Ethernet, Fast Ethernet, and
Gigabit Ethernet
In this presentation, we will take a closer look
at Ethernet, Fast Ethernet, and Gigabit Ethernets.
Ethernets
Collision Domains Should Be Small
The packets sent by the hosts on the same collision
domain may collide with each other.
2 Pairs of Wires in a UTP Cable Are
Always Used
2 pairs
No matter whether the device is a hub or a switch, 2 pairs of wires
in a UTP cable are always used to connect a host to the device.
Ethernet
MAC flow
in HalfDuplex
Mode
Switching Between Full and HalfDuplex Modes Is Simple
No matter whether full or half-duplex mode is used,
2 pairs of wires in a UTP cable are always used to
connect a host to a hub/switch port.
Full-Duplex Ethernet Is No Longer
Ethernet
• Ethernet is characterized by its CSMA/CD
MAC protocol when operating in halfduplex mode.
• In full-duplex mode, because CSMA/CD is
disabled and not used, actually full-duplex
Ethernet is no longer Ethernet.
– This is especially true for Gigabit Ethernet
because Gigabit Ethernet mostly operate in fullduplex mode.
Transmitted Bit Order
Little Endian: The least significant bit in a byte is
transmitted first.
Transmitted Byte Order
The bytes of a multi-byte field are sent from first
byte to last byte, with each byte sent in Little Endian
bit order.
Ethernet Address Format
Every vendor (e.g., 3COM) is assigned a vendor
block code. Therefore, every globally administered
address is globally unique.
IEEE 802.3 Ethernet Frame Format
If length is less
than or equal to
1500, it
represents the
length of data.
Otherwise, it
represents the
higher-level
protocol type.
Fast Ethernet
Encoding Method
Flow Controls Needed When
Ethernets Switches Are Used
• When a hub is used to connect all sending hosts
(half-duplex mode), Ethernet’s CSMA/CD MAC
algorithm is an effective flow/congestion control
(backoff) and retransmission (up to 16 times)
method.
• However, when a switch is used and full-duplex
mode is used, each host now has its own collision
domain and CSMA/CD no longer works. Frames
now may be lost due to buffer overflow inside the
switch.
– Hosts no longer know this type of “collision.”
Congestion control cannot automatically be performed.
– Also, retransmission cannot automatically be performed.
Backpressure Can Be Used In HalfDuplex Mode
• If a switch uses half-duplex mode to connect to
hosts, although each host has its own collision
domain, we can use some methods to ask sending
hosts to slow down their sending rates:
– Force collisions with incoming frames
– Make it appear as if the channel is busy
• If a switch uses full-duplex mode, the above two
methods won’t work. We need an explicit flow
control for switched full-duplex Ethernet switches.
IEEE 802.3x Flow Control
• Introduced for Fast and Gigabit Ethernets.
• Introduce special MAC-layer control packets
(PAUSE packets) to enable or disable frame
transmission.
• Thus a more generic MAC framework is defined
and an entity called (MAC Control) is introduced.
• MAC Control is responsible for generating,
sending, receiving, and performing the PAUSE
operation.
IEEE 802.3x Architecture
MAC Control Frame Format
PAUSE Function
• Implement a simple “stop-start” flow control
scheme.
• If a device wants to temporarily inhibit incoming
frames, it sends a PAUSE frame to the full-duplex
partner. This PAUSE frame contains a parameter
indicating the length of time the partner should
wait before sending more frames.
• If the device wants to cancel the timer at its
partner which is set up by its previous PAUSE
frame, the device can send another PAUSE frame
that contains a parameter of zero time.
PAUSE Frame Implementation
PAUSE frames have
higher priority over
normal data frames.
Must Stop Transmission in 512
Bit Time
• After receiving a
PAUSE frame, the
receiver has 512 bit
time to decode it.
• No more new
transmission is
allowed after this 512
bit time period.
Buffer Thresholds to Control When
to Issue PAUSE Frames
• The buffer space
(above the high water
mark and below the
low water mark)
should be larger than
link RTT * link BW.
• Otherwise, frames
may be dropped or
link utilization may
not be 100%.
Input-Buffered Switches Assumed
for IEEE 802.3x
• IEEE 802.3x is a link-level
flow control scheme.
• It uses the buffer
occupancy level of an
input port to determine
whether to inhibit frames
from incoming.
• Unfortunately, this scheme
does not work for outputbuffered switch. (Why?
because there will be no
queue in input ports.)
IEEE 802.3x May Cause Low Link
Utilization
• IEEE 802.3x is a link-level flow control applied to
input-buffered switches.
• If the input-buffered switch has the HOL blocking
problem, then using IEEE 802.3x will not further
lower the link utilization. (It is already low .)
• However, if the input port uses virtual output
queuing (one queue for each output port) to avoid
the HOL blocking problem, using IEEE 802.3x
may further lower the link utilization.
Low Link Utilization Example
Output port 1 is busy or blocked all the time.
Switch A
Switch B
P1 P2
P2
P2
P1, P1,
P1, P1
PAUSE
Output port 2 is idle.
After receiving the PAUSE frame, switch A cannot send any
more frame to switch B. However, sending P2 should be
allowed because P2 can be immediately forwarded when it
arrives at switch B.
Auto-Negotiation Eases Configuration
• Starting from Fast Ethernet
• For UTP cables
• Auto negotiation about duplex mode (half or full),
maximum speed (10 or 100 Mbps), and flow
control support.
• Avoid configuration headache.
– The same RJ45 jack can support 10 or 100Mbps and
half or full duplex mode, how do you know which
setting you should use when you plug your RJ45 into a
hub or a switch’s port?
– So, the best setting that you should choose is autonegotiation.
Auto-Negotiation Message Format
Auto-Negotiation Message Transfer
Because link speed is one of the parameter that needs
to negotiate, the auto-negotiation message cannot be
transferred at a chosen speed (e.g., 10 or 100 Mbps).
Instead, it is transmitted at a predetermined clock
rate.
Gigabit Ethernet (IEEE 802.3z)
Gigabit Ethernet Architecture
Encoding Method
1000BASE-X Physical Media
Shortwave Laser
Longwave laser
Fibers
Ethernet Technology Migration
Ethernet Technology Migration
Minimum Ethernet Frame Length
Needs to be Increased
• When link speed increases from 100 Mbps to
1000 Mbps, for half-duplex CSMA/CD to work
well (to be able to detect collision), we can:
– Reduce the network size by a factor of 10 to only 20
meter
• Not acceptable
– Or increase the minimum Ethernet frame length by
a factor of 10
• Network efficiency will decrease.
• Higher layer protocols needs to be modified.
Carrier Extension
• Gigabit Ethernet does not want to modify
application programs.
– We want application programs to still use 64 bytes as
the minimum frame size. Ideally, they should not need
to know whether they are using Ethernet, Fast Ethernet,
or Gigabit Ethernet.
• However, to be able to detect collision, the
minimum frame size must be enlarged from 64
bytes to 512 bytes (4096 bits).
• Carrier extension is introduced to insert padding at
the MAC layer so that every transmitted frame is
at least 512 bytes.
Carrier Extension
Frame Bursting
• Although carrier extension solves the
collision detection problem, the link
utilization can be very low.
– If every transmitted frame is a minimum-sized
frame, because in every transmitted 512 bytes,
there are only 64 useful bytes. The effective
link utilization is only 64/512 = 12.5%
• Frame bursting is introduced to improve
link utilization.
Frame Bursting
• If the source host has many small frames to send, it can send
them in a burst after a MAC arbitration.
• Except for the first frame, other frames sent in a burst need not
be carrier-extended.
• If the first frame is larger than 512 bytes, it need not be carrierextended.
• The burst can be as long as 8192 bytes to improve throughput.
However, the livelock problem is not solved!
Why Doing Carrier Extension for the
First Frame in a Burst?
• First, if there is no carrier extension, when collision
happens, the sending host may have sent out many small
frames. If the sending host resends all frames that have
been sent, two problems may occur:
– They may already be removed from the buffer.
– The receiver may receive duplicate frames.
• Second, if there is carrier extension and collision is
detected in this period, the sending host knows that only
the first frame should be retransmitted.
• Third, if there is carrier extension and there is no collision
detected in this period, the sending host can be assured
that no collision will happen in the future.
Gigabit Ethernet Mostly Operates in
Full-Duplex Mode
• Although half-duplex (hub) Gigabit Ethernet’s
MAC performance is academically interesting,
commercially they are rarely used.
– Under 1000 Mbps high speed, CSMA/CD’s performance
is bad when the number of participating hosts increases.
– Why not using full-duplex mode (switch) to get much
more bandwidth at slightly increased cost?
• Nowadays, almost all Gigabit products are switches.
– Since CSMA/CD is totally disabled, Gigabit Ethernet
actually no longer is Ethernet.
– The sending host just pumps its frames into the cable as
fast as it can.
Physical Layer Architecture
Various Line Coding Methods
Gigabit
Ethernet
uses NRZ.
Manchester
consumes
too much
Bandwidth.
Gigabit
Ethernet
cannot afford
to use it.
1000BASE-X Uses 8B/10B Block Code
• Unlike Manchester encoding, whose baud rate is
200% of the data rate, in 8B/10B, the baud rate is
only 125% of the data rate.
– Supporting 1000 Mbps is already difficult for some
media such as UTP. Asking UTP to support 2000 Mbps
baud rate is too much.
• In 8B/10B, an 8-bit byte is encoded into a 10 bit
code word.
– Provide sufficient signal transitions for clock recovery
at the receiver
– Allow easily detecting single or multi-bit errors
– The number of one and zero bits transmitted are almost
the same (DC balance).
8B/10B Block Coding
• No more 4 consecutive ones or zeros will be generated in a
10-bit code.
• The imbalance between the number of ones and zeros in a
10-bit code is at most 1.
1000BASE-X Physical Layer
1000BASE-X Connectors
• One jack is for transmission. The other jack is for reception.
• To connect two hosts together, you need to ‘cross-over’ the
transmission and reception jacks by yourself.
1000BASE-T over UTP
• Product is already available, but not too long ago.
(introduced to Taiwan just in March, 2001)
• Use four (all) pairs of wires in an UTP cable to together
support 1000 Mbps.
• The quality level of UTP cables can be CAT 5.
• The cost of a 1000BASE-T network adapter is much
cheaper than that of a 1000BASE-X network adapter.
(7,400 vs 25,000 NT dollars)
• We will purchase and test 1000BASE-T NIC soon (for
the “Design and implementation of high-speed network
systems” course.)