Network Working Group                                      Jeffrey Mogul
Request for Comments: 917                    Computer Science Department
                                                     Stanford University
                                                            October 1984

                            INTERNET SUBNETS


Status Of This Memo

   This RFC suggests a proposed protocol for the ARPA-Internet
   community, and requests discussion and suggestions for improvements.
   Distribution of this memo is unlimited.

Overview

   We discuss the utility of "subnets" of Internet networks, which are
   logically visible sub-sections of a single Internet network.  For
   administrative or technical reasons, many organizations have chosen
   to divide one Internet network into several subnets, instead of
   acquiring a set of Internet network numbers.

   We propose procedures for the use of subnets, and discuss approaches
   to solving the problems that arise, particularly that of routing.

Acknowledgment

   This proposal is the result of discussion with several other people.
   J. Noel Chiappa, Chris Kent, and Tim Mann, in particular, provided
   important suggestions.

1. Introduction

   The original view of the Internet universe was a two-level hierarchy:
   the top level the catenet as a whole, and the level below it a
   collection of "Internet Networks", each with its own Network Number.
   (We do not mean that the Internet has a hierarchical topology, but
   that the interpretation of addresses is hierarchical.)

   While this view has proved simple and powerful, a number of
   organizations have found it inadequate and have added a third level
   to the interpretation of Internet addresses.  In this view, a given
   Internet Network might (or might not) be divided into a collection of
   subnets.

   The original, two-level, view carries a strong presumption that, to a
   host on an Internet network, that network may be viewed as a single
   edge; to put it another way, the network may be treated as a "black
   box" to which a set of hosts is connected.  This is true of the




Mogul                                                           [Page 1]


RFC 917                                                     October 1984
Internet Subnets


   ARPANET, because the IMPs mask the use of specific links in that
   network.  It is also true of most local area network (LAN)
   technologies, such as Ethernet or ring networks.

   However, this presumption fails in many practical cases, because in
   moderately large organizations (e.g., Universities or companies with
   more than one building) it is often necessary to use more than one
   LAN cable to cover a "local area".  For example, at this writing
   there are eighteen such cables in use at Stanford University, with
   more planned.

   There are several reasons why an organization might use more than one
   cable to cover a campus:

      - Different technologies: Especially in a research environment,
        there may be more than one kind of LAN in use; e.g., an
        organization may have some equipment that supports Ethernet, and
        some that supports a ring network.

      - Limits of technologies: Most LAN technologies impose limits,
        based electrical parameters, on the number of hosts connected,
        and on the total length of the cable.  It is easy to exceed
        these limits, especially those on cable length.

      - Network congestion: It is possible for a small subset of the
        hosts on a LAN to monopolize most of the bandwidth.  A common
        solution to this problem is to divide the hosts into cliques of
        high mutual communication, and put these cliques on separate
        cables.

      - Point-to-Point links: Sometimes a "local area", such as a
        university campus, is split into two locations too far apart to
        connect using the preferred LAN technology.  In this case,
        high-speed point-to-point links might connect several LANs.

   An organization that has been forced to use more than one LAN has
   three choices for assigning Internet addresses:

      1. Acquire a distinct Internet network number for each cable.

      2. Use a single network number for the entire organization, but
         assign host numbers without regard to which LAN a host is on.
         (We will call this choice "transparent subnets".)

      3. Use a single network number, and partition the host address
         space by assigning subnet numbers to the LANs. ("Explicit
         subnets".)


Mogul                                                           [Page 2]


RFC 917                                                     October 1984
Internet Subnets


   Each of these approaches has disadvantages.  The first, although not
   requiring any new or modified protocols, does result in an explosion
   in the size of Internet routing tables.  Information about the
   internal details of local connectivity is propagated everywhere,
   although it is of little or no use outside the local organization.
   Especially as some current gateway implementations do not have much
   space for routing tables, it would be nice to avoid this problem.

   The second approach requires some convention or protocol that makes
   the collection of LANs appear to be a single Internet network.  For
   example, this can be done on LANs where each Internet address is
   translated to a hardware address using an Address Resolution Protocol
   (ARP), by having the bridges between the LANs intercept ARP requests
   for non-local targets.  However, it is not possible to do this for
   all LAN technologies, especially those where ARP protocols are not
   currently used, or if the LAN does not support broadcasts.  A more
   fundamental problem is that bridges must discover which LAN a host is
   on, perhaps by using a broadcast algorithm.  As the number of LANs
   grows, the cost of broadcasting grows as well; also, the size of
   translation caches required in the bridges grows with the total
   number of hosts in the network.

   The third approach addresses the key problem: existing standards
   assume that all hosts on an Internet local network are on a single
   cable.  The solution is to explicitly support subnets.  This does
   have a disadvantage, in that it is a modification of the Internet
   Protocol, and thus requires changes to IP implementations already in
   use (if these implementations are to be used on a subnetted network.)
   However, we believe that these changes are relatively minor, and once
   made, yield a simple and efficient solution to the problem.  Also,
   the approach we take in this document is to avoid any changes that
   would be incompatible with existing hosts on non-subnetted networks.

   Further, when appropriate design choices are made, it is possible for
   hosts which believe they are on a non-subnetted network to be used on
   a subnetted one, as will be explained later.  This is useful when it
   is not possible to modify some of the hosts to support subnets
   explicitly, or when a gradual transition is preferred.  Because of
   this, there seems little reason to use the second approach listed
   above.

   The rest of this document describes approaches to subnets of Internet
   Networks.






Mogul                                                           [Page 3]


RFC 917                                                     October 1984
Internet Subnets


   1.1. Terminology

      To avoid either ambiguity or prolixity, we will define a few
      terms, which will be used in the following sections:

      Catenet

         The collection of connected Internet Networks

      Network

         A single Internet network (that may or may not be divided into
         subnets.)

      Subnet

         A subnet of an Internet network.

      Network Number

         As in [8].

      Local Address

         The bits in an Internet address not used for the network
         number; also known as "rest field".

      Subnet Number

         A number identifying a subnet within a network.

      Subnet Field

         The bit field in an Internet address used for the subnet
         number.

      Host Field

         The bit field in an Internet address used for denoting a
         specific host.

      Gateway

         A node connected to two or more administratively distinct
         networks and/or subnets, to which hosts send datagrams to be
         forwarded.



Mogul                                                           [Page 4]


RFC 917                                                     October 1984
Internet Subnets


      Bridge

         A node connected to two or more administratively
         indistinguishable but physically distinct subnets, that
         automatically forwards datagrams when necessary, but whose
         existence is not know to other hosts.  Also called a "software
         repeater".

2. Standards for Subnet Addressing

   Following the division presented in [2], we observe that subnets are
   fundamentally an issue of addressing.  In this section, we first
   describe a proposal for interpretation of Internet Addressing to
   support subnets.  We then discuss the interaction between this
   address format and broadcasting; finally, we present a protocol for
   discovering what address interpretation is in use on a given network.

   2.1. Interpretation of Internet Addresses

      Suppose that an organization has been assigned an Internet network
      number, has further divided that network into a set of subnets,
      and wants to assign host addresses: how should this be done?
      Since there are minimal restrictions on the assignment of the
      "local address" part of the Internet address, several approaches
      have been proposed for representing the subnet number:

         1. Variable-width field: Any number of the bits of the local
            address part are used for the subnet number; the size of
            this field, although constant for a given network, varies
            from network to network.  If the field width is zero, then
            subnets are not in use.

         2. Fixed-width field: A specific number of bits (e.g., eight)
            is used for the subnet number, if subnets are in use.

         3. Self-encoding variable-width field: Just as the width (i.e.,
            class) of the network number field is encoded by its
            high-order bits, the width of the subnet field is similarly
            encoded.

         4. Self-encoding fixed-width field: A specific number of bits
            is is used for the subnet number.  Subnets are in use if the
            high-order bit of this field is one; otherwise, the entire
            local address part is used for host number.

      Since there seems to be no advantage in doing otherwise, all these
      schemes place the subnet field as the most significant field in


Mogul                                                           [Page 5]


RFC 917                                                     October 1984
Internet Subnets


      the local address part.  Also, since the local address part of a
      Class C address is so small, there is little reason to support
      subnets of other than Class A and Class B networks.

      What criteria can we use to choose one of these four schemes?
      First, do we want to use a self-encoding scheme; that is, should
      it be possible to tell from examining an Internet address if it
      refers to a subnetted network, without reference to any other
      information?

      One advantage to self-encoding is that it allows one to determine
      if a non-local network has been divided into subnets.  It is not
      clear that this would be of any use.  The principle advantage,
      however, is that no additional information is needed for an
      implementation to determine if two addresses are on the same
      subnet.  However, this can also be viewed as a disadvantage: it
      may cause problems for non-subnetted networks which have existing
      host numbers that use arbitrary bits in the local address part
      <1>.  In other words, it is useful to be able control whether a
      network is subnetted independently from the assignment of host
      addresses.  Another disadvantage of any self-encoding scheme is
      that it reduces the local address space by at least a factor of
      two.

      If a self-encoding scheme is not used, it is clear that a
      variable-width subnet field is appropriate.  Since there must in
      any case be some per-network "flag" to indicate if subnets are in
      use, the additional cost of using an integer (the subnet field
      width) instead of a boolean is negligible.  The advantage of using
      a variable-width subnet field is that it allows each organization
      to choose the best way to allocate relatively scarce bits of local
      address to subnet and host numbers.

      Our proposal, therefore, is that the Internet address be
      interpreted as:

         <network-number><subnet-number><host-number>

      where the <network-number> field is as in [8], the <host-number>
      field is at least one bit wide, and the width of the
      <subnet-number> field is constant for a given network. No further
      structure is required for the <subnet-number> or <host-number>
      fields.  If the width of the <subnet-number> field is zero, then
      the network is not subnetted (i.e., the interpretation of [8] is
      used.)




Mogul                                                           [Page 6]


RFC 917                                                     October 1984
Internet Subnets


      For example, on a Class A network with an eight bit wide subnet
      field, an address is broken down like this:

                           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
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |0|    NETWORK    |     SUBNET    |         Host number         |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

      We expect that, for reasons of simplicity and efficient
      implementation, that most organizations will choose a subnet field
      width that is a multiple of eight bits.  However, an
      implementation must be prepared to handle other possible widths.

      We reject the use of "recursive subnets", the division of the host
      field into "sub-subnet" and host parts, because:

         - There is no obvious need for a four-level hierarchy.

         - The number of bits available in an IP address is not large
           enough to make this useful in general.

         - The extra mechanism required is complex.

   2.2. Changes to Host Software to Support Subnets

      In most implementations of IP, there is  code in the module that
      handles outgoing packet that does something like:

         IF ip_net_number(packet.ip_dest) = ip_net_number(my_ip_addr)
             THEN
                 send_packet_locally(packet, packet.ip_dest)
             ELSE
                 send_packet_locally(packet,
                    gateway_to(ip_net_number(packet.ip_dest)))

      (If the code supports multiple connected networks, it will be more
      complicated, but this is irrelevant to the current discussion.)

      To support subnets, it is necessary to store one more 32-bit
      quantity, called my_ip_mask.  This is a bit-mask with bits set in
      the fields corresponding to the IP network number, and additional
      bits set corresponding to the subnet number field.  For example,
      on a Class A network using an eight-bit wide subnet field, the
      mask would be 255.255.0.0.

      The code then becomes:


Mogul                                                           [Page 7]


RFC 917                                                     October 1984
Internet Subnets


         IF bitwise_and(packet.ip_dest, my_ip_mask)
                          = bitwise_and(my_ip_addr, my_ip_mask)
             THEN
                 send_packet_locally(packet, packet.ip_dest)
             ELSE
                 send_packet_locally(packet,
                    gateway_to(bitwise_and(packet.ip_dest, my_ip_mask)))

      Of course, part of the expression in the conditionally can be
      pre-computed.

      It may or may not be necessary to modify the "gateway_to"
      function, so that it performs comparisons in the same way.

      To support multiply-connected hosts, the code can be changed to
      keep  the "my_ip_addr" and "my_ip_mask" quantities on a
      per-interface basis; the expression in the conditional must then
      be evaluated for each interface.

   2.3. Subnets and Broadcasting

      In the absence of subnets, there are only two kinds of broadcast
      possible within the Internet Protocol <2>: broadcast to all hosts
      on a specific network, or broadcast to all hosts on "this
      network"; the latter is useful when a host does not know what
      network it is on.

      When subnets are used, the situation becomes slightly more
      complicated.  First, the possibility now exists of broadcasting to
      a specific subnet.  Second, broadcasting to all the hosts on a
      subnetted network requires additional mechanism; in [6] the use of
      "Reverse Path Forwarding" [3] is proposed.  Finally, the
      interpretation of a broadcast to "this network" is that it should
      not be forwarded outside of the original subnet.

      Implementations must therefore recognize three kinds of broadcast
      addresses, in addition to their own host addresses:

      This physical network

         A destination address of all ones (255.255.255.255) causes the
         a datagram to be sent as a broadcast on the local physical
         network; it must not be forwarded by any gateway.






Mogul                                                           [Page 8]


RFC 917                                                     October 1984
Internet Subnets


      Specific network

         The destination address contains a valid network number; the
         local address part is all ones (e.g., 36.255.255.255).

      Specific subnet

         The destination address contains a valid network number and a
         valid subnet number; the host field is all ones (e.g.,
         36.40.255.255).

      For further discussion of Internet broadcasting, see [6].

      One factor that may aid in deciding whether to use subnets is that
      it is possible to broadcast to all hosts of a subnetted network
      with a single operation at the originating host.  It is not
      possible to broadcast, in one step, to the same set of hosts if
      they are on distinct networks.

   2.4. Determining the Width of the Subnet Field

      How can a host (or gateway) determine what subnet field width is
      in use on a network to which it is connected?  The problem is
      analogous to several other "bootstrapping" problems for Internet
      hosts: how a host determines its own address, and how it locates a
      gateway on its local network.  In all three cases, there are two
      basic solutions: "hardwired" information, and broadcast-based
      protocols.

      "Hardwired" information is that available to a host in isolation
      from a network.  It may be compiled-in, or (preferably) stored in
      a disk file.  However, for the increasingly common case of a
      diskless workstation that is bootloaded over a LAN, neither
      hard-wired solution is satisfactory.  Instead, since most LAN
      technology supports broadcasting, a better method is for the
      newly-booted host to broadcast a request for the necessary
      information.  For example, for the purpose of determining its
      Internet address, a host may use the "Reverse Address Resolution
      Protocol" [4].

      We propose to extend the ICMP protocol [9] by adding a new pair of
      ICMP message types, "Address Format Request" and "Address Format
      Reply", analogous to the "Information Request" and "Information
      Reply" ICMP messages.  These are described in detail in
      Appendix I.

      The intended use of these new ICMPs is that a host, when booting,


Mogul                                                           [Page 9]


RFC 917                                                     October 1984
Internet Subnets


      broadcast an "Address Format Request" message <3>.  A gateway (or
      a host acting in lieu of a gateway) that receives this message
      responds with an "Address Format Reply".  If there is no
      indication in the request which host sent it (i.e., the IP Source
      Address is zero), the reply is broadcast as well.  The requesting
      host will hear the response, and from it determine the width of
      the subnet field.

      Since there is only one possible value that can be sent in an
      "Address Format Reply" on any given LAN, there is no need for the
      requesting host to match the responses it hears against the
      request it sent; similarly, there is no problem if more than one
      gateway responds.  We assume that hosts reboot infrequently, so
      the broadcast load on a network from use of this protocol should
      be small.

      If a host is connected to more than one LAN, it must use this
      protocol on each, unless it can determine (from a response on one
      of the LANs) that several of the LANs are part of the same
      network, and thus must have the same subnet field width.

      One potential problem is what a host should do if it receives no
      response to its "Address Format Request", even after a reasonable
      number of tries.  Three interpretations can be placed on the
      situation:

         1. The local net exists in (permanent) isolation from all other
            nets.

         2. Subnets are not in use, and no host supports this ICMP
            request.

         3. All gateways on the local net are (temporarily) down.

      The first and second situations imply that the subnet field width
      is zero.  In the third situation, there is no way to determine
      what the proper value is; the safest choice is thus zero.
      Although this might later turn out to be wrong, it will not
      prevent transmissions that would otherwise succeed.  It is
      possible for a host to recover from a wrong choice: when a gateway
      comes up, it should broadcast an "Address Format Reply"; when a
      host receives such a message that disagrees with its guess, it
      should adjust its data structures to conform to the received
      value.  No host or gateway should send an "Address Format Reply"
      based on a "guessed" value.




Mogul                                                          [Page 10]


RFC 917                                                     October 1984
Internet Subnets


      Finally, note that no host is required to use this ICMP protocol
      to discover the subnet field width; it is perfectly reasonable for
      a host with non-volatile storage to use stored information.

3. Subnet Routing Methods

   One problem that faces all Internet hosts is how to determine a route
   to another host.  In the presence of subnets, this problem is only
   slightly modified.

   The use of subnets means that there are two levels to the routing
   process, instead of one.  If the destination host is on the same
   network as the source host, the routing decision involves only the
   subnet gateways between the hosts.  If the destination is on a
   different network, then the routing decision requires the choice both
   of a gateway out of the source host's network, and of a route within
   the network to that gateway.

   Fortunately, many hosts can ignore this distinction (and, in fact,
   ignore all routing choices) by using a "default" gateway as the
   initial route to all destinations, and relying on ICMP Host Redirect
   messages to define more appropriate routes.  However, this is not an
   efficient method for a gateway or for a multi-homed host, since a
   redirect may not make up for a poor initial choice of route.  Such
   hosts should use a routing information exchange protocol, but that is
   beyond the scope of this document; in any case, the problem arises
   even when subnets are not used.

   The problem for a singly-connected host is thus to find at least one
   neighbor gateway.  Again, there are basic two solutions to this: use
   hard-wired information, or use broadcasts.  We believe that the
   neighbor-gateway acquisition problem is the same with or without
   subnets, and thus the choice of solution is not affected by the use
   of subnets.

   However, one problem remains: a source host must determine if
   datagram to a given destination address must be sent via a gateway,
   or sent directly to the destination host.  In other words, is the
   destination host on the same physical network as the source?  This
   particular phase of the routing process is the only one that requires
   an implementation to be explicitly aware of subnets; in fact, if
   broadcasts are not used, it is the only place where an Internet
   implementation must be modified to support subnets.

   Because of this, it is possible to use some existing implementations
   without modification in the presence of subnets <4>.  For this to
   work, such implementations must:


Mogul                                                          [Page 11]


RFC 917                                                     October 1984
Internet Subnets


      - Be used only on singly-homed hosts, and not as a gateway.

      - Be used on a broadcast LAN.

      - Use an Address Resolution Protocol (ARP), such [7].

      - Not be required to maintain connections in the case of gateway
        crashes.

   In this case, one can modify the ARP server module in a subnet
   gateway so that when it receives an ARP request, it checks the target
   Internet address to see if it is along the best route to the target.
   If it is, it sends to the requesting host an ARP response indicating
   its own hardware address.  The requesting host thus believes that it
   knows the hardware address of the destination host, and sends packets
   to that address.  In fact, the packets are received by the gateway,
   and forwarded to the destination host by the usual means.

   This method requires some blurring of the layers in the gateways,
   since the ARP server and the Internet routing table would normally
   not have any contact.  In this respect, it is somewhat
   unsatisfactory.  Still, it is fairly easy to implement, and does not
   have significant performance costs.  One problem is that if the
   original gateway crashes, there is no way for the source host to
   choose an alternate route even if one exists; thus, a connection that
   might otherwise have been maintained will be broken.

   One should not confuse this method of "ARP-based subnetting" with the
   superficially similar use of ARP-based bridges.  ARP-based subnetting
   is based on the ability of a gateway to examine an IP address and
   deduce a route to the destination, based on explicit subnet topology.
   In other words, a small part of the routing decision has been moved
   from the source host into the gateway.  An ARP-based bridge, in
   contrast, must somehow locate each host without any assistance from a
   mapping between host address and topology.  Systems built out of
   ARP-based bridges should not be referred to as "subnetted".

   N.B.: the use of ARP-based subnetting is complicated by the use of
   broadcasts.  An ARP server [7] should never respond to a request
   whose target is a broadcast address.  Such a request can only come
   from a host that does not recognize the broadcast address as such,
   and so honoring it would almost certainly lead to a forwarding loop.
   If there are N such hosts on the physical network that do not
   recognize this address as a broadcast, then a packet sent with a
   Time-To-Live of T could potentially give rise to T**N spurious
   re-broadcasts.



Mogul                                                          [Page 12]


RFC 917                                                     October 1984
Internet Subnets


4. Case Studies

   In this section, we briefly sketch how subnets have been used by
   several organizations.

   4.1. Stanford University

      At Stanford, subnets were introduced initially for historical
      reasons.  Stanford had been using the Pup protocols [1] on a
      collection of several Experimental Ethernets [5] since 1979,
      several years before Internet protocols came into use.  There were
      a number of Pup gateways in service, and all hosts and gateways
      acquired and exchanged routing table information using a simple
      broadcast protocol.

      When the Internet Protocol was introduced, the decision was made
      to use an eight-bit wide subnet number; Internet subnet numbers
      were chosen to match the Pup network number of a given Ethernet,
      and the Pup host numbers (also eight bits) were used as the host
      field of the Internet address.

      The Pup-only gateways were then modified to forward Internet
      datagrams according to their Pup routing tables; they otherwise
      had no understanding of Internet packets and in fact did not
      adjust the Time-to-live field in the Internet header.  This seems
      to be acceptable, since bugs that caused forwarding loops have not
      appeared.  The Internet hosts that are multi-homed and thus can
      serve as gateways do adjust the Time-to-live field; since all of
      the currently also serve as Pup gateways, no additional routing
      information exchange protocol was needed.

      Internet host implementations were modified to understand subnets
      (in several different ways, but with identical effects).  Since
      all already had Pup implementations, the Internet routing tables
      were maintained by the same process that maintained the Pup
      routing tables, simply translating the Pup network numbers into
      Internet subnet numbers.

      When 10Mbit Ethernets were added, the gateways were modified to
      use the ARP-based scheme described in an earlier section; this
      allowed unmodified hosts to be used on the 10Mbit Ethernets.

      IP subnets have been in use since early 1982; currently, there are
      about 330 hosts, 18 subnets, and a similar number of subnet
      gateways in service.  Once the Pup-only gateways are converted to
      be true Internet gateways, an Internet-based routing exchange
      protocol will be introduced, and Pup will be phased out.


Mogul                                                          [Page 13]


RFC 917                                                     October 1984
Internet Subnets


   4.2. MIT

      MIT was the first IP site to accumulate a large collection of
      local network links.  Since this happened before network numbers
      were divided into classes, to have assigned each link at MIT its
      own IP network number would have used up a good portion of the
      available address space.  MIT decided to use one IP network
      number, and to manage the 24-bit "rest" field itself, by dividing
      it into three 8-bit fields; "subnet", "reserved, must be zero",
      and "host".   Since the CHAOS protocol already in use at MIT used
      an 8-bit subnet number field, it was possible to assign each link
      the same subnet number in both protocols.  The IP host field was
      set to 8 bits since most available local net hardware at that
      point used 8 bit addresses, as did the CHAOS protocol; it was felt
      that reserving some bits for the future was wise.

      The initial plan was to use a dynamic routing protocol between the
      IP subnet gateways; several such protocols have been mooted but
      nobody has bothered to implement one; static routing tables are
      still used.  It is likely that this change will finally be made
      soon.

      To solve the problem that imported IP software always needed
      modification to work in the subnetted environment, MIT searched
      for a model of operation that led to the least change in host IP
      software.  This led to a model where IP gateways send ICMP Host
      Redirects rather than Network Redirects.  All internal MIT IP
      gateways now do so.  With hosts that can maintain IP routing
      tables for non-local communication on a per host basis, this hides
      most of the subnet structure.  The "minimum adjustment" for host
      software to work correctly in both subnetted and non-subnetted
      environments is the bit-mask algorithm mentioned earlier.

      MIT has no immediate plans to move toward a single "approved"
      protocol; this is due partly to the degree of local autonomy and
      the amount of installed software, and partly to the lack of a
      single prominent industry standard.  Rather, the approach taken
      has been to provide a single set of physical links and packet
      switches, and to layer several "virtual" protocol nets atop the
      single set of links.  MIT has had some bad experiences with trying
      to exchange routing information between protocols and wrap one
      protocol in another; the general approach is to keep the protocols
      strictly separated except for sharing the basic hardware.  Using
      ARP to hide the subnet structure is not much in favor; it is felt
      that this overloads the address resolution operation.  In a
      complicated system (i.e. one with loops, and variant link speeds),



Mogul                                                          [Page 14]


RFC 917                                                     October 1984
Internet Subnets


      a more sophisticated information interchange will be needed
      between gateways; making this an explicit mechanism (but one
      insulated from the hosts) was felt to be best.

   4.3. Carnegie-Mellon University

      CMU uses a Class B network currently divided into 11 physical
      subnets (two 3Mbit Experimental Ethernets, seven 10Mbit Ethernets,
      and two ProNet rings.) Although host numbers are assigned so that
      all addresses with a given third octet will be on the same subnet
      (but not necessarily vice versa), this is essentially an
      administrative convenience.  No software currently knows the
      specifics of this allocation mechanism or depends on it to route
      between cables.

      Instead, an ARP-based bridge scheme is used.  When a host
      broadcasts an ARP request, all bridges which receive it cache the
      original protocol address mapping and then forward the request
      (after the appropriate adjustments) as an ARP broadcast request
      onto each of their other connected cables.  When a bridge receives
      a non-broadcast ARP reply with a target protocol address not its
      own, it consults its ARP cache to determine the cable onto which
      the reply should be forwarded.  The bridges thus attempt to
      transparently extend the ARP protocol into a heterogenous
      multi-cable environment.  They are therefore required to turn ARP
      broadcasts on a single cable into ARP broadcasts on all other
      connected cables even when they "know better".  This algorithm
      works only in the absence of cycles in the network connectivity
      graph (which is currently the case).  Work is underway to replace
      this simple-minded algorithm with a protocol implemented among the
      bridges, in support of redundant paths and to reduce the
      collective broadcast load.  The intent is to retain the ARP base
      and host transparency, if possible.

      Implementations supporting the 3Mbit Ethernet and 10Mb proNET ring
      at CMU use RFC-826 ARP (instead of some wired-in mapping such as
      simply using the 8-bit hardware address as the the fourth octet of
      the IP address).

      Since there are currently no redundant paths between cables, the
      issue of maintaining connections across bridge crashes is moot.
      With about 150 IP-capable hosts on the net, the bridge caches are
      still of reasonable size, and little bandwidth is devoted to ARP
      broadcast forwarding.

      CMU's network is likely to grow from its relatively small,
      singly-connected configuration centered within their CS/RI


Mogul                                                          [Page 15]


RFC 917                                                     October 1984
Internet Subnets


      facility to a campus-wide intra-departmental configuration with
      5000-10000 hosts and redundant connections between cables.  It is
      possible that the ARP-based bridge scheme will not scale to this
      size, and a system of explicit subnets may be required.  The
      medium-term goal, however, is an environment into which unmodified
      extant (especially 10Mb ethernet based) IP implementations can be
      imported; the intent is to stay with a host-transparent (thus
      ARP-based) routing mechanism as long as possible.  CMU is
      concerned that even if subnets become part of the IP standard they
      will not be widely implemented; this is the major obstacle to
      their use at CMU.






































Mogul                                                          [Page 16]


RFC 917                                                     October 1984
Internet Subnets


I. Address Format ICMP

   Address Format Request or Address Format Reply

       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
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |     Type      |      Code     |          Checksum             |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |           Identifier          |       Sequence Number         |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

      IP Fields:

         Addresses

            The address of the source in an address format request
            message will be the destination of the address format reply
            message.  To form an address format reply message, the
            source address of the request becomes the destination
            address of the reply, the source address of the reply is set
            to the replier's address, the type code changed to A2, the
            subnet field width inserted into the Code field, and the
            checksum recomputed.  However, if the source address in the
            request message is zero, then the destination address for
            the reply message should denote a broadcast.

      ICMP Fields:

         Type

            A1 for address format request message

            A2 for address format reply message

         Code

            0 for address format request message

            Width of subnet field, in bits, for address format reply
            message

         Checksum

            The checksum is the 16-bit one's complement of the one's




Mogul                                                          [Page 17]


RFC 917                                                     October 1984
Internet Subnets


            complement sum of the ICMP message starting with the ICMP
            Type.  For computing the checksum, the checksum field should
            be zero.  This checksum may be replaced in the future.

         Identifier

            An identifier to aid in matching request and replies, may be
            zero.

         Sequence Number

            A sequence number to aid in matching request and replies,
            may be zero.

      Description

         A gateway receiving an address format request should return it
         with the Code field set to the number of bits of Subnet number
         in IP addresses for the network to which the datagram was
         addressed.  If the request was broadcast, the destination
         network is "this network".  The Subnet field width may be from
         0 to (31 - N), where N is the width in bits of the IP net
         number field (i.e., 8, 16, or 24).

         If the requesting host does not know its own IP address, it may
         leave the source field zero; the reply should then be
         broadcast.  Since there is only one possible address format for
         a network, there is no need to match requests with replies.
         However, this approach should be avoided if at all possible,
         since it increases the superfluous broadcast load on the
         network.

            Type A1 may be received from a gateway or a host.

            Type A2 may be received from a gateway, or a host acting in
            lieu of a gateway.













Mogul                                                          [Page 18]


RFC 917                                                     October 1984
Internet Subnets


II. Examples

   For these examples, we assume that the requesting host has  address
   36.40.0.123, that there is a gateway at 36.40.0.62, and that on
   network 36.0.0.0, an 8-bit wide subnet field is in use.

   First, suppose that broadcasting is allowed, and that 36.40.0.123
   knows  its own address.  It sends the following datagram:

      Source address:          36.40.0.123
      Destination address:     36.255.255.255
      Protocol:                ICMP = 1
      Type:                    Address Format Request = A1
      Code:                    0

   36.40.0.62 will hear the datagram, and should respond with this
   datagram:

      Source address:          36.40.0.62
      Destination address:     36.40.0.123
      Protocol:                ICMP = 1
      Type:                    Address Format Reply = A2
      Code:                    8

   For the following examples, assume that address 255.255.255.255
   denotes "broadcast to this physical network", as described in [6].

   The previous example is inefficient, because it potentially
   broadcasts  the request on many subnets.  The most efficient method,
   and the one we recommend, is for a host to first discover its own
   address (perhaps  using the "Reverse ARP" protocol described in [4]),
   and then to send  the ICMP request to 255.255.255.255:

      Source address:          36.40.0.123
      Destination address:     255.255.255.255
      Protocol:                ICMP = 1
      Type:                    Address Format Request = A1
      Code:                    0

   The gateway can then respond directly to the requesting host.

   Suppose that 36.40.0.123 is a diskless workstation, and does not know
   even its own host number.  It could send the following datagram:






Mogul                                                          [Page 19]


RFC 917                                                     October 1984
Internet Subnets


      Source address:          0.0.0.0
      Destination address:     255.255.255.255
      Protocol:                ICMP = 1
      Type:                    Address Format Request = A1
      Code:                    0

   36.40.0.62 will hear the datagram, and should respond with this
   datagram:

      Source address:          36.40.0.62
      Destination address:     36.40.255.255
      Protocol:                ICMP = 1
      Type:                    Address Format Reply = A2
      Code:                    8

   Note that the gateway uses the narrowest possible broadcast to reply
   (i.e., sending the reply to 36.255.255.255 would mean that it is
   transmitted on many subnets, not just the one on which it is needed.)
   Even so, the overuse of broadcasts presents an unnecessary load to
   all hosts on the subnet, and so we recommend that use of the
   "anonymous" (0.0.0.0) source address be kept to a minimum.

   If  broadcasting is not allowed, we assume that hosts have wired-in
   information about neighbor gateways; thus, 36.40.0.123 might send
   this datagram:

      Source address:          36.40.0.123
      Destination address:     36.40.0.62
      Protocol:                ICMP = 1
      Type:                    Address Format Request = A1
      Code:                    0

   36.40.0.62 should respond exactly as in the previous case.
















Mogul                                                          [Page 20]


RFC 917                                                     October 1984
Internet Subnets


Notes

   <1>  For example, some host have addresses assigned by concatenating
        their Class A network number with the low-order 24 bits of a
        48-bit Ethernet hardware address.

   <2>  Our discussion of Internet broadcasting is based on [6].

   <3>  If broadcasting is not supported, them presumably a host "knows"
        the address of a neighbor gateway, and should send the ICMP to
        that gateway.

   <4>  This is what was referred to earlier as the coexistence of
        transparent and explicit subnets on a single network.



































Mogul                                                          [Page 21]


RFC 917                                                     October 1984
Internet Subnets


References

   1.  D.R. Boggs, J.F. Shoch, E.A. Taft, and R.M. Metcalfe. "Pup: An
       Internetwork Architecture."  IEEE Transactions on Communications
       COM-28, 4, pp612-624, April 1980.

   2.  David D. Clark.  Names, Addresses, Ports, and Routes.  RFC-814,
       MIT-LCS, July 1982.

   3.  Yogan K. Dalal and Robert M. Metcalfe. "Reverse Path Forwarding
       of Broadcast Packets."  Comm. ACM 21, 12, pp1040-1048, December
       1978.

   4.  Ross Finlayson, Timothy Mann, Jeffrey Mogul, Marvin Theimer. A
       Reverse Address Resolution Protocol. RFC-903, Stanford
       University, June 1984.

   5.  R.M. Metcalfe and D.R. Boggs. "Ethernet: Distributed Packet
       Switching for Local Computer Networks."  Comm. ACM 19, 7,
       pp395-404, July 1976.  Also CSL-75-7, Xerox Palo Alto Research
       Center, reprinted in CSL-80-2.

   6.  Jeffrey Mogul. Broadcasting Internet Datagrams. RFC-919, Stanford
       University, October 1984.

   7.  David Plummer. An Ethernet Address Resolution Protocol. RFC-826,
       Symbolics, September 1982.

   8.  Jon Postel. Internet Protocol. RFC-791, USC-ISI, September 1981.

   9.  Jon Postel. Internet Control Message Protocol. RFC-792, USC-ISI,
       September 1981.

















Mogul                                                          [Page 22]