Network Working Group                                      S. E. Deering
Request for Comments: 966                                 D. R. Cheriton
                                                     Stanford University
                                                           December 1985

                              Host Groups:
             A Multicast Extension to the Internet Protocol


1. Status of this Memo

   This RFC defines a model of service for Internet multicasting and
   proposes an extension to the Internet Protocol (IP) to support such a
   multicast service.  Discussion and suggestions for improvements are
   requested.  Distribution of this memo is unlimited.

2. Acknowledgements

   This memo was adapted from a paper [7] presented at the Ninth Data
   Communications Symposium.  This work was sponsored in part by the
   Defense Advanced Research Projects Agency under contract N00039-83-
   K-0431 and National Science Foundation Grant DCR-83-52048.

   The Internet task force on end-to-end protocols, headed by Bob
   Braden, has provided valuable input in the development of the host
   group model.

3. Introduction

   In this paper, we describe a model of multicast service we call host
   groups and propose this model as a way to support multicast in the
   DARPA Internet environment [14].  We argue that it is feasible to
   implement this facility as an extension of the existing "unicast" IP
   datagram model and mechanism.

   Multicast is the transmission of a datagram packet to a set of zero
   or more destination hosts in a network or internetwork, with a single
   address specifying the set of destination hosts.  For example, hosts
   A, B, C and D may be associated with multicast address X. On
   transmission, a packet with destination address X is delivered with
   datagram reliability to hosts A, B, C and D.

   Multicast has two primary uses, namely distributed binding and
   multi-destination delivery.  As a binding mechanism, multicast is a
   robust and often more efficient alternative to the use of name
   servers for finding a particular object or service when a particular
   host address is not known.  For example, in a distributed file
   system, all the file servers may be associated with one well-known
   multicast address.  To bind a file name to a particular server, a
   client sends a query packet containing the file name to the file
   server multicast address, for delivery to all the file servers.  The

   server that recognizes the file name then responds to the client,
   allowing subsequent interaction directly with that server host.  Even
   when name servers are employed, multicast can be used as the first
   step in the binding process, that is, finding a name server.

   Multi-destination delivery is useful to several applications,
   including:

      - distributed, replicated databases [6,9].

      - conferencing [11].

      - distributed parallel computation, including distributed
        gaming [2].

   Ideally, multicast transmission to a set of hosts is not more
   complicated or expensive for the sender than transmission to a single
   host.  Similarly, multicast transmission should not be more expensive
   for the networks and gateways than traversing the shortest path tree
   that connects the sending host to the hosts identified by the
   multicast address.

   Multicast, transmission to a set of hosts, is properly distinguished
   from broadcast, transmission to all hosts on a network or
   internetwork. Broadcast is not a generally useful facility since
   there are few reasons for communicating with all hosts.

   A variety of local network applications and systems make use of
   multicast.  For instance, the V distributed system [8] uses
   network-level multicast for implementing efficient operations on
   groups of processes spanning multiple machines.  Similar use is being
   made for replicated databases [6] and other distributed applications
   [4]. Providing multicast in the Internet environment would allow
   porting such local network distributed applications to the Internet,
   as well as making some existing Internet applications more robust and
   portable (by, for example, removing "wired-in" lists of addresses,
   such as gateway addresses).

   At present, an Internet application logically requiring multicast
   must send individually addressed packets to each recipient.  There
   are two problems with this approach.  Firstly, requiring the sending
   host to know the specific addresses of all the recipients defeats its
   use as a binding mechanism.  For example, a diskless workstation
   needs on boot to determine the network address of a disk server and
   it is undesirable to "wire in" specific network addresses.  With a
   multicast facility, the multicast address of the boot servers (or

   name servers that hold the addresses of the boot servers) can be
   well-known, allowing the workstation to transmit its initial queries
   to this address.

   Secondly, transmitting multiple copies of the same packet makes
   inefficient use of network bandwidth, gateway resources and sender
   resources.  For instance, the same packet may repeatedly traverse the
   same network links and pass through the same gateways.  Furthermore,
   the local network level cannot recognize multi-destination delivery
   to take advantage of multicast facilities that the underlying network
   technologies may provide.  For example, local-area bus, ring, or
   radio networks, as well as satellite-based wide-area networks, can
   provide efficient multicast delivery directly.  Besides using
   excessive communication resources, the use of multiple transmissions
   to effect multicast severely limits the amount of parallelism in
   transmission and processing that can be achieved compared to an
   integrated multicast facility.

   The next section describes the host group model of multicast service.
   Section 5 describes the extensions to IP to support the host group
   model.  Section 6 discusses the implementation of multicast within
   the networks and gateways making up the Internet.  Section 7 relates
   this model to other proposals.  Finally, we conclude with remarks on
   our experimental prototype implementation of host groups and comments
   on future directions for investigation.

4. The Host Group Model

   The Internet architecture defines a name space of individual host
   addresses.  The host group model extends that name space to include
   addresses of host groups.  A host group is a set of zero or more
   Internet hosts <1>.   When an IP packet is sent with a host group
   address as its destination, it is delivered with "best effort"
   datagram reliability to all members of that host group.

   The sender need not be a member of the destination group.  We refer
   to such a group as open, in contrast to a closed group where only
   members are allowed to send to the group.  We chose to provide open
   groups because they are more flexible and more consistent as an
   extension of conventional unicast models (even though they may harder
   to implement).

   Dynamic management of group membership provides flexible binding of
   Internet addresses to hosts.  Hosts may join and leave groups over
   time. A host may also belong to more than one group at a time.
   Finally, a host may belong to no groups at times, during which that
   host is unreachable within the Internet architecture.  In fact, a

   host need not have an individual Internet address at all.  Some hosts
   may only be associated with multi-host group addresses.  For
   instance, there may be no reason to contact an individual time server
   in the Internet, so time servers would not require individual
   addresses.

   Internet addresses are dynamically allocated for transient groups,
   groups that often last only as long as the execution of a single
   distributed program.  In addition, a range of host group identifiers
   is reserved for identifying permanent groups.  One use of permanent
   host groups identifiers is for host groups with standard logical
   meanings such as "name server group", "boot server group", "Internet
   monitor group", etc.

   In the current Internet architecture, addresses are bound to single
   hosts.  The host group model generalizes the binding of Internet
   addresses to hosts by allowing one address to bind to multiple hosts
   on multiple networks, more than one address to be bound (in part) to
   one host, and the binding of an address to host to be dynamic, i.e.
   possible to be modified under application control.  Within this more
   general model, the current architecture is supported as a special
   case, retaining its current semantics and implementation.

   The following subsections provide further details of the model.

   4.1. Host Group Management

      Dynamic binding of Internet addresses to hosts is managed by the
      following three operations which are made available to clients of
      the Internet Protocol <2>:

         CreateGroup ( type ) --> outcome, group-address, access-key

      requests the creation of a new transient host group with the
      invoking host as its only member.  The type argument specifies
      whether the group is restricted or unrestricted.  A restricted
      group restricts membership based on the access-key.  Only hosts
      presenting a valid host access-key are allowed to join.  All
      unrestricted host groups have a null access-key.  outcome
      indicates whether the request is approved or denied.  If it is
      approved, a new transient group address is returned in
      group-address.  access-key is the protection key (or password)
      associated with the new group.  This should fail only if there are
      no free transient group addresses.

         JoinGroup ( group-address, access-key ) --> outcome

      requests that the invoking host become a member of the identified
      host group (permanent or transient).  outcome indicates whether
      the request is approved or denied.  A request is denied if the
      access key is invalid.

         LeaveGroup ( group-address ) --> outcome

      requests that the invoking host be dropped from membership in the
      identified group (permanent or transient).  outcome indicates
      whether the request is approved or denied.

      There is no operation to destroy a transient host group because a
      transient host group is deemed to no longer exist when its
      membership goes to zero.

      Permanent host group addresses are allocated and published by
      Internet administrators, in the same way as well-known TCP and UDP
      port numbers.  That is, they are published in future editions of
      the "Assigned Numbers" document [17].

   4.2. Packet Transmission

      Transmission of a packet in the host group model is controlled by
      two parameters of scope, one being the destination internetwork
      address and the other being the "distance" to the destination
      host(s).  In particular,

         Send ( dest-address, source-address, data, distance )

      transmits the specified data in an internetwork datagram to the
      host(s) identified by dest-address that are within the specified
      distance.  The destination address is thus similar to conventional
      networks except that delivery may be to multiple hosts; the
      distance parameter requires further discussion.

      Distance may be measured in several ways, including number of
      network hops, time to deliver and what might be called
      administrative distance. Administrative distance refers to the
      distance between the administrations of two different networks.
      For example, in a company the networks of the research group and
      advanced development group might be considered quite close to each
      other, networks of the corporate management more distant, and
      networks of other companies much more distant.  One may wish to
      restrict a query to members within one's own administrative domain
      because servers outside that domain may not be trusted.
      Similarly, error reporting outside of an administrative domain may
      not be productive and may in fact be confusing.

      Besides limiting the scope of transmission, the distance parameter
      can be used to control the scope of multicast as a binding
      mechanism and to implement an expanding scope of search for a
      desired service.  For instance, to locate a name server familiar
      with a given name, one might check with nearby name servers and
      expand the distance (by incrementing the distance on
      retransmission) to include more distant name servers until the
      name is found.

      To reach all members of a group, a sender specifies the maximum
      value for the distance parameter.  This maximum must exceed the
      "diameter" of the Internet.

      Packet reception is the same as conventional architectures.  That
      is,

         Receive () --> dest-address, source-address, data

      returns the next internetwork datagram that is, or has been,
      received.

   4.3. Delivery Requirements

      We identify several requirements for the packet delivery mechanism
      that are essential to host groups being a useful and used
      facility.

      Firstly, given the predominance of broadcast local-area networks
      and the locality of communication to individual networks, the
      delivery mechanism must be able to exploit the hardware's
      capability for very efficient multicast within a single local-area
      network.

      Secondly, the delivery mechanism must scale in sophistication to
      efficient delivery across the Internet as it acquires high-speed
      wide-area communication links and higher performance gateways.
      The former are being provided by the introduction of high-speed
      satellite channels and long-haul fiber optic links.  The latter
      are made feasible by the falling cost of memory and processing
      power plus the increasing importance in controlling access to
      relatively unprotected local network environments.  A host group
      delivery mechanism must be able to take advantage of these trends
      as they materialize.

      Finally, the delivery mechanism must avoid "systematic errors" in
      delivery to members of the host group.  That is, a small number of
      repeated transmissions must result in delivery to all group

      members within the specified distance, unless a member is
      disconnected or has failed.  We refer to this property as
      coverage.  In general, most reliable protocols make this basic
      assumption for unicast delivery.  It is important to guarantee
      this assumption for multicast as well or else applications using
      multicast may fail in unexpected ways when coverage is not
      provided.  For efficiency, the multicast delivery mechanism should
      also avoid regularly delivering multiple copies of a packet to
      individual hosts.

      Failure notification is not viewed as an essential requirement,
      given the datagram semantics of delivery.  However, a host group
      extension to IP should provide "hint"-level failure notification
      as the natural extension of the failure notification for unicast.

5. Extensions to IP

   This section discusses the specific extensions to the DARPA Internet
   Protocol required to support the host group model.  The extensions
   need be implemented only on those hosts that wish to join host groups
   or send to host groups; existing implementations are not affected by
   the proposed changes.

   5.1. Group Addresses

      A portion of the 32-bit IP address space is reserved for host
      group addresses.  The range of group addresses is chosen to be
      easily recognized and to not conflict with existing individual
      addresses. Either Class A addresses with a distinguished
      (currently unused) network number or Class D addresses (those
      starting with 111) would be suitable. The range of group addresses
      is further subdivided into a set of permanent group addresses and
      a set of temporary group addresses.

      Host group addresses may be used in the same way as individual
      addresses in the source, destination, and options fields of IP
      datagrams.  An IP implementation adds to the list of its own
      individual addresses, the addresses of all groups to which it
      belongs.  The source addresses of locally originated datagrams are
      validated against the list, and incoming datagrams which are not
      destined to an address on the list are discarded.  The addresses
      on the list change dynamically as IP users create, join and leave
      groups.

   5.2. Group Management

      To support the group management operations of CreateGroup,
      JoinGroup and LeaveGroup, an IP module must interact with one or
      more multicast agents which reside in neighbouring gateways or
      other special-purpose hosts.  These interaction are handled by an
      Internet Group Management Protocol (IGMP) which, like ICMP [15],
      is an integral part of the IP implementation.  A proposed
      specification for IGMP is given in Appendix I.

   5.3. Multicast Delivery

      In order to transmit a datagram destined to a host group, an IP
      module must map the destination group address into a local network
      address.  As with individual IP addresses, the mapping algorithm
      is local-network- specific.  On networks that directly support
      multicast, the IP host group address is mapped to a local network
      multicast address that includes all local members of the host
      group plus one or more multicast agents.  For networks that do not
      directly support multicast, the mapping may be to a more general
      broadcast address, to a list of local unicast addresses, or
      perhaps to the address of a single machine that handles
      multi-destination relaying.

   5.4. Distance Control

      The existing Time to Live field in the IP header can be used for
      crude control over the delivery radius of multicast datagrams.  To
      provide finer-grain control, a new IP option is defined to specify
      the maximum delivery distance in "administrative units", such as
      "this network", "this department", "this company", "this country",
      etc.  The set of units and their encoding is to be determined.

6. Implementation

   In this section, we sketch a design for implementing the host group
   model within the Internet.  This description of the design is given
   to further support the feasibility of the host group model as well as
   point out some of the problems yet to be addressed.

   Implementation of host groups involves implementing a binding
   mechanism (binding Internet addresses to zero or more hosts) and a
   packet delivery mechanism (delivering a packet to each host to which
   its destination address binds).  This facility fits most naturally
   into the gateways of the Internet and the switching nodes of the
   constituent point-to-point networks (as opposed to separate machines)
   because multicast binding and delivery is a natural extension of the

   unicast binding and delivery (i.e. routing plus store-and-forward).
   That is, a multicast packet is routed and transmitted to multiple
   destinations, rather than to a single destination.

   In the following description, we start with a basic, simple
   implementation that provides coverage and then refine this mechanism
   with various optimizations to improve efficiency of delivery and
   group management.

   6.1. Basic Implementation

      A host group defines a network group, which is the set of networks
      containing current members of the host group.  When a packet is
      sent to a host group, a copy is delivered to each network in the
      corresponding network group.  Then, within each network, a copy is
      delivered to each host belonging to the group.

      To support such multicast delivery, every Internet gateway
      maintains the following data structures:

         - routing table: conventional Internet routing information,
           including the distance and direction to the nearest gateway
           on every network.

         - network membership table:  A set of records, one for every
           currently existing host group.  The network membership record
           for a group lists the network group, i.e. the networks that
           contain members of the group.

         - local host membership table:  A set of records, one for each
           host group that has members on directly attached networks.
           Each local host membership record indicates the local hosts
           that are members of the associated host group.  For networks
           that support multicast or broadcast, the record may contain
           only the local network-specific multicast address used by the
           group plus a count of local members.  Otherwise, local group
           members may be identified by a list of unicast addresses to
           be used in the software implementation of multicast within
           the network.

      A host invokes the multicast delivery service by sending a
      group-destined IP datagram to an immediate neighbour gateway (i.e.
      a gateway that is directly attached to the same network as the
      sending host).  Upon receiving a group-destined datagram from a
      directly attached network, a gateway looks up the network
      membership record corresponding to the destination address of the
      datagram.  For each of the networks listed in the membership

      record, the gateway consults its routing table.  If, according to
      the routing table, a member network is directly attached, the
      gateway transmits a copy of the datagram on that network, using
      the network-specific multicast address allocated for the group on
      that network.  For a member network that is not directly attached
      the gateway creates a copy of the datagram with an additional
      inter-gateway header identifying the destination network.  This
      inter-gateway datagram is forwarded to the nearest gateway on the
      destination network, using conventional store-and-forward routing
      techniques.  At the gateway on the destination network, the
      datagram is stripped of its inter-gateway header and transmitted
      to the group's multicast address on that network.  The datagram is
      dropped by the relaying gateways whenever it exceeds its distance
      limit.

      The network membership records and the network-specific multicast
      structures are updated in response to group management requests
      from hosts.  A host sends a request to create, join, or leave a
      group to an immediate neighbour gateway.  If the host requests
      creation of a group, a new network membership record is created by
      the serving gateway and distributed to all other gateways.  If the
      host is the first on its network to join a group, or if the host
      is the last on its network to leave a group, the group's network
      membership record is updated in all gateways.  The updates need
      not be performed atomically at all gateways, due to the datagram
      delivery semantics; hosts can tolerate misrouted and lost packets
      caused by temporary gateway inconsistencies, as long as the
      inconsistencies are resolved within normal host retransmission
      periods. In this respect, the network membership data is similar
      to the network reachability data maintained by conventional
      routing algorithms, and can be handled by similar mechanisms.

      In many cases, a host joins a group that already has members on
      the same network, or leaves a group that has remaining members on
      the same network.  This is then a local matter between the hosts
      and gateways on a single network:  only the local host membership
      table needs to be updated to include or exclude the host.

      This basic implementation strategy meets the delivery requirements
      stated at the end of Section 4.  However, it is far from optimal,
      in terms of either delivery efficiency or group management
      overhead. Below, we discuss some further refinements to the basic
      implementation.

   6.2. Multicast Routing Between Networks

      Multicast routing among the Internet gateways is similar to
      store-and-forward routing in a point-to-point network.  The main
      difference is that the links between the nodes (gateways) can be a
      mixture of broadcast and unicast-type networks with widely
      different throughput and delay characteristics.  In addition,
      packets are addressed to networks rather than hosts (at the
      gateway level).

      We intend to use the extended reverse path forwarding algorithm of
      Dalal and Metcalfe [10].  Although originally designed for
      broadcast, it is a simple and efficient technique that can serve
      well for multicast delivery if network membership records in each
      gateway are augmented with information from neighbouring gateways.
      This algorithm uses the source network identifier, rather than a
      destination network identifier to make routing decisions.  Since
      the source address of a datagram may be a group address, it cannot
      be used to identify the source network of the datagram; the first
      gateway must add a header specifying the source network.  This
      approach minimizes redundant transmissions when multiple
      destination networks are reachable across a common intergateway
      link, a problem with the basic implementation described above.

      Note that we eliminate from consideration techniques that fail to
      deliver along the branches of the shortest delay tree rooted at
      the source, such as Wall's center-based forwarding [16] because
      this compromises the meaning of the multicast distance parameter
      and detracts from multicast performance in general.  We also
      rejected the approach of having a multicast packet carry more than
      one network identifier in its inter-gateway header to indicate
      multiple destination networks because the resulting variable
      length headers would cause buffering and fragmentation problems in
      the gateways.

   6.3. Multicasting Within Networks

      A simple optimization within a network is to have the sender use
      the local multicast address of a host group for its initial
      transmission. This allows the local host group members to receive
      the transmission immediately along with the gateways (which must
      now "eavesdrop" on all multicast transmissions).  A gateway only
      forwards the datagram if the destination host group includes
      members on other networks.  This scheme reduces the cost to reach
      local group members to one packet transmission from two required

      in the basic implementation <3> so transmission to local members
      is basically as efficient as the local multicast support provided
      by the network.

      A similar opportunity for reducing packet traffic arises when a
      datagram must traverse a network to get from one gateway to
      another, and that network also holds members of the destination
      group.  Again, use of a network-specific multicast address which
      includes member hosts plus gateways can achieve the desired
      effect.  However, in this case, hosts must be prepared to accept
      datagrams that include an inter-gateway header or, alternatively,
      every datagram must include a spare field in its header for use by
      gateways in lieu of an additional inter-gateway header.

   6.4. Distributing Membership Information

      A refinement to host group membership maintenance is to store the
      host group membership record for a group only in those gateways
      that are directly connected to member networks.  Information about
      other groups is cached in the gateway only while it is required to
      route to those other groups.  When a gateway receives a datagram
      to be forwarded to a group for which it has no network membership
      record (which can only happen if the gateway is not directly
      connected to a member network), it takes the following action.
      The gateway assumes temporarily that the destination group has
      members on every network in the internetwork, except those
      directly attached to the sending gateway, and routes the datagram
      accordingly.  In the inter-gateway header of the outgoing packet,
      the gateway sets a bit indicating that it wishes to receive a copy
      of the network membership record for the destination host group.
      When such a datagram reaches a gateway on a member network, that
      gateway sends a copy of the membership record back to the
      requesting gateway and clears the copy request bit in the
      datagram.

      Copies of network membership records sent to gateways outside of a
      group's member networks are cached for use in subsequent
      transmissions by those gateways.  That raises the danger of a
      stale cache entry leading to systematic delivery failures.  To
      counter that problem, the inter-gateway header contains a field
      which is a hash value or checksum on the network membership record
      used to route the datagram.  Gateways on member networks compare
      the checksum on incoming datagrams with their up-to-date records.
      If the checksums don't match, an up-to-date copy of the record is
      returned to the gateway with the bad record.

      This caching strategy minimizes intergateway traffic for groups

      that are only used within one network or within the set of
      networks on which members reside, the expected common cases.
      Partial replication with caching also reduces the overhead for
      network traffic to disseminate updates and keep all copies
      consistent.  Finally, it also reduces the total space required in
      all the gateways to support a large number of host groups.

      We have not addressed here the problem of maintaining up-to-date,
      consistent network membership records within the set of gateways
      connected to members of a group.  This can be viewed as a
      distributed database problem which has been well studied in other
      contexts.  The loose consistency requirements on network
      membership records suggest that the techniques used in Grapevine
      [3] might be useful for this application.

7. Related Work

   The use of unreliable multicast by higher-level protocols and the
   implementation of multicast within various individual networks have
   been well-studied (see [7] for references and discussion).  However,
   there is relatively little published work on the use or
   implementation of internetwork multicasting.

   Boggs, in his thesis [4], describes a number of distributed
   applications that are impossible or very awkward to support without
   the flexible binding nature of broadcast addressing.  Although he
   recognizes that almost all of his applications would be best served
   by a multicast mechanism, he advocates the use of "directed
   broadcast" because it is easy to implement within many kinds of
   networks and can be extended across an internetwork without placing
   any new burden on internetwork gateways.  In RFC-919 [13], Mogul
   proposes adopting directed broadcast for the DARPA Internet.

   Broadcasting has the undesirable side effect of delivering packets to
   more hosts than necessary, thus incurring overhead on uninvolved
   parties and possibly creating security problems.  As more and more
   applications take advantage of broadcasting, the overhead on all
   hosts continues to rise.  Clearly, broadcast does not scale up to a
   large internetwork.  As an attempt to handle the scaling problem,
   directed broadcast is less attractive than true multicast because the
   set of hosts that can be reached by a single "send" operation is an
   artifact of the internetwork topology, rather than a grouping that is
   meaningful to the sender.

   In RFC-947 [12], Lebowitz and Mankins propose the use of broadcast

   repeaters that pick up broadcast datagrams from one network and relay
   them to other networks for broadcast there.  This technique is even
   less selective of its targets than Bogg's directed broadcast method.

   Aguilar [1] suggests allowing an IP datagram to carry multiple
   destination addresses, which are used by the gateways to route the
   datagram to each recipient.  Such a facility would alleviate some of
   the inefficiencies of sending individual datagrams to a group, but it
   would not be able to take advantage of local network multicast
   facilities. More seriously, Aguilar's scheme requires the sender to
   know the individual IP addresses of all members of the destination
   group and thus lacks the flexible binding nature of true multicast or
   broadcast.

8. Concluding Remarks

   We have described a model of multicast communication for the
   Internet. As an extension of the existing Internet architecture, it
   views unicast communication and time-to-live constraints as special
   cases of the more general form of communication arising with
   multicast.  We have argued that this model is implementable in the
   Internet and that it provides a powerful facility for a variety of
   applications.  In some cases, it provides a facility that is required
   for certain applications to work in the Internet environment.  In
   other cases, it provides a more efficient, robust and possibly more
   elegant way of implementing existing Internet applications.

   We are currently implementing a prototype host group facility as an
   extension of IP.  For practical reasons, this prototype implements
   all group management functions and multicast routing outside of the
   Internet gateways, in special hosts called multicast agents, which
   are similar to the broadcast repeaters of Lebowitz and Mankins.  The
   collection of multicast agents in effect provides a second gateway
   system on top of the existing Internet, for multicast purposes.  The
   major costs of this separation are redundancy of routing tables
   between gateways and multicast agents and the increased delay and
   unreliability of extra hops in the delivery path.  Much of the
   routing information in the multicast agents must be "wired-in"
   because they do not have access to the gateways' routing tables.
   However, this rudimentary implementation provides an environment for
   evaluating the interface to the multicast service and for
   investigating group management and multicast routing protocols for
   eventual use in the gateways.  It also serves as a testbed for
   porting multicast-based distributed applications to the Internet.

   For now, we are restricting group membership to local networks that
   already have a broadcast or multicast capability, such as the

   Ethernet. We feel that, in the future, any network that is to support
   hosts other than just gateways must have a multicast addressing mode.
   Efficient implementation of multicast within point-to-point or
   virtual circuit networks deserves investigation.

   A significant issue raised by the host group model is authentication
   and access control in the Internet.  Gateways must control which
   hosts can create and join host groups, presumably making their
   decision based on the identity of the requestor (thus requiring
   authentication) and permissions (access control lists).  This issue
   does not arise in conventional internetwork architectures because
   host addresses are administratively assigned with no notion of
   dynamic assignment and binding as provided by host groups.  We
   believe that access control should be recognized as a proper and
   necessary function of gateways so as to protect the hosts of local
   networks from general internetwork activity.  Thus, group access
   control can be subsumed as part of this more general mechanism,
   although more investigation of the general issue is called for.

   On a philosophical point, there has been considerable reluctance to
   make open use of multicast on local networks because it was
   network-specific and not provided across the Internet.  We were
   originally of that school.  However, we recognized that our "hidden"
   uses of multicast in the V distributed system were essential unless
   we resorted to dramatically poorer solutions - wired-in addresses.
   We also recognized, as described in this paper, that an adequate
   multicast facility for the Internet was feasible.  As a consequence,
   we now argue that multicast is an important and basic facility to
   provide in local networks and internetworks.  Higher levels of
   communication, including applications, should feel free to make use
   of this powerful facility. Networks and internetworks lacking
   multicast should be regarded as deficient relative to the future (and
   present) requirements of sophisticated distributed applications and
   communication systems.

Appendix I. Internet Group Management Protocol (IGMP)

   The Internet Group Management Protocol (IGMP) is used between IP
   hosts and their immediate neighbour multicast agents to support the
   allocation of temporary group addresses and the addition and deletion
   of members of a group.

   Like ICMP, IGMP is a required part of all IP implementations.  IGMP
   messages are encapsulated in IP datagrams, with an IP protocol number
   of 2.  IGMP messages are formatted similarly to ICMP messages and the
   different IGMP message types are given values distinct from ICMP
   message types, so that both protocols may share common implementation
   modules or, perhaps, be merged into a single protocol.

   IGMP interactions take the form of request-response transactions.  A
   request message is sent by hosts to the permanent group of all
   immediate neighbour multicast agents.  Multicast agents reply to the
   IP source address of a request.  If no reply is received within a
   (currently unspecified) timeout interval, a host retransmits its
   request, up to some (currently unspecified) maximum number of times.
   IGMP transactions are considered idempotent, so that multicast agents
   need not recognize and filter out duplicate requests nor buffer
   replies <4>.

   The IGMP message formats and procedures are defined below, in the
   style used in the ICMP specification.

   Create Group Request or Create Group Reply Message

       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        |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |                         Group Address                         |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |                                                               |
      +                         Access Key                            +
      |                                                               |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

      IP Fields:

      Addresses

         A Create Group Request message is sent with an individual IP
         address of the sending host as its source, and the well-known
         group address of the multicast agents as its destination.

         The corresponding Create Group Reply is sent with those two
         addresses reversed.

      IGMP Fields:

      Type

         101 for Create Group Request
         102 for Create Group Reply

      Code

         For a Create Group Request message, the Code field indicates if
         the group is to be restricted:

            0 = unrestricted
            1 = restricted

         For a Create Group Reply message, the Code field specifies the
         outcome of the request:

            0 = request approved
            1 = request denied, no resources

      Checksum

         The checksum is the 16-bit one's complement of the one's
         complement sum of the IGMP message starting with the IGMP 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 Reply messages.

      Sequence Number

         A sequence number to aid in matching Request and Reply
         messages.

      Group Address

         For a Create Group Request message, a value of 0.

         For a Create Group Reply message, either a newly allocated
         group address (if the request is approved) or a value of 0 (if
         denied).

      Access Key

         For a Create Group Request message, a value of 0.

         For a Create Group Reply message, either a pseudo-random 64-bit
         number (if the request for a restricted group is approved) or
         0.

      Description

         A Create Group Request message is sent to the the group of
         local multicast agents by a host wishing to allocate a new
         temporary group.

         If no Reply message is received within t seconds, the Request
         is retransmitted.  If no Reply is received after n
         transmissions, the request is deemed to have failed.

         The first Reply message to arrive, if any, specifies the
         outcome of the request.  The request may be denied because of
         lack of resources (e.g. no table space in gateways or all
         temporary addresses in use).

         If the request is approved, the requesting host is considered
         to be the first and only current member of the new host group.

         The Identifier and Sequence Number fields are used to match the
         Reply to the corresponding Request.  The multicast agents may
         choose to use these values to minimize the chance of allocating
         more than one new group for a single request, for example when
         a Reply is lost and a

         Request is retransmitted.  However, the multicast agents must
         be prepared to recover temporary group addresses without
         requiring explicit Leave Group Requests from all members; they
         may choose simply to allocate a new address for every
         retransmission and recover unused ones when needed <5>.

   Join Group Request or Join Group Reply Message

       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        |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |                         Group Address                         |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |                                                               |
      +                         Access Key                            +
      |                                                               |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

      IP Fields:

      Addresses

         A Join Group Request message is sent with an individual IP
         address of the sending host as its source, and the well-known
         group address of the multicast agents as its destination.

         The corresponding Join Group Reply is sent with those two
         addresses reversed.

      IGMP Fields:

      Type

         103 for Join Group Request
         104 for Join Group Reply

      Code

         For a Join Group Request message, the Code field contains 0.

         For a Join Group Reply message, the Code field specifies the
         outcome of the request:

            0 = request approved
            1 = request denied, no resources
            2 = request denied, invalid group address
            3 = request denied, invalid access key

      Checksum

         The checksum is the 16-bit one's complement of the one's
         complement sum of the IGMP message starting with the IGMP 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 Reply messages.

      Sequence Number

         A sequence number to aid in matching Request and Reply
         messages.

      Group Address

         For a Join Group Request message, a host group address.

         For a Join Group Reply message, the same group address as in
         the corresponding request.

      Access Key

         For a Join Group Request message, the access key allocated when
         the group was created (0 for unrestricted groups).

         For a Join Group Reply message, the same access key as in the
         corresponding request.

      Description

         A Join Group Request message is sent to the the group of local
         multicast agents by a host wishing to join a specified,
         existing group.  If no Reply message is received within t
         seconds, the Request is retransmitted.  If no reply is received
         after n transmissions, the request is deemed to have failed.

         The first Reply message to arrive, if any, specifies the
         outcome of the request.  The request may be denied because of
         an invalid access key, an invalid specified group address (e.g.
         non-existent group) or lack of resources (e.g. no table space
         in gateways).

         The Identifier and Sequence Number fields are used to match the
         Reply to the corresponding Request.  If a multicast agent

         receives a request from a host to join a group to which it
         already belongs, the agent approves the request, under the
         assumption that the request was a retransmission for a lost
         Reply.

   Leave Group Request or Leave Group Reply Message

       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        |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |                         Group Address                         |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

      IP Fields:

      Addresses

         A Leave Group Request message is sent with an individual IP
         address of the sending host as its source, and the well-known
         group address of the multicast agents as its destination.

         The corresponding Leave Group Reply is sent with those two
         addresses reversed.

      IGMP Fields:

      Type

         105 for Leave Group Request
         106 for Leave Group Reply

      Code

         For a Leave Group Request message, the Code field contains 0.

         For  Leave Group Reply message, the Code field specifies the
         outcome of the request:

            0 = request approved
            2 = request denied, invalid group address

      Checksum

         The checksum is the 16-bit one's complement of the one's
         complement sum of the IGMP message starting with the IGMP 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 Reply messages.

      Sequence Number

         A sequence number to aid in matching Request and Reply
         messages.

      Group Address

         For a Leave Group Request message, a host group address.

         For a Leave Group Reply message, the same group address as in
         the corresponding request.

      Description

         A Leave Group Request message is sent to the the group of local
         multicast agents by a host wishing to leave a specified,
         existing group.  If no Reply message is received within t
         seconds, the Request is retransmitted.  If no reply is received
         after n transmissions, the request is deemed to have succeeded.

         The first Reply message to arrive, if any, specifies the
         outcome of the request.  The request may be denied only if the
         specified group address is invalid (e.g. an individual rather
         than a group address.)

         The Identifier and Sequence Number fields are used to match the
         Reply to the corresponding Request, as with other ICMP
         transactions. If a multicast agent receives a request from a
         host to leave a group to which it does not belong, the agent
         approves the request, under the assumption that the request was
         a retransmission for a lost Reply.

Notes:

   <1>  In reality, Internet addresses (individual or group) are bound
        to network interfaces or network attachment points, not the host
        machines per se.

   <2>  In this procedure call notation, the arguments for an operation
        are listed in parentheses after the operation name, and the
        returned values, if any, are listed after a --> symbol.

   <3>  One unicast transmission from sender to gateway and one
        multicast transmission from gateway to local group members

   <4>  This protocol may eventually be replaced by a more general
        reliable transaction protocol designed for this type of
        client/server interaction, as suggested in RFC-955 [5].

   <5>  Multicast agents can use an ICMP Echo message to determine if a
        group has any current members.  The Echo message should be
        transmitted several times before deciding the group address is
        no longer in use.

   [1]   L. Aguilar. Datagram Routing for Internet Multicasting. In ACM
         SIGCOMM '84 Communications Architectures and Protocols, pages
         58-63. ACM, June, 1984.

   [2]   E. J. Berglund and D. R. Cheriton. Amaze: A distributed
         multi-player game program using the distributed V kernel. In
         Proceedings of the Fourth International Conference on
         Distributed Systems. IEEE, May, 1984.

   [3]   A. D. Birrell et al. Grapevine: an exercise in distributed
         computing. Communications of the ACM 25(4):260-274, April,
         1982.

   [4]   D. R. Boggs. Internet Broadcasting. PhD thesis, Stanford
         University, January, 1982.

   [5]   R. Braden. Towards a Transport Service for Transaction
         Processing Applications. Technical Report RFC-919, SRI Network
         Information Center, September, 1985.

   [6]   J-M. Chang. Simplifying Distributed Database Design by Using a
         Broadcast Network. In SIGMOD '84. ACM, June, 1984.

   [7]   D. R. Cheriton and S. E. Deering. Host Groups: A Multicast
         Extension for Datagram Internetworks. In Proceedings of the
         Ninth Data Communications Symposium. ACM/IEEE, September, 1985.

   [8]   D. R. Cheriton and W. Zwaenepoel. Distributed Process Groups in
         the V Kernel. ACM Transactions on Computer Systems 3(3), May,
         1985.

   [9]   F. Cristian et al. Atomic Broadcast: from simple message
         diffusion to Byzantine agreement. In 15th International
         Conference on Fault Tolerant Computing. , Ann Arbor, Michigan,
         June, 1985.

   [10]  Y. K. Dalal and R. M. Metcalfe. Reverse Path Forwarding of
         Broadcast Packets. Communications of the ACM 21(2):1040-1047,
         December, 1978.

   [11]  H. Forsdick. MMCF: A Multi-Media Conferencing Facility.
         personal communication.

   [12]  K. Lebowitz and D. Mankins. Multi-network Broadcasting within
         the Internet.Technical Report RFC-947, SRI Network Information
         Center, June, 1985.

   [13]  J. Mogul. Broadcasting Internet Datagrams. Technical Report
         RFC-919, SRI Network Information Center, October, 1984.

   [14]  J. Postel. Internet Protocol. Technical Report RFC-791, SRI
         Network Information Center, September, 1981.

   [15]  J. Postel. Internet Control Message Protocol. Technical Report
         RFC-792, SRI Network Information Center, September, 1981.

   [16]  D. W, Wall. Mechanisms for Broadcast and Selective Broadcast.
         Technical Report 190, Computer Systems Laboratory, Stanford
         University, June, 1980.

   [17]  J. K. Reynolds and J. Postel. Assigned Numbers. Technical
         Report RFC-960, SRI Network Information Center, September,
         1981.