Rfc1582
TitleExtensions to RIP to Support Demand Circuits
AuthorG. Meyer
DateFebruary 1994
Format:TXT, HTML
Status:PROPOSED STANDARD






Network Working Group                                           G. Meyer
Request for Comments: 1582                                Spider Systems
Category: Standards Track                                  February 1994


              Extensions to RIP to Support Demand Circuits

Status of this Memo

   This document specifies an Internet standards track protocol for the
   Internet community, and requests discussion and suggestions for
   improvements.  Please refer to the current edition of the "Internet
   Official Protocol Standards" (STD 1) for the standardization state
   and status of this protocol.  Distribution of this memo is unlimited.

Abstract

   Running routing protocols on connection oriented Public Data
   Networks, for example X.25 packet switched networks or ISDN, can be
   expensive if the standard form of periodic broadcasting of routing
   information is adhered to.  The high cost arises because a connection
   has to all practical intents and purposes be kept open to every
   destination to which routing information is to be sent, whether or
   not it is being used to carry user data.

   Routing information may also fail to be propagated if the number of
   destinations to which the routing information is to be sent exceeds
   the number of channels available to the router on the Wide Area
   Network (WAN).

   This memo defines a generalized modification which can be applied to
   Bellman-Ford (or distance vector) algorithm information broadcasting
   protocols, for example IP RIP, Netware RIP or Netware SAP, which
   overcomes the limitations of the traditional methods described above.

   The routing protocols support a purely triggered update mechanism on
   demand circuits on WANs.  The protocols run UNMODIFIED on LANs or
   fixed point-to-point links.

   Routing information is sent on the WAN when the routing database is
   modified by new routing information received from another interface.
   When this happens a (triggered) update is sent to a list of
   destinations on other WAN interfaces.  Because routing protocols are
   datagram based they are not guaranteed to be received by the peer
   router on the WAN.  An acknowledgement and retransmission mechanism
   is provided to ensure that routing updates are received.





RFC 1582                       Demand RIP                  February 1994


   The WAN circuit manager advises the routing applications on the
   reachability and non-reachability of destinations on the WAN.

Acknowledgements

   I would like to thank colleagues at Spider, in particular Richard
   Edmonstone, Tom Daniel and Alam Turland, Yakov Rekhter (IBM), Martha
   Steenstrup (BBN), and members of the RIP-2 working group of the IETF
   for stimulating discussions and comments which helped to clarify this
   memo.

Conventions

   The following language conventions are used in the items of
   specification in this document:

      o  MUST -- the item is an absolute requirement of the specification.
         MUST is only used where it is actually required for interoperation,
         not to try to impose a particular method on implementors
         where not required for interoperability.

      o  SHOULD -- the item should be followed for all but exceptional cir-
         cumstances.

      o  MAY or optional -- the item is truly optional and may be followed
         or ignored according to the needs of the implementor.

   The words "should" and "may" are also used, in lower case, in their
   more ordinary senses.

Table of Contents

      1. Introduction ...........................................  3
      2. Running a routing Protocol on the WAN ..................  4
          2.1. Overview .........................................  4
          2.2. Presumption of Reachability ......................  6
          2.3. WAN Router list ..................................  7
          2.4. Triggered Updates and Unreliable Delivery ........  8
          2.5. Guaranteeing delivery of Routing Updates .........  8
          2.6. The Routing Database .............................  9
          2.7. New Packet Types ................................. 10
          2.8. Fragmentation .................................... 12
          2.9. Preventing Queue Overload ........................ 13
      3. IP Routing Information Protocol Version 1 .............. 13
      4. IP Routing Information Protocol Version 2 .............. 16
      5. Netware Routing Information Protocol ................... 17
      6. Netware Service Advertising Protocol ................... 20
      7. Timers ................................................. 24



RFC 1582                       Demand RIP                  February 1994


          7.1. Database Timer ................................... 24
          7.2. Retransmission Timer ............................. 25
          7.3. Reassembly Timer ................................. 26
      8. Implementation Considerations ...........................27
      9. Security Considerations ................................ 27
     10. References ............................................. 28
     11. Author's Address ....................................... 29

1. Introduction

   Routers are used on connection oriented networks, such as X.25 packet
   switched networks and ISDN networks, to allow potential connectivity
   to a large number of remote destinations.  Circuits on the Wide Area
   Network (WAN) are established on demand and are relinquished when the
   traffic subsides.  Depending on the application, the connection
   between any two sites for user data might actually be short and
   relatively infrequent.

   Practical experience of routing shows that periodic 'broadcasting' of
   routing updates on the WAN is unsatisfactory on several counts:

   o  Running a routing protocol like RIP is expensive if the standard
      form of transmitting routing information to every next hop router
      every 30 seconds is adhered to.  The more routers there are
      wishing to exchange information the worse the problem.  If there
      are N routers on the WAN, N * (N - 1) routing updates are sent over
      N * (N - 1)/2 connections every broadcast period.

      The expense arises because a circuit has to be kept open to each
      destination to which routing information is to be sent.  Routing
      updates are sufficiently frequent that little benefit is obtainable
      on most networks by attempting to set up a call purely for
      the duration of the routing update. (There are often minimum call
      charges, or there is a charge to set up a call irrespective of
      what data is sent.)

      The option of reducing the 'broadcast' frequency, while reducing
      the cost, would make the system less responsive.

   o  The number of networks to be connected (N) on the WAN can easily
      exceed the number of simultaneous calls (M) which the interface
      can support.  If this happens the routing 'broadcast' will only
      reach M next hop routers, and (N - M) next hop routers will not
      receive the routing update.

      A basic rate ISDN interface can support 2 simultaneous calls, and
      even the number of logical channels most users subscribe to on an
      X.25 network is not large.  There is no fundamental reason why



RFC 1582                       Demand RIP                  February 1994


      routing protocols should restrict the user to routing between so
      few sites.

   o  Since there is no broadcast facility on the WAN, 'broadcasting' of
      routing information actually consists of sending the updates
      separately to all known locations.  This means that N routing
      updates are queued for transmission on the WAN link (in addition
      to any data which might be queued).

      Routers take a pragmatic view on queuing datagrams for the WAN.
      If the queue length gets too long, so that datagrams at the end of
      the queue would take too long be transmitted in a reasonable time
      (say 1 to 2 seconds) the router may start discarding them.  On an
      X.25 network, with slow line speeds (perhaps 9600 baud), it may
      not take too many routing updates to fulfill this condition if the
      link is also actively being used to carry user data.

   This memo addresses all the above problems.

   The approach taken is to modify the routing protocols so as to send
   information on the WAN only when there has been an update to the
   routing database OR a change in the reachability of a next hop router
   is indicated by the task which manages connections on the WAN.

   Because datagrams are not guaranteed to get through on all WAN media,
   an acknowledgement and retransmission system is required to provide
   reliability.

   This memo describes the modifications required for Bellman-Ford (or
   distance vector) algorithm information broadcasting protocols, such
   as IP RIP [1,2] or Netware RIP and SAP [3] on the WAN.  The protocols
   run unmodified on Local Area Networks (LANs) or fixed point-to-point
   links, and so interoperate transparently with implementations
   adhering to the original specifications.

2. Running Routing Protocols on the WAN

2.1 Overview

   Multiprotocol routers are used on connection oriented Wide Area
   Networks (WANs), such as X.25 packet switched networks and ISDN
   networks, to interconnect LANs.  By using the multiplexing properties
   of the underlying WAN technology, several LANs can be interconnected
   simultaneously through a single physical interface on the router.

   A circuit manager provides an interface between the connectionless
   network layers (IP, IPX, CLNP etc) and the connection oriented WAN
   (X.25 or ISDN).  Figure 1 shows a schematic representative stack



RFC 1582                       Demand RIP                  February 1994


   showing the relationship between routing protocols, the network
   layers, the circuit manager and the connection oriented WAN.

             --------------           ---------  ---------
             |    RIP     |           |  RIP  |  |  SAP  |
             --------------           ---------  ---------
                   |                      |          |
             --------------               |          |
             |    UDP     |               |          |
             --------------               |          |
                   |                      |          |
             --------------             ----------------
             |    IP      |             |     IPX      |
             --------------             ----------------
                   |                           |
             -------------------------------------------
             |             Circuit Manager             |
             -------------------------------------------
                              ||||||||||
                              ||||||||||
                      ---------------------------
                      |   Connection Oriented   |
                      |        WAN stack        |
                      ---------------------------

     A WAN circuit manager will support a variety of network layer
     protocols, on its upper interface.  On its lower interface, it
     may support one or more subnetworks.  A subnetwork may support a
     number of Virtual Circuits.


            Figure 1.   Representative Multiprotocol Router stack

   The router has a translation table which relates the network layer
   address of the next hop router to the physical address used to
   establish a Virtual Circuit (VC) to it.  Datagrams may be
   encapsulated in a header to distinguish the network layer protocol
   [5].

   The circuit manager takes datagrams from the connectionless network
   layer protocols and (if one is not currently available) opens a VC to
   the next hop router.  A VC can carry all traffic between two end-
   point routers for a given network layer protocol (or with appropriate
   encapsulation all network layer protocols).  An idle timer is used to
   close the VC when the datagrams stop arriving at the circuit manager.

   Running routing protocols on the WAN has traditionally consisted of
   making small modifications to the methods used on LANs.  Where



RFC 1582                       Demand RIP                  February 1994


   routing information would be broadcast periodically on a LAN
   interface, it is converted to a series of periodic updates sent to a
   list of addresses on the WAN.

   This memo targets two areas:

   o  Eliminating the overkill inherent in periodic transmission of
      routing updates.

   o  Overcoming the bandwidth limitations on the WAN: the number of
      simultaneous VCs to next hop routers and restricted data
      throughput which the WAN link can support.

   The first of these is overcome by transmitting routing updates
   (called routing responses) only when required:

   o  Firstly, when a specific request for a routing update has been
      received.

   o  Secondly, when the routing database is modified by new
      information from another interface.

      Update information received in this way is not normally
      propagated on other interfaces immediately, but is delayed for a
      few seconds to allow information from several updates to be
      grouped.

   o  Thirdly, when the circuit manager indicates that a destination
      has changed from an unreachable (circuit down) to a reachable
      (circuit up) state.

   Because of the inherent unreliability of a datagram based system,
   both routing requests and routing responses require acknowledgement,
   and retransmission in the event of NOT receiving an acknowledgement.

   To overcome the bandwidth limitations the routing application can
   perform a form of self-imposed flow control, to spread routing
   updates out over a period of time.

2.2 Presumption of Reachability

   If a routing update is received from a next hop router on the WAN,
   entries in the update are thereafter always considered to be
   reachable, unless proven otherwise:

   o  If in the normal course of routing datagrams, the circuit manager
      fails to establish a connection to the next hop router, it
      notifies the routing application that the next hop router is not



RFC 1582                       Demand RIP                  February 1994


      reachable through an internal circuit down message.

      The routing application then goes through a process of timing out
      database entries to make them unreachable in the routing sense.

   o  If the circuit manager is subsequently able to establish a
      connec tion to the next hop router, it will notify the routing
      applica tion that the next hop router is reachable through an
      internal circuit up message.

      The routing application will then exchange messages with the next
      hop router so as to re-prime their respective routing databases
      with up-to-date information.

   Handling of circuit up and circuit down messages requires that the
   circuit manager takes responsibility for establishing (or
   reestablishing) the connection in the event of a next hop router
   becoming unreachable.  A description of the processes the circuit
   manager adopts to perform this task is outside the scope of this
   memo.

2.3 WAN Router list

   The routing task MAY be provided with a list of routers to send
   routing updates to on the WAN.  It will comprise of the logical
   addresses of next hop routers for which the router has a logical to
   physical address mapping.  Entries in the list SHOULD be categorized
   (on a per-peer basis) as follows:

   o  Running the standard routing protocol, namely transmitting
      updates periodically with the packet formats used in LAN
      broadcasts.

      This option is supported to allow interoperability with existing
      routing implementations, and might also be appropriate if some
      of the destinations are using Permanent Virtual Circuits (PVCs)
      rather than SVCs.

   o  Running the triggered update routing protocol proposed in this
      memo.

   Omitting an address from both of these categories is equivalent to
   not running the routing protocols.

   If routing packets arrive from a destination not supporting the
   appropriate variant they MUST be discarded.





RFC 1582                       Demand RIP                  February 1994


2.4 Triggered Updates and Unreliable Delivery

   If triggered update information is sent to next hop routers on the
   WAN only once it can fail to arrive for one of the following reasons:

   o  A free VC resource might not be available, because of a
      restricted number of X.25 logical channels or ISDN B-channels.

   o  The transmit queue might be full - requiring the datagram to be
      discarded.

   o  The VC might be pre-empted (in favour of establishing a VC to
      another next hop router) while the datagram is in a queue,
      resulting in the queue being flushed and the datagram
      discarded.

   o  In cases where the method of transport is not guaranteed, for
      example with PPP where there is no acknowledgement and
      retransmission of HDLC frames, a corrupted frame will result in
      the loss of the datagram.

2.5 Guaranteeing delivery of Routing Updates

   To guarantee delivery of routing updates on the WAN an
   acknowledgement and retransmission scheme MUST be used:

   o  Send a routing update to a next hop router on the WAN.

   o  The other router responds with an acknowledgement packet.

      The original router receives the acknowledgement.

   o  Otherwise the original router retransmits the update until an
      acknowledgement is received.

      Retransmission timer values are covered in section 7.

      In cases where the routing database is modified before an
      acknowledgement is received a new routing update with an
      updated sequence number is sent out.  If an acknowledgement for
      the old routing update is received it is ignored.

   o  A router only updates its routing database when it receives a
      complete update, which may consist of several fragments.  Each
      fragment is individually acknowledged.

   The above mechanism caters for cases where the datagram is lost
   because of a frame error or is discarded because of an over-full



RFC 1582                       Demand RIP                  February 1994


   queue.  The routing update and acknowledgement will eventually both
   get through.

   In cases where the circuit manager cannot establish a connection, a
   mechanism is provided to allow the circuit manager to inform the
   routing task of the failure to make a connection so that it can
   suppress retransmissions until a circuit becomes available.

2.6 The Routing Database

   A requirement of using triggered updates for propagating routing
   information is that NO routing information ever gets LOST or
   DISCARDED.

   The routing database MUST adopt one of the following strategies:

   o  It must keep ALL alternative routing information it learns from
      any routing updates from the LAN and the WAN, so that if the
      best route disappears an alternative route (if available) can
      replace it as the new best route.

   o  If the amount of memory this consumes is problematic the routing
      application must keep SOME alternative routing information - say
      a best route and two alternatives.

      If the router ever has to discard routing information about a
      route it should note the fact.  If the routes that have been
      kept disappear because they have become unreachable, the router
      MUST issue a request on all interfaces to try and obtain
      discarded alternatives.

      It is recommended that the request is issued BEFORE all routes
      to a destination have been lost.

   Entries in the routing database can either be permanent or temporary.
   Entries learned from broadcasts on LANs are temporary. They will
   expire if not periodically refreshed by further broadcasts.

   Entries learned from a triggered response on the WAN are 'permanent'.
   They MUST not time out in the normal course of events.  The entries
   state MUST be changed to 'temporary' by the following events:

   o  The arrival of a routing update containing the entry set to
      unreachable.

      The normal hold down timer MUST be started, after which the
      entry disappears from the routing database.




RFC 1582                       Demand RIP                  February 1994


   o  The arrival of a routing update with the entry absent.

      If the hold down timer is not already running, the entry MUST be
      set to unreachable and the hold down timer started.

   o  A message sent from the circuit manager, to indicate that it
      failed to make a connection in normal running.

      The routing table MUST be scanned for all routes via that next
      hop router.  Aging of these routing entries MUST commence.  If
      the aging timer expires the entry MUST be set to unreachable and
      the hold down timer started.  If the hold down timer expires the
      entry disappears from the routing database.

   o  If the interface goes down, the circuit manager should indicate
      that all circuits on that interface have gone down.

   Database timer values are covered in section 7.

2.7 New Packet Types

   To support triggered updates, three new packet types MUST be
   supported:

   TRIGGERED REQUEST

             A request to the responding system to send all
             appropriate elements in its routing database.

             A triggered request is retransmitted at periodic
             intervals until a triggered response is received.

             Routing requests are transmitted in the following
             circumstances:

             o  Firstly when the router is powered on.

             o  Secondly when the circuit manager indicates a
                destination has been in an unreachable (circuit down)
                state for an extended period and changes to a
                reachable (circuit up) state.

             o  Thirdly in the event of all routing update fragments
                failing to arrive within a set period.

             o  It may also send triggered requests at other times to
                compensate for discarding non-optimal routing
                information.



RFC 1582                       Demand RIP                  February 1994


   TRIGGERED RESPONSE

             A message containing all appropriate elements of the
             routing database.  An appropriate element is an entry
             NOT learned from the interface to which the routing
             information is being sent out.  This is known as "split
             horizon".

             Stability is improved by adding "poisoned reverse" on
             routes learned from a destination.  This consists of also
             including some routes learned from a destination in
             routing updates sent back to that destination, but
             setting the routes as unreachable.  A route is only
             poisoned if it is the best route (rather than an inferior
             alternative route) in the database.

             A triggered response message may be sent in response to a
             triggered request, or it may be an update message issued
             because of a change in the routing database.

             A triggered response message MUST be sent in response to
             a triggered request message even if there are no routes
             to propagate.  This would be the case for a host which
             had a WAN interface only, but which wished to run the
             triggered update protocol.

             A triggered response is retransmitted at periodic
             intervals until a triggered acknowledgement is received.

   TRIGGERED ACKNOWLEDGEMENT

             A message sent in response to every triggered response
             packet received.

   Triggered response and triggered acknowledgement packets MUST contain
   additional fields for a sequence number, fragment number and number
   of fragments.

   If a triggered request or response is not acknowledged after 10
   retransmissions, routes to the destination should be marked as
   unreachable for the duration of a hold down timer before being
   deleted.

   The destination should then be polled at a lower frequency using
   triggered request packets.  When a triggered response is received,
   the router should prime the next hop router my sending its routing
   database through triggered response packets.




RFC 1582                       Demand RIP                  February 1994


   Strictly speaking polling should occur indefinitely to guarantee
   database integrity.  However the administrator MAY wish the router to
   cease polling after a few attempts believing that the lack of
   response is due to a mis-configuration of the next hop router.  The
   destination should be marked as NOT supporting the mechanism and no
   further routing messages should be sent to that destination.

   Before marking the destination as not supporting the mechanism, at
   least 5 triggered request polls (without acknowledgement) should be
   sent.

   If a destination marked as not supporting the mechanism, subsequently
   sends a valid 'triggered' message, the destination should be marked
   as supporting the mechanism once more (to allow for the next hop
   router's configuration being changed).  It should be sent a triggered
   request and a triggered response to obtain and propagate up-to-date
   routing information.

2.8 Fragmentation

   If a routing update is sufficiently large, the information MUST be
   fragmented over several triggered response packets:

   o  Each fragment MUST be individually acknowledged with a triggered
      acknowledgement packet.

      The sender of the routing update MUST periodically retransmit
      fragments which have not been acknowledged (or until the
      destination is marked as not supporting the mechanism).

   o  A router receiving fragments MUST re-assemble them before
      updating its routing database.

   o  If all fragments are not received within four times the
      retransmit period, they MUST be discarded.

      A triggered request packet MUST then be sent to the originator
      of the routing update.

      On receiving the triggered request packet, the originator of the
      routing update MUST retransmit ALL fragments.

   o  If a fragment with an updated sequence number is received, ALL
      fragments with the earlier sequence number MUST be discarded.

      An updated sequence number is defined as any sequence number
      that is different.  There is no concept of the value of the
      sequence number conveying its age.



RFC 1582                       Demand RIP                  February 1994


   Fragmentation timer values are covered in section 7.

2.9 Preventing Queue Overload

   In order to prevent too many routing messages being queued at a WAN
   interface, the routing task MAY operate a scheme whereby
   'broadcasting' of a triggered request or triggered response to a WAN
   interface is staggered.  All routing requests or routing responses
   are not sent to ALL next hop routers on the interface in a single
   batch:

   o  The routing task should limit the number of outstanding triggered
      request messages for which a triggered response has not been
      received.

   o  The routing task should limit the number of outstanding triggered
      response messages for which a triggered acknowledgement has not
      been received.

   As outstanding messages are appropriately acknowledged, further
   messages can be sent out to other next hop routers, until all next
   hop routers have been sent the message and have acknowledged it.

   The maximum number of outstanding messages transmitted without
   acknowledgement is a function of the link speed and the number of
   other routing protocols operating the triggered update mechanism.

   Messages should always be acknowledged immediately (even if it causes
   the limit to be exceeded), since a connection is almost certainly
   available.  This has the potential benefit of allowing the VC to
   close sooner (on its idle timer).

   Sending all triggered request fragments to a destination at once is
   also beneficial.

3. IP Routing Information Protocol Version 1

   This section should be read in conjunction with reference [1].

   IP RIP is a UDP-based protocol which generally sends and receives
   datagrams on UDP port number 520.

   To support the mechanism outlined in this proposal the packet format
   for RIP version 1 [1] is modified as shown in Figure 2.

   Every Routing Information Protocol datagram contains the following:





RFC 1582                       Demand RIP                  February 1994


   COMMAND   Commands supported in RIP Version 1 are: request (1),
             response (2), traceon (3), traceoff (4), SUN reserved (5).
             The fields sequence number, fragment number and number of
             fragments MUST NOT be included in packets with these
             command values.

             The following new commands (with values in brackets) are
             required:

       TRIGGERED REQUEST (6)

                 A request for the responding system to send all of its
                 routing database.

                 Only the first 4 octets of the packet format shown in
                 figure 2 are sent, since all routing information is
                 implied by this request type.

       TRIGGERED RESPONSE (7)

                 A message containing all of the sender's routing
                 database, excluding those entries learned from the
                 interface to which the routing information is being
                 sent.

                 This message may be sent in response to a triggered
                 request, or it may be an update message resulting
                 from a change in the routing database.

                 A triggered response message MUST be sent in response
                 to a triggered request message even if there are no
                 routes to propagate.  This would be the case for a
                 host which had a WAN interface only, but which wished
                 to run the triggered update protocol.

     0                   1                   2                   3 3
     0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
     +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
     | command (1)   | version (1)   |      must be zero (2)         |
     +---------------+---------------+-------------------------------+

        The following new fields are inserted for some commands

     0                   1                   2                   3 3
     0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
     +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
     |    sequence number (2)        | fragment (1)  |no of frags (1)|
     +-------------------------------+-------------------------------+



RFC 1582                       Demand RIP                  February 1994


          Followed by up to 25 routing entries (each 20 octets)

     0                   1                   2                   3 3
     0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
     +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
     | address family identifier (2) |      must be zero (2)         |
     +-------------------------------+-------------------------------+
     |                         IP address (4)                        |
     +---------------------------------------------------------------+
     |                        must be zero (4)                       |
     +---------------------------------------------------------------+
     |                        must be zero (4)                       |
     +---------------------------------------------------------------+
     |                          metric (4)                           |
     +---------------------------------------------------------------+
                                     .
                                     .

     The format of an IP RIP datagram in octets, with each tick mark
     representing one bit.    All fields are in network order.

     The four octets: sequence number (2), fragment number (1) and
     number of fragments (1) are not present in the original RIP
     specification.  They are only present if command takes the
     values 7 or 8.


          Figure 2.   IP Routing Information Protocol packet format


       TRIGGERED ACKNOWLEDGEMENT (8)

                 A message sent in response to every triggered response
                 packet received.

                 Only the first 8 octets of the packet format shown in
                 figure 2 are sent.

   VERSION   In this instance Version 1.

   SEQUENCE NUMBER

             This is a new field inserted if command takes the values 7
             or 8.

             The sequence number MUST be incremented every time updated
             information is sent out on a WAN.  The sequence number
             wraps round at 65535.



RFC 1582                       Demand RIP                  February 1994


             When a triggered acknowledgement is sent the sequence
             number is set to the same value as the triggered response
             packet being acknowledged.

             The sequence number MUST be identical over fragments.  If a
             fragment is retransmitted the sequence number MUST not
             change.

   FRAGMENT NUMBER

             The fragment number is one for the first fragment of a
             routing update, and is incremented for each subsequent
             fragment.  A fragment can contain up to 25 routing entries.

             When a triggered acknowledgement is sent the fragment
             number is set to the same value as the triggered response
             packet being acknowledged.

   NUMBER OF FRAGMENTS

             In a triggered response packet this indicates the number of
             packets required to complete the routing update.

             This field has no relevance for triggered acknowledgement
             packets so should be set to zero.

   For triggered response packets the rest of the datagram contains a
   list of destinations, with information about each.  Each entry in
   this list contains the address family identifier (2 for IP), a
   destination network or host, and the metric for it.  The packet
   format is intended to allow RIP to carry routing information for
   several different protocols, identifiable by the family identifier.

   The IP address is the usual Internet address, stored as 4 octets in
   network order.  The metric field contains a value between 1 and 15
   inclusive, specifying the current metric for the destination, or the
   value 16 (representing 'infinity'), which indicates that the
   destination is not reachable.  Each route sent by a router supersedes
   any previous route to the same destination from the same router.

   The maximum datagram size is 508 octets, excluding UDP and IP
   headers.

4. IP Routing Information Protocol Version 2

   An enhancement to IP RIP to include subnetting has recently become
   available [2].  This section only describes differences from that
   RFC.



RFC 1582                       Demand RIP                  February 1994


   The triggered update mechanism can be supported by including the
   triggered request (6), triggered response (7) and triggered
   acknowledgement (8) commands described in the previous section.

   The sequence number, fragment number and number of fragments fields
   are included in triggered response and triggered acknowledgement
   commands.

   The triggered request packet should also contain the 4 extra octets
   corresponding to the sequence number, fragment number and number of
   fragments fields - but set to zero.

   Because additional security information is included in RIP Version 2
   packets, this MUST be appended to the triggered request and triggered
   acknowledgement packets, as well as being present in the triggered
   response packet.

   The version number becomes 2.  Other aspects of packet layout follow
   reference [2].

5. Netware Routing Information Protocol

   This section should be read in conjunction with references [3], since
   it only describes differences from the specification.

   Netware [3] is the trade name of Novell Research's protocols for
   computer communication which are derived and extended from Xerox
   Network System's (XNS) protocols [4].

   Netware supports a mechanism that allows routers on an internetwork
   to exchange routing information using the Routing Information
   Protocol (RIP) which runs over the Internetwork Packet Exchange (IPX)
   protocol using socket number 453h.

   Netware RIP and IP RIP share a common heritage, in that they are both
   based on XNS RIP, but there is some divergence, mostly at the packet
   format level to reflect the differing addressing schemes.

   The triggered update mechanism can be applied to Netware RIP.  To
   support the mechanism outlined in this proposal the packet format for
   Netware RIP is modified as shown in Figure 3.

   Every datagram contains the following:








RFC 1582                       Demand RIP                  February 1994


   RIP OPERATION

             Operations supported in standard Netware RIP are: request
             (1) and response (2).

             The fields sequence number, fragment number and number of
             fragments MUST NOT be included in packets with these
             operation values.

             The following new operations are required (with values
             chosen to be the same as for IP RIP commands):

     0                   1         1
     0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5
     +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
     |       operation (2)           |
     +---------------+---------------+

        The following new fields are inserted for some operations

     0                   1                   2                   3 3
     0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
     +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
     |      sequence number (2)      | fragment (1)  |no of frags (1)|
     +-------------------------------+-------------------------------+

           Followed by up to 50 routing entries (each 8 octets)

     0                   1                   2                   3 3
     0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
     +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
     |                       network number (4)                      |
     +---------------------------------------------------------------+
     |       number of hops (2)      |      number of ticks (2)      |
     +---------------------------------------------------------------+
                                     .
                                     .

     The format of a Netware RIP datagram in octets, with each tick
     mark representing one bit.  All fields are in network order.

     The four octets: sequence number (2), fragment number (1) and
     number of fragments (1) are not present in the original RIP
     specification.  They are only present if operation takes the
     values 7 or 8.

        Figure 3.   Netware Routing Information Protocol packet format




RFC 1582                       Demand RIP                  February 1994


       TRIGGERED REQUEST (6)

                 A request for the responding system to send all of its
                 routing database.

                 Only the first 2 octets of the packet format shown in
                 figure 3 are sent, since all routing information is
                 implied by this request type.

       TRIGGERED RESPONSE (7)

                 A message containing all of the sender's routing
                 database, excluding those entries learned from the
                 interface to which the routing information is being
                 sent.

                 This message may be sent in response to a triggered
                 request, or it may be an update message resulting
                 from a change in the routing database.

                 A triggered response message MUST be sent in response
                 to a triggered request message even if there are no
                 routes to propagate.  This would be the case for a
                 host which had a WAN interface only, but which wished
                 to run the triggered update protocol.

       TRIGGERED ACKNOWLEDGEMENT (8)

                 A message sent in response to every triggered
                 response packet received.

                 Only the first 6 octets of the packet format shown in
                 figure 3 are sent.

   SEQUENCE NUMBER

             This is a new field inserted if operation takes the
             values 7 or 8.

             The sequence number MUST be incremented every time
             updated information is sent out on a WAN.  The sequence
             number wraps round at 65535.

             When a triggered acknowledgement is sent the sequence
             number is set to the same value as the triggered response
             packet being acknowledged.





RFC 1582                       Demand RIP                  February 1994


             The sequence number MUST be identical over fragments.  If
             a fragment is retransmitted the sequence number MUST not
             change.

   FRAGMENT NUMBER

             The fragment number is one for the first fragment of a
             routing update, and is incremented for each subsequent
             fragment.  A fragment can contain up to 50 routing entries.

             When a triggered acknowledgement is sent the fragment
             number is set to the same value as the triggered response
             packet being acknowledged.

   NUMBER OF FRAGMENTS

             In a triggered response packet this indicates the number
             of packets required to complete the routing update.

             This field has no relevance for triggered acknowledgement
             packets so should be set to zero.

   For triggered response packets the rest of the datagram contains a
   list of networks, with information about each.  Each entry in this
   list contains a destination network, and the number of hops and
   number of ticks for each.

   The maximum datagram size is 406 octets, excluding the IPX header (a
   further 30 octets).

6. Netware Service Advertising Protocol

   This section should be read in conjunction with references [3], since
   it only describes differences from the specification.

   Netware [3] also supports a mechanism that allows servers on an
   internetwork to advertise their services by name and type using the
   Service Advertising Protocol (SAP) which runs over the Internetwork
   Packet Exchange (IPX) protocol using socket number 452h.

   SAP operates on similar principals to running RIP.  Routers act as
   SAP agents, collecting service information from different networks
   and relay it to interested parties.

   To support the triggered update mechanism outlined in this proposal
   the packet format for Netware SAP is modified as shown in Figure 4.

   Every Service Advertising Protocol datagram contains the following:



RFC 1582                       Demand RIP                  February 1994


   SAP OPERATION

             Operations supported in standard Netware SAP are: general
             service query (1), general service response (2), nearest
             service query (3) and nearest service response (4).

             The fields sequence number, fragment number and number of
             fragments MUST NOT be included in packets with these
             operation values.

             The following new operations are required:

       TRIGGERED GENERAL SERVICE QUERY (6)

                 A request for the responding system to send the
                 identities of all servers of all types.

                 Only the first 2 octets of the packet format shown in
                 figure 4 are sent, since all service types are
                 implied by this request type.

     0                   1         1
     0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5
     +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
     |       operation (2)           |
     +---------------+---------------+

        The following new fields are inserted for some operations

     0                   1                   2                   3 3
     0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
     +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
     |      sequence number (2)      | fragment (1)  |no of frags (1)|
     +-------------------------------+-------------------------------+

















RFC 1582                       Demand RIP                  February 1994


           Followed by up to 8 service entries (each 66 octets)

     0                   1                   2                   3 3
     0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
     +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
     |                       Service Type (4)                        |
     +---------------------------------------------------------------+
     |                       Service Name (48)                       |
     +                                                               +
                                  .
     |                            .                                  |
     +---------------------------------------------------------------+
     |                       Network Address (4)                     |
     +---------------------------------------------------------------+
     |                        Node Address (6)                       |
     +                               +-------------------------------+
     |                               |      Socket Address (2)       |
     +---------------------------------------------------------------+
     |       Hops to Server (2)      |
     +-------------------------------+
                                     .
                                     .

     The format of a Netware SAP datagram in octets, with each tick
     mark representing one bit.  All fields are in network order.

     The four octets: sequence number (2), fragment number (1) and
     number of fragments (1) are not present in the original SAP
     specification.  They are only present if operation takes the
     values 7 or 8.


        Figure 4.   Netware Service Advertising Protocol packet format


       TRIGGERED GENERAL SERVICE RESPONSE (7)

                 A message containing all of the sender's services
                 table, excluding those entries learned from the
                 interface to which the service advertising
                 information is being sent out.

                 This message may be sent in response to a triggered
                 general service query, or it may be an update message
                 resulting from a change in the service advertising
                 database.





RFC 1582                       Demand RIP                  February 1994


                 A triggered general service response message MUST be
                 sent in response to a triggered general request
                 message even if there are no services to advertise.
                 This would be the case for a router with a LAN
                 network which had work stations but no servers on it.

       TRIGGERED GENERAL SERVICE ACKNOWLEDGEMENT (8)

                 A message sent in response to every triggered general
                 service response packet received.

                 Only the first 6 octets of the packet format shown in
                 figure 4 are sent.

   SEQUENCE NUMBER

             This is a new field inserted if operation takes the values
             7 or 8.

             The sequence number MUST be incremented every time updated
             information is sent out on a WAN.  The sequence number
             wraps round at 65535.

             When a triggered general service acknowledgement is sent
             the sequence number is set to the same value as the
             triggered general service response packet being
             acknowledged.

             The sequence number MUST be identical over fragments.  If
             a fragment is retransmitted the sequence number MUST not
             change.

   FRAGMENT NUMBER

             The fragment number is one for the first fragment of a
             triggered general service response update, and is
             incremented for each subsequent fragment.  A fragment can
             contain up to 8 service entries.

             When a triggered general service acknowledgement is sent,
             the fragment number is set to the same value as the
             triggered general service response packet being
             acknowledged.

   NUMBER OF FRAGMENTS

             In a triggered response packet this indicates the number of
             packets required to complete the service update.



RFC 1582                       Demand RIP                  February 1994


             This field has no relevance for triggered acknowledgement
             packets so should be set to zero.

   For triggered general service response packets the rest of the
   datagram contains a list of services, with information about each.
   Each entry in this list contains the service type, service name, full
   address (network, node and socket), and the number of hops to the
   server.

   The maximum datagram size is 534 octets, excluding the IPX header (a
   further 30 octets).

7. Timers

   A number of timers are supported to handle the triggered update
   mechanism:

   o  Database timers.

   o  Retransmission timer.

   o  Reassembly timer.

   In this section appropriate timer values for IP RIP are suggested.

   For other routing protocols, only the database timer should need to
   take different values.  The database timer values are chosen to match
   equivalent timer operation for using the protocol on a LAN.  The
   behaviour of a routing entry when a timer is running becomes
   indistinguishable from a routing entry learned from a broadcast
   update.

   Implementations MAY make timer values configurable - and hence
   different from the values suggested here - but interoperability
   requires that all timers on a sub-network should be the same in all
   routers.

7.1 Database Timers

   Routes learned by a triggered response command (7) are normally
   considered to be permanent - that is they do NOT time out unless
   activated by one of the following events:

   o  If the circuit manager indicates that a next hop router cannot be
      contacted, all routes learned from that next hop router should
      start timing out as if they had (just) been learned from a
      conventional response command (2).




RFC 1582                       Demand RIP                  February 1994


      Namely each route exists while the database entry timer is
      running and is advertised on other interfaces as if still
      present.  The route is then advertised as unreachable while a
      further hold down timer is allowed to expire, at which point the
      entry is deleted.

      If the circuit manager indicates that the next hop router can be
      contacted while the database entry timer is running, the routes
      are reinstated as permanent entries.

      If the database entry timer has expired and the circuit manager
      indicates that the next hop router is reachable, the routing
      application MUST issue a triggered request.  The routes will be
      reinstated on the basis of any triggered response packet(s)
      received.

   o  If a triggered response packet is received in which a route is
      marked unreachable, the hold down timer MUST be started and the
      entry is advertised as unreachable on other interfaces.  On
      expiry of the hold down timer the entry is deleted.

      If a triggered response packet is received in which an existing
      route is ABSENT, the hold down timer MUST also be started and
      the entry is advertised as unreachable on other interfaces.  On
      expiry of the hold down timer the entry is deleted.

   For IP RIP the hold down timer should always run for 120 seconds, to
   be consistent with RIP usage on broadcast networks.  The database
   entry timer should by default run for 180 seconds.  The network can
   be made more responsive by reducing the database entry timer value.
   However, making this timer too short can lead to network
   instabilities.  The duration of the database entry timer allows a
   period of grace in which contention for network resources can be
   resolved by the circuit manager.

7.2 Retransmission Timer

   The routing task runs a retransmission timer:

   o  When a triggered request is sent it will be retransmitted
      periodically while a triggered response packet is not received.

   o  When a triggered response is sent a note of the sequence number
      and fragment number(s) of the routing update is kept.

      Fragments will be retransmitted at periodic intervals while a
      triggered acknowledgement packet is not received for the
      appropriate fragment.



RFC 1582                       Demand RIP                  February 1994


   With call set up time on the WAN being of the order of a second, a
   value of 5 seconds for the retransmission timer is appropriate.

   If no response is received after 10 retransmissions, routes via the
   next hop router are marked as unreachable, the hold down timer MUST
   be started and the entry is advertised as unreachable on other
   interfaces.  On expiry of the hold down timer the entry is deleted.

   The next hop router is then polled using a triggered request packet
   at 60 second intervals.  If a response is received the routers should
   exchange routing information using triggered response packets.

   It may not be desirable to poll indefinitely, since a lack of
   response (when a circuit is up) is most likely caused by incorrect
   configuration of the next hop router.  An administrator definable
   number of polls (5 or greater) should be provided.

   If the circuit manager indicates that the next hop router is
   unreachable, the retransmission is suppressed until the circuit
   manager indicates that the next hop router is reachable once more.
   Counting of the number of retransmissions continues from where it
   left off prior to the circuit down indication.

7.3 Reassembly Timer

   When a router receives a triggered response update it MUST
   acknowledge each fragment.  If the routing update is fragmented over
   more than one packet, the receiving router MUST store the fragments
   until ALL fragments are received.

   On receiving the first fragment a timer should be started.  If all
   fragments of the routing update are not received within that period
   they are discarded - and a triggered request is sent back to the
   originator (with retransmissions if necessary).  The originator MUST
   then resend ALL triggered response fragments.

   The reassembly timer should be set to four times the value of the
   retransmission timer.  With a suggested retransmission timer value of
   5 seconds, the suggested reassembly timer value SHOULD be 20 seconds.

   Implementations MAY allow the reassembly timer and retransmission
   timer to be configurable (in the 1:4 ratio), but interoperability
   will be compromised on WANs where all participating routers DO NOT
   support the same values for these timers.

   Fragments MUST also be discarded if a new fragment with a different
   sequence number is received.  A triggered request MUST not be sent in
   this instance.



RFC 1582                       Demand RIP                  February 1994


8. Implementation Considerations

   In the implementation described in this memo, it is assumed that
   there is a close binding between the circuit manager and the routing
   applications - that they are in some way the same 'program'.  This is
   not necessarily true of all products which are routers.

   In particular there are UNIX host implementations in which the
   routing application is distinct from the kernel, where the circuit
   manager is likely to be installed.  In such systems it is possible to
   stop (or crash) the routing applications independently of what is
   happening in the kernel.

   Other implementations might have the circuit manager on a separate
   card which again may give the circuit manager a life of its own.

   In implementations where the applications and circuit manager have
   independent lives, a keep-alive mechanism MUST be provided between
   the applications and the circuit manager, so that if the application
   or network layer dies and is subsequently re-started they can
   resynchronize their state tables.

   Ideally, when an application dies, the circuit manager should close
   all existing VCs appropriate to the application and make no further
   outgoing calls and reject incoming calls until the application is
   running again.

   If the circuit manager is using some form of encapsulation, several
   applications may be sharing the same VC.  If this is the case the
   circuit manager may wish to filter out datagrams for the appropriate
   network layer if only one of the applications is affected.  But this
   is not an ideal solution.

   Conversely if the application believes the circuit manager has died,
   it should mark all routes via the circuit manager as unreachable and
   advertise them on other interfaces for the duration of the hold down
   timer before deleting them.

9. Security Considerations

   Security is provided my a number of aspects:

   o  The circuit manager is required to be provided with a list of
      physical addresses to enable it to establish a call to the next
      hop router on an X.25 SVC or ISDN B-channel.

      The circuit manager SHOULD only allow incoming calls to be
      accepted from the same well defined list of routers.



RFC 1582                       Demand RIP                  February 1994


      Elsewhere in the system there will be a set of logical address
      and physical address tuples to enable the network protocols to
      run over the correct circuit.  This may be a lookup table, or in
      some instances there may be an algorithmic conversion between
      the two addresses.

   o  The routing (or service advertising) task MUST be provided with a
      list of logical addresses to which triggered updates are to be
      sent on the WAN.  The list MAY be a subset of the list of next
      hop routers maintained by the circuit manager.

      There MAY also be a separate list of next hop routers to which
      traditional broadcasts of routing (or service advertising)
      updates should be sent.  Next hop routers omitted from either
      list are assumed to be not participating in routing (or service
      advertising) updates.

      The list (or lists) doubles as a list of routers from which
      routing updates are allowed to be received from the WAN.  Any
      routing information received from a router not in the
      appropriate list MUST be discarded.

10. References

   [1] Hedrick. C., "Routing Information Protocol", STD 34, RFC 1058,
       Rutgers University, June 1988.

   [2] Malkin. G., "RIP Version 2 - Carrying Additional Information",
       RFC 1388, Xylogics, January 1993.

   [3] Novell Incorporated., "IPX Router Specification", Version 1.10,
       October 1992.

   [4] Xerox Corporation., "Internet Transport Protocols", Xerox System
       Integration Standard XSIS 028112, December 1981.

   [5] Malis. A., Robinson. D., and R. Ullmann, "Multiprotocol
       Interconnect on X.25 and ISDN in the Packet Mode", RFC 1356, BBN
       Communications, Computervision Systems Integration, Process
       Software Corporation, August 1992.











RFC 1582                       Demand RIP                  February 1994


11. Author's Address

       Gerry Meyer
       Spider Systems
       Stanwell Street
       Edinburgh EH6 5NG
       Scotland, UK

       Phone: (UK) 31 554 9424
       Fax:   (UK) 31 554 0649
       EMail: gerry@spider.co.uk