💾 Archived View for gemini.bortzmeyer.org › rfc-mirror › rfc1029.txt captured on 2023-12-28 at 18:59:19.

View Raw

More Information

⬅️ Previous capture (2021-11-30)

-=-=-=-=-=-=-

Network Working Group                                           G. Parr
Request For Comments: 1029                         University of Ulster
                                                               May 1988


        A MORE FAULT TOLERANT APPROACH TO ADDRESS RESOLUTION FOR
                    A MULTI-LAN SYSTEM OF ETHERNETS

STATUS OF THIS MEMO

   This memo discusses an extension to a Bridge Protocol to detect and
   disclose changes in neighbouring host address parameters in a Multi-
   LAN system of Ethernets.  The problem is one which is appearing more
   and more regularly as the interconnected systems grow larger on
   Campuses and in Commercial Institutions.  This RFC suggests a
   protocol enhancement for the Internet community, and requests
   discussion and suggestions for improvements.  Distribution of this
   memo is unlimited.

ABSTRACT

   Executing a protocol P, a sending host S decides, through P's routing
   mechanism, that it wants to transmit to a target host T located
   somewhere on a connected piece of 10Mbit Ethernet cable which
   conforms to IEEE 802.3.  To actually transmit the Ethernet packet, a
   48 bit Ethernet/hardware address must be generated.  The addresses
   assigned to hosts within protocol P are not always compatible with
   the corresponding Ethernet address (being different address space
   byte orderings or values).  A protocol is presented which allows
   dynamic distribution of the information required to build tables that
   translate a host's address in protocol P's address space into a 48
   bit Ethernet address.  An extension is incorporated to allow such a
   protocol to be flexible enough to exist in a Transparent Bridge, or
   generic Host.  The capability of the Bridge to detect host reboot
   conditions in a multi-LAN environment is also discussed, emphasising
   particularly the effect on channel bandwidth.  To illustrate the
   operation of the protocol mechanisms, the Internet Protocol (IP) is
   used as a benchmark [6], [8].  Part 1 presents an introduction to
   Address Resolution, whilst Part 2 discusses a reboot detection
   process.

DEFINITIONS:

      CATENET: a group of IP networks linked together
      IP     : Internet Protocol






Parr                                                            [Page 1]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


                                 PART 1

INTRODUCTION

   In the Ethernet, while all packets are broadcast, the hardware
   interface selects only those with either the explicit hardware
   broadcast address or the individual hardware address of this
   interface.  Packets which do not have one of these two addresses are
   rejected by the interface and do not get passed to the host software.
   This saves a great deal of otherwise wasted effort by the host
   software having to examine packets and reject them.  If the interface
   hardware selected packets to pass to the host software by means of
   the protocol address, there would be no need for any translation from
   protocol to Ethernet address.  Although it is very important to
   minimize the number of packets which each host must examine, so
   reducing especially needless inspections, use of the hardware
   broadcast address should be confined to those situations where it is
   uniquely beneficial.  Perhaps if one were designing a new local
   network one could eliminate the need for an address translation, but
   in the real world of existing networks it fills a very important
   purpose.  A rare use of the broadcast hardware address, which avoids
   putting any processing load on the other hosts of the Ethernet, is
   where hosts obtain the information they need to use the specific and
   individual hardware addresses to exchange most of their packets.

REASONING BEHIND ADDRESS RESOLUTION

   The process of converting from the logical host address to the
   physical Ethernet address has been termed ADDRESS RESOLUTION, and has
   prompted research into a method which can be easily interfaced,
   whilst at the same time remaining portable.

   The Ethernet requires 48 bit addresses on the physical cable [11] due
   to the fact that the manufacturers of the LAN interface controllers
   assign a unique 48 bit address during production.  Of course, Network
   Managers do not want to be bothered using this address to identify
   the destination at the higher-level.  Rather, they would prefer to
   assign their logical names to the hosts within their supervision, and
   allow some lower level protocol to perform a resolving operation.
   Most of these logical protocol addresses are not 48 bits long, nor do
   they necessarily have any relationship to the 48 bit address space.

   For example, IP addresses have a 32 bit address space [6], thus
   giving rise to the need to distribute dynamically the correspondences
   between a <PROTOCOLTYPE,PROTOCOL-ADDRESS> pair, and a 48 bit Ethernet
   address.





Parr                                                            [Page 2]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


EXAMPLE ARP OPERATION

   Here is a review of the operation of ARP as defined in RFC-826 [5].
   Let hosts X and Y exist on the same Ethernet cable.  They have
   physical Ethernet addresses EA(X), and EA(Y), and DoD Internet
   addresses IPA(X), and IPA(Y).  Let the Ethernet type of Internet be
   ET(IP).  Host X begins an application, and sooner or later wishes to
   communicate an Internet packet to host Y.  Host X has knowledge of
   the Internet address of Y, i.e., (IPA(Y)), and informs the lower
   level that it wishes to talk to IPA(Y).  The lower-level subsequently
   consults the ARP Module (ARM) to convert <ET(IP),IPA(Y)> into a 48
   bit Ethernet address but because X has not talked to Y previously, it
   does not have this information in its Translation Cache (TC).  It
   discards (or queues) the Internet packet, and creates a new Address
   Resolution packet with:

       PACKET FIELD             VALUE ASSIGNED

        HRDTYP                   ETHERNET

        PROTYP                   ET(IP)

        HRDLEN                   length (EA(X))

        PROTLEN                  length (IPA(X))

        ARPOPC                   REQUEST

        SOURCE HWR               EA(X)

        SOURCE PROT              IPA(X)

        TARGET HWR               don't know

        TARGET PROT              IPA(Y)

   It then broadcasts this packet to all hosts on the connecting cable.
   Host Y picks up this packet and determines that it understands the
   hardware type (Ethernet), that it speaks the indicated protocol
   (Internet), and that the packet is for it, that is, TARGET PROTOCOL
   ADDRESS = IPA(Y).  Replacing any previous entry, it enters the
   information that <ET(IP),IPA(X) translates to EA(X).  It then learns
   that this is an ARREQ packet, so it swaps fields, placing EA(Y) in
   the new sender Ethernet address field SOURCE HARDWARE ADDRESS, EA(X)
   as TARGET HARDWARE ADDRESS, IPA(X) as TARGET PROTOCOL ADDRESS, IPA(Y)
   as SOURCE PROTOCOL ADDRESS, and sets the opcode to REPLY.  The packet
   is then sent with direct routing address information to EA(X).  Thus,
   Y now knows how to send to X, but X still doesn't know EA(Y).



Parr                                                            [Page 3]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


   When X receives the ARREP packet from Y, it gets the address
   information into its translation cache ET(IP),IPA(Y)>-->EA(Y),
   notices that it is a REPLY, and discards the packet (i.e., disposes
   of the dynamic packet buffer).  However, if the original Internet
   Module packet had been queued, it could have been accessed and given
   the full addressing information from the translation cache.
   Alternatively, had it been discarded, the higher level would have
   succeeded on a subsequent attempt, and the Internet packet would be
   transmitted immediately.

OBTAINING GREATER NETWORKING RANGE

   There are many benefits to be gained in dividing a large multiuser
   network into smaller, more manageable networks.  These include : Data
   Security; Overall Network Reliability; Performance Enhancement; not
   to mention the most obvious: Greater Networking Range.  In some
   network technologies, cable length may be stipulated not to exceed a
   certain range due to electrical limitations.  By installing a Bridge,
   this restriction is effectively eliminated.  An important
   consideration is the effect the induced Bridge delays will have on
   the protocol timeouts in operation on each LAN/Subnet.  Careful
   analysis of upper bounds on timeouts would have to be made in order
   to gain full benefit from the increased range.  In the case of
   Ethernet the following system parameters exist [11], [12]:

        - the bus bandwidth is 10Mbit/s

        - the maximum node-to-node cable length is 1500 m

        - the maximum point-to-point link cable length is 1000 m

        - the maximum number of repeaters between two nodes is two

        - the worst case end-to-end bus propagation delay is 22.5 us

        - the jam time after collision is 32bit

        - the minimum interframe time is 9.6 us

        - the slot size is 512 bit = 51.2 us

   Once a decision has being taken to subnet, the resulting subLANs may
   be connected by including a Bridge to link them together and
   providing a protocol which makes the collection of subnets appear as
   a single network.  The basic idea of the Bridge providing 'repeater'
   facilities would not suffice in this application.  Moreover, the
   Bridge would have to have further 'intelligence' to enable it to
   select those packets which are destined for remote networks based on



Parr                                                            [Page 4]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


   the protocol address of the target host.  Thereby preventing it from
   forwarding packets needlessly that will not be accepted.  If this
   procedure was not adhered to, the channel bandwidth on the remote
   networks would be inundated with packets, causing local valid traffic
   to backoff and the efficiency of the respective networks to rapidly
   decrease.

   One problem fundamental to the operation of the Bridge is how it
   discovers on which LAN a particular host is interfaced.  If there are
   only two LANs in the system, each will have a dedicated cache at the
   Bridge, and when a packet is received at the particular interface,
   the source host's address parameters are entered in the respective
   LAN cache.  However, when we consider a Multi-LAN environment, the
   procedure becomes more complicated.

   ___
    |
    |-----h3
    |                                            E4
    |-----hq                            |-----------------------|
    |                _                             |        |
    |-----hx        | | B1                         |        |
    |---------------| |                            |        |
    |-----h1        |_|                            |        |
    |                |     h19                     |        |      ______
    |                |    |                       | |        -----|______|  B4
    |                |    |                       | | B3              |
    |-----he       |-------------------| E2       |_|                 |
    |                    |                         |                  |
    |-----h5             |                         |                  |
    |                    |                         |                  |
    |                   ---                ---     |                  |
   ---                  | |                 |-------                  |
   E1                   | | B2              |                         |
                        | |-----------------|                         |
                        ---                 |                         |
                                            |          |---------------------
                                           ---                              |
                                            E3                              |
                                                                            |
                      FIGURE 1.  A MULTI-LAN TOPOLOGY


   In the normal set-up, whenever B3 or B4 would receive a packet on E4,
   they would both update the caches on their E4 interface.  In
   addition, a method must be provided to permit B4 to distinguish
   between packets arriving on E4 from E1, E2, E3, and those which
   actually originated on E4.



Parr                                                            [Page 5]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


   This is so that packets can be categorized as being of remote or
   local source and processed accordingly.  The most obvious solution is
   for each Bridge to act as an AGENT and plug in its address as the
   source of any packets it cascades to a remote network, instead of the
   packet being cascaded with its original source address.  At Bridge
   boot, it may issue a broadcast request for all locally connected
   hosts/devices to return their local network protocol addresses.  On
   subsequent receipt of this information, the Bridge could then update
   the cache for each of its interfaces so that it would now have a base
   from which to perform future operations.

   The alternative to this automatic procedure is to permit manual
   intervention in the Bridge software which could be activated by the
   network manager in order to key in the addresses of the hosts
   connected to each LAN interface.

   Thus, having provided a means for the Bridge to obtain the original
   state of the LAN addresses when it boots, how then does the Bridge
   distinguish the arrival of a new host on the locally connected system
   from transmissions which were sent from a remote source and cascaded
   by an adjacent Bridge?  Two approaches are currently under
   consideration to solve this problem, namely Explicit Subnets, and
   Transparent Subnets [4], [7], [9], [14].

   In the Explicit Subnet approach, the location of the host in the
   system is important.  The address of the host in the protocol suite
   will reflect which subnet the host is interfaced to.  Consequently
   the protocol address space is divided into a three level hierarchy of
   <network,subnet,host>.  Within the Internet there are five addressing
   divisions in operation [10], classes A, B, C, D, and E.  Classes D
   and E relate to an addressing technique that will be used for
   management of multi-casting groups and will not be discussed here.
   With such a structure, it is possible to provide an address mask at
   each interface so that received packets may have their source address
   fields examined and compared with the address mask of this LAN.  In
   so doing, the component which is being verified is actually the
   subnet address.  If the masking operation is successful the source
   must exist on this LAN, otherwise it must be remote.

   With the Transparent scheme, the first time a newly booted host
   'speaks' it will be looking for addressing information (probably
   using BOOTSTRAP [1], RARP [2] or ARP [5]).  Accordingly, the Bridge
   will detect these respective requests and be in a position to perform
   operations on the address parameters.  The current approach in
   Transparent Subnetting is that before any such requests can be
   cascaded by the Bridge to an adjacent LAN, that Bridge will place its
   interface address parameters into the source address fields, thus
   acting as the AGENT.  Therefore, this Bridge will 'see' either



Parr                                                            [Page 6]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


   packets arriving from the remote Bridge address, or local packets.
   By virtue of the RARP/ARP operation, which hosts perform when they
   first come up, any hi-level packets received on to the network not
   having the bridge address, and not having a mapping in the cache for
   that LAN, can be considered as being remote.

   Currently, there is a move toward the Transparent subnet proposal
   originally described by Postel [7].  This has been due mainly to
   practical problems of incompatible implementations from different
   vendors, and the restrictions that the Explicit address space place
   on the adaptability of the system to change (class C addresses are
   not flexible enough for the Explicit scheme).  It is also the opinion
   of the Author of this paper that the Agent technique adopted by the
   Bridges could have shortcomings in a dynamic environment which would
   be detrimental to its operation; for example, where the bridges
   themselves relocate or crash, or in the management of the "Agent For
   Who" cache at the bridge.  Insofar as Loop Resolution and
   SelfStabilization after failure are Bridge problems that need to be
   addressed, it is strongly felt their satisfactory solution will be
   supported by elimination of the Agent technique [13].

BRIDGE OPERATIONS

   Referring to figure 1, assume that at some stage during its
   processing [E1H3] wishes to communicate with [E2H19].  [E1H3] obtains
   knowledge of the Internet address of [E2H19] from its translation
   cache, but will not require the knowledge that [E2H19] exists on a
   completely different subnet.  [E1H3] calls its Internet Module to
   transmit the packet.  As detailed, the usual procedure of passing
   control to its ARM is performed in an attempt to obtain a
   translation.  If we assume that [E1H3], and [E2H19] have not talked
   before, the ARM in [E1H3] will not be able to resolve the addresses
   on the first attempt.

   In such a case, an ARREQ packet is assembled and broadcast to all
   hosts on the network [E1].  The packet traverses the cable and is
   eventually picked up by the (B1) Bridge Address Resolution Module
   (BARM), whereupon it determines whether or not it should intervene in
   the request.  If the target is determined as remote (i.e., having no
   match in the local cache), the BARM examines its Global Translation
   Cache (GTC) to determine if it has an entry for <protocol,[E2H19]>.
   Should a mapping be obtained at the Bridge, there is no need for the
   broadcast REQUEST packet to be cascaded on to the remote network
   [E2].  It is therefore assumed that the entries in the GTC reflect
   the most current addressing information.  A match thus obtained, the
   original ARREQ packet buffer is adapted as required and returned
   directly to [E1H3] via the Bridges hardware interface IFE1.




Parr                                                            [Page 7]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


   On the other hand, should the Bridges' GTC have no information on
   [E2H19], the BARM would have to perform the following steps:

      1.  drop the current ARREQ from [E1H3],

      2.  create its own ARREQ using the Bridge source addresses
          and copy the target_internet_addr from the original
          [E1H3] ARREQ packet,

      3.  broadcast the ARREQ on network E2 via network interface
          IFE2, and go into a timeout awaiting a REPLY.

   Should this timeout period expire, a number of retries will be
   permitted under control of the BARM.  Alternatively, if a REPLY is
   received within the timeout interval, then the BARM will update its
   GTC.  The ARM of [E1H3] next will attempt to transmit another ARREQ,
   but this time a mapping will be obtained at the BARM'S GTC, and the
   appropriate REPLY will be returned.

   Part 1 has described the state of the art of the behaviour of Address
   Resolution.  Part 2 now extends the study to the more serious problem
   of rebooting hosts in a multi-LAN system of Ethernets, and the
   effects such changes have on the integrity of state information held
   in ARP caches and routing tables.

                                 PART 2

THE CAPTURE OF REBOOTS

   Because Address Resolution packets are broadcast, all hosts on the
   connecting cable including the Transparent Bridge will pick them up
   and determine what they are.  Referring to figure 1, it may well be
   the case that a host on E1 wishes to communicate with a fellow host
   on the same physical ether.  Hence, if Hx wishes to talk to Hw on the
   same ether, but has not done so previously, it will broadcast an
   Address Resolution packet in the normal fashion.  The Bridge will
   also 'see' the packet as it passes by, and will act as described
   above, unless that is, there is some method of preventing it doing
   so; there is no point in the Bridge invoking its ARM, and wasting
   processing time if the problem is going to be resolved locally.

   It may occur however, that H1 wants to communicate with H5.  If
   however, H5 has not talked with anyone before (i.e., it has been
   "dormant"), H1 will issue an ARREQ.  The Bridge will not know that H5
   is local because it won't have been entered in the local address
   cache from previous conversations.  To avoid broadcasting an ARREQ to
   all networks/subnets, one way around this problem is to set up the
   contents of the local cache at Bridge startup time.  Therefore, the



Parr                                                            [Page 8]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


   Bridge will already know not to intervene.  Thus, if the Bridge (with
   2 nets) finds that a particular IP destination address is not in the
   local cache of interface 1, it would have to examine its GTC and scan
   it for a mapping.  Should no mapping be obtained at interface 2, one
   of two possibilities exist:

        1. the target host doesn't exist locally

        2. the caches are corrupt (the eventuality of this should
           be negligible!)

   If it is assumed that each of the translation caches contains have
   the most recent addressing information regarding its own domain of
   the network then, in this example, if the Bridge does not get a
   mapping at the GTC it would appear that the host must exist remotely
   from E1, and E2.

   Such a conclusion would ignore cases in which a host unplugs from a
   particular hardware interface and plugs into another hardware
   interface, or where logical names are reassigned to different
   interfaces due to host user change.  Either of these events could
   happen had the host being accessed on E2, which would mean that a
   REBOOT has taken place.

   Anticipating these possiblities local caches are essential.  In
   normal operation, the Bridge will process and forward IP packets
   received from one network, and destined for another.  If the Bridge
   picks up an ARREQ, it will first look for a mapping in its GTC before
   discarding the original ARREQ, and transmitting its own to the remote
   network.  In any case, the Bridge will always examine the local cache
   entries at the receiving interface, so that it may determine if the
   target address is local or remote.  When the Bridge first scans the
   local cache, it does so with the source IP address as the key.  If no
   mapping is retrieved, it then scans the GTC with the same key.
   Should a mapping now be obtained, it remains for the Bridge to insert
   the source IP into the local cache, where it has either been
   previously deleted or corrupted.

   However, if the source IP exists in the respective local cache, the
   validity of the source Ethernet address should also be verified by
   examining the respective entry in the GTC.  A scan of the GTC is then
   performed with <protocol,source_prot_addr> as the key.  If a mapping
   is retrieved, the respective <et_addr> should be checked against the
   source Ethernet address in the packet header.  If the addresses do
   not match, then we have uncovered a Hardware Reboot condition (i.e.,
   a change in Ethernet ID).  On the other hand, should the scan of the
   GTC with <protocol,source_prot_addr> fail to obtain a mapping, then
   the Bridge would scan the GTC with the current Ethernet address in



Parr                                                            [Page 9]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


   the packet header.  If this obtains a mapping, then a Protocol Reboot
   condition (i.e., change in logical ID) has been detected.

   In the next section, the implications of these forms of 'Reboot' are
   discussed.

REBOOT SCENARIO

   In normal operation, packets will uneventfully traverse each subnet
   either as complete Internet packets, broadcast ARREQ's, or direct
   ARREP's.  The Bridge attached to each subnet will 'hear', and 'see'
   all packets as they travel past its connected interfaces.  Because of
   the existence of the local caches at each interface, the Bridge can
   decide whether or not to intervene.  In general circumstances, each
   host on the Catenet will have a translation cache containing
   <protocol,source_prot_addr,source_et_addr> entries for all packets it
   has observed.  Most of these entries will have been due to processing
   ARREQ packets, which were broadcast, and by receiving REPLY packets.
   In accordance with the foregoing , the Bridge will have a cache
   attached to each subnet interface containing entries for protocol
   addresses.

   Within the Bridge's Global Translation Cache (GTC) will be entries of
   all <protocol,source_prot_addr,source_hrd_addr> triplets relating to
   valid hosts which have been recognised.  If we assume that we have
   just connected up a Catenet such as that illustrated in figure 1,
   then at power-up no stations will have knowledge about their
   neighbours.  If the Bridges are to remain transparent, the
   translation caches at each host will be totally empty.  The only
   addressing details that will be in existence will be the protocol
   addresses stored in the local caches of the Bridges.

   The hosts subsequently begin to run applications and will want to
   communicate with one another.  The first ARREQ is broadcast on the
   respective subnet and all hosts, including the Bridge's interface to
   the subnet, will pick it up and store the details.  If, for example,
   Hx issues an ARREQ for Hq, the Bridge will not intervene since there
   is no need (providing no reboot has occurred at Hq).  However, if Hx
   wishes to talk with Hz, B1 will determine that the target IP in the
   respective ARREQ does not exist in the local cache of IFE1, so it
   will examine the GTC, with the <protocol,target_prot_addr> of Hw as
   the key.

   It is assumed that there will be a timeout mechanism in operation at
   the source of any packet.  In addition, the Bridge may also place the
   target address in a 'search list' of currently sought hosts, so as to
   prevent ARREQs from different sources being cascaded for the same
   target.  Under these conditions, Hx may re-issue its original ARREQ,



Parr                                                           [Page 10]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


   but will be ignored until the host Hw has replied to the ARREQ
   transmitted by the Bridge.

NORMAL RUNNING STATE

   Assuming that a few ARP's have been issued, IP packets will start
   traversing the Catenet with full addressing information.  Again, the
   Bridges will 'see' all the packets.  If we extend the situation one
   step further, and assume that several conversations have taken place
   across the Catenet, there will be entries in the translation caches
   of the hosts concerned, regarding the
   <protocol,target_prot_addr,target_hrd_addr> triplets of those hosts
   with which the conversations took place.  The Bridges also, will have
   details in their GTC's for packets which they cascaded.

   If a host is relocated, any connections initiated by that host will
   still work, provided that its own translation cache is cleared when
   it does physically move.  However, any connections subsequently
   initiated to it by other hosts on the Catenet will have no particular
   reason to know to discard their old translation for that host.
   Ideally, 48 bit Ethernet addresses will be unique and fixed for all
   time.

RECOGNITION OF THESE REBOOT CONDITIONS

   With reference to figure 1, assume that for some reason a fault
   occurs on the hardware interface of <E1He>.  The result of this is
   that a new interface is installed with a newly acquired hardware
   address.  When <E1He> is powered up, the previous contents of its
   translation cache are cleared and it has no recollection of local, or
   remote host addresses.  Accordingly, <E1He> begins to issue ARREQ's
   to hosts it requires.  Whenever <E1He> transmits its first ARREQ, it
   could be termed a 'HELLO PACKET', since everyone on the subnet can
   pick up the packet, and store the relevant information in their
   translation caches.  Within hosts, a mapping will be found on the old
   <protocol,source_prot_addr> pair, and the current <et_addr> of the
   packet header will replace whatever is entered in the translation
   cache.

   At this point it would be easy for each host with an entry to
   recognise the Hardware Reboot situation and inform the subnet with a
   respective broadcast reboot packet.  But allowing such a procedure
   would be extremly inefficient on the broadcast medium, and would
   drastically outweigh any improvements in performance which might be
   obtained in the long term.  In any case, given the fact that the
   ARREQ is broadcast, all stations on the subnet will recognise the
   reboot.  The important point to consider is the effect such a reboot
   will have on subsequent conversations which are initiated remotely.



Parr                                                           [Page 11]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


   Can redundant transmissions be thwarted before they tie up processing
   time on hosts en-route to the rebooted target?  How these
   difficulties are resolved is critical to the level of performance
   obtained in a Catenet configuration.  Since it is not optimal for
   hosts to inform the system of a reboot, it is left to the Bridge.
   Whenever the Bridge receives a packet, be it IP, or ARP, it examines
   the source address parameters in the packet header, in the hope of
   detecting any incompatibilities between them and the entries in its
   caches.  There are three distinct possibilities, namely, a difference
   in the 48 bit hardware address only, a difference in the protocol
   address, and two completely new addresses.  If an incompatibility is
   discovered, a "REBOOT" packet is constructed and issued on all remote
   interfaces containing the appropiate information, allowing Bridges to
   update their GTC's and generic hosts their ARP caches.

   The structure of the Reboot packet is as depicted in figure 2.

    0                   1                   2                   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
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   | P A C K E T     O P C O D E   |REB OPC|      S O U R C E      |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |        H A R D W A R E            A D D R E S S               |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |       S O U R C E   P R O T O C O L     A D D R E S S         |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |     M U L T I C A S T   T A R G E T    H A R D W A R E        |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |    A D D R E S S      |   M U L T I C A S T     T A R G E T   |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |   P R O T O C O L     |
   +-+-+-+-+-+-+-+-+-+-+-+-+

          ---------> NEXT FOLLOWS A VARIANT FIELD ON REBOOT  OPCODE

   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |  O L D         S O U R C E        H A R D W A R E             |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |  A D D R E S S        |
   +-+-+-+-+-+-+-+-+-+-+-+-+

    OR

   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
   |  O L D     S O U R C E    P R O T O C O L      A D D R E S S  |
   +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

                          FIGURE 2. REBOOT PACKET



Parr                                                           [Page 12]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


   The following definitions apply:

        PACKET FIELD              VALUE

        OPCODE                    REBOOT

        REBOOT OPCODE             HARDWARE

        REBOOT OPCODE             PROTOCOL

   The format is then as follows:

        48 bit broadcast Ethernet address for the destination,

        48 bit Ethernet address of source Bridge,

        16 bit Protocol type = PACKET OPCODE - REBOOT.


   For completeness and error checking it may be an advantage to have a
   field which specifies the length of addresses in the Ethernet and
   protocol address spaces.  Thus, the Reboot packet structure contains
   the following:

   FIELD          FIELD SIZE                    DESCRIPTION

   HRDLEN          4 bit             byte length of Ethernet address

   PROTLEN         4 bit             byte length of Protocol address


   SOURCE
   PROTOCOL
   ADDRESS        32 bit            current protocol address of host

   TARGET
   PROTOCOL
   ADDRESS        32 bit           broadcast target protocol address

   REBOOT
   OPCODE          4 bit            will be either PROTOCOL or HARDWARE


   if   PROTOCOL       32 bit         old protocol address

   else HARDWARE       48 bit         old hardware  address





Parr                                                           [Page 13]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


   As shown, depending on the REBOOT-OPCODE, the structure will continue
   with either the 48 bit old hardware address or the 32 bit old
   protocol address.  The choice of a variant packet structure is for
   reasons of curtailing the size of the packet to the fields that are
   truely necessary in each situation.  From this Reboot packet
   structure, the process of generating such a packet can be considered.
   When the Bridge algorithm detects a reboot, it should create a reboot
   packet structure containing the relevant addressing information and
   subsequently multicast it on the interface(s) which access(es) the
   remote subnet(s).  The decision as to which interface(s) is/are
   local, and which is/are remote, can be resolved automatically
   whenever a packet is received.  With respect to this packet transfer
   the receive interface at the Bridge becomes local, and all others are
   tagged as remote.

   Thus, hosts on the subnet remote from the reboot are informed of the
   situation immediately as it is detected by the Bridge.  In the
   Catenet configuration illustrated in fig 1, this will have the effect
   of updating the Translation Cache within each host, whenever it
   receives the packet.  If for example, <E4Hw> reboots under hardware,
   B3 will detect this occurance.  There is no reason for the subnets
   E1, E2, E3 to be aware of this episode.  In normal operation, B3 will
   recognise the reboot from the first ARREQ issued from <E4Hw>.  With
   this reboot detection facility, B3 will be in a position to inform
   the hosts on E1, E2, and E3.  B3 can then create and issue the Reboot
   packet via its interface with E3.  When B3 picks it up, it will
   update its own caches and subsequently cascade the packet onto E2,
   where it will be passed on to E1 via B1.

ARGUMENTS FOR REBOOT PACKETS

   It is envisaged that introducing Reboot packets, will serve to
   enhance the bandwidth achievable within a Catenet system.  Problems
   of addressing 'dead' hosts will no longer exist in a correctly
   functioning configuration.  Translation Caches will have on hand the
   most recent addressing information available, which should also serve
   to enhance the performance of the routing strategy in operation.
   Multiple, redundant processing of packets destined for 'dead' hosts
   will be avoided.  Weighing this against the processing involved with
   a single multicast of Reboot packets, it is expected that the latter
   will be is the most economically viable in relation to the long-term
   traffic presented to the system.

CONCLUSION

   It appears that reboots are becoming increasingly common on internet
   networks.  Many sites use Personal Computers (PC) as terminals and
   the typical way to finish a session is to switch them off!  With the



Parr                                                           [Page 14]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


   increasing popularity of multitasking Operating Systems on these
   types of machines, problems are more likely to occur, particularly
   when the PCs are diskless, or participating in a distributed file
   system of some kind.  Given the importance of correct addressing in
   communications networks running Ethernet, it is anticipated the
   reboot mechanism described will serve to improve the correctness and
   validity of the protocol/network address mappings which may be stored
   in the translation caches.  To this degree, simulation is expected to
   show that the volume of invalid traffic will decrease, to the benefit
   of hosts, Bridges and servers alike.  Likewise, ratification of the
   routing policy is anticipated and since redundant/obsolete packets
   will be thwarted, the efficient utilization of available channel
   bandwidth across the catenet is also expected to improve.  Thus,
   effectively increasing Catenet throughput for 'valid' packets, and
   therefore enhancing the level of service provided to the end users.

   It is obvious that the proposed scheme implies the alteration of the
   packet processing code in Bridges/Gateways.  The point to remember is
   the increased favour with which larger, more complex Multi-LAN
   systems of Ethernets are being received.  The recent adaption of
   extra telephone cables to serve as the transmission media for the
   Ethernet can only result in installation costs being reduced, therein
   making the Ethernet more attractive within large corporate buildings,
   etc.  It is sensible to suggest that the probability of host address
   re-assignment shall increase in proportion to the number of physical
   systems attached, component failure rate (for whatever reason),
   relocation of resources, and the size and turnover of the workforce
   (i.e., people moving from one room to another).  Simulation
   experiments are currently being developed to analyse the resultant
   traffic patterns under this scheme, and it is hoped to highlight
   thresholds where adoption of the scheme becomes a necessity.

   In addition, the Author is currently extending the boundaries of this
   problem to encompass the reboot, or relocation of Bridges themselves.
   Involved with this are the phenomena of loop resolution, load sharing
   and duplicate packet suppression.  It is envisaged that a Self-
   Stabilizationg Bridge Protocol will result that will be more "light-
   weight" than those adhering to the Spanning Tree Algorithm.

   The Author would appreciate feedback/comments on this RFC.  My
   network address is: CBAD13%UCVAX.ULSTER.AC.UK@CUNYVM.CUNY.EDU.

ACKNOWLEDGEMENTS

   The Author acknowledges with gratitute the help and comments
   contributed by Mr. Piotr Bielkowitz (Supervisor) of the Computing
   Science Department, and the time devoted my Mr. Raymond Robinson for
   painstakingly preparing the first draft of this paper on 'Pagemaker'.



Parr                                                           [Page 15]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


   Thanks are due also to Dr. M. W. A. Smith of Information Systems for
   his assistance.  Finally, this work was supported under a grant from
   the Department of Education for Northern Ireland of which the Author
   is extremely grateful.

REFERENCES

   [1]  Croft, Bill, and John Gilmore, "Bootstrap Protocol", RFC-951,
        Stanford University, September 1985.

   [2]  Finlayson, Mann, Mogul, and Theimer, "A Reverse Address
        Resolution Protocol", RFC-903, Computer Science Dept, Stanford
        University, June 1984.

   [3]  Lorimer, Alan, and Jim Reid, "ARP Information Communique",
        Computer Science Dept, Strathclyde University, 1987.

   [4]  Mogul, Jeffrey, "Internet Subnets", RFC-917, Computer Science
        Dept, Stanford University, October 1984.

   [5]  Plummer, David, "An Ethernet Address Resolution Protocol", RFC-
        826, MIT, November 1982.

   [6]  Postel, Jon, "DARPA Internet Program Protocol Specification",
        RFC-791, USC/Information Sciences Institute, September 1981.

   [7]  Postel, Jon, "Multi-LAN Address Resolution", RFC-925,
        USC/Information Sciences Institute, October 1984.

   [8]  Postel, Jon, Carl Sunshine, and Danny Cohen, "The ARPA Internet
        Protocol", Computer Networks, no. 5, pp. 261-271, 1981.

   [9]  Postel, Jon, and Jeff Mogul, "Internet Standard Subnetting
        Procedure", RFC-950, USC/Information Sciences Institute and
        Stanford University, August 1985.

   [10] Reynolds, Joyce, and Jon Postel, "Assigned Numbers", RFC-1010,
        USC/Information Sciences Institute, May 1987.

   [11] "The Ethernet: a local area network, data link layer and
        physical layer specification", Version 1.0 DEC, Intel and Xerox
        Corporations, USA 30 September 1980).

   [12] Hughes, H.D., and L. Li, "Simulation model of an Ethernet",
        Computer Performance, Vol 3, no. 4, December 1982.

   [13] Parr, Gerald P., "Address Resolution For An Intelligent
        Filtering Bridge Running On A Subnetted Ethernet System", ACM



Parr                                                           [Page 16]

RFC 1029           Fault Tolerant ARP for Multi-LANs            May 1988


        SIGCOMM Computer Communication Review, (July/August 1987), vol.
        17, no. 3.

   [14] Smoot, Carl-Mitchell, and John S. Quarterman, "Using ARP to
        Implement Transparent Subnet Gateways", RFC-1027, Texas Internet
        Consulting, October 1987.













































Parr                                                           [Page 17]