PART 5: LAN Interface Cards

Definition
The network interface card (NIC) is the piece of hardware that
fits inside your computer to provide the physical connection to
the network. Every computer attached to a LAN uses one sort of
network interface card or another. In most cases, the card fits
directly into the expansion bus of the computer. In some cases,
the card will be part of a separate unit to which the computer
attaches through a serial or parallel connection.

The interface card takes data from your PC, puts it into the
appropriate format and sends it over the cable to another LAN
interface card. This card receives the data, puts it into a form
the PC understands and sends it to the PC. Simple, right?

Wrong. To get one byte of data from here to there hundreds of
things must happen. Buffers must be checked. Requests must be
acknowledged. Sessions must be established. Tokens must be sent.
Collisions must be detected. The list can seem endless.

Luckily, the work of the interface card can be broken down into
eight tasks: host-card communications; buffering; packet
formation; parallel-serial conversion; encoding/decoding; cable
access; handshaking and transmission/reception. These are the
steps taken to get data from the memory of one computer to the
memory of another computer.

Preparing for Transmission
PC-NIC Communications. There are three ways to move data from the
PC's memory to the network card and back again: DMA, I/O mapping
and shared memory.

Shared memory is just that. Part of host memory is shared by the
network interface card's processor. This is a very fast method of
transfer, since no buffering on the card is required. Both the
card and the PC do their work on the data in the same place, so
no transfer is necessary.

DMA is short for Direct Memory Access. All Intel-based computers
come with something called a DMA controller. It takes care of the
transfer of data from an input/output device to the PC's main
memory so the PC's main microprocessor, or CPU, does not have to.
For DMA transfer, the controller or processor on the interface
card sends a signal to the CPU indicating it needs DMA. The CPU
then relinquishes control of the PC bus to the DMA controller.
(The bus is the piece of the computer that connects other parts,
e.g. memory and processor).

Once the DMA controller has command of the bus, it begins to take
the data from the card and place it directly in memory. It can do
this because it has been informed by the CPU of the appropriate
memory address at which to begin putting data in memory. After
all the data is in memory, the DMA controller returns control of
the bus to the CPU and tells it how much data has been put in
memory. Of course, the whole process takes fractions of a second.

There are two types of I/O transfers depending on the PC and
peripheral. The two most important are memory-mapped I/O and
program I/O. In a memory-mapped I/O transfer, the host CPU
assigns some of its memory space to the I/O device, in this case
the network interface card. So out of the possible 640K bytes
available, some amount, say 12K bytes, are given to the network
card. This memory is then treated as if it were main memory. No
special instructions in the CPU are needed to get data from the
card since it is like taking data from one part of main memory to
another.

With program I/O, the CPU is given a set of special instructions
to handle the input/output functions. These instructions can be
built into the chip or come with software. To send data, a
request is sent from the network interface card to the CPU. The
CPU then moves the data from the card over the bus to main
memory.

Shared memory is the fastest method of moving data between the
network interface card and the PC, but it not often used for
expense and execution reasons. The advantage of DMA is that it
removes work from the CPU, so it can perform other functions
while data transfer is taking place. The disadvantage is the CPU
cannot access memory while the DMA controller is working. I/O
mapping doesn't remove work from the CPU and it also takes up
memory, but it can be faster than DMA.

Different NICs use different types of NIC-PC communications.
Yours will probably use DMA or program I/O. Experts disagree on
which is better. If you really need to tell the difference,
you'll have to experiment with your own applications.

Buffering. Most NICs use a buffer. The buffer is a storage place
that holds data as it is moving into and out of the NIC. The
purpose of the buffer is simple: to make up for inherent delays
in transmission. To do this, a buffer temporarily holds data,
either for transmission onto the network cable or for transfer
into the PC. While in the buffer, data may be acted on,
"packetized" or "depacketized," or it may simply sit while the
NIC handles other things.

A buffer is needed because some parts of data transfer are slower
than others. Data usually comes into the card faster than it can
be converted from serial or parallel, depacketized, read and
sent. This is true in both directions.

Some NICs do not have buffers. Instead they use PC RAM. This can
be less expensive, but usually slower. It also takes up precious
memory.

Packet formation. This is the most important job of the network
and it is almost always done by the NIC. Packets are the units of
transmission used on most LANs. Files and messages for
transmission are broken up into packets as they are sent. At the
other end the packets are put together to reform the original
file or message.

A packet has three sections: header, data and trailer. The header
includes an alert to signal "packet on the way," the packet's
source address, destination address and clock information to
synchronize transmission. In some networks, headers also have
preambles used for various purposes, like setting up parameters
for transmission. They can also have a control field to direct
the packet through the network, a byte count and a message type
field.

The data section is just that, the data being sent, for example,
the numbers in a spreadsheet or words in a document. On some
networks, the data section of a packet is as large as 12K bytes.
On Ethernet, it is 4K bytes. Most networks fall between 1K and 4K
bytes.

The trailer has an error checking part called a cyclical
redundancy check (CRC). It is a number which is the result of a
mathematical calculation done on the packet by the sending NIC.
When the packet arrives at its destination, the same mathematical
calculation is repeated. If the result is the same, no errors
occurred in transmission -- all the ones and zeros were in the
right place. If the numbers don't match, an error occurred and
the packet is sent again. The trailer, like the header, can hold
other information.

TRANSMISSION
Parallel-serial conversion. Data comes from the PC in parallel
form, eight bits at a time. It must travel over the cable in
serial form, one bit at a time. And vice versa. Thus, the network
interface card must perform the conversion between the two forms.
Usually this is done by a controller on the card.

Encoding/decoding.
Once a packet is formed and changed from parallel to serial, it
is ready for sending over the line. To do this it must be
encoded; it must be converted into a series of electrical pulses
that convey information.

Most network interface cards use Manchester encoding. Serial data
is divided into things called bit periods. Each of these periods
(actual fractions of seconds) is divided in half. The two halves
together represent a bit. From the first half to the second half
of each bit period there is a change in the polarity of the
signal, from positive to negative, or vice versa. There must be a
change during each bit period because the change represents the
data. A change from negative to positive represents a 1. A change
from positive to negative represents a 0. Or vice versa depending
on the network.

Of course, it is these 1's and 0's that represent data. That's
how digital data is actually sent using electrical impulses.

Cable access. 
Before data can be sent, however, the network interface card must
have access to the cable. Not all cards can send at once. If they
do, their transmissions will collide and be lost.

The various access methods used were discussed last month. Token
Ring and Arcnet LANs use an electronic token to grant network
access. Ethernet lets any workstation transmit at will and then
look for collisions to see if it must transmit again. The entire
protocol for the access method, all the circuitry and firmware
(software written into hardware), resides on the network access
card. This is its main job, getting data onto the cable in
reliable form.

Handshaking. 
After getting data from the PC, formatting it, encoding it and
getting cable access, the network interface card has just one
more task before it can send data: handshaking. In order to send
data successfully, there must be a second NIC card waiting to
receive it. To make sure, there is a short period of
communication between two cards before data is sent. During this
period, the parameters for the upcoming communication are decided
upon through negotiation.

During negotiation, the transmitting card sends the parameters it
wants to use. The receiving card answers with its parameters. The
card with the slower, smaller, less complicated parameters always
wins. That's because more sophisticated cards can "lower"
themselves while less sophisticated cards can't "raise"
themselves.

Negotiation sets the maximum packet size to be sent, how many
packets before an answer, timer values, acknowledge time outs
(how long to wait for an answer), buffer sizes, etc.

Transmission/reception. 
Finally, everything is set. The only thing left is for the
transceiver to put the data on the cable. The transceiver gives
the data power to make it down the line. It actually puts the
electrical signal out over the cable, making sure the data can
get to the next card, the next repeater, amplifier or bridge.

At the other end, a transceiver is waiting to accept the signal
and begin the whole process in reverse, from modulated signal
through decoding, serial- parallel conversion and depacketizing
to PC-readable data.

PICK A CARD
Most people look only at performance when they choose a network
interface card. Some will also consider the access method and
topology used. This is fine when choosing the type of network
interface card. Ethernet tends to be better for bursty networks.
Its topology makes sense in scientific environments. Token Ring
and Arcnet do better with constant traffic. Their topologies are
better in offices. But there is more to choosing a network
interface card than performance and access method.

The most important consideration in buying a network interface is
reliability.  It doesn't matter how fast your network card is if
it doesn't work, causes errors, loses packets, drops the line,
etc. There is nothing more frustrating than having to isolate
network hardware problems. Moreover, once found, replacing a
network interface card means opening up a PC, setting dip
switches and possibly reconfiguring software. Look for a network
interface card that will work forever. This means talking to
users and installers.

Other considerations when buying a network interface card center
on your application. Figure out what kind of traffic your network
will be sending. Different topologies, different access methods
and different cabling schemes will guide your choice of network
interface card. So will the network operating system you plan on
using. But we'll get to that hairy subject in next month's
column.

-- by Aaron Brenner