express a variety of
   routing policies, and make explicit the entire route to the
   destination.  With aggregation, this is a superset of the domains
   that form the path to the destination.



















Estrin, Rekhter & Hotz                                         [Page 22]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


4.0 Source-demand routing (SDR)



   Inter-domain routers participating in the SDR component forward
   packets according to routing information computed and installed by
   the domain that originates the traffic (source routing domain).

   It is important to realize that requiring route installation by the
   source routing domain is not a matter of choice, but rather a
   necessity.  If a particular route is used by a small number of
   domains (perhaps only one) then it is more appropriate to have the
   source compute and install the special route instead of burdening the
   intermediate nodes with the task of looking for and selecting a route
   with the specialized requirements.  In addition, if the demand for
   the route is unpredictable, and thus can be determined only by the
   source, it should be up to the source to install the route.

   In general, information that is used by source routing domains for
   computing source-demand routes reflects administrative (but not
   operational) status of the routing facilities (i.e., configured
   topology and policy) [Footnote: If SDR uses NR information then
   operational status could be considered in some route selection.].
   Consequently, it is possible for a source routing domain to compute a
   route that is not operational at route installation time.  The SDR
   component attempts to notify the source domain of failures when route
   installation is attempted.  Similarly, the SDR component provides
   mechanisms for the source routing domain to be notified of failures
   along previously-installed active routes.  In other words, the SDR
   component performs routing that is adaptive to topological changes;
   however, the adaptability is achieved as a consequence of the route
   installation and route management mechanisms.  This is different from
   the NR component, where status changes are propagated and
   incorporated by nodes as soon as possible.  Therefore, to allow
   faster adaptation to changes in the operational status of routing
   facilities, the SDR component allows the source domain to switch to a
   route computed by the NR component, if failure along the source-
   demand route is detected (either during the route installation phase,
   or after the route is installed), and if policy permits use of the NR
   route.

   The NR component will group domains into confederations to achieve
   its scaling goals (see [IDRP91]).  In contrast, SDR will allow an
   AD-level route to be used by an individual domain without allowing
   use by the entire confederation to which the domain belongs.
   Similarly, a single transit domain may support a policy or special
   TOS that is not supported by other domains in its confederation(s).
   In other words, the architecture uses SDR to support non-
   hierarchical, non-aggregated policies where and when needed.
   Consequently, SDR by itself does not have the scaling properties of



Estrin, Rekhter & Hotz                                         [Page 23]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   NR.  In compensation, SDR does not require a complete, global domain
   map if portions of the world are never traversed or communicated
   with.  As a result of the looser routing structure, SDR does not
   guarantee that a participating source routing domain will always have
   sufficient information to compute a route to a destination.  In
   addition, if the domain does have sufficient information, it is
   possible that the quantity may be large enough to preclude storage
   and/or route computation in a timely fashion.  However, despite the
   lack of guarantees, it is a goal of the architecture to provide
   efficient methods whereby sources can obtain the information needed
   to compute desired routes [Footnote: The primary goal of policy or
   TOS routing is to compute a route that satisfies a set of specialized
   requirements, and these requirements take precedence over optimality.
   In other words, even if a routing domain that participates in SDR or
   NR has sufficient information to compute a route, given a particular
   set of requirements, the architecture does not guarantee that the
   computed route is optimal.].

   Essential to SDR is the assumption that the routes installed on
   demand will be used sparingly.  The architecture assumes that at any
   given moment the set of all source-demand routes installed in an
   internet forms a small fraction of the total number of source-demand
   routes that can potentially be installed by all the routing domains.
   It is an assumption of the architecture that the number of routes
   installed in a BR by the SDR component should be on the order of log
   N (where N is the total number of routing domains in the Internet),
   so that the scaling properties of the SDR component are comparable
   with the scaling properties of the NR component.  The NR component
   achieves this property as a result of hierarchy.

   Note that the above requirement does not imply that only a few
   domains can participate in SDR, or that routes installed by the SDR
   component must have short life times.  What the requirement does
   imply, is that the product of the number of routes specified by
   domains that participate in SDR, times the average SDR-route life
   time, is bounded.  For example, the architecture allows either a
   small number of SDR routes with relatively long average life times,
   or a large number of SDR routes with relatively short average life
   times.  But the architecture clearly prohibits a large number of SDR
   routes with relatively long average life times.  The number of SDR
   routes is a function of the number of domains using SDR routes and
   the number of routes used per source domain.

   In summary, SDR is well suited for traffic that (1) is not widely-
   used enough (or is not sufficiently predictable or steady) to justify
   computation and maintenance by the NR component, and (2) whose
   duration is significantly longer than the time it takes to perform
   the route installation procedure.



Estrin, Rekhter & Hotz                                         [Page 24]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   The architecture does not require all domains in the Internet to
   participate in SDR.  Therefore, issues of scalability (with respect
   to the size of the Internet) become less crucial (though still
   important) to the SDR component.  Instead, the primary focus of the
   SDR component is shifted towards the ability to compute routes that
   satisfy specialized requirements, where we assume that the total
   number of domains requiring special routes simultaneously through the
   same part of the network is small relative to the total population.

4.1 Path Vector vs. Link State for SDR



   It is feasible to use either a distance vector or link state method
   of route computation along with source routing.  One could imagine,
   for instance, a protocol like BGP in which the source uses the full
   AD path information it receives in routing updates to create a source
   route. Such a protocol could address some of the deficiencies
   identified with distance vector, hop-by-hop designs.  However, we opt
   against further discussion of such a protocol because there is less
   to gain by using source routing without also using a link state
   scheme.  The power of source routing, in the context of inter-AD
   policy routing, is in giving the source control over the entire
   route.  This goal cannot be realized fully when intermediate nodes
   control which legal routes are advertised to neighbors, and therefore
   to a source.

   In other words, intermediate nodes should be able to preclude the use
   of a route by expressing a transit policy, but if a route is not
   precluded (i.e.,  is legal according to all ADs in the route), the
   route should be made available to the source independent of an
   intermediate domain's preferences for how its own traffic flows.

   Therefore, the SDR component employs an IDPR-like architecture in
   which link-state style updates are distributed with explicit policy
   terms included in each update along with the advertising node's
   connectivity.

4.2 Distribution of Routing Information



   By using a hop-by-hop NR component based on PV to complement the
   source-routing SDR component, we have alleviated the pressure to
   aggregate SDR forwarding information; the large percentage of inter-
   domain traffic carried, simultaneously, by any particular border
   router will be forwarded using aggregated NR forwarding information.
   However, the use of NR does not address the other major scaling
   problem associated with SDR: that of distributing, storing, and
   computing over a complete domain-level topology map.  In this section
   we describe promising opportunities for improving the scaling
   properties of SDR routing information distribution, storage, and



Estrin, Rekhter & Hotz                                         [Page 25]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   computation.

   Note that we do not propose to solve this problem in the same way
   that we solve it for NR.  A priori abstraction will not be employed
   since different domains may require different methods of abstracting
   the same routing information.  For example, if we aggregate routing
   information of domains that do not share the same policy and TOS
   characteristics (i.e., services), then outside of the aggregate, only
   those services that are offered by all domains in the aggregate will
   be advertised.  In order to locate special routes, SDR only uses
   aggregates when the component domains (and in turn the aggregate)
   advertise the required TOS and policy descriptions.  When the
   required TOS or policy characteristics are not offered by an
   aggregate, full information about the component domains is used to
   construct a route through those domains that do support the
   particular characteristics.  Consequently, we need some other, more
   flexible, means of reducing the amount of information distributed and
   held.  We address two issues in turn: distribution of configured
   topology and policy information, and distribution of dynamic status
   information.

4.2.1 Configured Information



   Information about the existence of inter-domain links, and policies
   maintained by domains, changes slowly over time.  This is referred to
   as configured information.  In the current IDPR specification
   complete, global, configuration information is kept by a route server
   in each domain.  Route servers (RS) are the entities that compute
   source routes.  On startup a RS can download the connectivity
   database from a neighbor RS; as domains, inter-domain links, or
   policies change, the changes are flooded to a RS in each domain.

   We have not yet specified the exact mechanisms for distributing
   configured connectivity information for SDR.  However, unlike the
   current IDPR specification, the SDR component will not flood all
   configured information globally.  Several alternate methods for
   organizing and distributing information are under investigation.

   Configured information may be regularly distributed via an out-of-
   band channel, e.g., CD/ROM.  In a similar vein, this information
   could be posted in several well-known locations for retrieval, e.g.,
   via FTP.  Between these "major" updates, aggregated collections of
   changes may be flooded globally.  Moreover, limited flooding (e.g.,
   by hop-count) could be used as appropriate to the "importance" of the
   change; while a policy change in a major backbone may still be
   flooded globally, a new inter-regional link may be flooded only
   within those regions, and information about an additional link to a
   non-transit domain may not be available until the next regularly-



Estrin, Rekhter & Hotz                                         [Page 26]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   scheduled "major" distribution.

   Changes that are not distributed as they occur will not necessarily
   be discovered.  However, a route server may learn pertinent
   information by direct query of remote servers, or through error
   messages resulting from traffic sent along failed routes.  Complete
   global flooding may be avoided by using some combination of these
   mechanisms.

   Even if an initial implementation uses a simple global flood, we must
   study the problem of structuring connectivity information such that
   it can be retrieved or distributed in a more selective manner, while
   still allowing sources to discover desired routes.  For example, we
   imagine RSs requesting filtered information from each other.  How the
   RSs should define filters that will get enough information to find
   special routes, while also effectively limiting the information, is
   an open question.  Again, the question is how to effectively
   anticipate and describe what information is needed in advance of
   computing the route.

   The essential dilemma is that networks are not organized in a nicely
   geographical or topologically consistent manner (e.g., it is not
   effective to ask for all networks going east-west that are within a
   certain north-south region of the target), hence a source domain does
   not know what information it needs (or should ask for) until it
   searches for, and discovers, the actual path.  Even with a central
   database, techniques are needed to structure configuration
   information so that the potential paths that are most likely to be
   useful are explored first, thereby reducing the time required for
   route computation.

   One promising approach organizes information using route fragments
   (partial paths) [Footnote: Route fragments were first suggested by
   Dave Clark and Noel Chiappa.].  Although the number of route
   fragments grows faster than the number of domains (at least O(N^2)),
   we can selectively choose those that will be useful to compute
   routes.  In particular, for each stub domain, fragments would be
   constructed to several well-known backbones [Footnote: Route
   fragments may be computed by a destination's route server and either
   made available via information service queries or global flooding.
   In addition, NR computed routes may be used as SDR route fragments.].
   Among its benefits, this approach aggregates domain information in a
   manner useful for computing source-routes, and provides an index,
   namely the destination, which facilitates on-demand reference and
   retrieval of information pertinent to a particular route computation.
   At this point, it is not clear how route fragments will affect SDR's
   ability to discover non-hierarchical routes.




Estrin, Rekhter & Hotz                                         [Page 27]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


4.2.2 Dynamic Status Information



   Assuming a technique for global or partial distribution of configured
   information, a second issue is whether, and how, to distribute
   dynamic status information (i.e., whether an inter-domain connection
   is up or down).

   In the current version of IDPR, dynamic status information is flooded
   globally in addition to configuration information.  We propose to
   distribute status information based strictly on locality.  First,
   dynamic information will be advertised within a small hop-count
   radius.  This simple and low-overhead mechanism exploits topological
   locality.  In addition to flooding status updates to nearby nodes, we
   also want to provide more accurate route information for long
   distance communications that entails more than a few network hops.
   Reverse path update (RPU) is a mechanism for sending dynamic status
   information to nodes that are outside the k-hop radius used for
   updates, but that nevertheless would obtain better service (fewer
   failed setups) by having access to the dynamic information [Estrin-
   etal91].

   RPU uses the existing active routes (represented by installed setup
   state or by a cache of the most recent source routes sent via the
   node in question) as a hint for distribution of event notifications.
   Instead of reporting only the status of the route being used, RPU
   reports the status of the domain's other inter-domain connections.
   If source routing exhibits route locality, the source is more likely
   to use other routes going through the node in question; in any case
   the overhead of the information about other links will be minimal.

   In this way, sources will receive status information from regions of
   the network through which they maintain active routes, even if those
   regions are more than k hops away.  Using such a scheme, k could be
   small to maximize efficiency, and RPU could be used to reduce the
   incidence of failed routes resulting from inaccurate status
   information.  This will be useful if long-path communication exhibits
   route locality with respect to regions that are closer to the
   destination (and therefore outside the k hop radius of flooded
   information).  In such situations, flooding information to the source
   of the long route would be inefficient because k would have to be
   equal to the length of the route, and in almost all cases, the
   percentage of nodes that would use the information decreases
   significantly with larger k.

4.3 Source-Demand Route Management



   SDR may be built either on top of the network layer supported by the
   NR component, or in parallel with it.  SDR forwarding will be



Estrin, Rekhter & Hotz                                         [Page 28]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   supported via two techniques: loose source-routing and route setup.

   The first technique, loose source-routing, would allow the originator
   of a packet to specify a sequence of domains that the packet should
   traverse on its path to a destination.  Forwarding such a packet
   within a domain, or even between domains within a confederation,
   would be left to intra-domain routing.  This avoids per-connection
   state and supports transaction traffic.

   The second technique, route setup, will be based on mechanisms
   developed for IDPR and described in [IDPR90].  It is well suited to
   conversations that persist significantly longer than a round-trip-
   time.  The setup protocol defines packet formats and the processing
   of route installation request packets (i.e, setup packets).  When a
   source generates a setup packet, the first border router along the
   specified source route checks the setup request, and if accepted,
   installs routing information; this information includes a path ID,
   the previous and next hops, and whatever other accounting-related
   information the particular domain requires.  The setup packet is
   passed on to the next BR in the domain-level source route, and the
   same procedure is carried out [Footnote: The setup packet may be
   forwarded optimistically, i.e., before checks are completed, to
   reduce latency.].  When the setup packet reaches the destination, an
   accept message is propagated back hop by hop, and each BR en route
   activates its routing information.  Subsequent data packets traveling
   along the same path to the destination include a path ID in the
   packet.  That path ID is used to locate the appropriate next-hop
   information for each packet.

   Border routers that support both the NR and the SDR components, must
   be able to determine what forwarding mechanism to use.  That is, when
   presented with a network layer PDU, such a BR should be able to make
   an unambiguous decision about whether forwarding of that PDU should
   be handled by the NR or the SDR component.  Discrimination mechanisms
   are dependent on whether the new network layer introduced by the SDR
   component is built on top of, or in parallel with, the network layers
   supported by the NR component.  Once the discrimination is made,
   packets that have to be forwarded via routes installed by the SDR
   component are forwarded to the exit port associated with the
   particular Path ID in the packet header.  Packets that have to be
   forwarded via routes installed by the NR component are forwarded to
   the exit port associated with the particular destination and Type of
   Service parameters (if present) in their packet headers.

   Next, we describe the primary differences between the IDPR setup
   procedure previously specified, and the procedure we propose to
   develop for this hybrid architecture.




Estrin, Rekhter & Hotz                                         [Page 29]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   During route installation, if a BR on the path finds that the
   remainder of the indicated route from the BR to the destination is
   identical to the NR route from the BR to the destination, then the BR
   can turn off the SDR route at that point and map it onto the NR
   route.  For this to occur, the specifications of the SDR route must
   completely match those of the NR route.  In addition, the entire
   forward route must be equivalent (i.e., the remaining hops to the
   destination).

   Moreover, if the NR route changes during the course of an active SDR
   route, and the new NR route does not match the SDR route, then the
   SDR route must be installed for the remainder of the way to the
   destination.  Consequently, when an SDR route is mapped onto an NR
   route, the original setup packet must be saved.  A packet traveling
   from a source to destination may therefore traverse both an SDR and
   an NR route segment; however, a packet will not traverse another SDR
   segment after traveling over an NR segment.  However, during
   transient periods packets could traverse the wrong route and
   therefore this must be an optional and controllable feature.

   A source can also request notification if a previously-down link or
   node returns to operation some time after a requested route setup
   fails.  If a BR on the route discovers that the requested next-hop BR
   is not available, the BR can add the source to a notification list
   and when the next-hop BR becomes reachable, a notification can be
   sent back to the source.  This provides a means of flushing out bad
   news when it is no longer true.  For example, a domain might decide
   to route through a secondary route when its preferred route fails;
   the notification mechanism would inform the source in a timely manner
   when its preferred route is available again.

   A third option addresses adaptation after route installation.  During
   packet forwarding along an active SDR route, if a BR finds that the
   SDR route has failed, it may redirect the traffic along an existing
   NR route to the destination.  This adaptation is allowed only if use
   of the NR route does not violate policy; for example, it may provide
   a less desirable type of service.  This is done only if the source
   selects the option at route setup time.  It is also up to the source
   whether it is to be notified of such actions.

   When a SDR route does fail, the detecting BR sends notification to
   the source(s) of the active routes that are affected.  Optionally,
   the detecting BR may include additional information about the state
   of other BRs in the same domain.  In particular, the BR can include
   its domain's most recent "update" indicating that domain's inter-
   domain links and policy.  This can be helpful to the extent there is
   communication locality; i.e., if alternative routes might be used
   that traverse the domain in question, but avoid the failed BR.



Estrin, Rekhter & Hotz                                         [Page 30]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   In summary, when a route is first installed, the source has several
   options (which are represented by flags in the route setup packet):

     1. If an NR route is available that satisfies all local policy
        and TOS, then use it.  Otherwise...

     2. Indicate whether the source wants to allow the setup to
        default to a NR route if the SDR route setup fails.



     3. Request notification of mapping to a NR route.



     4. Request additional configured information on failure.



     5. Request addition to a notification list for resource
        re-availability.



     6. Allow data packets to be rerouted to a NR route when failure
        happens after setup (so long  as no policy is violated).



     7. Request notification of a reroute of data packets.



     8. Request additional configured information on failure notice
        when the route is active.



     9. Request addition to a notification list if an active route
        fails.



























Estrin, Rekhter & Hotz                                         [Page 31]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


5.0 The Unified Architecture



   In addition to further evaluation and implementation of the proposed
   architecture, future research must investigate opportunities for
   increased unification of the two components of our architecture.  We
   are investigating several opportunities for additional commonality:

     1. Routing Information Base:
        Perhaps a single RIB could be shared by both NR and SDR.
        NR routes can be represented as a directed graph labeled
        with flags (on the nodes or links) corresponding to the
        generic transit constraints.  SDR requires that this graph
        be augmented by links with non-generic policies that have
        been discovered and maintained for computing special routes;
        in addition, special policy flags may be added to links
        already maintained by the NR component.

     2. Information Distribution:
        The NR path vectors could include address(es) of repositories
        for SDR-update information for each AD (or confederation) to
        assist the SDR component in retrieving selective information
        on demand.  For domains with minimal policies, where the space
        required for policy information is smaller than the space
        required for a repository address (e.g., if the policies for
        the domain listed are all wildcard), the NR path vectors could
        include a flag to that effect.

     3. Packet Forwarding:
        We should consider replacing the current IDPR-style network
        layer (which contains a global path identifier used in
        forwarding data packets to the next policy gateway on an
        IDPR route)  with a standard header (e.g., IP or CLNP),
        augmented with some option fields.  This would  unify the
        packet header parsing and forwarding functions for SDR and NR,
        and possibly eliminate some encapsulation overhead.

     4. Reachability Information:
        Currently IDRP distributes network reachability information
        within updates, whereas IDPR only distributes domain
        reachability information.  IDPR uses a domain name service
        function to map network numbers to domain numbers; the latter
        is needed to make the routing decision.   We should consider
        obtaining the network reachability and domain information in
        a unified manner.







Estrin, Rekhter & Hotz                                         [Page 32]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


5.1 Applicability to Various Network Layer Protocols



   The proposed architecture is designed to accommodate such existing
   network layer protocols as IP ([Postel81]), CLNP ([ISO-473-88]), and
   ST-II ([ST2-90]).  In addition, we intend for this architecture to
   support future network layer mechanisms, e.g., Clark and Jacobson's
   proposal or Braden and Casner's Integrated Services IP.  However on
   principal we can not make sweeping guarantees in advance of the
   mechanisms themselves.  In any case, not all of the mentioned
   protocols will be able to utilize all of the capabilities provided by
   the architecture.  For instance, unless the increase in the number of
   different types of services offered is matched by the ability of a
   particular network layer protocol to unambiguously express requests
   for such different types of services, the capability of the
   architecture to support routing in the presence of a large number of
   different types of service is largely academic.  That is, not all
   components of the architecture will have equal importance for
   different network layer protocols.  On the other hand, this
   architecture is designed to serve the future global internetworking
   environment.  The extensive research and development currently
   underway to implement and evaluate network mechanisms for different
   types of service suggests that future networks will offer such
   services.

   One of the fundamental issues in the proposed architecture is the
   issue of single versus multiple protocols.  The architecture does not
   make any assumptions about whether each network layer is going to
   have its own inter-domain routing protocol, or a single inter-domain
   routing protocol will be able to cover multiple network layers
   [Footnote: Similar issue already arose with respect to the intra-
   domain routing protocol, which generated sufficient amount of
   controversy within the Internet community.  It is our opinion, that
   the issue of single versus multiple protocols is more complex for the
   inter-domain routing than for the intra-domain routing.].  That is,
   the proposed architecture can be realized either by a single inter-
   domain routing protocol covering multiple network layers, or by
   multiple inter-domain routing protocols (with the same architecture)
   tailored to a specific network layer [Footnote: If the single
   protocol strategy is adopted, then it is likely that IDRP will be
   used as a base for the NR component.  Since presently IDRP is
   targeted towards CLNP, further work is needed to augment it to
   support IP and ST-II.  If the multiple protocol strategy is adopted,
   then it is likely that BGP will be used as a base for the NR
   component for IP, and IDRP will be used as a base for the NR
   component for CLNP.  Further work is needed to specify protocol in
   support for the NR component for ST-II.  Additional work may be
   needed to specify new features that may be added to BGP.].




Estrin, Rekhter & Hotz                                         [Page 33]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


5.2 Transition



   The proposed architecture is not intended for full deployment in the
   short term future.  We are proposing this architecture as a goal
   towards which we can begin guiding our operational and research
   investment over the next 5 years.

   At the same time, the architecture does not require wholesale
   overhaul of the existing Internet.  The NR component may be phased in
   gradually.  For example, the NR component for IP may be phased in by
   replacing existing EGP-2 routing with BGP routing.  Once the NR
   component is in place, it can be augmented by the facilities provided
   by the SDR component.

   The most critical components of the architecture needed to support
   SDR include route installation and packet forwarding in the routers
   that support SDR.  Participation as a transit routing domain requires
   that the domain can distribute local configuration information (LCI)
   and that some of its routers implement the route installation and
   route management protocols.  Participation as a source requires that
   the domain have access to a RS to compute routes, and that the source
   domain has a router that implements the route installation and route
   management protocols.  In addition, a network management entity must
   describe local configuration information and send it to the central
   repository(ies).  A collection and distribution mechanism must be put
   in place, even if it is centralized.

























Estrin, Rekhter & Hotz                                         [Page 34]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


6.0 Conclusions and Future Work



   In summary, the proposed architecture combines hop-by-hop path-
   vector, and source-routed link-state, protocols, and uses each for
   that which it is best suited: NR uses PV and multiple, flexible,
   levels of confederations to support efficient routing of generic
   packets over generic routes; SDR uses LS computation over a database
   of configured and dynamic information to route special traffic over
   special routes.  In the past, the community has viewed these two as
   mutually exclusive; to the contrary, they are quite complementary and
   it is fortunate that we, as a community, have pursued both paths in
   parallel.  Together these two approaches will flexibly and
   efficiently support TOS and policy routing in very large global
   internets.

   It is now time to consider the issues associated with combining and
   integrating the two.  We must go back and look at both architectures
   and their constituent protocols, eliminate redundancies, fill in new
   holes, and provide seamless integration.

7.0 Acknowledgments



   We would like to thank Hans-Werner Braun (San Diego Supercomputer
   Center), Lee Breslau (USC), Scott Brim (Cornell University), Tony Li
   (cisco Systems), Doug Montgomery (NIST), Tassos Nakassis (NIST),
   Martha Steenstrup (BBN), and Daniel Zappala (USC) for their comments
   on a previous draft.
























Estrin, Rekhter & Hotz                                         [Page 35]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


8.0 References



   [ANSI 87-150R]  "Intermediate System to Intermediate System Intra-
   Domain Routing Exchange Protocol", ANSI X3S3.3/87-150R.

   [BGP 91]  Lougheed, K., and Y. Rekhter, "A Border Gateway Protocol 3
   (BGP-3)", RFC 1267, cisco Systems, T.J. Watson Research Center, IBM
   Corp., October 1991.

   [Breslau-Estrin 91]  Breslau, L., and D. Estrin, "Design and
   Evaluation of Inter-Domain Policy Routing Protocols", To appear in
   Journal  of Internetworking Research and Experience, 1991.  (Earlier
   version appeared in ACM Sigcomm 1990.)

   [Clark 90]  Clark, D., "Policy Routing in Internetworks", Journal of
   Internetworking Research and Experience, Vol.  1, pp. 35-52, 1990.

   [Dijkstra 59]  Dijkstra, E., "A Note on Two Problems in Connection
   with Graphs", Numer. Math., Vol.  1, 1959, pp. 269-271.

   [ECMA89]  "Inter-Domain Intermediate Systems Routing", Draft
   Technical Report ECMA TR/ISR, ECMA/TC32-TG 10/89/56, May 1989.

   [EGP]  Rosen, E., "Exterior Gateway Protocol (EGP)", RFC 827, BBN,
   October 1982.

   [Estrin 89]  Estrin, D., "Policy Requirements for Inter
   Administrative Domain Routing", RFC 1125, USC Computer Science
   Department, November 1989.

   [Estrin-etal91]  Estrin, D., Breslau, L., and L. Zhang, "Protocol
   Mechanisms for Adaptive Routing in Global Multimedia Internets",
   University of Southern California, Computer Science Department
   Technical Report, CS-SYS-91-04, November 1991.

   [Hedrick 88]  Hedrick, C., "Routing Information Protocol", RFC 1058,
   Rutgers University, June 1988.

   [Honig 90]  Honig, J., Katz, D., Mathis, M., Rekhter, Y., and J. Yu,
   "Application of the Border Gateway Protocol in the Internet", RFC
   1164, Cornell Univ. Theory Center, Merit/NSFNET, Pittsburgh
   Supercomputing Center, T.J. Watson Research Center, IBM Corp., June
   1990.

   [IDPR90]  Steenstrup, M., "Inter-Domain Policy Routing Protocol
   Specification and Usage: Version 1", Work in Progress, February 1991.





Estrin, Rekhter & Hotz                                         [Page 36]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


   [IDRP91]  "Intermediate System to Intermediate System Inter-domain
   Routeing Exchange Protocol", ISO/IEC/ JTC1/SC6 CD10747.

   [ISIS10589]  "Information Processing Systems - Telecommunications and
   Information Exchange between Systems - Intermediate System to
   Intermediate System Intra-Domain Routing Exchange Protocol for use in
   Conjunction with the protocol for providing the Connectionless-mode
   Network Service (ISO 8473)", ISO/IEC 10589.

   [ISO-473 88]  "Protocol for providing the connectionless-mode network
   service", ISO 8473, 1988.

   [Jaffee 82]  Jaffee, J., and F. Moss, "A Responsive Distributed
   Routing Algorithm for Computer Networks", IEEE Transactions on
   Communications, July 1982.

   [Little 89]  Little, M., "Goals and Functional Requirements for
   Inter-Autonomous System Routing", RFC 1126, SAIC, October 1989.

   [Oran 89]  Oran, D., "Expert's Paper: The Relationship between
   Addressing and Routeing", ISO/JTC1/SC6/WG2, 1989.

   [OSPF]  Moy, J., "The Open Shortest Path First (OSPF) Specification",
   RFC 1131, Proteon, October 1989.

   [Postel 81]  Postel, J., "Internet Protocol", RFC 791, DARPA,
   September 1981.

   [Rekhter 91]  Rekhter, Y., "IDRP protocol analysis: storage
   complexity", IBM Research Report RC17298(#76515), October 1991.

   [Shin87] Shin, K., and M. Chen, "Performance Analysis of Distributed
   Routing Strategies Free of Ping-Pong-Type Looping", IEEE Transactions
   on Computers, February 1987.

   [ST2-90]  Topolcic, C., "Experimental Internet Stream Protocol,
   version 2 (ST II)", RFC 1190, CIP Working Group, October 1990.

   [Zaumen 91] Zaumen, W., and J. Garcia-Luna-Aceves, "Dynamics of Link
   State and Loop-free Distance-Vector Routing Algorithms", ACM Sigcomm
   '91, Zurich, Switzerland, September 1991.

   [Zhang 91] Zhang, L., "Virtual Clock: A New Traffic Control Algorithm
   for Packet Switching Networks".







Estrin, Rekhter & Hotz                                         [Page 37]

RFC 1322       A Unified Approach to Inter-Domain Routing       May 1992


Security Considerations

   Security issues are not discussed in this memo.

Authors' Addresses



   Deborah Estrin
   University of Southern California
   Computer Science Department, MC 0782
   Los Angeles, California 90089-0782

   Phone: (310) 740-4524
   EMail: estrin@usc.edu


   Yakov Rekhter
   IBM T.J. Watson Research Center
   P.O. Box 218
   Yorktown Heights, New York 10598

   Phone: (914) 945-3896
   EMail: yakov@ibm.com


   Steven Hotz
   University of Southern California
   Computer Science Department, MC 0782
   Los Angeles, California 90089-0782

   Phone: (310) 822-1511
   EMail: hotz@usc.edu




















Estrin, Rekhter & Hotz                                         [Page 38]