hedrick@topaz.rutgers.edu (Charles Hedrick) (07/03/87)
Now that we know how to open connections, let's get back to the
applications programs. As mentioned earlier, once TCP has opened a
connection, we have something that might as well be a simple wire.
All the hard parts are handled by TCP and IP. However we still need
some agreement as to what we send over this connection. In effect
this is simply an agreement on what set of commands the application
will understand, and the format in which they are to be sent.
Generally, what is sent is a combination of commands and data. They
use context to differentiate. For example, the mail protocol works
like this: Your mail program opens a connection to the mail server at
the other end. Your program gives it your machine's name, the sender
of the message, and the recipients you want it sent to. It then sends
a command saying that it is starting the message. At that point, the
other end stops treating what it sees as commands, and starts
accepting the message. Your end then starts sending the text of the
message. At the end of the message, a special mark is sent (a dot in
the first column). After that, both ends understand that your program
is again sending commands. This is the simplest way to do things, and
the one that most applications use.
File transfer is somewhat more complex. The file transfer protocol
involves two different connections. It starts out just like mail.
The user's program sends commands like "log me in as this user", "here
is my password", "send me the file with this name". However once the
command to send data is sent, a second connection is opened for the
data itself. It would certainly be possible to send the data on the
same connection, as mail does. However file transfers often take a
long time. The designers of the file transfer protocol wanted to
allow the user to continue issuing commands while the transfer is
going on. For example, the user might make an inquiry, or he might
abort the transfer. Thus the designers felt it was best to use a
separate connection for the data and leave the original command
connection for commands. (It is also possible to open command
connections to two different computers, and tell them to send a file
from one to the other. In that case, the data couldn't go over the
command connection.)
Remote terminal connections use another mechanism still. For remote
logins, there is just one connection. It normally sends data. When
it is necessary to send a command (e.g. to set the terminal type or to
change some mode), a special character is used to indicate that the
next character is a command. If the user happens to type that special
character as data, two of them are sent.
We are not going to describe the application protocols in detail in
this document. It's better to read the RFC's yourself. However there
are a couple of common conventions used by applications that will be
described here. First, the common network representation: TCP/IP is
intended to be usable on any computer. Unfortunately, not all
computers agree on how data is represented. There are differences in
character codes (ASCII vs. EBCDIC), in end of line conventions
(carriage return, line feed, or a representation using counts), and in
whether terminals expect characters to be sent individually or a line
at a time. In order to allow computers of different kinds to
communicate, each applications protocol defines a standard
14
representation. Note that TCP and IP do not care about the
representation. TCP simply sends octets. However the programs at
both ends have to agree on how the octets are to be interpreted. The
RFC for each application specifies the standard representation for
that application. Normally it is "net ASCII". This uses ASCII
characters, with end of line denoted by a carriage return followed by
a line feed. For remote login, there is also a definition of a
"standard terminal", which turns out to be a half-duplex terminal with
echoing happening on the local machine. Most applications also make
provisions for the two computers to agree on other representations
that they may find more convenient. For example, PDP-10's have 36-bit
words. There is a way that two PDP-10's can agree to send a 36-bit
binary file. Similarly, two systems that prefer full-duplex terminal
conversations can agree on that. However each application has a
standard representation, which every machine must support.
3.1 An example application: SMTP
In order to give a bit better idea what is involved in the application
protocols, I'm going to show an example of SMTP, which is the mail
protocol. (SMTP is "simple mail transfer protocol.) We assume that a
computer called TOPAZ.RUTGERS.EDU wants to send the following message.
Date: Sat, 27 Jun 87 13:26:31 EDT
From: hedrick@topaz.rutgers.edu
To: levy@red.rutgers.edu
Subject: meeting
Let's get together Monday at 1pm.
First, note that the format of the message itself is described by an
Internet standard (RFC 822). The standard specifies the fact that the
message must be transmitted as net ASCII (i.e. it must be ASCII, with
carriage return/linefeed to delimit lines). It also describes the
general structure, as a group of header lines, then a blank line, and
then the body of the message. Finally, it describes the syntax of the
header lines in detail. Generally they consist of a keyword and then
a value.
Note that the addressee is indicated as LEVY@RED.RUTGERS.EDU.
Initially, addresses were simply "person at machine". However recent
standards have made things more flexible. There are now provisions
for systems to handle other systems' mail. This can allow automatic
forwarding on behalf of computers not connected to the Internet. It
can be used to direct mail for a number of systems to one central mail
server. Indeed there is no requirement that an actual computer by the
name of RED.RUTGERS.EDU even exist. The name servers could be set up
so that you mail to department names, and each department's mail is
routed automatically to an appropriate computer. It is also possible
that the part before the @ is something other than a user name. It is
possible for programs to be set up to process mail. There are also
provisions to handle mailing lists, and generic names such as
15
"postmaster" or "operator".
The way the message is to be sent to another system is described by
RFC's 821 and 974. The program that is going to be doing the sending
asks the name server several queries to determine where to route the
message. The first query is to find out which machines handle mail
for the name RED.RUTGERS.EDU. In this case, the server replies that
RED.RUTGERS.EDU handles its own mail. The program then asks for the
address of RED.RUTGERS.EDU, which is 128.6.4.2. Then the mail program
opens a TCP connection to port 25 on 128.6.4.2. Port 25 is the
well-known socket used for receiving mail. Once this connection is
established, the mail program starts sending commands. Here is a
typical conversation. Each line is labelled as to whether it is from
TOPAZ or RED. Note that TOPAZ initiated the connection:
RED 220 RED.RUTGERS.EDU SMTP Service at 29 Jun 87 05:17:18 EDT
TOPAZ HELO topaz.rutgers.edu
RED 250 RED.RUTGERS.EDU - Hello, TOPAZ.RUTGERS.EDU
TOPAZ MAIL From:<hedrick@topaz.rutgers.edu>
RED 250 MAIL accepted
TOPAZ RCPT To:<levy@red.rutgers.edu>
RED 250 Recipient accepted
TOPAZ DATA
RED 354 Start mail input; end with <CRLF>.<CRLF>
TOPAZ Date: Sat, 27 Jun 87 13:26:31 EDT
TOPAZ From: hedrick@topaz.rutgers.edu
TOPAZ To: levy@red.rutgers.edu
TOPAZ Subject: meeting
TOPAZ
TOPAZ Let's get together Monday at 1pm.
TOPAZ .
RED 250 OK
TOPAZ QUIT
RED 221 RED.RUTGERS.EDU Service closing transmission channel
First, note that commands all use normal text. This is typical of the
Internet standards. Many of the protocols use standard ASCII
commands. This makes it easy to watch what is going on and to
diagnose problems. For example, the mail program keeps a log of each
conversation. If something goes wrong, the log file can simply be
mailed to the postmaster. Since it is normal text, he can see what
was going on. It also allows a human to interact directly with the
mail server, for testing. (Some newer protocols are complex enough
that this is not practical. The commands would have to have a syntax
that would require a significant parser. Thus there is a tendency for
newer protocols to use binary formats. Generally they are structured
like C or Pascal record structures.) Second, note that the responses
all begin with numbers. This is also typical of Internet protocols.
The allowable responses are defined in the protocol. The numbers
allow the user program to respond unambiguously. The rest of the
response is text, which is normally for use by any human who may be
watching or looking at a log. It has no effect on the operation of
the programs. (However there is one point at which the protocol uses
part of the text of the response.) The commands themselves simply
allow the mail program on one end to tell the mail server the
16
information it needs to know in order to deliver the message. In this
case, the mail server could get the information by looking at the
message itself. But for more complex cases, that would not be safe.
Every session must begin with a HELO, which gives the name of the
system that initiated the connection. Then the sender and recipients
are specified. (There can be more than one RCPT command, if there are
several recipients.) Finally the data itself is sent. Note that the
text of the message is terminated by a line containing just a period.
(If such a line appears in the message, the period is doubled.) After
the message is accepted, the sender can send another message, or
terminate the session as in the example above.
Generally, there is a pattern to the response numbers. The protocol
defines the specific set of responses that can be sent as answers to
any given command. However programs that don't want to analyze them
in detail can just look at the first digit. In general, responses
that begin with a 2 indicate success. Those that begin with 3
indicate that some further action is needed, as shown above. 4 and 5
indicate errors. 4 is a "temporary" error, such as a disk filling.
The message should be saved, and tried again later. 5 is a permanent
error, such as a non-existent recipient. The message should be
returned to the sender with an error message.
(For more details about the protocols mentioned in this section, see
RFC's 821/822 for mail, RFC 959 for file transfer, and RFC's 854/855
for remote logins. For the well-known port numbers, see the current
edition of Assigned Numbers, and possibly RFC 814.)
4. Protocols other than TCP: UDP and ICMP
So far, we have described only connections that use TCP. Recall that
TCP is responsible for breaking up messages into datagrams, and
reassembling them properly. However in many applications, we have
messages that will always fit in a single datagram. An example is
name lookup. When a user attempts to make a connection to another
system, he will generally specify the system by name, rather than
Internet address. His system has to translate that name to an address
before it can do anything. Generally, only a few systems have the
database used to translate names to addresses. So the user's system
will want to send a query to one of the systems that has the database.
This query is going to be very short. It will certainly fit in one
datagram. So will the answer. Thus it seems silly to use TCP. Of
course TCP does more than just break things up into datagrams. It
also makes sure that the data arrives, resending datagrams where
necessary. But for a question that fits in a single datagram, we
don't need all the complexity of TCP to do this. If we don't get an
answer after a few seconds, we can just ask again. For applications
like this, there are alternatives to TCP.
The most common alternative is UDP ("user datagram protocol"). UDP is
designed for applications where you don't need to put sequences of
datagrams together. It fits into the system much like TCP. There is
17
a UDP header. The network software puts the UDP header on the front
of your data, just as it would put a TCP header on the front of your
data. Then UDP sends the data to IP, which adds the IP header,
putting UDP's protocol number in the protocol field instead of TCP's
protocol number. However UDP doesn't do as much as TCP does. It
doesn't split data into multiple datagrams. It doesn't keep track of
what it has sent so it can resend if necessary. About all that UDP
provides is port numbers, so that several programs can use UDP at
once. UDP port numbers are used just like TCP port numbers. There
are well-known port numbers for servers that use UDP. Note that the
UDP header is shorter than a TCP header. It still has source and
destination port numbers, and a checksum, but that's about it. No
sequence number, since it is not needed. UDP is used by the protocols
that handle name lookups (see IEN 116, RFC 882, and RFC 883), and a
number of similar protocols.
Another alternative protocol is ICMP ("Internet control message
protocol"). ICMP is used for error messages, and other messages
intended for the TCP/IP software itself, rather than any particular
user program. For example, if you attempt to connect to a host, your
system may get back an ICMP message saying "host unreachable". ICMP
can also be used to find out some information about the network. See
RFC 792 for details of ICMP. ICMP is similar to UDP, in that it
handles messages that fit in one datagram. However it is even simpler
than UDP. It doesn't even have port numbers in its header. Since all
ICMP messages are interpreted by the network software itself, no port
numbers are needed to say where a ICMP message is supposed to go.
5. Keeping track of names and information: the domain system
As we indicated earlier, the network software generally needs a 32-bit
Internet address in order to open a connection or send a datagram.
However users prefer to deal with computer names rather than numbers.
Thus there is a database that allows the software to look up a name
and find the corresponding number. When the Internet was small, this
was easy. Each system would have a file that listed all of the other
systems, giving both their name and number. There are now too many
computers for this approach to be practical. Thus these files have
been replaced by a set of name servers that keep track of host names
and the corresponding Internet addresses. (In fact these servers are
somewhat more general than that. This is just one kind of information
stored in the domain system.) Note that a set of interlocking servers
are used, rather than a single central one. There are now so many
different institutions connected to the Internet that it would be
impractical for them to notify a central authority whenever they
installed or moved a computer. Thus naming authority is delegated to
individual institutions. The name servers form a tree, corresponding
to institutional structure. The names themselves follow a similar
structure. A typical example is the name BORAX.LCS.MIT.EDU. This is
a computer at the Laboratory for Computer Science (LCS) at MIT. In
order to find its Internet address, you might potentially have to
consult 4 different servers. First, you would ask a central server
18
(called the root) where the EDU server is. EDU is a server that keeps
track of educational institutions. The root server would give you the
names and Internet addresses of several servers for EDU. (There are
several servers at each level, to allow for the possibly that one
might be down.) You would then ask EDU where the server for MIT is.
Again, it would give you names and Internet addresses of several
servers for MIT. Generally, not all of those servers would be at MIT,
to allow for the possibility of a general power failure at MIT. Then
you would ask MIT where the server for LCS is, and finally you would
ask one of the LCS servers about BORAX. The final result would be the
Internet address for BORAX.LCS.MIT.EDU. Each of these levels is
referred to as a "domain". The entire name, BORAX.LCS.MIT.EDU, is
called a "domain name". (So are the names of the higher-level
domains, such as LCS.MIT.EDU, MIT.EDU, and EDU.)
Fortunately, you don't really have to go through all of this most of
the time. First of all, the root name servers also happen to be the
name servers for the top-level domains such as EDU. Thus a single
query to a root server will get you to MIT. Second, software
generally remembers answers that it got before. So once we look up a
name at LCS.MIT.EDU, our software remembers where to find servers for
LCS.MIT.EDU, MIT.EDU, and EDU. It also remembers the translation of
BORAX.LCS.MIT.EDU. Each of these pieces of information has a "time to
live" associated with it. Typically this is a few days. After that,
the information expires and has to be looked up again. This allows
institutions to change things.
The domain system is not limited to finding out Internet addresses.
Each domain name is a node in a database. The node can have records
that define a number of different properties. Examples are Internet
address, computer type, and a list of services provided by a computer.
A program can ask for a specific piece of information, or all
information about a given name. It is possible for a node in the
database to be marked as an "alias" (or nickname) for another node.
It is also possible to use the domain system to store information
about users, mailing lists, or other objects.
There is an Internet standard defining the operation of these
databases, as well as the protocols used to make queries of them.
Every network utility has to be able to make such queries, since this
is now the official way to evaluate host names. Generally utilities
will talk to a server on their own system. This server will take care
of contacting the other servers for them. This keeps down the amount
of code that has to be in each application program.
The domain system is particularly important for handling computer
mail. There are entry types to define what computer handles mail for
a given name, to specify where an individual is to receive mail, and
to define mailing lists.
(See RFC's 882, 883, and 973 for specifications of the domain system.
RFC 974 defines the use of the domain system in sending mail.)
19
6. Routing
The description above indicated that the IP implementation is
responsible for getting datagrams to the destination indicated by the
destination address, but little was said about how this would be done.
The task of finding how to get a datagram to its destination is
referred to as "routing". In fact many of the details depend upon the
particular implementation. However some general things can be said.
First, it is necessary to understand the model on which IP is based.
IP assumes that a system is attached to some local network. We assume
that the system can send datagrams to any other system on its own
network. (In the case of Ethernet, it simply finds the Ethernet
address of the destination system, and puts the datagram out on the
Ethernet.) The problem comes when a system is asked to send a
datagram to a system on a different network. This problem is handled
by gateways. A gateway is a system that connects a network with one
or more other networks. Gateways are often normal computers that
happen to have more than one network interface. For example, we have
a Unix machine that has two different Ethernet interfaces. Thus it is
connected to networks 128.6.4 and 128.6.3. This machine can act as a
gateway between those two networks. The software on that machine must
be set up so that it will forward datagrams from one network to the
other. That is, if a machine on network 128.6.4 sends a datagram to
the gateway, and the datagram is addressed to a machine on network
128.6.3, the gateway will forward the datagram to the destination.
Major communications centers often have gateways that connect a number
of different networks. (In many cases, special-purpose gateway
systems provide better performance or reliability than general-purpose
systems acting as gateways. A number of vendors sell such systems.)
Routing in IP is based entirely upon the network number of the
destination address. Each computer has a table of network numbers.
For each network number, a gateway is listed. This is the gateway to
be used to get to that network. Note that the gateway doesn't have to
connect directly to the network. It just has to be the best place to
go to get there. For example at Rutgers, our interface to NSFnet is
at the John von Neuman Supercomputer Center (JvNC). Our connection to
JvNC is via a high-speed serial line connected to a gateway whose
address is 128.6.3.12. Systems on net 128.6.3 will list 128.6.3.12 as
the gateway for many off-campus networks. However systems on net
128.6.4 will list 128.6.4.1 as the gateway to those same off-campus
networks. 128.6.4.1 is the gateway between networks 128.6.4 and
128.6.3, so it is the first step in getting to JvNC.
When a computer wants to send a datagram, it first checks to see if
the destination address is on the system's own local network. If so,
the datagram can be sent directly. Otherwise, the system expects to
find an entry for the network that the destination address is on. The
datagram is sent to the gateway listed in that entry. This table can
get quite big. For example, the Internet now includes several hundred
individual networks. Thus various strategies have been developed to
reduce the size of the routing table. One strategy is to depend upon
"default routes". Often, there is only one gateway out of a network.
20
This gateway might connect a local Ethernet to a campus-wide backbone
network. In that case, we don't need to have a separate entry for
every network in the world. We simply define that gateway as a
"default". When no specific route is found for a datagram, the
datagram is sent to the default gateway. A default gateway can even
be used when there are several gateways on a network. There are
provisions for gateways to send a message saying "I'm not the best
gateway -- use this one instead." (The message is sent via ICMP. See
RFC 792.) Most network software is designed to use these messages to
add entries to their routing tables. Suppose network 128.6.4 has two
gateways, 128.6.4.59 and 128.6.4.1. 128.6.4.59 leads to several other
internal Rutgers networks. 128.6.4.1 leads indirectly to the NSFnet.
Suppose we set 128.6.4.59 as a default gateway, and have no other
routing table entries. Now what happens when we need to send a
datagram to MIT? MIT is network 18. Since we have no entry for
network 18, the datagram will be sent to the default, 128.6.4.59. As
it happens, this gateway is the wrong one. So it will forward the
datagram to 128.6.4.1. But it will also send back an error saying in
effect: "to get to network 18, use 128.6.4.1". Our software will then
add an entry to the routing table. Any future datagrams to MIT will
then go directly to 128.6.4.1. (The error message is sent using the
ICMP protocol. The message type is called "ICMP redirect.")
Most IP experts recommend that individual computers should not try to
keep track of the entire network. Instead, they should start with
default gateways, and let the gateways tell them the routes, as just
described. However this doesn't say how the gateways should find out
about the routes. The gateways can't depend upon this strategy. They
have to have fairly complete routing tables. For this, some sort of
routing protocol is needed. A routing protocol is simply a technique
for the gateways to find each other, and keep up to date about the
best way to get to every network. RFC 1009 contains a review of
gateway design and routing. However rip.doc is probably a better
introduction to the subject. It contains some tutorial material, and
a detailed description of the most commonly-used routing protocol.
7. Details about Internet addresses: subnets and broadcasting
As indicated earlier, Internet addresses are 32-bit numbers, normally
written as 4 octets (in decimal), e.g. 128.6.4.7. There are actually
3 different types of address. The problem is that the address has to
indicate both the network and the host within the network. It was
felt that eventually there would be lots of networks. Many of them
would be small, but probably 24 bits would be needed to represent all
the IP networks. It was also felt that some very big networks might
need 24 bits to represent all of their hosts. This would seem to lead
to 48 bit addresses. But the designers really wanted to use 32 bit
addresses. So they adopted a kludge. The assumption is that most of
the networks will be small. So they set up three different ranges of
address. Addresses beginning with 1 to 126 use only the first octet
for the network number. The other three octets are available for the
host number. Thus 24 bits are available for hosts. These numbers are
21
used for large networks. But there can only be 126 of these very big
networks. The Arpanet is one, and there are a few large commercial
networks. But few normal organizations get one of these "class A"
addresses. For normal large organizations, "class B" addresses are
used. Class B addresses use the first two octets for the network
number. Thus network numbers are 128.1 through 191.254. (We avoid 0
and 255, for reasons that we see below. We also avoid addresses
beginning with 127, because that is used by some systems for special
purposes.) The last two octets are available for host addesses,
giving 16 bits of host address. This allows for 64516 computers,
which should be enough for most organizations. (It is possible to get
more than one class B address, if you run out.) Finally, class C
addresses use three octets, in the range 192.1.1 to 223.254.254.
These allow only 254 hosts on each network, but there can be lots of
these networks. Addresses above 223 are reserved for future use, as
class D and E (which are currently not defined).
Many large organizations find it convenient to divide their network
number into "subnets". For example, Rutgers has been assigned a class
B address, 128.6. We find it convenient to use the third octet of the
address to indicate which Ethernet a host is on. This division has no
significance outside of Rutgers. A computer at another institution
would treat all datagrams addressed to 128.6 the same way. They would
not look at the third octet of the address. Thus computers outside
Rutgers would not have different routes for 128.6.4 or 128.6.5. But
inside Rutgers, we treat 128.6.4 and 128.6.5 as separate networks. In
effect, gateways inside Rutgers have separate entries for each Rutgers
subnet, whereas gateways outside Rutgers just have one entry for
128.6. Note that we could do exactly the same thing by using a
separate class C address for each Ethernet. As far as Rutgers is
concerned, it would be just as convenient for us to have a number of
class C addresses. However using class C addresses would make things
inconvenient for the rest of the world. Every institution that wanted
to talk to us would have to have a separate entry for each one of our
networks. If every institution did this, there would be far too many
networks for any reasonable gateway to keep track of. By subdividing
a class B network, we hide our internal structure from everyone else,
and save them trouble. This subnet strategy requires special
provisions in the network software. It is described in RFC 950.
0 and 255 have special meanings. 0 is reserved for machines that
don't know their address. In certain circumstances it is possible for
a machine not to know the number of the network it is on, or even its
own host address. For example, 0.0.0.23 would be a machine that knew
it was host number 23, but didn't know on what network.
255 is used for "broadcast". A broadcast is a message that you want
every system on the network to see. Broadcasts are used in some
situations where you don't know who to talk to. For example, suppose
you need to look up a host name and get its Internet address.
Sometimes you don't know the address of the nearest name server. In
that case, you might send the request as a broadcast. There are also
cases where a number of systems are interested in information. It is
then less expensive to send a single broadcast than to send datagrams
individually to each host that is interested in the information. In
22
order to send a broadcast, you use an address that is made by using
your network address, with all ones in the part of the address where
the host number goes. For example, if you are on network 128.6.4, you
would use 128.6.4.255 for broadcasts. How this is actually
implemented depends upon the medium. It is not possible to send
broadcasts on the Arpanet, or on point to point lines. However it is
possible on an Ethernet. If you use an Ethernet address with all its
bits on (all ones), every machine on the Ethernet is supposed to look
at that datagram.
Although the official broadcast address for network 128.6.4 is now
128.6.4.255, there are some other addresses that may be treated as
broadcasts by certain implementations. For convenience, the standard
also allows 255.255.255.255 to be used. This refers to all hosts on
the local network. It is often simpler to use 255.255.255.255 instead
of finding out the network number for the local network and forming a
broadcast address such as 128.6.4.255. In addition, certain older
implementations may use 0 instead of 255 to form the broadcast
address. Such implementations would use 128.6.4.0 instead of
128.6.4.255 as the broadcast address on network 128.6.4. Finally,
certain older implementations may not understand about subnets. Thus
they consider the network number to be 128.6. In that case, they will
assume a broadcast address of 128.6.255.255 or 128.6.0.0. Until
support for broadcasts is implemented properly, it can be a somewhat
dangerous feature to use.
Because 0 and 255 are used for unknown and broadcast addresses, normal
hosts should never be given addresses containing 0 or 255. Addresses
should never begin with 0, 127, or any number above 223. Addresses
violating these rules are sometimes referred to as "Martians", because
of rumors that the Central University of Mars is using network 225.
8. Datagram fragmentation and reassembly
TCP/IP is designed for use with many different kinds of network.
Unfortunately, network designers do not agree about how big packets
can be. Ethernet packets can be 1500 octets long. Arpanet packets
have a maximum of around 1000 octets. Some very fast networks have
much larger packet sizes. At first, you might think that IP should
simply settle on the smallest possible size. Unfortunately, this
would cause serious performance problems. When transferring large
files, big packets are far more efficient than small ones. So we want
to be able to use the largest packet size possible. But we also want
to be able to handle networks with small limits. There are two
provisions for this. First, TCP has the ability to "negotiate" about
datagram size. When a TCP connection first opens, both ends can send
the maximum datagram size they can handle. The smaller of these
numbers is used for the rest of the connection. This allows two
implementations that can handle big datagrams to use them, but also
lets them talk to implementations that can't handle them. However
this doesn't completely solve the problem. The most serious problem
is that the two ends don't necessarily know about all of the steps in
23
between. For example, when sending data between Rutgers and Berkeley,
it is likely that both computers will be on Ethernets. Thus they will
both be prepared to handle 1500-octet datagrams. However the
connection will at some point end up going over the Arpanet. It can't
handle packets of that size. For this reason, there are provisions to
split datagrams up into pieces. (This is referred to as
"fragmentation".) The IP header contains fields indicating the a
datagram has been split, and enough information to let the pieces be
put back together. If a gateway connects an Ethernet to the Arpanet,
it must be prepared to take 1500-octet Ethernet packets and split them
into pieces that will fit on the Arpanet. Furthermore, every host
implementation of TCP/IP must be prepared to accept pieces and put
them back together. This is referred to as "reassembly".
TCP/IP implementations differ in the approach they take to deciding on
datagram size. It is fairly common for implementations to use
576-byte datagrams whenever they can't verify that the entire path is
able to handle larger packets. This rather conservative strategy is
used because of the number of implementations with bugs in the code to
reassemble fragments. Implementors often try to avoid ever having
fragmentation occur. Different implementors take different approaches
to deciding when it is safe to use large datagrams. Some use them
only for the local network. Others will use them for any network on
the same campus. 576 bytes is a "safe" size, which every
implementation must support.
9. Ethernet encapsulation: ARP
There was a brief discussion earlier about what IP datagrams look like
on an Ethernet. The discussion showed the Ethernet header and
checksum. However it left one hole: It didn't say how to figure out
what Ethernet address to use when you want to talk to a given Internet
address. In fact, there is a separate protocol for this, called ARP
("address resolution protocol"). (Note by the way that ARP is not an
IP protocol. That is, the ARP datagrams do not have IP headers.)
Suppose you are on system 128.6.4.194 and you want to connect to
system 128.6.4.7. Your system will first verify that 128.6.4.7 is on
the same network, so it can talk directly via Ethernet. Then it will
look up 128.6.4.7 in its ARP table, to see if it already knows the
Ethernet address. If so, it will stick on an Ethernet header, and
send the packet. But suppose this system is not in the ARP table.
There is no way to send the packet, because you need the Ethernet
address. So it uses the ARP protocol to send an ARP request.
Essentially an ARP request says "I need the Ethernet address for
128.6.4.7". Every system listens to ARP requests. When a system sees
an ARP request for itself, it is required to respond. So 128.6.4.7
will see the request, and will respond with an ARP reply saying in
effect "128.6.4.7 is 8:0:20:1:56:34". (Recall that Ethernet addresses
are 48 bits. This is 6 octets. Ethernet addresses are conventionally
shown in hex, using the punctuation shown.) Your system will save
this information in its ARP table, so future packets will go directly.
Most systems treat the ARP table as a cache, and clear entries in it
24
if they have not been used in a certain period of time.
Note by the way that ARP requests must be sent as "broadcasts". There
is no way that an ARP request can be sent directly to the right
system. After all, the whole reason for sending an ARP request is
that you don't know the Ethernet address. So an Ethernet address of
all ones is used, i.e. ff:ff:ff:ff:ff:ff. By convention, every
machine on the Ethernet is required to pay attention to packets with
this as an address. So every system sees every ARP requests. They
all look to see whether the request is for their own address. If so,
they respond. If not, they could just ignore it. (Some hosts will
use ARP requests to update their knowledge about other hosts on the
network, even if the request isn't for them.) Note that packets whose
IP address indicates broadcast (e.g. 255.255.255.255 or 128.6.4.255)
are also sent with an Ethernet address that is all ones.
10. Getting more information
This directory contains documents describing the major protocols.
There are literally hundreds of documents, so we have chosen the ones
that seem most important. Internet standards are called RFC's. RFC
stands for Request for Comment. A proposed standard is initially
issued as a proposal, and given an RFC number. When it is finally
accepted, it is added to Official Internet Protocols, but it is still
referred to by the RFC number. We have also included two IEN's.
(IEN's used to be a separate classification for more informal
documents. This classification no longer exists -- RFC's are now used
for all official Internet documents, and a mailing list is used for
more informal reports.) The convention is that whenever an RFC is
revised, the revised version gets a new number. This is fine for most
purposes, but it causes problems with two documents: Assigned Numbers
and Official Internet Protocols. These documents are being revised
all the time, so the RFC number keeps changing. You will have to look
in rfc-index.txt to find the number of the latest edition. Anyone who
is seriously interested in TCP/IP should read the RFC describing IP
(791). RFC 1009 is also useful. It is a specification for gateways
to be used by NSFnet. As such, it contains an overview of a lot of
the TCP/IP technology. You should probably also read the description
of at least one of the application protocols, just to get a feel for
the way things work. Mail is probably a good one (821/822). TCP
(793) is of course a very basic specification. However the spec is
fairly complex, so you should only read this when you have the time
and patience to think about it carefully. Fortunately, the author of
the major RFC's (Jon Postel) is a very good writer. The TCP RFC is
far easier to read than you would expect, given the complexity of what
it is describing. You can look at the other RFC's as you become
curious about their subject matter.
Here is a list of the documents you are more likely to want:
rfc-index list of all RFC's
25
rfc1012 somewhat fuller list of all RFC's
rfc1011 Official Protocols. It's useful to scan this to see
what tasks protocols have been built for. This defines
which RFC's are actual standards, as opposed to
requests for comments.
rfc1010 Assigned Numbers. If you are working with TCP/IP, you
will probably want a hardcopy of this as a reference.
It's not very exciting to read. It lists all the
offically defined well-known ports and lots of other
things.
rfc1009 NSFnet gateway specifications. A good overview of IP
routing and gateway technology.
rfc1001/2 netBIOS: networking for PC's
rfc973 update on domains
rfc959 FTP (file transfer)
rfc950 subnets
rfc937 POP2: protocol for reading mail on PC's
rfc894 how IP is to be put on Ethernet, see also rfc825
rfc882/3 domains (the database used to go from host names to
Internet address and back -- also used to handle UUCP
these days). See also rfc973
rfc854/5 telnet - protocol for remote logins
rfc826 ARP - protocol for finding out Ethernet addresses
rfc821/2 mail
rfc814 names and ports - general concepts behind well-known
ports
rfc793 TCP
rfc792 ICMP
rfc791 IP
rfc768 UDP
rip.doc details of the most commonly-used routing protocol
ien-116 old name server (still needed by several kinds of
system)
ien-48 the Catenet model, general description of the
26
philosophy behind TCP/IP
The following documents are somewhat more specialized.
rfc813 window and acknowledgement strategies in TCP
rfc815 datagram reassembly techniques
rfc816 fault isolation and resolution techniques
rfc817 modularity and efficiency in implementation
rfc879 the maximum segment size option in TCP
rfc896 congestion control
rfc827,888,904,975,985
EGP and related issues
To those of you who may be reading this document remotely instead of
at Rutgers: The most important RFC's have been collected into a
three-volume set, the DDN Protocol Handbook. It is available from the
DDN Network Information Center, SRI International, 333 Ravenswood
Avenue, Menlo Park, California 94025 (telephone: 800-235-3155). You
should be able to get them via anonymous FTP from sri-nic.arpa. File
names are:
RFC's:
rfc:rfc-index.txt
rfc:rfcxxx.txt
IEN's:
ien:ien-index.txt
ien:ien-xxx.txt
rip.doc is available by anonymous FTP from topaz.rutgers.edu, as
/pub/tcp-ip-docs/rip.doc.
Sites with access to UUCP but not FTP may be able to retreive them via
UUCP from UUCP host rutgers. The file names would be
RFC's:
/topaz/pub/pub/tcp-ip-docs/rfc-index.txt
/topaz/pub/pub/tcp-ip-docs/rfcxxx.txt
IEN's:
/topaz/pub/pub/tcp-ip-docs/ien-index.txt
/topaz/pub/pub/tcp-ip-docs/ien-xxx.txt
/topaz/pub/pub/tcp-ip-docs/rip.doc
Note that SRI-NIC has the entire set of RFC's and IEN's, but rutgers
and topaz have only those specifically mentioned above.
27