Network Working Group                                         H-W. Braun
Request for Comments:  1104                                 Merit/NSFNET
                                                               June 1989


                     Models of Policy Based Routing

1. Status of this Memo

   The purpose of this RFC is to outline a variety of models for policy
   based routing.  The relative benefits of the different approaches are
   reviewed.  Discussions and comments are explicitly encouraged to move
   toward the best policy based routing model that scales well within a
   large internetworking environment.

   Distribution of this memo is unlimited.

2. Acknowledgements

   Specific thanks go to Yakov Rekhter (IBM Research), Milo Medin
   (NASA), Susan Hares (Merit/NSFNET), Jessica Yu (Merit/NSFNET) and
   Dave Katz (Merit/NSFNET) for extensively contributing to and
   reviewing this document.

3. Overview

   To evaluate the methods and models for policy based routing, it is
   necessary to investigate the context into which the model is to be
   used, as there are a variety of different methods to introduce
   policies.  Most frequently the following three models are referenced:

       Policy based distribution of routing information
       Policy based packet filtering/forwarding
       Policy based dynamic allocation of network resources (e.g.,
       bandwidth, buffers, etc.)

   The relative properties of those methods need to be evaluated to find
   their merits for a specific application.  In some cases, more than
   one method needs to be implemented.

   While comparing different models for policy based routing, it is
   important to realize that specific models have been designed to
   satisfy a certain set of requirements.  For different models these
   requirements may or may not overlap.  Even if they overlap, they may
   have a different degree of granularity.  In the first model, the
   requirements can be formulated at the Administrative Domain or
   network number level.  In the second model, the requirements can be
   formulated at the end system level or probably even at the level of

   individual users.  In the third model, the requirements need to be
   formulated at both the end system and local router level, as well as
   at the level of Routing Domains and Administrative Domains.

   Each of these models looks at the power of policy based routing in a
   different way.  They may be implemented separately or in combination
   with other methods.  The model to describe policy based dynamic
   allocation of network resources is orthogonal to the model of policy
   based distribution of routing information.  However, in an actual
   implementation each of these models may interact.

   It is important to realize that the use of a policy based scheme for
   individual network applications requires that the actual effects as
   well as the interaction of multiple methods need to be determined
   ahead of time by policy.

   While uncontrolled dynamic routing and allocation of resources may
   have a better real time behavior, the use of policy based routing
   will provide a predictable, stable result based on the desires of the
   administrator.  In a production network, it is imperative to provide
   continuously consistent and acceptable services.

4. Policy based distribution of routing information

   Goals:

      The goal of this model is to enforce certain flows by means of
      policy based distribution of routing information.  This
      enforcement allows control over who can and who can not use
      specific network resources.

      Enforcement is done at the network or Administrative Domain (AD)
      level - macroscopic policies.

   Description:

      A good example of policy based routing based on the distribution
      of routing information is the NSFNET with its interfaces to mid-
      level networks [1], [2].  At the interface into the NSFNET, the
      routing information is authenticated and controlled by four means:

         1. Routing peer authentication based on the source address.

         2. Verification of the Administrative Domain identification
            (currently EGP Autonomous System numbers).

         3. Verification of Internet network numbers which are
            advertised via the routing peer.

         4. Control of metrics via a Routing Policy Data Base for the
            announced Internet network numbers to allow for primary
            paths to the NSFNET as well as for paths of a lesser
            degree.

      At the interfaces that pass routing traffic out of the NSFNET, the
      NSS routing code authenticates the router acting as an EGP peer by
      its address as well as the Administrative Domain identification
      (Autonomous System Number).

      Outbound announcements of network numbers via the EGP protocol are
      controlled on the basis of Administrative Domains or individual
      network numbers by the NSFNET Routing Policy Data Base.

      The NSFNET routing policy implementation has been in place since
      July 1988 and the NSFNET community has significant experience with
      its application.

      Another example of policy controlled dissimination of routing
      information is a method proposed for ESNET in [3].

   Benefits:

      A major merit of the control of routing information flow is that
      it enables the engineering of large wide area networks and allows
      for a more meshed environment than would be possible without tight
      control.  Resource allocation in a non-hostile environment is
      possible by filtering specific network numbers or Administrative
      Domains on a per need basis.  Another important benefit of this
      scheme is that it allows for network policy control with virtually
      no performance degradation, as only the routing packets themselves
      are relevant for policy control.  Routing tables are generated as
      a result of these interactions.  This means that this scheme
      imposes only very little impact on packet switching performance at
      large.

   Concerns:

      Policy based routing information distribution does not address
      packet based filtering.  An example is the inability to prevent
      malicious attacks by introduced source routed IP packets.  While
      resource allocation is possible, it extends largely to filtering
      on network numbers or whole Administrative Domains, but it would
      not extend to end systems or individual users.

   Costs:

      Policy based routing in the NSFNET is implemented in a series of

      configuration files.  These configuration files are in turn
      generated from a routing information database.  The careful
      creation of this routing information database requires knowledge
      of the Internet at large.  Because the Internet is changing
      constantly, the upkeep of this routing information database is a
      continuous requirement.  However, the effort of collecting and
      maintaining an accurate view of the Internet at large can be
      distributed.

      Since policy controlled distribution of routing information allows
      for filtering on the basis of network numbers or Administrative
      Domains, the routing information database only needs to collect
      information for the more than 1300 networks within the Internet
      today.

5. Policy based packet filtering/forwarding

   Goals:

      The goal of the model of policy based packet filtering/forwarding
      is to allow the enforcement of certain flows of network traffic on
      a per packet basis.  This enforcement allows the network
      administrator to control who can and who can not use specific
      network resources.

      Enforcement may be done at the end system or even individual user
      level - microscopic policies.

   Description:

      An example of packet/flow based policies is outlined in [4].  In a
      generic sense, policy based packet filtering/forwarding allows
      very tight control of the distribution of packet traffic.  An
      implemented example of policy based filtering/forwarding is a
      protection mechanism built into the NSFNET NSS structure, whereby
      the nodes can protect themselves against packets targeted at the
      NSFNET itself by filtering according to IP destination. While this
      feature has so far not been enabled, it is fully implemented and
      can be turned on within a matter of seconds.

   Benefits:

      The principal merit of this scheme is that it allows the
      enforcement of packet policies and resource allocation down to
      individual end systems and perhaps even individual end users.  It
      does not address a sane distribution of routing information.  If
      policies are contained in the packets themselves it could identify
      users, resulting in the ability of users to move between

      locations.

   Concerns:

      The major concern would be the potentially significant impact on
      the performance of the routers, as, at least for tight policy
      enforcements, each packet to be forwarded would need to be
      verified against a policy data base.  This limitation makes the
      application of this scheme questionable using current Internet
      technology, but it may be very applicable to circuit switched
      environments (with source-routed IP packets being similar to a
      circuit switched environment).  Another difficulty could be the
      sheer number of potential policies to be enforced, which could
      result in a very high administrative effort.  This could result
      from the creation of policies at the per-user level.  Furthermore,
      the overhead of carrying policy information in potentially every
      packet could result in additional burdens on resource
      availabilities.  This again is more applicable to connection-
      oriented networks, such as public data networks, where the policy
      would only need to be verified at the call setup time.  It is an
      open question how well packet based policies will scale in a large
      and non homogeneous Internet environment, where policies may be
      created by all of the participants.  These creations of policy
      types of services may have to be doable in real time.

      Scaling may require hierarchy.  Hierarchy may conflict with
      arbitrary Type of Service (TOS) routing, which is one of the
      benefits of this model.

   Costs of implementation:

      A large scale implemention of packet based policy routing would
      require a routing information base that would contain information
      down to the end system level and possibly end users.  If one would
      assume that for each of the 1300 networks there is an average of
      200 end systems, this would result in over 260000 end systems
      Internet wide.  Each end system in turn could further contribute
      some information on the type of traffic desired, including types
      of service (issues like agency network selection), potentially on
      a per-user basis.  The effort for the routing policy data base
      could be immense, in particular if there is a scaling requirement
      towards a variety of policies for backbones, mid-level networks,
      campus networks, subnets, hosts, and users.  The administration of
      this "packet routing" database could be distributed.  However,
      with a fully distributed database of this size several consistency
      checks would have to be built into the system.

6. Policy based dynamic allocation of network resources (e.g.,
      bandwidth, buffers, etc.).

   Goals:

      Flexible and economical allocation of network resources based on
      current needs and certain policies.  Policies may be formulated at
      the network or Administrative Domain (AD) levels.  It is also
      possible to formulate policies which will regulate resource
      allocation for different types of traffic (e.g., Telnet, FTP,
      precedence indicators, network control traffic).

      Enforcement of policy based allocation of network resources might
      be implemented within the following parts of the network:

          routers for networks and Administrative Domain (AD) levels
          circuit switches for networks
          end systems establishing network connections

   Description:

      Policy based allocation of bandwidth could allow the modulation of
      the circuits of the networking infrastructure according to real
      time needs.  Assuming that available resources are limited towards
      an upper bound, the allocation of bandwidth would need to be
      controlled by policy.  One example might be a single end system
      that may or may not be allowed to, perhaps even automatically,
      take resources away from other end systems or users.  An example
      of dynamic bandwidth allocation is the currently implemented
      circuit switched IDNX component of the NSFNET, as well as the MCI
      Digital Reconfiguration Service (DRS) which is planned for the
      NSFNET later this year.

      Another model for resource allocation occurs at the packet level,
      where the allocation is controlled by multiple packet queues.
      This could allow for precedence queuing, with preferences based on
      some type of service and preferred forwarding of recognized
      critical data, such as network monitoring, control and routing.
      An example can be found in the NSFNET, where the NSFNET nodes
      prefer traffic affiliated with the NSFNET backbone network number
      over all other traffic, to allow for predictable passing of
      routing information as well as effective network monitoring and
      control.  At the other end of the spectrum, an implementation
      could also allow for queues of most deferrable traffic (such as
      large background file transfers).

   Benefits:

      Dynamic allocation of bandwidth could allow for a truly flexible
      environment where the networking infrastructure could create
      bandwidth on a per need basis.  This could result in significant
      cost reductions during times when little bandwidth is needed.
      This method could potentially accommodate real time transient high
      bandwidth requirements, potentially by reducing the bandwidth
      available to other parts of the infrastructure.  A positive aspect
      is that the bandwidth allocation could be protocol independent,
      with no impact on routing protocols or packet forwarding
      performance.

      Policy based allocation of bandwidth can provide a predictable
      dynamic environment.  The rules about allocation of bandwidth at
      the circuit level or at the packet level need to be determined by
      a consistent and predictable policy, so that other networks or
      Administrative Domains can tune their allocation of networking
      resources at the same time.

   Concerns:

      The policies involved in making dynamic bandwidth allocation in a
      largely packet switching environment possible are still in the
      development phase.  Even the technical implications of
      infrastructure reconfiguration in result of events happening on a
      higher level still requires additional research.

      A policy based allocation of bandwidth could tune the network to
      good performance, but could cause networks located in other
      Administrative Domains to pass traffic poorly.  It is important
      that network resource policy information for a network be
      discussed within the context of its Administrative Domain.
      Administrative Domains need to discuss their network resource
      allocation policies with other Administrative Domains.

      The technical problem of sharing network resource policy
      information could be solved by a making a "network resource policy
      information" database available to all administrators of networks
      and Administrative Domains.  However, the political problems
      involved in creating a network resource policy with impact on
      multiple Administrative Domains does still require additional
      study.

7. Discussion

   Both the first and the second model of policy based routing are
   similar in the sense that their goal is to enforce certain flows.

   This enforcement allows the control of access to scarce network
   resources (if the resource is not scarce, there is no performance
   reason to control access to it).  The major difference is the level
   of enforcement: macroscopic level versus microscopic level control.

   Associated with the enforcement for a certain network resource is the
   cost.  If this cost is higher than the cost required to make a
   particular resource less scarce, then the feasibility of enforcement
   may be questionable.

   If portions of the Internet find that microscopic enforcement of
   policy is necessary, then this will need to be implementable without
   significant performance degradation to the networking environment at
   large.  Local policies within specific Routing Domains or
   Administrative Domains should not affect global Internet traffic or
   routing.  Policies within Administrative Domains which act as traffic
   transit systems (such as the NSFNET) should not be affected by
   policies a single network imposes for its local benefit.

   Some models of policy routing are trying to deal with cases where
   network resources require rather complex usage policies.  One of
   scenarios in [4] is one in which a specific agency may have some
   network resource (in the example it is a link) which is sometimes
   underutilized.  The goal is to sell this resource to other agencies
   during the underutilization period to recover expenses.  This
   situation is equivalent to the problem of finding optimum routes,
   with respect to a certain TOS, in the presence of network resources
   (e.g., links) with variable characteristics.  Any proposed solution
   to this problem should address such issues as network and route
   stability.  More feasibility study is necessary for the whole
   approach where links used for global communication are also subject
   to arbitrary local policies.  An alternative approach would be to
   reconfigure the network topology so that underutilized links will be
   dropped and possibly returned to the phone company.  This is
   comparable to what the NSFNET is planning on doing with the MCI
   Digital Reconfiguration Service (DRS).  A DRS model may appear
   cleaner and more easy to implement than a complicated model like the
   one outlined in [4].

   The models for policy based routing emphasize that careful
   engineering of the Internet needs to decided upon the profile of
   traffic during normal times, outage periods, and peak loads.  This
   type of engineering is not a new requirement.  However, there could
   potentially be a significant benefit in deciding these policies ahead
   of time and using policy based routing to implement specific routing
   policies.

8. Accounting vs. Policy Based Routing

   Quite often Accounting and Policy Based Routing are discussed
   together.  While the application of both Accounting and Policy Based
   Routing is to control access to scarce network resources, these are
   separate (but related) issues.

   The chief difference between Accounting and Policy Based Routing is
   that Accounting combines history information with policy information
   to track network usage for various purposes.  Accounting information
   may in turn drive policy mechanisms (for instance, one could imagine
   a policy limiting a certain organization to a fixed aggregate
   percentage of dynamically shared bandwidth).  Conversely, policy
   information may affect accounting issues.  Network accounting
   typically involves route information (at any level from AD to end
   system) and volume information (packet, octet counts).

   Accounting may be implemented in conjunction with any of the policy
   models mentioned above.  Similar to the microscopic versus
   macroscopic policies, accounting may be classified into different
   levels.  One may collect accounting data at the AD level, network
   level, host level, or even at the individual user level.  However,
   since accounting may be organized hierarchically, microscopic
   accounting may be supported at the network or host level, while
   macroscopic accounting may be supported at the network or AD level.
   An example might be the amount of traffic passed at the interface
   between the NSFNET and a mid-level network or between a mid-level
   network and a campus.  Furthermore, the NSFNET has facilities
   implemented to allow for accounting of traffic trends from individual
   network numbers as well as application-specific information.

   Full-blown accounting schemes suffer the same types of concerns
   previously discussed, with the added complication of potentially
   large amounts of additional data gathered that must be reliably
   retrieved.  As pointed out in [4], policy issues may impact the way
   accounting data is collected (one administration billing for packets
   that were then dropped in the network of another administration).
   Microscopic accounting may not scale well in a large internet.

   Furthermore, from the standpoint of billing, it is not clear that the
   services provided at the network layer map well to the sorts of
   services that network consumers are willing to pay for.  In the
   telephone network (as well as public data networks), users pay for
   end-to-end service and expect good quality service in terms of error
   rate and delay (and may be unwilling to pay for service that is
   viewed as unacceptable).  In an internetworking environment, the
   heterogeneous administrative environment combined with the lack of
   end-to-end control may make this approach infeasible.

   Lightweight approaches to accounting can be used (with less impact)
   when specific, limited goals are set.  One suggested approach
   involves monitoring traffic patterns.  If a pattern of abuse (e.g.,
   unauthorized use) develops, an accounting system could track this and
   allow corrective action to be taken, by changing routing policy or
   imposing access control (blocking hosts or nets).  Note that this is
   much less intrusive into the packet forwarding aspects of the
   routers, but requires distribution of a policy database that the
   accounting system can use to reduce the raw information.  Because
   this approach is statistical in nature, it may be slow to react.

9. References

   [1] Rekhter, Y., "EGP and Policy Based Routing in the New NSFNET
       Backbone", RFC 1092, IBM Research, February 1989.

   [2] Braun, H-W., "The NSFNET Routing Architecture", RFC 1093,
       Merit/NSFNET Project, February 1989.

   [3] Collins, M., and R. Nitzan, "ESNET Routing", DRAFT Version 1.0,
       LLNL, May 1989.

   [4] Clark, D., "Policy Routing in Internet Protocols", RFC 1102,
       M.I.T. Laboratory for Computer Science, May 1989.

Author's Address

   Hans-Werner Braun
   Merit Computer Network
   University of Michigan
   1075 Beal Avenue
   Ann Arbor, Michigan 48109

   Telephone:      313 763-4897
   Fax:            313 747-3745
   EMail:          hwb@merit.edu