RFC 1009
Requirements for Internet gateways
Network Working Group R. Braden Request for Comments: 1009 J. Postel Obsoletes: 985 ISI June 1987 Requirements for Internet Gateways Status of this Memo This document is a formal statement of the requirements to be met by gateways used in the Internet system. As such, it is an official specification for the Internet community. Distribution of this memo is unlimited. This RFC summarizes the requirements for gateways to be used between networks supporting the Internet protocols. While it was written specifically to support National Science Foundation research programs, the requirements are stated in a general context and are applicable throughout the Internet community. The purpose of this document is to present guidance for vendors offering gateway products that might be used or adapted for use in an Internet application. It enumerates the protocols required and gives references to RFCs and other documents describing the current specifications. In a number of cases the specifications are evolving and may contain ambiguous or incomplete information. In these cases further discussion giving specific guidance is included in this document. Specific policy issues relevant to the NSF scientific networking community are summarized in an Appendix. As other specifications are updated this document will be revised. Vendors are encouraged to maintain contact with the Internet research community. 1. Introduction The following material is intended as an introduction and background for those unfamiliar with the Internet architecture and the Internet gateway model. General background and discussion on the Internet architecture and supporting protocol suite can be found in the DDN Protocol Handbook [25] and ARPANET Information Brochure [26], see also [19, 28, 30, 31]. The Internet protocol architecture was originally developed under DARPA sponsorship to meet both military and civilian communication requirements [32]. The Internet system presently supports a variety of government and government-sponsored operational and research activities. In particular, the National Science Foundation (NSF) is building a major extension to the Internet to provide user access to
national supercomputer centers and other national scientific
resources, and to provide a computer networking capability to a large
number of universities and colleges.
In this document there are many terms that may be obscure to one
unfamiliar with the Internet protocols. There is not much to be done
about that but to learn, so dive in. There are a few terms that are
much abused in general discussion but are carefully and intentionally
used in this document. These few terms are defined here.
Packet A packet is the unit of transmission on a physical
network.
Datagram A datagram is the unit of transmission in the IP
protocol. To cross a particular network a datagram is
encapsulated inside a packet.
Router A router is a switch that receives data transmission
units from input interfaces and, depending on the
addresses in those units, routes them to the
appropriate output interfaces. There can be routers
at different levels of protocol. For example,
Interface Message Processors (IMPs) are packet-level
routers.
Gateway In the Internet documentation generally, and in this
document specifically, a gateway is an IP-level
router. In the Internet community the term has a long
history of this usage [32].
1.1. The DARPA Internet Architecture
1.1.1. Internet Protocols
The Internet system consists of a number of interconnected
packet networks supporting communication among host computers
using the Internet protocols. These protocols include the
Internet Protocol (IP), the Internet Control Message Protocol
(ICMP), the Transmission Control Protocol (TCP), and
application protocols depending upon them [22].
All Internet protocols use IP as the basic data transport
mechanism. IP [1,31] is a datagram, or connectionless,
internetwork service and includes provision for addressing,
type-of-service specification, fragmentation and reassembly,
and security information. ICMP [2] is considered an integral
part of IP, although it is architecturally layered upon IP.
ICMP provides error reporting, flow control and first-hop
gateway redirection.
Reliable data delivery is provided in the Internet protocol
suite by transport-level protocols such as the Transmission
Control Protocol (TCP), which provides end-end retransmission,
resequencing and connection control. Transport-level
connectionless service is provided by the User Datagram
Protocol (UDP).
1.1.2. Networks and Gateways
The constituent networks of the Internet system are required
only to provide packet (connectionless) transport. This
requires only delivery of individual packets. According to the
IP service specification, datagrams can be delivered out of
order, be lost or duplicated and/or contain errors. Reasonable
performance of the protocols that use IP (e.g., TCP) requires
an IP datagram loss rate of less than 5%. In those networks
providing connection-oriented service, the extra reliability
provided by virtual circuits enhances the end-end robustness of
the system, but is not necessary for Internet operation.
Constituent networks may generally be divided into two classes:
* Local-Area Networks (LANs)
LANs may have a variety of designs, typically based upon
buss, ring, or star topologies. In general, a LAN will
cover a small geographical area (e.g., a single building or
plant site) and provide high bandwidth with low delays.
* Wide-Area Networks (WANs)
Geographically-dispersed hosts and LANs are interconnected
by wide-area networks, also called long-haul networks.
These networks may have a complex internal structure of
lines and packet-routers (typified by ARPANET), or they may
be as simple as point-to-point lines.
In the Internet model, constituent networks are connected
together by IP datagram forwarders which are called "gateways"
or "IP routers". In this document, every use of the term
"gateway" is equivalent to "IP router". In current practice,
gateways are normally realized with packet-switching software
executing on a general-purpose CPU, but special-purpose
hardware may also be used (and may be required for future
higher-throughput gateways).
A gateway is connected to two or more networks, appearing to
each of these networks as a connected host. Thus, it has a
physical interface and an IP address on each of the connected
networks. Forwarding an IP datagram generally requires the
gateway to choose the address of the next-hop gateway or (for
the final hop) the destination host. This choice, called
"routing", depends upon a routing data-base within the gateway.
This routing data-base should be maintained dynamically to
reflect the current topology of the Internet system; a gateway
normally accomplishes this by participating in distributed
routing and reachability algorithms with other gateways.
Gateways provide datagram transport only, and they seek to
minimize the state information necessary to sustain this
service in the interest of routing flexibility and robustness.
Routing devices may also operate at the network level; in this
memo we will call such devices MAC routers (informally called
"level-2 routers", and also called "bridges"). The name
derives from the fact that MAC routers base their routing
decision on the addresses in the MAC headers; e.g., in IEEE
802.3 networks, a MAC router bases its decision on the 48-bit
addresses in the MAC header. Network segments which are
connected by MAC routers share the same IP network number,
i.e., they logically form a single IP network.
Another variation on the simple model of networks connected
with gateways sometimes occurs: a set of gateways may be
interconnected with only serial lines, to effectively form a
network in which the routing is performed at the internetwork
(IP) level rather than the network level.
1.1.3. Autonomous Systems
For technical, managerial, and sometimes political reasons, the
gateways of the Internet system are grouped into collections
called "autonomous systems" [35]. The gateways included in a
single autonomous system (AS) are expected to:
* Be under the control of a single operations and
maintenance (O&M) organization;
* Employ common routing protocols among themselves, to
maintain their routing data-bases dynamically.
A number of different dynamic routing protocols have been
developed (see Section 4.1); the particular choice of routing
protocol within a single AS is generically called an interior
gateway protocol or IGP.
An IP datagram may have to traverse the gateways of two or more
ASs to reach its destination, and the ASs must provide each
other with topology information to allow such forwarding. The
Exterior Gateway Protocol (EGP) is used for this purpose,
between gateways of different autonomous systems.
1.1.4. Addresses and Subnets
An IP datagram carries 32-bit source and destination addresses,
each of which is partitioned into two parts -- a constituent
network number and a host number on that network.
Symbolically:
IP-address ::= { <Network-number>, <Host-number> }
To finally deliver the datagram, the last gateway in its path
must map the host-number (or "rest") part of an IP address into
the physical address of a host connection to the constituent
network.
This simple notion has been extended by the concept of
"subnets", which were introduced in order to allow arbitrary
complexity of interconnected LAN structures within an
organization, while insulating the Internet system against
explosive growth in network numbers and routing complexity.
Subnets essentially provide a two-level hierarchical routing
structure for the Internet system. The subnet extension,
described in RFC-950 [21], is now a required part of the
Internet architecture. The basic idea is to partition the
<host number> field into two parts: a subnet number, and a true
host number on that subnet.
IP-address ::=
{ <Network-number>, <Subnet-number>, <Host-number> }
The interconnected LANs of an organization will be given the
same network number but different subnet numbers. The
distinction between the subnets of such a subnetted network
must not be visible outside that network. Thus, wide-area
routing in the rest of the Internet will be based only upon the
<Network-number> part of the IP destination address; gateways
outside the network will lump <Subnet-number> and <Host-number>
together to form an uninterpreted "rest" part of the 32-bit IP
address. Within the subnetted network, the local gateways must
route on the basis of an extended network number:
{ <Network-number>, <Subnet-number> }.
The bit positions containing this extended network number are
indicated by a 32-bit mask called the "subnet mask" [21]; it is
recommended but not required that the <Subnet-number> bits be
contiguous and fall between the <Network-number> and the
<Host-number> fields. No subnet should be assigned the value
zero or -1 (all one bits).
Flexible use of the available address space will be
increasingly important in coping with the anticipated growth of
the Internet. Thus, we allow a particular subnetted network to
use more than one subnet mask. Several campuses with very
large LAN configurations are also creating nested hierarchies
of subnets, sub-subnets, etc.
There are special considerations for the gateway when a
connected network provides a broadcast or multicast capability;
these will be discussed later.
1.2. The Internet Gateway Model
There are two basic models for interconnecting local-area networks
and wide-area (or long-haul) networks in the Internet. In the
first, the local-area network is assigned a network number and all
gateways in the Internet must know how to route to that network.
In the second, the local-area network shares (a small part of) the
address space of the wide-area network. Gateways that support
this second model are called "address sharing gateways" or
"transparent gateways". The focus of this memo is on gateways
that support the first model, but this is not intended to exclude
the use of transparent gateways.
1.2.1. Internet Gateways
An Internet gateway is an IP-level router that performs the
following functions:
1. Conforms to specific Internet protocols specified in
this document, including the Internet Protocol (IP),
Internet Control Message Protocol (ICMP), and others as
necessary. See Section 2 (Protocols Required).
2. Interfaces to two or more packet networks. For each
connected network the gateway must implement the
functions required by that network. These functions
typically include:
a. encapsulating and decapsulating the IP datagrams with
the connected network framing (e.g., an Ethernet
header and checksum);
b. sending and receiving IP datagrams up to the maximum
size supported by that network, this size is the
network's "Maximum Transmission Unit" or "MTU";
c. translating the IP destination address into an
appropriate network-level address for the connected
network (e.g., an Ethernet hardware address);
d. responding to the network flow control and error
indication, if any.
See Section 3 (Constituent Network Interface), for
details on particular constituent network interfaces.
3. Receives and forwards Internet datagrams. Important
issues are buffer management, congestion control, and
fairness. See Section 4 (Gateway Algorithms).
a. Recognizes various error conditions and generates
ICMP error and information messages as required.
b. Drops datagrams whose time-to-live fields have
reached zero.
c. Fragments datagrams when necessary to fit into the
MTU of the next network.
4. Chooses a next-hop destination for each IP datagram,
based on the information in its routing data-base. See
Section 4 (Gateway Algorithms).
5. Supports an interior gateway protocol (IGP) to carry out
distributed routing and reachability algorithms with the
other gateways in the same autonomous system. In
addition, some gateways will need to support the
Exterior Gateway Protocol (EGP) to exchange topological
information with other autonomous systems. See
Section 4 (Gateway Algorithms).
6. Provides system support facilities, including loading,
debugging, status reporting, exception reporting and
control. See Section 5 (Operation and Maintenance).
1.2.2. Embedded Gateways
A gateway may be a stand-alone computer system, dedicated to
its IP router functions. Alternatively, it is possible to
embed gateway functionality within a host operating system
which supports connections to two or more networks. The
best-known example of an operating system with embedded gateway
code is the Berkeley BSD system. The embedded gateway feature
seems to make internetting easy, but it has a number of hidden
pitfalls:
1. If a host has only a single constituent-network
interface, it should not act as a gateway.
For example, hosts with embedded gateway code that
gratuitously forward broadcast packets or datagrams on
the same net often cause packet avalanches.
2. If a (multihomed) host acts as a gateway, it must
implement ALL the relevant gateway requirements
contained in this document.
For example, the routing protocol issues (see Sections
2.6 and 4.1) and the control and monitoring problems are
as hard and important for embedded gateways as for
stand-alone gateways.
Since Internet gateway requirements and
specifications may change independently of operating
system changes, an administration that operates an
embedded gateway in the Internet is strongly advised
to have an ability to maintain and update the gateway
code (e.g., this might require gateway code source).
3. Once a host runs embedded gateway code, it becomes part
of the Internet system. Thus, errors in software or
configuration of such a host can hinder communication
between other hosts. As a consequence, the host
administrator must lose some autonomy.
In many circumstances, a host administrator will need to
disable gateway coded embedded in the operating system,
and any embedded gateway code must be organized so it
can be easily disabled.
4. If a host running embedded gateway code is concurrently
used for other services, the O&M (operation and
maintenance) requirements for the two modes of use may
be in serious conflict.
For example, gateway O&M will in many cases be performed
remotely by an operations center; this may require
privileged system access which the host administrator
would not normally want to distribute.
1.2.3. Transparent Gateways
The basic idea of a transparent gateway is that the hosts on
the local-area network behind such a gateway share the address
space of the wide-area network in front of the gateway. In
certain situations this is a very useful approach and the
limitations do not present significant drawbacks.
The words "in front" and "behind" indicate one of the
limitations of this approach: this model of interconnection is
suitable only for a geographically (and topologically) limited
stub environment. It requires that there be some form of
logical addressing in the network level addressing of the
wide-area network (that is, all the IP addresses in the local
environment map to a few (usually one) physical address in the
wide-area network, in a way consistent with the { IP address
<-> network address } mapping used throughout the wide-area
network).
Multihoming is possible on one wide-area network, but may
present routing problems if the interfaces are geographically
or topologically separated. Multihoming on two (or more)
wide-area networks is a problem due to the confusion of
addresses.
The behavior that hosts see from other hosts in what is
apparently the same network may differ if the transparent
gateway cannot fully emulate the normal wide-area network
service. For example, if there were a transparent gateway
between the ARPANET and an Ethernet, a remote host would not
receive a Destination Dead message [3] if it sent a datagram to
an Ethernet host that was powered off.
1.3. Gateway Characteristics Every Internet gateway must perform the functions listed above. However, a vendor will have many choices on power, complexity, and features for a particular gateway product. It may be helpful to observe that the Internet system is neither homogeneous nor fully-connected. For reasons of technology and geography, it is growing into a global-interconnect system plus a "fringe" of LANs around the "edge". * The global-interconnect system is comprised of a number of wide-area networks to which are attached gateways of several ASs; there are relatively few hosts connected directly to it. The global-interconnect system includes the ARPANET, the NSFNET "backbone", the various NSF regional and consortium networks, other ARPA sponsored networks such as the SATNET and the WBNET, and the DCA sponsored MILNET. It is anticipated that additional networks sponsored by these and other agencies (such as NASA and DOE) will join the global-interconnect system. * Most hosts are connected to LANs, and many organizations have clusters of LANs interconnected by local gateways. Each such cluster is connected by gateways at one or more points into the global-interconnect system. If it is connected at only one point, a LAN is known as a "stub" network. Gateways in the global-interconnect system generally require: * Advanced routing and forwarding algorithms These gateways need routing algorithms which are highly dynamic and also offer type-of-service routing. Congestion is still not a completely resolved issue [24]. Improvements to the current situation will be implemented soon, as the research community is actively working on these issues. * High availability These gateways need to be highly reliable, providing 24 hour a day, 7 days a week service. In case of failure, they must recover quickly. * Advanced O&M features These gateways will typically be operated remotely from a regional or national monitoring center. In their
interconnect role, they will need to provide sophisticated
means for monitoring and measuring traffic and other events
and for diagnosing faults.
* High performance
Although long-haul lines in the Internet today are most
frequently 56 Kbps, DS1 lines (1.5 Mbps) are of increasing
importance, and even higher speeds are likely in the future.
Full-duplex operation is provided at any of these speeds.
The average size of Internet datagrams is rather small, of
the order of 100 bytes. At DS1 line speeds, the
per-datagram processing capability of the gateways, rather
than the line speed, is likely to be the bottleneck. To
fill a DS1 line with average-sized Internet datagrams, a
gateway would need to pass -- receive, route, and send --
2,000 datagrams per second per interface. That is, a
gateway which supported 3 DS1 lines and and Ethernet
interface would need to be able to pass a dazzling 2,000
datagrams per second in each direction on each of the
interfaces, or a aggregate throughput of 8,000 datagrams per
second, in order to fully utilize DS1 lines. This is beyond
the capability of current gateways.
Note: some vendors count input and output operations
separately in datagrams per second figures; for these
vendors, the above example would imply 16,000 datagrams
per second !
Gateways used in the "LAN fringe" (e.g., campus networks) will
generally have to meet less stringent requirements for
performance, availability, and maintenance. These may be high or
medium-performance devices, probably competitively procured from
several different vendors and operated by an internal organization
(e.g., a campus computing center). The design of these gateways
should emphasize low average delay and good burst performance,
together with delay and type-of-service sensitive resource
management. In this environment, there will be less formal O&M,
more hand-crafted static configurations for special cases, and
more need for inter-operation with gateways of other vendors. The
routing mechanism will need to be very flexible, but need not be
so highly dynamic as in the global-interconnect system.
It is important to realize that Internet gateways normally operate
in an unattended mode, but that equipment and software faults can
have a wide-spread (sometimes global) effect. In any environment,
a gateway must be highly robust and able to operate, possibly in a
degraded state, under conditions of extreme congestion or failure
of network resources.
Even though the Internet system is not fully-interconnected, many
parts of the system do need to have redundant connectivity. A
rich connectivity allows reliable service despite failures of
communication lines and gateways, and it can also improve service
by shortening Internet paths and by providing additional capacity.
The engineering tradeoff between cost and reliability must be made
for each component of the Internet system.
2. Protocols Required in Gateways The Internet architecture uses datagram gateways to interconnect constituent networks. This section describes the various protocols which a gateway needs to implement. 2.1. Internet Protocol (IP) IP is the basic datagram protocol used in the Internet system [19, 31]. It is described in RFC-791 [1] and also in MIL-STD-1777 [5] as clarified by RFC-963 [36] ([1] and [5] are intended to describe the same standard, but in quite different words). The subnet extension is described in RFC-950 [21]. With respect to current gateway requirements the following IP features can be ignored, although they may be required in the future: Type of Service field, Security option, and Stream ID option. However, if recognized, the interpretation of these quantities must conform to the standard specification. It is important for gateways to implement both the Loose and Strict Source Route options. The Record Route and Timestamp options are useful diagnostic tools and must be supported in all gateways. The Internet model requires that a gateway be able to fragment datagrams as necessary to match the MTU of the network to which they are being forwarded, but reassembly of fragmented datagrams is generally left to the destination hosts. Therefore, a gateway will not perform reassembly on datagrams it forwards. However, a gateway will generally receive some IP datagrams addressed to itself; for example, these may be ICMP Request/Reply messages, routing update messages (see Sections 2.3 and 2.6), or for monitoring and control (see Section 5). For these datagrams, the gateway will be functioning as a destination host, so it must implement IP reassembly in case the datagrams have been fragmented by some transit gateway. The destination gateway must have a reassembly buffer which is at least as large as the maximum of the MTU values for its network interfaces and 576. Note also that it is possible for a particular protocol implemented by a host or gateway to require a lower bound on reassembly buffer size which is larger than 576. Finally, a datagram which is addressed to a gateway may use any of that gateway's IP addresses as destination address, regardless of which interface the datagram enters. There are five classes of IP addresses: Class A through Class E [23]. Of these, Class D and Class E addresses are
reserved for experimental use. A gateway which is not
participating in these experiments must ignore all datagrams with
a Class D or Class E destination IP address. ICMP Destination
Unreachable or ICMP Redirect messages must not result from
receiving such datagrams.
There are certain special cases for IP addresses, defined in the
latest Assigned Numbers document [23]. These special cases can be
concisely summarized using the earlier notation for an IP address:
IP-address ::= { <Network-number>, <Host-number> }
or
IP-address ::= { <Network-number>, <Subnet-number>,
<Host-number> }
if we also use the notation "-1" to mean the field contains all 1
bits. Some common special cases are as follows:
(a) {0, 0}
This host on this network. Can only be used as a source
address (see note later).
(b) {0, <Host-number>}
Specified host on this network. Can only be used as a
source address.
(c) { -1, -1}
Limited broadcast. Can only be used as a destination
address, and a datagram with this address must never be
forwarded outside the (sub-)net of the source.
(d) {<Network-number>, -1}
Directed broadcast to specified network. Can only be used
as a destination address.
(e) {<Network-number>, <Subnet-number>, -1}
Directed broadcast to specified subnet. Can only be used as
a destination address.
(f) {<Network-number>, -1, -1}
Directed broadcast to all subnets of specified subnetted
network. Can only be used as a destination address.
(g) {127, <any>}
Internal host loopback address. Should never appear outside
a host.
The following two are conventional notation for network numbers,
and do not really represent IP addresses. They can never be used
in an IP datagram header as an IP source or destination address.
(h) {<Network-number>, 0}
Specified network (no host).
(i) {<Network-number>, <Subnet-number>, 0}
Specified subnet (no host).
Note also that the IP broadcast address, which has primary
application to Ethernets and similar technologies that support an
inherent broadcast function, has an all-ones value in the host
field of the IP address. Some early implementations chose the
all-zeros value for this purpose, which is not in conformance with
the specification [23, 49, 50].
2.2. Internet Control Message Protocol (ICMP)
ICMP is an auxiliary protocol used to convey advice and error
messages and is described in RFC-792 [2].
We will discuss issues arising from gateway handling of particular
ICMP messages. The ICMP messages are grouped into two classes:
error messages and information messages. ICMP error messages are
never sent about ICMP error messages, nor about broadcast or
multicast datagrams.
The ICMP error messages are: Destination Unreachable, Redirect,
Source Quench, Time Exceeded, and Parameter Problem.
The ICMP information messages are: Echo, Information,
Timestamp, and Address Mask.
2.2.1. Destination Unreachable The distinction between subnets of a subnetted network, which depends on the address mask described in RFC-950 [21], must not be visible outside that network. This distinction is important in the case of the ICMP Destination Unreachable message. The ICMP Destination Unreachable message is sent by a gateway in response to a datagram which it cannot forward because the destination is unreachable or down. The gateway chooses one of the following two types of Destination Unreachable messages to send: * Net Unreachable * Host Unreachable Net unreachable implies that an intermediate gateway was unable to forward a datagram, as its routing data-base gave no next hop for the datagram, or all paths were down. Host Unreachable implies that the destination network was reachable, but that a gateway on that network was unable to reach the destination host. This might occur if the particular destination network was able to determine that the desired host was unreachable or down. It might also occur when the destination host was on a subnetted network and no path was available through the subnets of this network to the destination. Gateways should send Host Unreachable messages whenever other hosts on the same destination network might be reachable; otherwise, the source host may erroneously conclude that ALL hosts on the network are unreachable, and that may not be the case. 2.2.2. Redirect The ICMP Redirect message is sent by a gateway to a host on the same network, in order to change the gateway used by the host for routing certain datagrams. A choice of four types of Redirect messages is available to specify datagrams destined for a particular host or network, and possibly with a particular type-of-service. If the directly-connected network is not subnetted, a gateway can normally send a network Redirect which applies to all hosts on a specified remote network. Using a network rather than a host Redirect may economize slightly on network traffic and on host routing table storage. However, the saving is not significant, and subnets create an ambiguity about the subnet
mask to be used to interpret a network Redirect. In a general
subnet environment, it is difficult to specify precisely the
cases in which network Redirects can be used.
Therefore, it is recommended that a gateway send only host (or
host and type-of-service) Redirects.
2.2.3. Source Quench
All gateways must contain code for sending ICMP Source Quench
messages when they are forced to drop IP datagrams due to
congestion. Although the Source Quench mechanism is known to
be an imperfect means for Internet congestion control, and
research towards more effective means is in progress, Source
Quench is considered to be too valuable to omit from production
gateways.
There is some argument that the Source Quench should be sent
before the gateway is forced to drop datagrams [62]. For
example, a parameter X could be established and set to have
Source Quench sent when only X buffers remain. Or, a parameter
Y could be established and set to have Source Quench sent when
only Y per cent of the buffers remain.
Two problems for a gateway sending Source Quench are: (1) the
consumption of bandwidth on the reverse path, and (2) the use
of gateway CPU time. To ameliorate these problems, a gateway
must be prepared to limit the frequency with which it sends
Source Quench messages. This may be on the basis of a count
(e.g., only send a Source Quench for every N dropped datagrams
overall or per given source host), or on the basis of a time
(e.g., send a Source Quench to a given source host or overall
at most once per T millseconds). The parameters (e.g., N or T)
must be settable as part of the configuration of the gateway;
furthermore, there should be some configuration setting which
disables sending Source Quenches. These configuration
parameters, including disabling, should ideally be specifiable
separately for each network interface.
Note that a gateway itself may receive a Source Quench as the
result of sending a datagram targeted to another gateway. Such
datagrams might be an EGP update, for example.
2.2.4. Time Exceeded
The ICMP Time Exceeded message may be sent when a gateway
discards a datagram due to the TTL being reduced to zero. It
may also be sent by a gateway if the fragments of a datagram
addressed to the gateway itself cannot be reassembled before
the time limit.
2.2.5. Parameter Problem
The ICMP Parameter Problem message may be sent to the source
host for any problem not specifically covered by another ICMP
message.
2.2.6. Address Mask
Host and gateway implementations are expected to support the
ICMP Address Mask messages described in RFC-950 [21].
2.2.7. Timestamp
The ICMP Timestamp message has proven to be useful for
diagnosing Internet problems. The preferred form for a
timestamp value, the "standard value", is in milliseconds since
midnight GMT. However, it may be difficult to provide this
value with millisecond resolution. For example, many systems
use clocks which update only at line frequency, 50 or 60 times
per second. Therefore, some latitude is allowed in a
"standard" value:
* The value must be updated at a frequency of at least 30
times per second (i.e., at most five low-order bits of
the value may be undefined).
* The origin of the value must be within a few minutes of
midnight, i.e., the accuracy with which operators
customarily set CPU clocks.
To meet the second condition for a stand-alone gateway, it will
be necessary to query some time server host when the gateway is
booted or restarted. It is recommended that the UDP Time
Server Protocol [44] be used for this purpose. A more advanced
implementation would use NTP (Network Time Protocol) [45] to
achieve nearly millisecond clock synchronization; however, this
is not required.
Even if a gateway is unable to establish its time origin, it
ought to provide a "non-standard" timestamp value (i.e., with
the non-standard bit set), as a time in milliseconds from
system startup.
New gateways, especially those expecting to operate at T1 or
higher speeds, are expected to have at least millisecond
clocks.
2.2.8. Information Request/Reply
The Information Request/Reply pair was intended to support
self-configuring systems such as diskless workstations, to
allow them to discover their IP network numbers at boot time.
However, the Reverse ARP (RARP) protocol [15] provides a better
mechanism for a host to use to discover its own IP address, and
RARP is recommended for this purpose. Information
Request/Reply need not be implemented in a gateway.
2.2.9. Echo Request/Reply
A gateway must implement ICMP Echo, since it has proven to be
an extremely useful diagnostic tool. A gateway must be
prepared to receive, reassemble, and echo an ICMP Echo Request
datagram at least as large as the maximum of 576 and the MTU's
of all of the connected networks. See the discussion of IP
reassembly in gateways, Section 2.1.
The following rules resolve the question of the use of IP
source routes in Echo Request and Reply datagrams. Suppose a
gateway D receives an ICMP Echo Request addressed to itself
from host S.
1. If the Echo Request contained no source route, D should
send an Echo Reply back to S using its normal routing
rules. As a result, the Echo Reply may take a different
path than the Request; however, in any case, the pair
will sample the complete round-trip path which any other
higher-level protocol (e.g., TCP) would use for its data
and ACK segments between S and D.
2. If the Echo Request did contain a source route, D should
send an Echo Reply back to S using as a source route the
return route built up in the source-routing option of
the Echo Request.
2.3. Exterior Gateway Protocol (EGP) EGP is the protocol used to exchange reachability information between Autonomous Systems of gateways, and is defined in RFC-904 [11]. See also RFC-827 [51], RFC-888 [46], and RFC-975 [27] for background information. The most widely used EGP implementation is described in RFC-911 [13]. When a dynamic routing algorithm is operated in the gateways of an Autonomous System (AS), the routing data-base must be coupled to the EGP implementation. This coupling should ensure that, when a net is determined to be unreachable by the routing algorithm, the net will not be declared reachable to other ASs via EGP. This requirement is designed to minimize spurious traffic to "black holes" and to ensure fair utilization of the resources on other systems. The present EGP specification defines a model with serious limitations, most importantly a restriction against propagating "third party" EGP information in order to prevent long-lived routing loops [27]. This effectively limits EGP to a two-level hierarchy; the top level is formed by the "core" AS, while the lower level is composed of those ASs which are direct neighbor gateways to the core AS. In practice, in the current Internet, nearly all of the "core gateways" are connected to the ARPANET, while the lower level is composed of those ASs which are directly gatewayed to the ARPANET or MILNET. RFC-975 [27] suggested one way to generalize EGP to lessen these topology restrictions; it has not been adopted as an official specification, although its ideas are finding their way into the new EGP developments. There are efforts underway in the research community to develop an EGP generalization which will remove these restrictions. In EGP, there is no standard interpretation (i.e., metric) for the distance fields in the update messages, so distances are comparable only among gateways of the same AS. In using EGP data, a gateway should compare the distances among gateways of the same AS and prefer a route to that gateway which has the smallest distance value. The values to be announced in the distance fields for particular networks within the local AS should be a gateway configuration parameter; by suitable choice of these values, it will be possible to arrange primary and backup paths from other AS's. There are other EGP parameters, such as polling intervals, which also need to be set in the gateway configuration.
When routing updates become large they must be transmitted in
parts. One strategy is to use IP fragmentation, another is to
explicitly send the routing information in sections. The Internet
Engineering Task Force is currently preparing a recommendation on
this and other EGP engineering issues.
2.4. Address Resolution Protocol (ARP)
ARP is an auxiliary protocol used to perform dynamic address
translation between LAN hardware addresses and Internet addresses,
and is described in RFC-826 [4].
ARP depends upon local network broadcast. In normal ARP usage,
the initiating host broadcasts an ARP Request carrying a target IP
address; the corresponding target host, recognizing its own IP
address, sends back an ARP Reply containing its own hardware
interface address.
A variation on this procedure, called "proxy ARP", has been used
by gateways attached to broadcast LANs [14]. The gateway sends an
ARP Reply specifying its interface address in response to an ARP
Request for a target IP address which is not on the
directly-connected network but for which the gateway offers an
appropriate route. By observing ARP and proxy ARP traffic, a
gateway may accumulate a routing data-base [14].
Proxy ARP (also known in some quarters as "promiscuous ARP" or
"the ARP hack") is useful for routing datagrams from hosts which
do not implement the standard Internet routing rules fully -- for
example, host implementations which predate the introduction of
subnetting. Proxy ARP for subnetting is discussed in detail in
RFC-925 [14].
Reverse ARP (RARP) allows a host to map an Ethernet interface
address into an IP address [15]. RARP is intended to allow a
self-configuring host to learn its own IP address from a server at
boot time.
2.5. Constituent Network Access Protocols
See Section 3.
2.6. Interior Gateway Protocols Distributed routing algorithms continue to be the subject of research and engineering, and it is likely that advances will be made over the next several years. A good algorithm needs to respond rapidly to real changes in Internet connectivity, yet be stable and insensitive to transients. It needs to synchronize the distributed data-base across gateways of its Autonomous System rapidly (to avoid routing loops), while consuming only a small fraction of the available bandwidth. Distributed routing algorithms are commonly broken down into the following three components: A. An algorithm to assign a "length" to each Internet path. The "length" may be a simple count of hops (1, or infinity if the path is broken), or an administratively-assigned cost, or some dynamically-measured cost (usually an average delay). In order to determine a path length, each gateway must at least test whether each of its neighbors is reachable; for this purpose, there must be a "reachability" or "neighbor up/down" protocol. B. An algorithm to compute the shortest path(s) to a given destination. C. A gateway-gateway protocol used to exchange path length and routing information among gateways. The most commonly-used IGPs in Internet gateways are as follows. 2.6.1. Gateway-to-Gateway Protocol (GGP) GGP was designed and implemented by BBN for the first experimental Internet gateways [41]. It is still in use in the BBN LSI/11 gateways, but is regarded as having serious drawbacks [58]. GGP is based upon an algorithm used in the early ARPANET IMPs and later replaced by SPF (see below). GGP is a "min-hop" algorithm, i.e., its length measure is simply the number of network hops between gateway pairs. It implements a distributed shortest-path algorithm, which requires global convergence of the routing tables after a change in topology or connectivity. Each gateway sends a GGP
routing update only to its neighbors, but each update includes
an entry for every known network, where each entry contains the
hop count from the gateway sending the update.
2.6.2. Shortest-Path-First (SPF) Protocols
SPF [40] is the name for a class of routing algorithms based on
a shortest-path algorithm of Dijkstra. The current ARPANET
routing algorithm is SPF, and the BBN Butterfly gateways also
use SPF. Its characteristics are considered superior to
GGP [58].
Under SPF, the routing data-base is replicated rather than
distributed. Each gateway will have its own copy of the same
data-base, containing the entire Internet topology and the
lengths of every path. Since each gateway has all the routing
data and runs a shortest-path algorithm locally, there is no
problem of global convergence of a distributed algorithm, as in
GGP. To build this replicated data-base, a gateway sends SPF
routing updates to ALL other gateways; these updates only list
the distances to each of the gateway's neighbors, making them
much smaller than GGP updates. The algorithm used to
distribute SPF routing updates involves reliable flooding.
2.6.3. Routing Information (RIP)
RIP is the name often used for a class of routing protocols
based upon the Xerox PUP and XNS routing protocols. These are
relatively simple, and are widely available because they are
incorporated in the embedded gateway code of Berkeley BSD
systems. Because of this simplicity, RIP protocols have come
the closest of any to being an "Open IGP", i.e., a protocol
which can be used between different vendors' gateways.
Unfortunately, there is no standard, and in fact not even a
good document, for RIP.
As in GGP, gateways using RIP periodically broadcast their
routing data-base to their neighbor gateways, and use a
hop-count as the metric.
A fixed value of the hop-count (normally 16) is defined to be
"infinity", i.e., network unreachable. A RIP implementation
must include measures to avoid both the slow-convergence
phenomen called "counting to infinity" and the formation of
routing loops. One such measure is a "hold-down" rule. This
rule establishes a period of time (typically 60 seconds) during
which a gateway will ignore new routing information about a
given network, once the gateway has learned that network is
unreachable (has hop-count "infinity"). The hold-down period
must be settable in the gateway configuration; if gateways with
different hold-down periods are using RIP in the same
Autonomous System, routing loops are a distinct possibility.
In general, the hold-down period is chosen large enough to
allow time for unreachable status to propagate to all gateways
in the AS.
2.6.4. Hello
The "Fuzzball" software for an LSI/11 developed by Dave Mills
incorporated an IGP called the "Hello" protocol [39]. This IGP
is mentioned here because the Fuzzballs have been widely used
in Internet experimentation, and because they have served as a
testbed for many new routing ideas.
2.7. Monitoring Protocols
See Section 5 of this document.
2.8. Internet Group Management Protocol (IGMP)
An extension to the IP protocol has been defined to provide
Internet-wide multicasting, i.e., delivery of copies of the same
IP datagram to a set of Internet hosts [47, 48]. This delivery is
to be performed by processes known as "multicasting agents", which
reside either in a host on each net or (preferably) in the
gateways.
The set of hosts to which a datagram is delivered is called a
"host group", and there is a host-agent protocol called IGMP,
which a host uses to join, leave, or create a group. Each host
group is distinguished by a Class D IP address.
This multicasting mechanism and its IGMP protocol are currently
experimental; implementation in vendor gateways would be premature
at this time. A datagram containing a Class D IP address must be
dropped, with no ICMP error message.
(next page on part 2)
