Computer Communication Networks

Fundamentals Circuit-Switching Packet-Switching Virtual-Circuit-Switching Excursion Bibliography

Data Communications and Networking

COSC 264Introduction to Communication Networks

Dr. Andreas Willig1

Dr. Muhammad Asad Arfeen2

1Dept. of Computer Science and Software EngineeringUniversity of Canterbury, Christchurch

2Dept. of Computer and Information Systems EngineeringNED University of Engineering & Technology, Karachi

UoC, 2014


Outline

Fundamentals

Circuit-Switching

Packet-Switching

Virtual-Circuit-Switching

Excursion


About this Module

• Goals:

• Understand the reasons for creating networks• Know major design issues• Understand basic operation modes of networks

• This module is based on [9, Chap. 10, 11] and [6]


Outline

Fundamentals

Circuit-Switching

Packet-Switching


Excursion


Fundamentals of Networks

• So far we have mostly studied communication between two

stations, using a single channel

• How can we communicate in a larger population of users /

stations / terminals? Say: all persons in a country?

• Fundamental aim of networks:

• Provide good “reachability” and “service quality” to users. . .• . . . but do so at reasonable costs for the users . . .• . . . and giving revenue to network providers

• Networks are getting more useful to users as the number

of users increases

Key question

How to build and operate large-scale networks?


A First Look on Network Design

• We first consider a very simple model of a network: as asimple communication graph:

• Nodes represent stations / switching elements . . .• Edges represent direct communication links

• Design goal: given a number N of stations, create links so

as to make the graph connected!

QuestionWhat are desirable additional objectives for such a graph and

what are important constraints?


First Option: Fully Meshed Network

• Each of N stations has separate

link/cable to each other station

• Total number of links:N(N−1)

2

• Advantage: cutting one link does

not disable communication with

other partners

• Disadvantage: costs!!

• Disadvantage: poor scalability!!


Second Option: A Minimum Spanning Tree

• Total number of links: N − 1

• Advantage: minimal # of links to

enable communication between all

nodes

• Disadvantage: one link failure can

partition network

• Disadvantage: stations see

signals for communications they

are not involved in – security

issue!

• Disadvantage: stations need the

ability to “forward” foreign

communications – additional

complexity


Third Option: A Star Network

• Total number of links: N

• One additional network element:

a switching element

• Advantage: one link failure does

not partition remaining network

• Disadvantage: switching element

is now single point of failure


Fourth Option: A Partially Meshed Network

• Total number of links: between N

andN(N−1)

2

• One link failure does not partition

network

• One switching element might fail

without compromising

connectivity (much)


A First Conclusion

ConclusionBy carefully designing partially meshed networks we can

achieve a good balance between costs (# of links, # of

switching elements) and network properties like resilience

against link failures or switching element failures


A More Realistic Model

• Links are labeled with available

capacities (e.g. measured in

Mbit/s)

• Links could also be labeled with

other quantities, e.g. delays,

transmission costs, . . . , or any

combination of these

• End nodes are labeled as A, B,

C, . . .

• Missing in this figure, but

nonetheless present is the

generated traffic demand


A More Realistic Model (2)

• Suppose that nodes E and C

want to communicate at a

sustained rate of 3 Mbit/s

• A possible route (which has to

be identified!!) is shown


A More Realistic Model (3)

• Now suppose that nodes D and

B also want to communicate at a

sustained rate of 3 MBit/s

• Again, a suitable route has to be

identified

• Why this (the green) one?

• What is implicitly assumed here?

• How would route setup

considerations change if the

communication flows are “bursty”

instead of sustained?

• Do you know of an example of

“bursty” traffic?


Important Tasks in Networking

• A primary task in any network is to identify suitable routes

• Subject to constraints on capacities, allowable delay, . . .• Respectful to existing traffic flows• Remember setting up the route from D to B• This is called routing

• A second task is to avoid traffic overload situations inspecific network areas, or at least to react properly to them

• This is called congestion control• The switching element close to B and C will likely not

accept much more traffic• Do you know of a real-life example of congestion control?


Important Tasks in Networking (2)

• A third important task is to avoid overflowing a receiverwith data from the sender

• This is called flow control

• A fourth important task is to deal with errors:

• Transmitted data can be garbled or lost =⇒ error control• Stations, switching elements or links fail =⇒ resiliency


Further Important Aspects of Networking

• Network management (short-to-medium term)

• React to link / station failures• Monitor network usage and performance to identify

bottlenecks, problems, . . .• Collect information necessary for billing and accounting

• Network design and strategy (long term)• Decide on deployment of new links / switching elements

• Often as reaction to changing user communities / usage

patterns, or new applications

• Decide on additional traffic control mechanisms

• Example: blocking P2P traffic

• Decide on billing / pricing schemes


Some Common Communication Patterns

• Unicast

• Broadcast

• Multicast


Communication Patterns – Unicast

• Only two nodes in the network involved

• One is the transmitter, the other the receiver, but nodes

can also have both roles

• Examples: phone connections, viewing a web page


Communication Patterns – Broadcast

• One node as sender, all other as receivers

• Examples: Radio, TV, . . .


Communication Patterns – Multicast

• One node as sender, several, but not all others as

receivers

• Often, in multicast groups all nodes can act as sender

• Example: Internet chat, phone conferences


Traffic: A Key Network Design Factor

• Any network is expected to carry a certain class of traffic

• Examples:

• The POTS (plain old telephone system) carries voice traffic• The Internet carries WWW, P2P, interactive video, . . .

Important point

The design of a network is strongly influenced by the traffic it is

supposed to carry!


Outline

Fundamentals

Circuit-Switching

Packet-Switching


Excursion


Voice/CBR Traffic

• In modern telephone systems voice is transmitted digitally,the analog voice signal is A/D-converted with fixedsampling rate and resolution

• ISDN: 8 kHz sampling rate, 8 bit resolution =⇒ 64 kbit/s• Cellular phone systems add voice compression, e.g. the

voice coders in GSM generate between 4.75 and 12 kbit/s

• The voice data is generated at a fixed rate

• This is not always true, some systems transmit nothing

when speaker is not active, we then have on-off behaviour

• This is called a continuous bit rate (CBR) data stream

• Other examples of CBR data: CBR video, periodic sensor

measurements, . . .


Voice/CBR Traffic (2)

• This fixed rate must be supported by the network,

otherwise users get dissatisfied with voice quality

• Network must support bounded round-trip delay, otherwiseusers get annoyed (boundary is at ≈ 250 ms)

• See [5] for information on “acceptable” voice transmission

• Phone calls last about three minutes on average


Circuit Switching

• A good way to support voice / CBR traffic is to set up adedicated connection or circuit between end points for theduration of the connection

• Think of this as a “private cable”

• The lifetime of a connection encompasses three phases:

• Connection setup: identify route, set aside resources

(buffers, processing capacity, bandwidth) in switching

elements and links, so that resources are guaranteed• Connection usage: use the established connection to

transmit CBR data – the pre-reserved resources guarantee

that this connection is not influenced by other connections• Connection teardown: free the reserved resources• How do you trigger these steps in the POTS?

• Switching elements in CS networks are called switches


Circuit Switching – Important Properties

• A routing decision is made only once (at connection setup)

and never/rarely modified

• A connection has its resources guaranteed

• Any bandwidth not used by a connection cannot be

re-used by other connections, depending on data rate

dynamics this can result in poor utilization

• Connection setup takes time, it does not pay off when only

very little data needs to be transmitted

• Connection setup may fail when no route or not enough

resources are available in the network

• Most complex step is connection setup, connection usage

is a low-complexity task for switches


Circuit Switching – Admission Control

• An important part of connection setup is that switching

elements check whether enough resources are available

for the new connection without compromising the

resources already granted to existing connections

• The outcome can be positive or negative, result is

signalled to user wishing to set up connection

• This procedure is called admission control


A Closer Look at Switches

(compare [9, Fig. 10.7])

• Switches are conceptually

separated into two logical units

• Switch controller performs

complex processing during

connection setup (e.g. routing,

signaling, resource reservation,

switching fabric configuration)

• Switching fabric is used by

established connections, it

forwards data from inputs to

outputs


Switches – Signaling

• Signaling: control information exchanged to set up or tear

down a circuit

• In the POTS certain signaling events between your phone

and the next office / switch (e.g. off-hook, dialing) use the

same channel as the speech data would use

• Nowadays, inter-switch signaling often uses another

channel / network than the speech data

DefinitionIn in-band signaling user-data and signaling data share the

same communications channel, in out-of-band signaling

user-data and signaling data are transmitted on different

channels / networks.


Space Switching Fabrics


• Fabric has N inputs, N outputs

• Signal pathes are physically

separated from each other

(divided in space)

• For each of the N2 possible

connections a small switch exists

at a crosspoint, which is closed

when a connection is

established, otherwise it is open

• For each input line at most one

crosspoint can be closed

• Problem: space switches for

large N are costly!!!

• Problem: crosspoint utilization is

1/N at most


Space Switching Fabrics (2)


• Fabric has N inputs, N outputs,

but is composed of several

interconnected smaller fabrics

• This kind of design is called a

three-stage space switch

• The number of crosspoints is

much smaller than N2

• During connection setup a free

path through switching fabrics

has to be found


Space Switching Fabrics (3)

DefinitionA switch is non-blocking when at all times a free output port

can be reached from any input port. A switch is blocking when

it does not have this property.

• The (full) space switch is non-blocking

• The three-stage space switch is blocking

• Find an example!!• What would happen if we remove the middle stage?


Switches – Further Remarks

• Design of three-stage space switches must fulfillconstraints on blocking probability

• Depends on assumptions about the call arrival process• Much literature available (e.g. [2])

• There are also other types of switches, e.g. Time-Division

Switches: switching fabric uses high-speed link with time

subdivided into time-slots, one connection receives regular

time slots


Outline

Fundamentals

Circuit-Switching

Packet-Switching


Excursion


Data Traffic

• Many data applications naturally have time-varying rates

• Called Variable-Bit-Rate (VBR) or bursty traffic

• WWW is an example: users alternate between clicking and

thinking/reading, a significant fraction of time no data is

transferred at all


Data Traffic over Circuit-Switched Networks

• CS-networks are not well suited to VBR traffic:

• Reserved rate is not well utilized during traffic pauses• Reserved rate might be too small during traffic peaks• No re-use of underutilized connections by other

connections

ConclusionA more flexible networking mechanism is needed for data

traffic!!!


Packet Switching

• Data flows are segmented into packets

• Packets are basic unit of transmission, not connections

• A packet consists of:

• A packet header containing meta-information about the

packet, e.g. address fields (see below)• The packet payload• Possibly a packet trailer for error detection / correction

• Packets are transmitted individually and independently

• There is no notion of a connection, packets can be sent

immediately without having to set up any state / resource

reservation in the network

• Analogy: letter transfer in postal network, envelopes

correspond to packet headers

• The Internet is a packet-switched network!


Packet Switching – Consequences

• Let us call a sequence of packets between the same

source-destination-pair a flow

• Each packet is routed individually, different packets in thesame flow can take different routes

• Each switching element (called router in packet-switched

networks) makes a routing decision for each packet• Each packet must include information facilitating routing,

e.g. header fields for source and destination address• Routers do not attempt to identify flows, nor do they store

any per-flow state• Packets do not necessarily arrive in the same order as they

have been sent (packet reordering) [3]

• Many flows can share a link, bandwidth not utilized by one

flow can be used by others

=⇒ This is called statistical multiplexing


Packet Switching – Consequences (2)

• Lack of resource reservation means: there are no(t many)guarantees for packet delivery

• Internet/IP “best effort” service: packet is delivered – maybe• IPs lack of guarantees is compensated (in parts) by TCP

Important Point

Routers in packet-switched networks perform more complex

processing during information transfer than switching fabrics in

circuit-switched networks!


Packet Switching – Congestion

• Since flow data rates and routes often cannot be predicted

in advance, routers buffer some packets to prevent packet

dropping in temporary overload situations

• Routers only have a finite amount of memory, and when

overload situation sustains, packet dropping is inevitable,

this is called congestion

• Important question: which packets to drop?

• Congestion control schemes either try to avoidcongestion or to deal with it in a graceful manner, e.g. by:

• Subject packet flows to admission control• Sending signals to traffic sources to reduce data rates• Make good decisions about which packets to drop (e.g.

based on priorities) – applications must help here!!• Modify the pricing during congestion


Packet Switching – Choice of Packet Size

• Packet overheads (header, trailer) have a fixed size

• Payload size is variable (within bounds)

• Tradeoff:

• Small payload size leads to high overhead• Small payload size leads to reduced susceptibility to errors

• Packet size limits can be technology- or application-driven

• Example: too long packets might block important alarm

messages for an unacceptably long time (blocking times)


A Closer Look at Routers



A Closer Look at Routers (2)

• For each new packet router must make a routing decision

• Generally, routing decision can be based on many factors:

• Source address, destination address• Packet priorities, type of data carried (video, web, . . . )• Link states of various candidate outputs• Queue lengths at various candidate outputs• Reachability of destination at various candidate outputs

• But: this is too complex for high-performance routers

processing millions of packets per second


A Closer Look at Routers (3)

• Traditional Internet approach:

• Routing is based on destination addresses only• Routers consult a forwarding table, giving for each

destination address (range) the identity of the outgoing link

• Forwarding table lookup can be made fast [10]

• Forwarding table is populated as a result of a routingprotocol, which is executed by the routers control unit

collaboratively with other routers• Routing protocol operates at relatively large timescales

(seconds, minutes), so the forwarding table is stable for

many packets in succession

• This distinction between routing and forwarding is an

example of separation between mechanism and strategy


Outline

Fundamentals

Circuit-Switching

Packet-Switching


Excursion


State of Affairs

• Circuit-switching:

• Can give guaranteed bandwidths• No statistical multiplexing, no re-use of resources• Data forwarding is low-complexity operation for switches• Routing is done only once

• Packet-switching:

• Cannot (easily) give any guarantees• Allows statistical multiplexing• Data forwarding is higher-complexity operation for routers• Routing is done for every packet

• Can we marry these approaches and have their benefits

without having their problems?


State of Affairs (2)

No!!


State of Affairs (3)

. . . at least not fully . . .



• Virtual circuit switching is a combination of packet- and

circuit-switching

• Major characteristics:

• It is connection-oriented, i.e. a connection needs to be

established before data transfer can commence• Information in a connection is transmitted in packets• A connection can accommodate VBR traffic• Statistical multiplexing capabilities are offered!• All packets in a connection follow the same path, packet

processing in switches is simpler than for routers• Often there are no (strict) bandwidth guarantees for a

connection

• VCS technologies: ATM [9, Chap. 11], [7], [4], (G)MPLS

(specified amongst others in RFCs 3031, 3032)


Simplified Timeline for Circuit-Switching


Simplified Timeline for Packet-Switching


Simplified Timeline for Virtual-Circuit-Switching


VC Setup and Usage

• The approach to set up and use virtual circuits is common

in many VCS technologies

• We will explain here a caricature of the label setup process

in MPLS, a similar technique has been adopted in ATM

• Given: a source station S, a destination D and a set of

intermediate switches s1, s2, . . . , sn along the route

between S and D

• A label is a link-local unique connection identifier of fixed

width (e.g. 20 bit in MPLS shim header)


Idealized VC Connection Setup

• The following procedure establishes a path and a reverse

path between S and D

• Station S starts the process, allocates an un-used label l0for link S ↔ s1 and includes this in a setup-request packet

• Switch s1 now:

• determines its incoming link for this request, here: i1• determines the next switch towards D (here: s2) and

outgoing link to it, here: o1

• allocates an un-used label l1 for the link o1

• re-writes the label l0 in the setup-request by l1• stores the quadruple (l0, i1, o1, l1) in a forwarding table

• Switches s2, . . . , sn behave similarly


Idealized VC Connection Setup (2)

• Destination receives setup-request with label ln and:

• stores it as pertaining to connection between S and D• creates a setup-response, sends it to switch sn with label ln

• Switch sn behaves as follows:

• It determines the incoming link, which here is just on

• It consults the forwarding table with key (on, ln), finds the

“output port” in and label ln−1

• It re-writes the label ln in the setup-response packet to ln−1

and forwards the packet over link in back to switch sn−1

• Switches sn−1, sn−1, . . . .s1 behave similarly

• Upon arrival of the setup-response at S the connection is

established


VC Connection Usage

• Suppose S wants to transmit a packet to D

• S prepends the label l0 to the packet and sends it to s1

• Switch s1 now:

• determines the incoming port i1 for that packet• consults the forwarding table with key (i1, l0) and finds

output port o1 and output label l1• rewrites the label l0 in the packet with l1, and• forwards the packet via output o1.

• The other switches behave similarly

Key point

The switches perform a very simple operation, often much

simpler than what routers have to do!!!


ATM – Asynchronous Transfer Mode

• Developed in the 80’s and 90’s of the last century [7], [4]

• It is defined in ITU-T standards, e.g. I.150, I.432

• Nowadays mainly used in high-speed backbone networks

• Aim: integrated broadband network for voice- and data

• Major characteristics:

• Virtual-circuit-switched• Transmission based on fixed-size packets called cells• Different types of connections are offered (CBR real-time,

VBR real-time and non-real-time), some of them with

network-internal resource reservation

=⇒ Quality-of-Service support!• ATM performs admissission control


ATM Cell

• This is an ATM cell at boundary

between user terminal and first

switch – in-network cells have no

GFC, but a longer VPI

• GFC: control traffic flow rates

• PT: distinguishes between user-

and management cells, which

may or may not have

experienced congestion

• CLP = Cell Loss Priority (guides

which cells should be preferrably

discarded upon congestion)

• VPI and VCI together make up a

label similar to the MPLS label

• ATM switches can also operate

on the VPI field only

• No payload checksum??


ATM – Physical Layer

• Data rates in I.432: 622.08, 155.52, 51.84, 25.6 Mbps

• ATM is a point-to-point technology

• Standard I.432 specifies two transmission schemes

• Cell-based Physical Layer:

• An ATM interface emits cells continuously• Problem: receiver must identify cell boundaries

(synchronization)• Receiver “hunts” bit-per-bit for bit-subsequence where HEC

fits to previous 32 bits• Once a matching HEC is found, synchronization is achieved• Bit errors in header do not only invalidate cell, but could

also lead to loss of synchronization

• In an SDH-based physical layer ATM cells are embeddedinto SDH frames

• SDH/SONET is an optical transmission technology with a

TDMA structure [8]


Packet- or Cell-Switches

• Cell switches operate on fixed-size packets, called cells

(like in ATM)

• Packet switches / routers operate on variable-length

packets (like IP routers have to)

• But: many modern IP routers internally work as cell

switches by splitting IP packets into a sequence of cells

• All following explanations refer to cell-switching, but can be

equally applied to packet switching

• The following is partly based on [6, Sec. 1.2]


Cell-Switches

• We now discuss a few issues in

cell switching, considering an

example

• A cell switch has N input lines, N

output lines and a switching

matrix in between

• Switch operates in time-slotted

manner, slot-time is equal to

cell-transmission time

• Cell arrivals and cell departures

are synchronized to slot

boundaries


Cell-Switches (2)

• Our example is a 4 × 4 switch

(i.e. N = 4)

• Inputs and outputs are numbered

from 1 to 4, respectively

• Inputs and outputs have the

same rate

• Cells are marked e.g. as a3,

where a identifies the cell and 3

identifies output link

• Assumption: switch is

non-blocking


Cell-Switches – The Need for Queues

• In each time slot at most one cell can arrive per input link

• These cells can have destination conflicts: up to N arriving

cells can be destined for the same output during one slot

• Output serves only one cell, where to store excess cells?

• Two approaches:

• Input queueing: input links are equipped with FIFO

buffers, and head-of-queue cell is transferred to its output

only when it is the only cell for the given output, or if it has

been randomly selected among all cells for the same output• Output queueing: output links are equipped with FIFO

buffers, all input cells towards same output are transferred

in parallel


Cell Switches – Output Queueing

• Implementation of output queueing requires that buffer

memory of the output can accommodate up to N cells

during a slot time

• Limitations in memory speed create bounds for N

ConclusionOutput-queued switches have limited scalability!!


Cell Switches – Input Queueing

• In this example cell c4 (the only

cell for output 4!) has to wait until

cells a1 and b3 have been

transferred – but a1 suffers from

destination conflicts

• Called head-of-line (HOL)

blocking

• Requirements in memory access

speed are much more relaxed

than for output-queueing


Excursion: Throughput of Input-Queued Switches

• HOL blocking reduces throughput – by how much?

• Assumptions (from [6, Chap. 10]):• Input links are saturated, i.e. always have a new cell• Output links of each cell are selected randomly (uniform

distribution, independent of all other cells)

• TN ∈ [0, 1] denotes the throughput, i.e. average number of

cells leaving any output for an N × N switch

• Result:

limN→∞

TN = 2 −√

2 ≈ 0.586

• This indicates switch capacity, i.e. maximum cell arrival

rate for which average queue lengths are bounded

• Output-queued switches achieve TN = 1 for any technically

feasible N

• Question: can you think of an improvement of

input-queueing?


Outline

Fundamentals

Circuit-Switching

Packet-Switching


Excursion


An Unspoken Assumption

So far we have implicitly assumed that:

Assumption

Switching elements / routers do not modify the user data –

either the data is forwarded as it is or it is thrown away entirely!!

The relatively new field of network coding modifies this

assumption by letting routers compute new packets out of two

or more other packets [1], [11], [12]


Network Coding Example – Traditional Approach

• Given: two wireless stations A

and B and a router in between

• A and B cannot hear each other,

but both can hear the router (and

vice versa)

• A wants to transmit a message x

to B, and B wants to transmit

message y to A

• x and y have the same length

• Router simply forwards packets,

four packets required in total


Network Coding Example – Network Coding

• Instead of simply forwarding the

received packets, the router

computes a new packet as x

XOR y and broadcasts the

resulting packet to A and B

• A can compute y as x XOR (x

XOR y), i.e. by XOR’ing his own

packet with the one broadcast by

the router

• Similarly, B can compute x

• Only three packets required!


Rudolf Ahlswede, Ning Cai, Shuo-Yen Robert Li, and Raymond W. Yeung.Network information flow.IEEE Transactions on Information Theory, 46(4):1204–1216, July 2000.

John C. Bellamy.Digital Telephony.John Wiley and Sons, Chichester, UK, third edition, 2000.

Jon C. R. Bennett, Craig Partridge, and Nicholas Shectman.Packet Reordering is Not Pathological Network Behaviour.IEEE/ACM Transactions on Networking, 7(6):789–798, December 1999.

Martin DePrycker.Asynchronous Transfer Mode - The solution for Broadband-ISDN.Prentice Hall, New York , London, second edition, 1994.

Olivier Hersent, David Gurle, and Jean-Pierre Petit.IP Telephony – Packet-based multimedia communications systems.Addison-Wesley, Harlow / England, London, 2000.

Anurag Kumar, D. Manjunath, and Joy Kuri.Communication Networking – An Analytical Approach.Morgan Kaufmann Publishers, San Francisco, 2004.

Raif O. Onvural.Asynchronous Transfer Mode Networks - Performance Issues.Artech House, Boston, London, second edition, 1995.

Mike Sexton and Andy Reid.Broadband Networking – ATM, SDH and SONET.


Artech House, Boston, 1997.

William Stallings.Data and Computer Communications.Prentice Hall, Englewood Cliffs, New Jersey, fourth edition, 2006.

George Varghese.Network Algorithmics: An Interdisciplinary Approach to Designing FastNetworked Devices.Morgan Kaufmann Publishers, San Francisco, 2004.

Raymond W. Yeung.Information Theory and Network Coding.Springer, New York, 2008.

Raymond W. Yeung, Shuo-Yen Robert Li, Ning Cai, and Zhen Zhang.Network Coding Theory.now Publishers Inc., Boston / Delft, 2006.

Protocol Layering Elements of Service and Protocol Design Bibliography


COSC 264Network Architectures and Protocol Basics

Dr. Andreas Willig1




UoC, 2014


Outline

Protocol Layering

The Concept of Layering

The OSI Reference Model

The TCP/IP Reference Model

Elements of Service and Protocol Design

Service Primitives

A few Standard Protocol Mechanisms


About this Module

• We look at architectures for packet-switched networks

• Goals:

• Understand protocol layering and two reference models• Understand the concepts of services, protocols and the

relationship between both

• This module is based on [8, Chap. 2], [4]

• Further references: [3], [2], [9], [1], [5]


Outline

Protocol Layering



Networking Software

• The Internet and POTS are among the most complex

technical systems, they require vast amounts of software

• Structuring principles help to organize networkingsoftware to achieve:

• Modularity and software re-use• Independence of network technologies (Transparency)• Separation of concerns• Correctness

Layering

A key structuring principle for networking software is layering:

the functionality is decomposed into a chain of layers so that

layer N offers services (through an interface) to layer N + 1

and itself is only allowed to use services offered by layer N − 1.


Outline

Protocol Layering






Layering Concepts


Layering Concepts (2)

• A layer N offers an N-service interface

• The next higher layer N + 1 is only allowed to use the N-interface, but not any ofthe lower interfaces (e.g. the N − 1 interface) – this applies to all layers!

• The N-interface offers services at service access points (SAP)

• The N-interface can offer several SAPs, this allows to multiplex between differentlayer N + 1 protocols or different layer N + 1 “connections” or “sessions”

• Example: POTS offers voice and fax services, uses telephone sockets as SAPs


Layering Concepts (3)

• The layer N-service is implemented through an N-protocol

• The N-protocol makes direct use of N − 1 services

• The N-protocol makes no assumption whatsoever on what is on layer N + 1

• It exchanges protocol data units (PDUs) with a peer N-protocol entity – itconstructs these PDUs itself and hands them over to its local N − 1-layer todeliver them to peer N-protocol entity (which in turn receives it from its localN − 1 layer)


General Layout of a PDU

• The N-PDU is constructed by the N-protocol entity

• It carries the data handed over by layer N + 1 for transmission, also referred toas user data or N-SDU (service data unit)

• The sending N-protocol entity adds an N-protocol header which carries controlinformation (e.g. sequence numbers, addresses, flags) important for theN-protocol but not the receiving N + 1 layer

• It might furthermore add an N-protocol trailer (usually a checksum)

• The receiving N-protocol entity removes the N header and trailer and hands overthe N + 1 data to its local layer N + 1 entity


Layered PDU Processing

• An N-PDU is treated as payload / user data by the N − 1 layer

• Each layer adds own header and trailer before handing down to lower layer

• Receiving layer removes its header / trailer before handing payload to upper layer

• How would you support efficient PDU processing in operating systems?


About Interfaces

• Interfaces specify a service that a certain layer offers

• Example:

• The socket interface offers reliable, in-sequence and

byte-oriented data transfer through an interface somewhat

resembling a file system interface• The TCP/IP protocol stack implements this service• Applications just use the socket interface and are not

concerned with the operation of the TCP protocol

Important Point

Standardized interfaces allow higher layers to ignore the

operation and properties of lower layers

• Several (not only networking) standard documents tend to

specify interfaces, and not procedures / algorithms


Cross-Layer Approaches

• How many layers should a protocol stack have?

• Having many layers . . .

• . . . can lead to clean and modular software• . . . incurs more overhead (processing, headers)

• Existing protocol stack architectures have no more than

five to seven layers, some have less

• Recently, cross-layer designs have become popular:

• Non-neighbored layers are allowed to exchange information

and use their services• Done for performance reasons• But: can lead to unwanted interactions between layers [5]


Outline

Protocol Layering






The OSI Seven Layer Model

• OSI = Open Systems Interconnection

• Set of standards and protocols createdby ISO

• See [9]

• The model was not commerciallysuccessful, but helped greatly to clarifynetworking architectures anddiscussion


The OSI Seven Layer Model – A Second View

• Lowest three layers exchange PDUs between physically connected hosts

• Upper four layers exchange protocol messages between end hosts (perhapsover several intermediate nodes, called routers)

• This already hints at a network architecture where end nodes are interconnectedthrough routers!


OSI RM – Physical Layer

• Concerned with transmission of digital data (e.g. bits,

bytes) over a physical medium

• Often involves specification of:

• Cable types (wired) or frequencies / bandwidth (wireless)• Connectors• Electrical specifications• Modulation / demodulation and signal specification• Carrier- or bit synchronization methods


OSI RM – Link Layer

• Task: (reliable) transfer of messages over physical link

• Link layer messages are often called frames

• Often involves specification of:• Framing:

• delineation of frame start and end• choice of frame size

• Error control (e.g. coding or retransmission-based)

• Coding is also often regarded as a PHY functionality

• Medium access control

• distributes right to send on shared channel to several

participants• often considered as a separate “sub-layer” of link layer

• Flow control

• Avoid overwhelming a slow receiver with too much data


OSI RM – Network Layer

• Concerned with:

• Providing a link technology-independent abstraction of

entire network to higher layers• Addressing and routing• End-to-end delivery of messages

• Network- and higher-layer messages are called packets


• Addressing formats• Exchange of routing information and route computation• Depending on technology: establishment, maintenance and

teardown of connections


OSI RM – Transport Layer

• Concerned with:

• (reliable, in-sequence, transparent) end-to-end data transfer• programming abstractions (interface) to higher layers


• Error-control procedures (Question: why again?)• Flow control procedures• Congestion control procedures

• Protect network against overloading• Can also be considered a network-layer issue


OSI RM – Session and Representation Layer

• Session layer:

• Concerned with establishing communication sessions

between applications• A session can involve several transport layer connections in

parallel or sequentially• A session might control the way in which two partners

interact, for example enforce that partners speak

alternatingly

• Representation layer:

• Translates between different representations of data types

used on different end hosts• Example: host A uses low-endian ints, host B big-endian


OSI RM – Application Layer

• Contains application support functions or functionalities

used in many applications

• Examples:

• File transfer services• Directory services• Transaction processing support (e.g. two-phase commit)


Outline

Protocol Layering







• This model is used in the Internet

• The Internet follows the so-calledend-to-end principle: Layers 3 andbelow are kept simple, mostcomplexity resides in transport layer


The TCP/IP Reference Model – A Second View

• This reference model also uses a network architecture where end nodes (calledhosts) are interconnected through routers!


The Application Layer

• Consists of applications using services of transport layer

• Accesses transport layer through socket interface

• There are well-known application-layer protocols, e.g.:

• SMTP (email)• HTTP (web)• FTP (file transfer)• RTP (real-time video and audio)


The Transport Layer

• Provides end-to-end communications to applications

• Offers its services through socket interface

• Standard transport layer protocols:

• TCP: reliable, in-sequence byte-stream transfer• UDP: unreliable, un-ordered message transfer

but other protocols can be used as well (e.g. SCTP)

• SAPs are called ports, used for application multiplexing

• Several applications / processes can use transport service• One application is bound to one port• Ports are identified by numbers• The PDUs generated by TCP / UDP are called segments• TCP / UDP segments include the port number• TCP / UDP receiver delivers incoming segment to the

application denoted by the port number


The Transport Layer (2)

• TCP has mechanisms for:

• Error control (retransmission-based) and in-order delivery• Flow control• Congestion control

• UDP has none of these features

• TCP and UDP hand over segments to the Internet layer


The Internet Layer

• This is a key part of the TCP/IP reference model

• Uses IP (Internet Protocol), its PDUs are called datagrams

• All higher-layer segments are encapsulated in datagrams

• The IP protocol:

• specifies an addressing scheme• provides end-to-end delivery of datagrams (forwarding)• does not specify how routing is done, left to dedicated

protocols• has no mechanisms for error, flow and congestion control• can send IP datagram over any network interface


The Internet Layer (2)

• “Everything over IP, IP over everything”


The Physical and Network Interface Layer

• The physical layer is similar to the PHY of the OSI RM

• The Network Interface Layer:

• Accepts IP datagrams and delivers them over physical link• Receives IP datagrams and delivers them to local IP layer• Includes medium access control, framing, address

resolution• Might also include link-layer error- and flow control


End-to-End Principle

• The Internet layer protocol (IP) is very simple

• Transport protocols run only in the hosts

• This is the end-to-end principle [7]:

• Keep routers simple• Realize reliability, sequencing etc. only in end hosts


Outline

Protocol Layering



The Five Elements of a Protocol

• Compare [4]

• A protocol specification should explain:

• the service provided• the assumptions about the environment in which it operates• the vocabulary of messages (PDUs) used by the protocol• the encoding (binary representation) of the messages• the procedure rules guarding the consistency of exchanges

of messages and service primitives

• The rules are the hardest to get right

Important Point

A protocol implements a service. Higher layers only use the

service and are not exposed to the internals of the protocol!


Why is Protocol Engineering so complex?

• Network protocols run as a distributed system of entities

(computers, routers) exchanging messages

• Networking protocols need mechanisms to deal with:

• Loss of messages• Failure of network links• Crash of entities• Large differences in processing speeds among entities• Incompatible data representations• Equipment from different vendors• Random delays in message transmission• Errors induced into messages (e.g. by the channel)• . . .

• Many of these are non-issues in software running on single

computer!


Protocol Engineering

• Major steps in protocol engineering

• Design of service and protocol• Initial performance evaluation• Formal specification of service and protocol design

• Often as finite state automata, e.g. in SDL language

• Validation: is the (formal) design doing the right things?

• Does it fulfill the service?• Presence of deadlocks? Liveness?• Can undesirable situations occur?

• Protocol implementation• Verification and testing: does the implementation

correspond to the formal specification?• Performance tuning, e.g. optimization of parameters

• See [6], [4]


Outline

Protocol Layering


Service Primitives



Service Providers and Service Users

• An N-protocol implements an N-service

• Stated differently: the N-protocol is the N-service provider!

• An N+1-protocol (or the application) is the N-service user

• Service provider and user:

• talk to each other through service primitives• have to obey rules in the usage of services

• Example: before a telephone can use any “send voice data”

service, it must have used “connection setup” service before

• Stated differently: service provider and user also run a

protocol for exchanging service primitives!

• Standard service primitives for a service S:

• S.request

• S.indication

• S.response

• S.confirmation


Confirmed Service

• Service user at A issues anS.request service primitive,possibly carrying user data

• The service provider for S (aprotocol) generates one or morePDUs and sends them to host B

• Service user at B is informedabout A’s service requestthrough an S.indication primitive

• Service user at B preparesresponse (possibly with data),gives it to local service providerthrough S.response

• B’s response is made known toA’s service user throughS.confirm primitive

• Key point: response comes fromB’s service user!

• Do you know an example?


Unconfirmed Service

• Service user at A issues anS.request primitive

• Service provider for S generatesone or more PDUs and sendsthem to host B

• Service user at B is informedthrough an S.indication primitive

• Service user at A has no cluewhether service request hasreached B



Confirmed Delivery Service

• Roughly similar to confirmedservice

• Key difference: it is B’s serviceprovider generating a response,not B’s service user!

• Thus, A’s service user has noinformation about the behaviourof B’s service user



Confirmed Transmission Service

• Roughly similar to theunconfirmed service

• Key difference: service user at Agets confirmation that any PDUsrelated to its service requesthave indeed been sent



Outline

Protocol Layering


Service Primitives



Multiplexing

• Multiplexing allows to transmitdata from several N SAPs over asingle N − 1 SAP

• When several N SAPs are usedin parallel, the N protocol entityneeds to make schedulingdecisions to decide which NSAP to serve next

• Sending N entity needs toinclude an SAP identifier into theN PDU to allow receiver entity todeliver an incoming N-PDU tothe right SAP

• Example: TCP supports severalSAPs through port numbers,port numbers are part of TCPheader


Splitting

• An N-entity can transmit datareceived from higher layers viaN-SAP over several N − 1 SAPs

• Allows transmission of data overseveral channels to increasethroughput and / or reliabilitythrough parallel transmission

• N-entity needs to makescheduling decisions on whichN − 1 SAP to use for a givenPDU

• Additional mechanisms forsequencing might becomenecessary


Fragmentation and Reassembly

• PDUs often have a limited size –on the lower layers this is usuallyfor physical reasons

• To make PDU sizes transparentto higher layers, an N-layer canaccept large N-SDUs andpartition the data into severalN-PDUs (fragments), eachhaving own header, and transmitthem separately

• Fragments must be numbered toallow receiver correctre-assembly

• Question: How should thereceiver deal with losses ofPDUs?

• Disadvantage: higher overhead


Blocking and Deblocking

• Sometimes higher layersproduce very small N-SDUs

• Instead of putting each N-SDUinto separate N-PDU, transmitterwaits until several N-SDUs arepresent (blocking) and putsthem into one N-PDU to saveoverhead

• Receiver entity decomposesreceived N-PDU (deblocking)and delivers several N-SDUs tohigher layers, this requiresmarkers in the N-PDUseparating the N-SDUs

• Question: when should senderstop collecting N-SDUs andsend an N-PDU?


Sequence Numbers

• An N-entity can maintain a sequence number

• For each newly constructed PDU the sequence number is

written into the N-PDU header, afterwards the sequence

number is incremented

• Sequence numbers allow the receiver to:

• Detect duplicate PDUs (and drop them)• Detect lost PDUs (and possibly request their retransmission

from sender)• Put N-PDUs back in the right order when network has

reordered them

• Implementation issues:

• Sequence number space is finite, wrapovers need to be

handled• Choice of initial sequence number


Mung Chiang, Steven H. Low, A. Robert Calderbank, and John C. Doyle.Layering as Optimization Decomposition: A Mathematical Theory of NetworkArchitectures.Proceedings of the IEEE, 95(1):255–312, January 2007.

Douglas E. Comer.Internetworking with TCP/IP – Principles, Protocols and Architecture, volume 1.Prentice Hall, Upper Saddle River, New Jersey, fifth edition, 2006.

John Day.Patterns in Network Architecture – A Return to Fundamentals.Prentice Hall, Upper Saddle River, New Jersey, 2008.

Gerard J. Holzmann.Design and Validation of Computer Protocols.Prentice Hall, Englewood Cliffs, 1992.

Vikas Kawadia and P. R. Kumar.A Cautionary Perspective on Cross-Layer Design.IEEE Wireless Communications, 12(1):3–11, February 2005.

Miroslav Popovic.Communication Protocol Engineering.CRC Press, Boca Raton, Florida, 2006.

Jerome H. Saltzer, David P. Reed, and David D. Clark.End-to-end arguments in system design.ACM Transactions on Computer Systems, 2(4):277–288, November 1984.



Hubert Zimmermann.OSI Reference Model–The ISO Model of Architecture for Open SystemsInterconnection.IEEE Transactions on Communications, 28(4):425–432, April 1980.

LANs MAC Bridges, Switches Bibliography


COSC 264Local Area Networks

Introduction

Dr. Andreas Willig1




UoC, 2014


Outline

LANs

Introduction

LAN Protocol Architecture

Topologies

MAC

Fundamentals

Orthogonal Schemes: FDMA, TMDA, SDMA, CDMA

Random Access Protocols

Other Schemes

Bridges, Switches

Repeaters and Hubs

Bridges and Switches


Outline

LANs

MAC

Bridges, Switches


Preliminaries

• The following slides are based mainly on [44], [38]

• The older, but still very good book [7] covers some related

topics, book is available in pdf format here:

http://web.mit.edu/dimitrib/www/datanets.html

http://web.mit.edu/dimitrib/www/datanets.html


Outline

LANs

Introduction


Topologies

MAC

Bridges, Switches


Local Area Networks (LANs)

• LANs are packet-switched networks

• Packets are often called frames in LAN context

• They have limited geographical extension, usually ≤ 1 km

• Offer a shared transmission medium to multiple stations

• Often controlled by only one owner / administrative entity

• Offer low cost for station attachment

• Support higher rates than usually experienced over

wide-area networks

• Some application areas:

• Connect desktop computers to share files, emails, . . .• Allow several computers to share printers, file servers, . . .• Interactive video or telephony between local users


LANs vs. WANs

• WAN = Wide Area Network:

• have national, continental or global geographical extension• typically controlled by several administrative entities• Often use high-capacity fibers for long-haul links

• WAN Examples:

• Internet• POTS

• In the Internet, LANs are an elementary unit

• Internet = Network of Networks!• LANs are attached to Routers, Routers are interconnected

via other LANs or via point-to-point connections


Outline

LANs

Introduction


Topologies

MAC

Bridges, Switches



• LAN standards typically specify the following layers:

• Physical layer (PHY), specifying transmission media,

network adapters, data rates and network topologies• Medium Access Control (MAC) sublayer, specifying how

stations share the transmission medium• Optional: Link layer (often called logical link control),

providing error-control, flow-control, etc.

• This is the protocol architecture followed by the IEEE

standards, which are the dominant LAN standards


The IEEE 802 Standards Series

• IEEE = Institute of Electrical and Electronics Engineers

• IEEE is a professional association, not a national or

international standards body

• Nonetheless, ISO adopted several IEEE standards

• Some important standards series (there are many more):

• IEEE 802.1: Overview, Bridging, Network Management,• IEEE 802.2: Logical Link Control (LLC) [35]• IEEE 802.3: Ethernet [20]• IEEE 802.11: Wireless LAN [22]• IEEE 802.15: Wireless Personal Area Network (WPAN),

incl. Bluetooth [21], IEEE 802.15.4 [29], . . .

See http://standards.ieee.org/getieee802

http://standards.ieee.org/getieee802


The IEEE 802 Standards Series (2)


The PHY

• The physical layer is responsible for:

• Encoding / modulation and demodulation / decoding• Preamble generation and removal• Determination of frame boundaries

• Preambles:

• allow receiver to acquire carrier- and bit-/symbol

synchronization• Use a well-known symbol pattern

• Some mechanisms for frame-boundary determination:

• Surrounding idle times on the medium• Unique start- and end-delimiters in frames• Length fields• Coding violations

• PHY specification in IEEE standards includes transmission

media, connectors, data rates, used signals, . . .


The PHY – Example Frame Structure

(from IEEE 802.11, [22, Fig. 15-1]):

• SYNC field is preamble

• Start-frame-delimiter (SFD) indicatesstart of PHY frame

• Signal field indicates modulationused in MPDU part

• Service field reserved for futureuse

• Length field indicates length offrame (in ms)

• CRC covers only PHY header


The PHY – Transmission Media

• Various transmission media, wired and wireless, are used

in IEEE 802 standards

• Wired transmission media:

• Twisted-Pair Cabling (e.g. switched Ethernet), up to 10 Gb/s

supported in Ethernet• Baseband coaxial cable (e.g. in classical Ethernet)• Broadband coaxial cable (e.g. cable TV)• Optical fibers, commercially available dense WDM systems

support up to 80 channels with 10 - 100 Gb/s each [33]

• Wireless transmission media:

• Radio-frequencies (e.g. 2.4 GHz or 5.2 / 5.75 GHz bands),

IEEE 802.11n supports up to 600 Mb/s• Infrared, IrDA supported 1 Gb/s over very short distances


The MAC

• MAC will be covered shortly ...


Logical Link Control – LLC

• LLC: link layer protocol specification for IEEE 802.x

standards

• LLC focuses on frame transmission to direct, single-hop

neighbours

• Major responsibilities:

• Error control• Flow control• Framing• Service provisioning to higher layers


LLC – Error Control

• Frames can be corrupted due to:

• Thermal noise in the receiver• Too low signal strength at the receiver• Collisions on the channel• Interference from external transmitters• Faulty processing• . . .

• Many applications don’t tolerate erroneously received data

• Countermeasure: Error-control [32], [30], [7], [31]

• Error control will be discussed in more detail later


LLC – Flow Control

• Goal: protecting a slow receiver from having the sendertransmit data at a too high rate

• Receiver could have a slow processor, little memory, . . .

• A few approaches:• Explicit signaling:

• Binary signaling: “You may / may not send me data”, for

example the RR/RNR (Receiver Ready, Receiver Not

Ready) frames in HDLC• Credit-based flow control: “You can send me up to hundred

bytes before i start to drop”, used for example in TCP

• Implicit signaling:

• Delay or suppress error-control feedback


Outline

LANs

Introduction


Topologies

MAC

Bridges, Switches


Bus Topology

• Stations are attached via tap line to bus

• Bus is a broadcast medium, signals are heard by all stations

• Parallel transmissions collide, all packets are garbled

• Frames include destination address field to indicate intended receiver

• Terminating resistances prevent signal reflections

• Bus length usually limited to a few dozens / hundreds of m, e.g. to

prevent excessive signal attenuation

• By cascading several buses (using branch points) more complex

broadcast toplogogies can be built (e.g. trees)

• Examples: classical Ethernet, Token Bus (IEEE 802.4 [17])


Ring Topology

• Network consists of stations, relaysand point-to-point links betweenrelays

• Relays pass through frames almostimmediately, they act as repeaters,they also copy signals to attachedstation and inject station’s signalsinto ring

• Data flows in well-defined direction

• Ring size often too small to carrymore than one frame at a time,access to ring regulated by MACscheme

• Examples: Token Ring (IEEE 802.5[18]), Fiber Distributed Data Interface(FDDI, ANSI Std. X3T9.5), ResilientPacket Ring (IEEE 802.17)


Star Topology

• Stations are attached to a centralunit, often with separate cables fortransmit and receive directions

• Central unit can act either as arepeater or as a bridge/switch

• A repeater copies incoming signalsto all other outputs

• A bridge/switch copies incomingframe only to output wheredestination is

• Discussed in more depth later


Wireless Topologies

• Wireless transmission media are much more prone toerrors than wired media [39], [52]

• Bit-error rates of up to 10−2 are not uncommon

• Wireless links are influenced by propagation environment

• Doors, moving faces in the surrounding, walls, . . .

Important Property

Wireless transmission channels are time-varying, the links are

volatile and unpredictable, there is usually no static wireless

topology


Outline

LANs

MAC

Bridges, Switches


Outline

LANs

MAC

Fundamentals



Other Schemes

Bridges, Switches


Medium Access Control (MAC)

• MAC protocols are schemes which allow a number of

users to coordinate the access to a common channel

• MAC protocols are a vast subject, some references (barely

the tip of the iceberg!) are: [14], [41], [34], [10], [53], [47],

[2], [26], [28], [11], [48], [25], [49], [50] [4], [6], [3], [12]

• The MAC layer is often regarded as a separate sub-layer

between PHY and link-layer

• This view is supported by the fact that the MAC has a

distinguished task not covered by any other layer

• MAC protocols are heavily influenced by the properties of

the underlying transmission medium


MAC Definition

• We are given:

• A number of users / stations wishing to communicate• A shared communications channel / resource that can only

be used by one station at a time• No other means for information exchange between stations

DefinitionMAC protocols are rules by which distributed stations

coordinate access to a common channel to share it efficiently

and in a manner satisfying given performance requirements

Example: 100 blind persons in a room – how to distribute the

right to talk?


Important Assumptions

• The shared channel is a broadcast medium, i.e.transmission of one station is heard by all other stations

• Not necessarily true for wireless transmission media

• In case of parallel transmissions all contendingtransmissions are garbled, i.e. cannot be reliably decoded

• Not necessarily true for wireless transmission media• Often not true for CDMA systems, also not in OFDMA


MAC Design Desiderata

• Small medium access delay: time between arrival of packetto empty station and start of successful transmission

• Depends on overheads: collisions, waiting times, . . .• For lightly loaded medium a small access delay is desirable• Hard real-time applications require bounded access delay

• In real-time applications: support for packet priorities, i.e.distinction between important and less important packets

• Local priorities: station makes local decisions, but A’s

important packets can be blocked by B’s unimportant ones• Global priorities: all stations reach consensus (how?) about

which station has most important packet

• Fairness and fair re-use of unused resources

• Efficiency: low overhead, high throughput

• Stability: increasing load should not decrease throughput

Note: you usually do not get all of these at the same time . . .


MAC vs Duplexing

• Channel duplexing schemes occur naturally when

full-duplex operation is required, e.g. in voice conversations

• These schemes allow a station to separate its transmitted

signals from its received signals

• Difference between duplexing and the MAC problem:

• MAC coordinates transmissions of multiple users• Duplexing coordinates parallel transmission and reception

for a single user

• But: MAC mechanisms can be used to implement

duplexing


Outline

LANs

MAC

Fundamentals



Other Schemes

Bridges, Switches


Orthogonal Schemes

• In orthogonal schemes the behavior of one station does

not influence the behavior / throughput / transmission

success / . . . of other stations

• The four main (mostly) orthogonal schemes are:

• FDMA = Frequency Division Multiple Access• TDMA = Time Division Multiple Access• SDMA = Space Division Multiple Access• CDMA = Code Division Multiple Access

• We will not discuss SDMA and CDMA


Frequency Division Multiple Access (FDMA)


FDMA (2)

• The given channel bandwidth is subdivided into N

sub-channels

• Between the sub-channels and at the fringe of the channelthere are guard bands:

• Reduction of adjacent-channel interference

• A sub-channel is exclusively assigned to a station i on a

long-term basis for transmission of data, no other station

is allowed to transmit on this channel

• To receive data, a station must:

• Either possess one separate receiver for each channel, or• have a single tunable receiver that must be switched to a

specific channel before data can be received on it

• Problems: coordination/rendez-vous, tuning times


FDMA (3)

• If totally available bandwidth is Bbs,station iisassigned≈ 1N of

B on a long-term basis (neglecting guard bands)

• Medium access delay for a new packet arriving to an

empty station i is always zero, since i can start

transmission immediately without risk of collision

• If a packet has size BN bits, its transmission takes exactly

one second, i.e.:

Transmission Delay = 1

where transmission delay is the time until transmission of a

frame completes (measured after arrival to empty station)


FDMA – Advantages

• N stations can transmit in parallel

• There is no need for time synchronization between the N

transmitters


FDMA – Disadvantages

• Need for N receivers or tunable receivers increases

system complexity

• Frequency synchronization required

• There is no re-use, i.e. channels unused by one station

cannot be used by others

ConclusionFDMA is good for CBR but bad for VBR traffic


Time Division Multiple Access (TDMA)


TDMA (2)

• Each station uses the whole frequency band (except someguard bands at the fringe of the spectrum), but only atcertain times:

• Time is subdivided into superframes of duration TSF

• Each superframe is subdivided into N time-slots• There are short guard times between time slots• One or more time slots are exclusively and long-term

assigned to a station i for transmission

• Stations must be time-synchronized to avoid overlapping

transmissions, guard times are required to compensate

(small) synchronization errors


Access and Transmission Delay in TDMA

• Neglecting guard times, each station gets the full channel

bandwidth Bbsforafractionof 1N of time

• Assume that:

• station i owns one time slot• TSF = 1 second• a time-slot suffices to transmit B

Nbits

• a packet of BN

bits arrives at random time to empty station i

• Medium access delay = waiting time until station i ’s next

slot starts

Access Delay =TSF

2= 0.5s


Access and Transmission Delay in TDMA (2)

• The time to transmit the packet is 1N seconds

• Assuming no channel errors we have:

Transmission Delay = Access Delay +1

Ns

= 0.5s +1

Ns ≤ 1s

for N > 2 this is a true inequality

ConclusionIn this example in TDMA we start later and finish sooner than

with FDMA!!


TDMA – Advantages and Disadvantages

• Advantages:

• It is easier to achieve asymmetric bandwidth assignments

in TDMA than in FDMA: using multiple time-slots is much

simpler than transmitting on multiple frequencies in parallel• TDMA tends to have better transmission delays than FDMA• No tunable receivers required

• Disadvantages:

• Tight time-synchronization between stations required• High expected access delay even in otherwise idle systems• Not possible to re-use unused time slots

ConclusionTDMA is good for CBR but bad for VBR traffic


Orthogonal Schemes – Discussion

• These schemes separate users perfectly from each other

• Typically they assign resources (frequency, time) to users

on longer timescales, which may be permanent or for the

duration of a call

• In so-called demand-assignment schemes resources are

also exclusively assigned to stations, but on much shorter

timescales (e.g. duration of a data burst in data traffic) [24]

ConclusionIn orthogonal schemes resource (de)allocation is considered a

rare event, in demand-assignment (DA) schemes not. DA

schemes must be much more efficient in signaling resource

(de)allocation.


Outline

LANs

MAC

Fundamentals



Other Schemes

Bridges, Switches



• Random Access protocols:

• do not attempt to reserve channel resources for longer time• do not require a central station or (much) shared state• do not access the medium at predictable times• often have low complexity (e.g. only little signaling, if any)• typically involve some random element

• Random access protocols are used standalone and alsoas building blocks for more complex protocols, e.g.:

• ALOHA / slotted ALOHA is used for signaling bandwidth

requests in demand-assignment protocols• In GSM a mobile uses slotted ALOHA to request call setup

Important Point

Random access protocols accept risk of collisions to save

coordination overhead and have overall improved efficiency!


ALOHA / Slotted ALOHA

• ALOHA [1] is one of the earliest MAC protocols, developed

≈ 1970 at the University of Hawaii

• Assumptions:

• N uncoordinated transmitters• One receiver (e.g. a base station)• If two packets overlap at receiver, a collision occurs


The Pure ALOHA Protocol

• When a new packet arrives at an empty station:

• a checksum is computed and appended to the frame• the frame is then transmitted immediately, there is no

coordination with other stations• an acknowledgement timer is started

• The receiver sends an immediate ack upon successful

reception of a packet – upon collisions or transmission

errors it remains quiet

• If the transmitter receives an ack, the frame is removed

and the ack timer is canceled


The Pure ALOHA Protocol (2)

• When ack timer expires, transmitter enters backoff mode:

• The transmitter chooses a random backoff time• It waits for this time without further action• At backoff timer expiry the frame is re-transmitted• The ack timer is set again, backoff mode is left• Question: why is the backoff time chosen randomly?

• When the number of failed trials exceeds a threshold, the

frame is dropped

• The precise choice of random distribution for backoff timesis critical for delay, throughput and stability! [13], [27]

• Often the random distribution depends on the number of

subsequent collisions seen by the frame• An example backoff strategy will be discussed later!

• When a new packet arrives to backlogged station, it is

stored in queue and served after all previous packets


Advantages of Pure ALOHA

• Quite simple to implement

• If network load is small:

• new frames are sent immediately =⇒ access delay is zero• they can use the full channel bandwidth• and the probability of collision at the receiver is low

ConclusionFor low network loads most packets can have the minimum

possible transmission delay


Disadvantages of Pure ALOHA

• Consider stations A and B sending frames of same length:

• When B starts its frame during the (two frame times long)

vulnerability period of A’s frame, the frames collide

• When more stations are active / load is increased, the

collision probability increases

• ALOHA cannot distinguish between collisions and channel

errors destroying a frame


Throughput

• Suppose all packets have the same length, packet transmission time is τ

• Define througput as the (average) number of successfully received packetsduring time τ in the absence of channel errors

• Which of these (idealized) curves is the throughput of ALOHA? And which onethe throughput of FDMA? How about the other curves?


Slotted ALOHA

• Slotted ALOHA is similar to ALOHA, but:

• Time is subdivided into time slots• A time slot is sufficient to accommodate frame

transmission, two round-trip times and ack transmission• All stations are time-synchronized• Any frame transmission has to start at slot boundary

ConclusionThe vulnerability period is reduced to one time-slot, slotted ALOHA

has better throughput


The CSMA-Family of Protocols

• CSMA = Carrier Sense Multiple Access [25], [49]

• Common assumption: all stations can determine the stateof the medium (almost) instantaneously:

• Busy: at least one station is currently transmitting• Idle: no station is transmitting

This operation is called carrier-sensing (CS) or

clear-channel assessment (CCA)

• Common approach: Listen-before-talk• Before station transmits a frame, it performs CS operation• If channel is busy, station defers transmission according to

one of several possible strategies• The maximum number of deferrals (or backoffs) a station

might experience for a frame is often bounded

• CSMA protocols do not eliminate collisions completely, but

reduce their rate or their impact


The CSMA-Family of Protocols (2)

• CSMA approach is especially useful when mediumpropagation time is small compared to packet length, sinceother stations notice transmission quickly after it started

• Typically satisfied in LANs, propagation delay is small• By this, collisions can only occur when two stations start

transmitting at almost the same time (time difference

smaller than propagation delay)

• When propagation time is large compared to packet length,the sender might have already stopped transmission whenreceiver senses busy carrier for first time

• Example: multi-access satellite configurations• Here LBT is almost useless, ALOHA is reasonable


Nonpersistent CSMA

• If a station senses a busy medium, it:

• draws a backoff time from a given random distribution• defers from channel activities during backoff time, and• then senses the medium again and starts over

• If the station detects an idle medium, it starts transmitting

immediately

• In case of a collision again a backoff time is chosen andprocess starts over

• Question: how to diagnose collisions?

• Performance problem: with high probability a medium is

idle for some time after transmission has finished, this

lowers utilization


p-persistent CSMA

• Be p ∈ (0, 1) a parameter known to all stations

• If a station senses a busy medium, it defers until the end of

the ongoing transmission, when medium becomes idle

• A station divides time on idle medium in small time slots• At the beginning of a time slot a station performs a random

experiment: with probability p it starts transmission, with

probability 1 − p it defers for one further slot• When station defers, it checks medium during remaining

slot time: when another station started transmission, station

waits for end of this transmission and starts over• Time slot just large enough to accommodate these activities

• In case of collision, process starts over

• Question: How would you choose p?

• Performance problem: again, medium will be idle for some

time after transmission has finished


1-persistent CSMA / CSMA-CD

• CSMA-CD was chosen for classical Ethernet

• If a station senses a busy medium, it defers until end of

ongoing transmission

• When medium becomes idle, station sends unconditionally

• This avoids idle times after previous transmission• But if two or more stations start, we surely have collision

and we need collision resolution procedure

• While transmitting, sender tests channel for collisions

• In case of a collision:

• Transmission is aborted• A jamming signal is sent to inform all stations about collision• A collision resolution procedure is started, e.g.:

• backoff schemes (used in Ethernet and WLANs)• tree algorithms [8]


Outline

LANs

MAC

Fundamentals



Other Schemes

Bridges, Switches


Polling Protocols

• Abstract view on polling systems:

• One central station / base station / hub• N clients / stations, each having a packet queue• The hub has two basic tasks:

• Query state of packet queues (e.g. # backlogged packets)• Grant bandwidth to stations based on results

• Querying a station should be less costly than data

transmission (e.g. in terms of bandwidth)

• Polling protocols are used to support time-bounded

services or minimum guaranteed bandwidth services e.g.

in the IEEE 802.11 WLAN w/ HCCA

• References: [15], [9], [16], [42], [43], [45], [46]


Polling Protocols (2)


Polling Protocols (3)

• Two logical channels are used for querying stations and

data transfer, but both can be mapped to one physical

channel

• A station must be known to the hub in order to be polled

=⇒ a registration protocol (e.g. ALOHA) is needed!

• Polling protocols can differ in:

• Polling sequence: in which sequence are stations polled• Service types: may a station send one or multiple packets

after it has been granted bandwidth?• Querying mechanisms: how is a station / a group of

stations queried?


Polling Protocols – Polling Sequences

• Round robin: all stations are polled one after another in a

circular fashion

=⇒ fair to all stations

• Table-driven polling: hub has an arbitrary list specifyingpolling sequence; if the end of the list is reached, the hubstarts over

• Allows uneven bandwidth distributions• Inflexible when traffic demands change• Often used in hard real-time systems


Polling Protocols – Service Types

• k -limited / time-limited service: a station is allowed to

transmit at most k packets or for at most t seconds

• Exhaustive service: a station may transmit packets as long

as its queue is nonempty

• Gated service: a station may send as many packets as arein the queue in the moment the grant / bandwidthassignment is received

• packets arriving during a stations service have to wait for

the next round


Polling Protocols – Querying Mechanisms

• Separate polling:

• the hub sends short poll packet to one station i• if station has nonempty queue, it starts transmitting• if station has empty queue, it remains quiet or returns extra

NULL packet to hub

• Separate polling has significant but constant overhead,

causes comparably long medium access delays for low

loads


Polling Protocols – Querying Mechanisms (2)

• Group testing is discussed in [51], [5], [8]

• Instead of sending a poll request to one station at a time, it

is sent to a group of stations

• Any station from group may answer with request packet:

• If no station answers, hub tests the next group• if one station answers, hub grants bandwidth to this station• if more than one station answers, a collision resolution

procedure is invoked; examples:

• query each station in the group separately• split group in two and group-test each sub-group separately

• Group testing shortens medium access delays in case of

light network load if groups are large



• Piggybacking: if a station sends a data packet, it indicates

its request for further bandwidth in the packets header

=⇒ no need for extra signalling packets!

• Piggybacking is often used as an additional mechanism to

one of the other mechanisms

• Piggybacking alone is not sufficient, since a station cannot

request bandwidth for new packets arriving to empty queue



• The query mechanisms described so far have twoimportant properties:

• Querying is initiated by the base station (except

piggybacking)• They are deterministic: for each station there is an upper

bound on the time before it can send its request



• Reservation protocols [40]: another frequent case is thatquerying is initiated by a station:

• station sends reservation request packet to hub• hub either accepts or rejects the request – indicated by

appropriate response packets• one example setup is that all stations have to share a

common signalling channel, e.g. using the ALOHA protocol,

while the data channel can be used exclusively

• Reservation protocols work well if signalling is a rare event

• Examples: MASCARA [36] and DQRUMA [24]


Token-Passing Protocols

• Assumption: stations are attached to a broadcast medium

• The right to initiate data transmissions is passed between

stations using special token frames, no central station

• After receiving a token a station may send data for limitedtime, afterwards passes token to its successor

• In absence of errors this guarantees a bounded medium

access time, required in real-time applications

• The stations form a logical token passing ring whichconstitutes predecessor / successor relationships

• All stations must have consistent view on logical ring

• Organization of the logical ring requires ring maintenance

mechanisms (involving special control frames)


Token-Passing Protocols – Problems

• Maintaining a logical ring can be hairy, if:

• not all stations can hear each other (partially meshed

topology)• mobility is involved (variable topology)

• Loss of token / control frames can create severe problems

[23], [55]

• Token-passing protocols tend to be very complex


Outline

LANs

MAC

Bridges, Switches


Coupling LANs

• It is often required to couple existing LANs, e.g. upon

merging different departments

• Coupling requires special coupling devices

• Types of coupling devices:

• PHY layer: repeaters, hubs• MAC layer: bridges, layer-2 switches• Network layer: router• Application layer: gateway

• See [37], [54]


Outline

LANs

MAC

Bridges, Switches

Repeaters and Hubs



Repeaters

• A repeater amplifies a signal on the analog level

• Any noise present in the signal is amplified as well• Repeaters add their own noise• Repeaters are not at all visible to any protocol or

modulation scheme• They can create slight delay (order of µs and less)


Regenerating Repeaters

• A regenerating repeater demodulates an incoming signalsymbol-per-symbol and modulates it again

• No interpretation whatsoever of protocol fields is done• Especially, no error checking / error correction is done,

regeneration can introduce errors


Hubs

• A hub is a centralized repeater, itbroadcasts signals incoming on oneport to all other ports

• No interpretation of the incomingframe is done, none of its fields isevaluated

• All stations are attached with onetransmit and one receive line

• A hub creates a logical bus on aphysical star

• Hubs can be cascaded

• It may be regenerative or not

• Question: Advantage over bus?


Outline

LANs

MAC

Bridges, Switches

Repeaters and Hubs



Bridges

• Bridges interconnect LANs on the MAC layer

Important Point

Bridges understand and interpret fields related to the MAC

protocol (e.g. address fields), repeaters / hubs do not!

• Nowadays they mostly connect LANs of the same type (i.e.

Ethernet – Ethernet), but bridges connecting LANs of

different types (e.g. Ethernet – Token Ring) also exist(ed)

• We focus on bridges interconnecting broadcast Ethernets

• Important: Ethernet frames carry an Ethernet source

address and an Ethernet destination address in their frame

header

• A bridge can connect several LANs

• Bridges can be cascaded


Basic Operation

• When bridge receives frame fromLAN A, it checks the frame forcorrectness, buffers it and checks theMAC destination address (dst)

• dst on LAN A: bridge does nothing

• dst on LAN B or dst unknown:bridge transmits frame on LAN B,following the rules of the MACprotocol!

• Same in direction B → A

• Bridge does not modify any frame,nor does it encapsulate them

• Stations need not be aware of thepresence of bridges (transparency)


Basic Operation (2)

• How does bridge know which stationis in which LAN?

• From reading a frames sourceaddress field (src) the bridge canlearn on which bridge interface asource can be reached

• When a bridge receives a frame withdst not having been observed sofar, it unconditionally re-transmits theframe on all interfaces exceptincoming one

• The latter approach is dangerouswhen several bridges are used andloops are present (see below)


Some Reasons to Use Bridges

• Reliability: by keeping LANs separated and only

interconnected by a bridge, failures in one LAN do not

affect others

• Performance: by carefully evaluating addresses, bridgescan confine traffic local to one LAN to that very LAN,enabling parallel local transmissions in different LANs

• Repeaters / hubs cannot do this “traffic separation”!

• Security: similarly, traffic local to one LAN cannot be

eavesdropped in the other LAN


Excursion: Encapsulating Bridges

• Bridges can be interconnectedthrough third-party link, e.g. serialline, microwave link, the Internet, . . .

• Third-party link requires own frameformat

• Approach: a bridge encapsulates aframe from LAN A into a frameappropriate for third-party link

• Receiving bridge decapsulatesframe and puts it on LAN B

• When is this useful?


A Larger Network Example

(Example taken from [44, Fig. 15.10])

• Packets between Stn1 and Stn7traverse two bridges

• Between Stn1 and Stn5 two differentpathes exist!

• Providing different pathes is useful toprovide fault-tolerance andload-balancing

• Only one bridge (Br1 or Br7) shouldforward packet from Stn1 to Stn5 toavoid duplicate packets on LAN E

• We have a routing problem!


A Larger Network Example (2)


• Fixed routing approach: each bridgepossesses for each incominginterface a table indicating whether aframe to dst should be forwardedand to which outgoing interface / LAN

• Problem: table needs to berecomputed and re-distributed uponevery change in topology

• Does not scale well to largeinstallations


Spanning Tree Approach

• The spanning tree approach addresses automatic

construction and maintenance of forwarding tables by

bridges in larger LAN installations

• It is specified in IEEE 802.1D [19]

• Algorithm consists of three elements:

• Frame forwarding• Address learning• Loop resolution


Spanning Tree Approach – Frame Forwarding

• For each port / attached LAN a bridge maintains twoinformations:

• A forwarding table• A flag indicating if port is in blocking or forwarding state

• Forwarding table contains:

• all MAC addresses which can be reached (directly or

indirectly) by sending to this port• A timer for each stored MAC address

• Example: for bridge Br2 (see previous slides)

• all stations in LANs A, B, D, and E on upper port• all stations in LANs C, F and G on lower port


Spanning Tree Approach – Frame Forwarding (2)

• Suppose a bridge receives frame to dst on port x

• Bridge checks forwarding tables on all other ports than x

• If dst is not found, bridge sends frame to all ports in

forwarding state except x

• If dst is found on port y and port y is in forwarding state,

the frame is sent to port y , otherwise (y not in forwarding

state) the frame is dropped


Spanning Tree Approach – Address Learning

• Suppose a frame from station src arrives on port x

• The bridge then checks:

• If no entry for src exists in the forwarding table for x , then it

is added and a timer is started• If already an entry for src exists on port x and the timer is

running, the timer is canceled and re-started

• If the timer for src at port x expires, the entry for src and

its timer are deleted from forwarding table of x

• This usage of timers is called soft state!

• Why is this done?• Soft-state is a fairly common mechanism in many protocols!


Spanning Tree Approach – Loop Resolution


• Suppose initially all forwarding tablesare empty

• At time t0 station Stn1 sends a frameto Stn2 using LAN A

• Both bridges receive frame in paralleland forward to LAN B, bridge Br1 attime t1 and bridge Br2 at time t2 > t1

• At time t1 bridge Br2 receives aframe with Stn1’s source address onLAN B, at time t2 the same happensto Br1

• Effect: Stn2 receives the same frametwo times

• Effect: Br1 and Br2 have Stn1included in forwarding tables for bothports!!

• Question: what happens if next Stn2sends a frame to Stn1?


Spanning Tree Approach – Loop Resolution (2)

• To avoid this kind of loops, IEEE 802.1D specifies the

spanning tree protocol

• Approach:

• Each bridge is equipped with an individual MAC address• A cost value is administratively assigned to each bridge• Bridges run a dedicated protocol among each other,

exchanging information about network topology• When topology is fully discovered, a minimum-weight

(related to per-bridge costs) spanning tree is computed

• In the computation LANs are taken as vertices and bridges

are taken as edges in the graph!!

• Spanning-tree computation can result in “removal” of

possible connections, the corresponding bridge ports are

set to state blocking• Tree is re-calculated upon changes in topology


Layer-2 Switches

• A switch is a centralized element,forwarding frames only to thecorrect output port

• Stations are attached to switch viapoint-to-point links with separatetransmit/receive lines (full-duplex)

• No broadcast medium anymore!

• A switch is able to process frames todistinct destinations in parallel,switches can therefore increasenetwork capacity as compared toLANs with broadcast medium

• Frames arriving in parallel for thesame destination are buffered(output buffering)

• Switches are transparent to stations

• Nowadays almost all Ethernetinstallations use switches


Layer-2 Switches (2)

• Difference to a hub:

• A hub builds a broadcast medium, only one station can

transmit at a time without collisions• A switch can accept up to N parallel transmissions, where

N is the number of stations in the LAN• Each attached station can receive packets at the full

medium capacity

Important Point

Switches remove the “shared medium” assumption and the

need for a MAC, but now the stations contend for the resources

of the switch (switching capacity, buffer memory)!


Layer-2 Switches (3)

• Operation modes of switches:

• Store-and-forward: switch receives a packet fully, checks

frame for correctness, reads off destination address,

determines outgoing port and transmits packet there• Cut-through-switch: switch starts forwarding to output

port already after having read the address field (appears

very early in most frame formats!), no check for frame

correctness

• Switches can be cascaded and more complex

infrastructures (including loops) can be built

• Switches often incorporate the same loop removal

technique as bridges


Norman Abramson.Development of the ALOHANET.IEEE Transactions on Information Theory, 31(2):119–123, March 1985.

Norman Abramson, editor.Multiple Access Communications – Foundations for Emerging Technologies.IEEE Press, New York, 1993.

Norman Abramson.Multiple Access in Wireless Digital Networks.Proceedings of the IEEE, 82(9):1360–1370, September 1994.

Ian F. Akyildiz, Janise McNair, Loren Carrasco, and Ramon Puigjaner.Medium access control protocols for multimedia traffic in wireless networks.IEEE Network Magazine, 13(4):39–47, 1999.

Mostafa H. Ammar and George N. Rouskas.On the performance of protocols for collecting responses over a multiple-accesschannel.IEEE Transactions on Communications, 43(2):412–420, February 1995.

Guiseppe Anastasi, Luciano Lenzini, Enzo Mingozzi, Andreas Hettich, andAndreas Krämling.Mac protocols for wideband wireless local access: Evolution towards wirelessatm.IEEE Personal Communications, 5(5):53–64, October 1998.


D. Bertsekas and R. Gallager.Data Networks.Prentice Hall, Englewood Cliffs, New Jersey, 1987.

J. I. Capetanakis.Tree Algorithm for Packet Broadcast Channels.IEEE Transactions on Information Theory, 25(5):505–515, September 1979.

Charles Chien, Mani B. Srivastava, Rajeev Jain, Paul Lettieri, Vipin Aggarwal,and Robert Sternowski.Adaptive Radio for Multimedia Wireless Links.IEEE Journal on Selected Areas in Communications, 17(5):793–813, May 1999.

Lou Dellaverson and Wendy Dellaverson.Distributed channel access on wireless atm links.IEEE Communications Magazine, 35(11):110–113, November 1997.

Andras Farago, Andrew D. Myers, Violet R. Syrotiuk, and Gergely V. Zaruba.Meta-MAC Protocols: Automatic Combination of MAC Protocols to OptimizePerformance for Unknown Conditions.IEEE Journal on Selected Areas in Communications, 18(9):1670–1681,September 2000.

Robert G. Gallager.A Perspective on Multiaccess Channels.IEEE Transactions on Information Theory, 31(2):124–142, March 1985.


Jonathan Goodman, Albert G. Greenberg, Neal Madras, and Peter March.Stability of binary exponential backoff.Journal of the ACM, 35(3):579–602, 988.

Ajay Chandra V. Gummalla and John O. Limb.Wireless medium access control protocols.IEEE Communications Surveys and Tutorials, 3(2):2–15, 2000.http://www.comsoc.org/pubs/surveys.

Boudewijn R. Haverkort.Performance of Computer Communication Systems – A Model Based Approach.John Wiley and Sons, Chichester / New York, 1998.

A. Hoffmann, R. J. Haines, and A. H. Aghvami.Performance analysis of a token based MAC protocol with asymmetric pollingstrategy (’TOPO’) for indoor radio local area networks under channel outageconditions.In Proc. International Conference on Communications (ICC), pages 1306–1311,New Orleans, Louisiana, 1994. IEEE.

IEEE.802.4 Token-passing Bus Access Method, 1985.

IEEE.802.5 Token Ring Access Method and Physical Layer Specifications, 1985.

IEEE Computer Society.


802.1D – IEEE Standard for Local and Metropolitan Area Networks – MediaAccess Control (MAC) Bridges, June 2004.

IEEE Computer Society.IEEE Standard for Information technology – Telecommunications and informationexchange between systems – Local and metropolitan area networks – Specificrequirements – Part 3: Carrier sense multiple access with collision detection(CSMA/CD) access method and physical layer specifications, December 2005.

IEEE Computer Society – Sponsored by the LAN/MAN Standards Committee.IEEE Standard for Information technology – Telecommunications and informationexchange between systems – Local and metropolitan area networks – Specificrequirements – Part 15.1: Wireless Medium Access Control (MAC) and PhysicalLayer (PHY) Specifications for wireless personal area networks (WPANs), June2005.

IEEE Computer Society, sponsored by the LAN/MAN Standards Committee.IEEE Standard for Information technology – Telecommunications and InformationExchange between Systems – Local and Metropolitan Area Networks – SpecificRequirements – Part 11: Wireless LAN Medium Access Control (MAC) andPhysical Layer (PHY) Specifications, 2007.

Hong ju Moon, Hong Seong Park, Sang Chul Ahn, and Wook Hyun Kwon.Performance Degradation of the IEEE 802.4 Token Bus Network in a NoisyEnvironment.Computer Communications, 21:547–557, 1998.


Mark J. Karol, Z. Liu, and K.Y. Eng.An efficient demand-assignment multiple access protocol for wireless (atm)networks.Wireless Networks, 1(3), 1995.

Leonard Kleinrock and Fouad A. Tobagi.Packet switching in radio channels: Part I carrier sense multiple access modelsand their throughput-/delay-characteristic.IEEE Transactions on Communications, 23(12):1400–1416, 1975.

J. F. Kurose, M. Schwartz, and Y. Yemini.Multiple-access protocols and time-constrained communication.ACM Computing Surveys, 16:43–70, March 1984.

Byung-Jae Kwak, Nah-Oak Song, and Leonard E. Miller.Performance Analysis of Exponential Backoff.IEEE/ACM Transactions on Networking, 13(2):343–355, April 2005.

S. S. Lam.Multiaccess protocols in computer communications. volume I: Principles.In W. Chon, editor, Principles of Communication and Network Protocols, pages114–155. Prentice-Hall, Englewood Cliffs, NJ, 1983.

LAN/MAN Standards Committee of the IEEE Computer Society.IEEE Standard for Information technology – Telecommunications and informationexchange between systems – Local and metropolitan area networks – Specificrequirements – Part 15.4: Wireless Medium Access Control (MAC) and Physical


Layer (PHY) Specifications for Low Rate Wireless Personal Area Networks(LR-WPANs), September 2006.revision of 2006.

Shu Lin and Daniel J. Costello.Error Control Coding.Prentice-Hall, Englewood Cliffs, New Jersey, second edition, 2004.

Shu Lin, Daniel J. Costello, and Michael J. Miller.Automatic-Repeat-Request Error-Control Schemes.IEEE Communications Magazine, 22(12):5–17, December 1984.

Hang Liu, Hairuo Ma, Magda El Zarki, and Sanjay Gupta.Error control schemes for networks: An overview.MONET – Mobile Networks and Applications, 2(2):167–182, 1997.

Biswanath Mukherjee.Optical WDM Networks.Optical Networks Series. Springer, New York, 2006.

Andrew D. Myers and Stefano Basagni.Wireless media access control.In Ivan Stojmenovic, editor, Handbook of Wireless Networks and MobileComputing, pages 119–143. John Wiley & Sons, New York, 2002.

The Editors of IEEE 802.IEEE 802.2, ISO/IEC 8802-2: Local Area Networks: Logical Link Control, 1989.


Nikos Passas, Sarantis Paskalis, Dimitri Vali, and Lazaros Merakos.Quality-of-service-oriented medium access control for wireless atm networks.IEEE Communications Magazine, 35(11):42–50, November 1997.

Radia Perlman.Interconnections Second Edition – Bridges, Routers, Switches andInternetworking Protocols.Addison-Wesley, Reading, Massachusetts, 1999.

Larry L. Peterson and Bruce S. Davie.Computer Networks – A Systems Approach.Morgan Kaufmann, San Francisco, fourth edition, 2007.

Theodore S. Rappaport.Wireless Communications – Principles and Practice.Prentice Hall, Upper Saddle River, NJ, USA, 2002.

Izhak Rubin.Access-Control Disciplines for Multi-Access Communication Channels:Reservation and TDMA Schemes.IEEE Transactions on Information Theory, 25(5):516–536, September 1979.

Izhak Rubin.Multiple access methods for communications networks.In Jerry D. Gibson, editor, The Communications Handbook, pages 622–649. CRCPress / IEEE Press, Boca Raton, Florida, 1996.


Izhak Rubin and L. F. M. de Moraes.Message Delay Analysis for Polling and Token Multiple-Access Schemes forLocal Communication Networks.IEEE Journal on Selected Areas in Communications, 1(5):935–947, 1983.

Oran Sharon and Eitan Altman.An efficient polling mac for wireless lans.IEEE/ACM Transactions on Networking, 9(4):439–451, August 2001.


Hideaki Takagi.Analysis of Polling Systems.MIT Press, Cambridge, Massachusetts, 1986.

Hideaki Takagi.Queueing analysis of polling models: an update.In Hideaki Takagi, editor, Stochastic Analysis of Computer and CommunicationSystems, pages 267–318. Elsevier, Amsterdam, 1990.

Andrew S. Tanenbaum.Computer Networks.Prentice-Hall, Englewood Cliffs, New Jersey, third edition, 1997.

Fouad A. Tobagi.


Multiaccess protocols in packet communications systems.IEEE Transactions on Communications, 28:468–488, 1980.

Fouad A. Tobagi and Leonard Kleinrock.Packet switching in radio channels: Part II the hidden terminal problem in csmaand busy-tone solutions.IEEE Transactions on Communications, 23(12):1417–1433, 1975.

Fouad A. Tobagi and Leonard Kleinrock.Packet switching in radio channels: Part III – polling and (dynamic) split-channelreservation multiple access.IEEE Transactions on Communications, 24(8):832–845, August 1976.

Don Towsley and J. K. Wolf.On adaptive tree polling algorithms.IEEE Transactions on Communications, 32(12):1294–1298, 1984.

David Tse and Pramod Viswanath.Fundamentals of Wireless Communications.Cambridge University Press, Cambridge, UK, 2005.

Harmen R. van As.Media access techniques: The evolution towards terabit/s LANs and MANs.Computer Networks and ISDN Systems, 26:603–656, 1994.

George Varghese and Radia Perlman.Transparent Interconnection of Incompatible Local Area Networks Using Bridges.


IEEE Journal on Selected Areas in Communications, 8(1):1565–1575, January1990.

Andreas Willig and Adam Wolisz.Ring stability of the PROFIBUS token passing protocol over error prone links.IEEE Transactions on Industrial Electronics, 48(5):1025–1033, October 2001.

Introduction PHY Half-Duplex Ethernet Bibliography


COSC 264Local Area Networks

Ethernet

Dr. Andreas Willig1




UoC, 2014


Outline

Introduction

PHY

Half-Duplex Ethernet


Outline

Introduction

PHY



Ethernet

• Ethernet is a packet-switched LAN technology

• Ethernet became the dominating wired LAN technology,

early competitors like Token-Ring vanished

• Standardized by IEEE ([1] and various amendments)

• Its name refers to the “Ether” that was believed to be the

necessary medium for electromagnetic wave propagation

• It offers:

• High data rates (10 Mbps, 100 Mbps, 1 Gbps, 10 Gbps, 100

Gbps)• Cheap and mature components• Large market penetration, many different vendors

• These slides are mainly based on [3, Chap. 16]


Brief History of Ethernet

• 1973: Developed at Xerox PARC by Metcalfe / Boggs

• 1976: first publication by Metcalfe / Boggs

• 1980: IEEE 802 working group adopted Ethernet

• 1981: IEEE 802.3 standard working group established

• 1983: 10Base2 Cheapernet standard published

• 1985: published as ISO standard 8802/3

• 1991: 10BaseT standard, adopting twisted-pair cabling

• 1995: Fast Ethernet standard with 100 Mbps published

• 1999: Gigabit Ethernet approved

• 2002: 10 Gb Ethernet approved

• 2012: 40 and 100 Gb Ethernet approved


Frame format

• The original Ethernet had a slightly different frame format

• Minimum payload size is 46 bytes, smaller messages must be padded, i.e. filledup with zeros

• Maximum payload size is 1500 bytes

• SOF = Start of Frame

• FCS = Frame Check Sequence, it is a CRC checksum


Ethernet Addresses

• Address fields are 48 bits long

• Each Ethernet adapter has its own address, typically

burned into adapter hardware

• Addresses are required to be globally unique

• Each vendor gets own address range and assigns unique

addresses from that range• Nowadays adapters with programmable address available

• Address representation as six colon-separated bytes in

hexadecimal representation, e.g.:

00:0c:29:10:fb:f3

• Special addresses and address ranges:

• Broadcast address: FF:FF:FF:FF:FF:FF• Addresses with first bit set to 1 but different from broadcast

address are multicast addresses• Addresses with first bit set to 0 are unicast addresses


The Length/Type Field

Type field Protocol

0x0800 IPv4

0x0806 ARP

0x809B AppleTalk

0x86DD IPv6

0x8863 PPPoE Discovery

0x8864 PPPoE Session

• When the L/T field carries value ≥

0x0600 it is interpreted as a typefield, indicating the higher-layerprotocol that is encapsulated in theEthernet frame

• The type field therefore providesprotocol multiplexing

• When the L/T field carries a value ≤

1500, it indicates the length of thepayload, then the type field isassumed to be in the first two bytesof the payload


Outline

Introduction

PHY



Ethernet PHYs

• The Ethernet standard specifies several different physicallayers, using different transmission media and adapters:

• Coaxial cables• Twisted pair cables• Fiber cables

• The standard also specifies several transmission rates: 10

Mbps, 100 Mbps, 1 Gbps, 10 Gbps, 40 Gbps, 100 Gbps

• Notation:

<Data rate> <Signaling> <Max. Segment Length>

where:

• Data rate is given in Mbps• Maximum segment length is given in hundreds of meters,

or uses letters for special types of media• Signaling always BASE, referring to baseband transmission


Topologies

• Ethernet supports fundamentally two types of topologies:

• Broadcast topologies• Switched Ethernet


Broadcast Topologies

• All stations share capacity of the medium, MAC is needed

• Stations operate in half-duplex mode

• In Ethernet:

• Bus topology (with coaxial cable)• Hub-based topology, using different types of cables

• Only used in the 10 Mbps PHYs

• Referred to as half-duplex Ethernet in the following


Switched Topologies

• Stations are attached to (cascaded) switches viapoint-to-point links, no MAC needed on those links

• The MAC is nonetheless “used”, since it is present anyway

• Stations operate in full-duplex mode

• Parallel transmissions possible, contention occurs for

resources in switches (e.g. buffer memory)

• Used for 100 Mbps and higher PHYs

• Referred to as switched Ethernet in the following


Ethernet PHYs – A Few Examples

• 10BASE5: 10 Mbps, 500 m segment length (coaxial cable)

• 10BASE-T: 10 Mbps, using unshielded twisted-pair (UTP)

cables in star topology

• 10BASE-F: 10 Mbps, fibre-optic cables with star topology

(repeaters as central elements)

• 100BASE-TX: 100 Mbps, shielded twisted-pair (STP)

cabling, star topology, or Category 5 UTP

There are many, many more PHY specifications . . .


Ethernet PHYs – Channel Coding Examples

• The 10 Mbps Ethernet variants use baseband transmissionand Manchester coding

• self-clocking!

• The 100 Mbps variant (100BASE-X) uses basebandtransmission and a more efficient 4b5b channel code:

• Blocks of four user bits are mapped to blocks of five bits• The 4b → 5b mapping is constructed so that sufficiently

many transitions between 0 and 1 are present, allowing

receiver to track symbol-/bit synchronization

• One 1 Gbps variant (1000BASE-X) uses 8b10b channel

code

• There are also other channel codes in use


Ethernet PHYs – Summary

• There is broadcast-/half-duplex Ethernet and switched

Ethernet

• There is a great selection of media types and speeds to

choose from


Outline

Introduction

PHY



Half-duplex MAC Protocol

• Half-duplex Ethernet uses CSMA/CD

• Carrier-Sense-Multiple-Access with Collision-Detection• This is a 1-persistent CSMA with specific collision

resolution procedure

• Major assumptions:

• Parallel transmissions lead to collisions• Stations monitor the medium for collision while transmitting

• Ethernet implements this by checking the voltage on the

medium, a collision is inferred when this voltage exceeds the

voltage applied by the transmitter alone


MAC Protocol

1. Suppose a new packet arrives at the MAC of a station, set

coll to zero

2. Station performs a carrier-sense operation

3. When the medium is idle, transmission starts immediately

4. When medium is busy:

4.1 listen until channel becomes idle again

4.2 start transmitting

5. While transmitting, check for collision

6. If collision is detected:

6.1 Abort frame transmission, send a short jamming signal

6.2 Increase counter coll, counting # of subsequent collisions

6.3 If coll > 16 then drop frame and set coll to zero

6.4 Wait for random time (the backoff time), then go to step 2

7. If no collision is detected, transmit until end of frame, set

coll to zero


Computation of Backoff Time

• Draw a random integer number uniformly from the interval

[

0, 1, . . . , 2min 10,coll− 1

]

which is called the backoff window

• The actual backoff time is computed by multiplying this

integer with the pre-defined slot time

• Slot time is common parameter, just large enough to cover

maximum round-trip time and small processing delays

Important Point

The backoff window size and therefore the average backoff time

doubles after each collision, until 10 collisions have been

observed! This is called the (truncated) binary exponential

backoff algorithm!

Question: why this algorithm?


Minimum Frame Size / Slot Time

• Minimum frame size (64 bytes for 10 Mbps Ethernet) is

chosen to ensure that a transmitter can indeed detect a

collision on a cable of given length

• Rationale: a frame must be long enough to still being

transmitted after the maximum round-trip delay plus small

processing times introduced by repeaters / hubs

• Requirement for maximum round-trip delay can be

understood from worst-case example (next slide)

Important Point

Smaller slot times / minimal frame sizes can be achieved by

restricting the maximum length of an Ethernet segment!


Minimum Frame Size (2)


Half-duplex MAC Protocol – Important properties

• Very short medium access delays under light load

• Under very heavy load and for many stations the collisionrate increases

• Rule of thumb: Ethernet should not be operated beyond 50

- 60% load!

• Important problem: the capture effect [2]


Half-duplex MAC Protocol – Capture effect

• Suppose two stations A and B have many packets in their

queue and start at the same time with collision counters

nA = nB = 0

• They start at the same time, experience a collision and

both set nA = nB = 1

• Station A draws a backoff slot of zero, B draws one

• As a result:

• A wins contention and starts transmitting• A sets nA = 0, whereas nB = 1

• Now A has another packet, both stations transmit after A

has finished its first packet, and experience a collision


Half-duplex MAC Protocol – Capture effect cont’d

• Result:

• A sets nA = 1• B sets nB = 2• Station A has therefore higher likelihood to win the next

contention cycle

• If A wins indeed, it sets nA = 0 whereas nB = 2

• This is repeated, after each iteration the likelihood of A

winning the contention increases

• After 16 trials node B throws away the packet

=⇒ frame delivery cannot be guaranteed!!!

QuestionCan you imagine a protocol modification that circumvents this

problem?


IEEE Computer Society.IEEE Standard for Information technology – Telecommunications and informationexchange between systems – Local and metropolitan area networks – Specificrequirements – Part 3: Carrier sense multiple access with collision detection(CSMA/CD) access method and physical layer specifications, December 2005.

K. K. Ramakrishnan and Henry Yang.The ethernet capture effect: Analysis and solution.In Proc. 19th Conference on Local Computer Networks (LCN’94), Minneapolis,USA, October 1994.


Introduction Error-Detecting Codes, Checksums Error-Correcting Codes ARQ Methods Bibliography


COSC 264Error Control

Dr. Andreas Willig1




UoC, 2014


Outline

Introduction

Error-Detecting Codes, Checksums

Introduction

Arithmetic in GF(2)

CRC Checksums

Error-Correcting Codes

Fundamental Considerations

Block Codes

Further Comments

ARQ Methods

Alternating Bit Protocol

Goback-N

Selective Repeat


About This Module

• Goals:

• Understand that bit- and packet-errors are a fact of life in

virtually all networks• Know the two fundamental types of error-control strategies

• Literature:

• Books on coding: [13], [17], [18], [4], [20], [16]• Interesting papers on coding: [5], [7], [23], [28]• ARQ and general error control: [15], [12], [11], [14]• This module is in parts based on [13]


Outline

Introduction



ARQ Methods


Some Reasons for Transmission Errors

• Individual bits in a packet can be wrong because of

thermal noise coupled with weak signal strength,

interference, deliberate jamming, . . .

• A crashing router / switch / station looses all packets

currently in its memory

• Packets are routed in the wrong direction

• Packets are dropped because of congestion

• Packets are dropped because of insufficient resources at

receiver

Important Point

Transmission errors are a fact of life and present a significant

design challenge for a network and a distributed application!


Three Ways to Deal With Errors

1. Correct them:

• Requires ability to detect presence of errors• This is the domain of open- and closed-loop error control

2. Ignore them:

• A few wrong pixels in a JPEG image won’t hurt very often

3. Conceal them:

• . . . or you replace them by average of neighbored pixels• Error concealment works only for certain types of user data


Types of Errors

• On the link-layer level:

• Packet fully lost (e.g. due to failure in preamble acquisition)• Packet truncated (e.g. loss of synchronization)• Bits in packet are flipped

• On the end-to-end level in a wide-area network:

• Packets are lost• Packets are duplicated• Packets are delayed for very long time• Packets are reordered• Bits in packet are flipped


Error Control Schemes

• Error-control schemes try to detect and correct errors

• They are typically employed:

• On the link-layer, protecting single-hop transmissions• On the transport-layer, protecting multi-hop transmissions

• They all involve, in one way or the other, redundancy!

QuestionSuppose in a wide-area multihop network we have error-control

on every link. Is this sufficient or do we additionally need

end-to-end error control mechanisms?


Open-Loop Error Control

• Transmitter does not receive any feedback from the

receiver about transmission outcome

• Major option for transmitter: equip packets with redundant

data, allowing receiver to correct bit errors

• This is called error-control coding or forward error

correction (FEC)

• Often applied in situations where feedback is infeasible:

• Real-time voice (POTS!) and video applications• When very long links are involved (Space probes)!


Closed-Loop Error Control

• ARQ = Automatic Repeat reQuest [3], [11], [15], [26]

• Here transmitter gets (and uses!!) feedback from receiver

• Basic idea:

• Transmitter augments each packet with a checksum• Receiver uses checksum to check the integrity of a received

packet and provides transmitter with appropriate feedback• If needed, the transmitter re-transmits packet

• Hybrid-ARQ schemes couple ARQ with coding, e.g.:

• Couple ARQ with a coding scheme that is active all the time• Transmit first packet without coding, but upon receiving

negative feedback switch on coding, increase coding

strength / redundancy with every further retransmission


Packet Error Rate vs. Bit Error Rate

• Suppose the channel only flips bits

• Each bit is flipped independently ofothers with probability p ∈ (0, 1),called bit error rate (BER)

• A packet with n bits is regarded aserroneous when at least one bit isflipped

• The packet error rate (PER) is thus:

P(p, n) = 1 − (1 − p)n

• Figure shows PER for varying BER(computed as 10b , with b being theBER exponent on x-axis) for packetswith 100 bits overhead (header,trailer) and varying # of user bits


Packet Error Rate vs. Bit Error Rate – Observations

• For fixed BER the PER increases with packet size

• Longer packets are more likely to be erroneous• Shorter packets have relatively more overhead

• For fixed packet size PER increases with increasing BER

• When BER becomes too bad, almost no packet is

transmitted correctly


Error Characteristics

• The error model used before (bits are independent and

have same error probability) is extremely simplistic

• In reality errors are often bursty:

• In an error burst of order x two successive erroneous bits

are separated by less than x bits• Example: when a wireless transmitter transmits at 11 Mbps

and an interference burst of 200 µs occurs, 2200 bits are

affected by the burst

• The practically observed error rates depend on thetransmission medium

• Fibre-optical cables have BERs in the order 10−15 and less• With wireless links BERs in the order of 10−3 . . . 10−2 can

be frequently observed

• Error modeling is a vast field (e.g. [1], [10], [8], [27])


Outline

Introduction



ARQ Methods


Outline

Introduction


Introduction

Arithmetic in GF(2)

CRC Checksums


ARQ Methods


Introduction

• With error-detection coding redundancy is added to apacket so that:

• Certain error patterns can be detected reliably• Other error patterns can be detected with high probability• No information about the position of errors in a packet can

be inferred (and hence no correction can be done)

• Error-detecting codes serve as a packet checksum andare used in ARQ schemes:

• Tx computes checksum as function of header and data• Tx attaches checksum to packet and transmits both• Receiver computes own checksum (with same algorithm)

over (received) header and data• Receiver compares own checksum with the one received• If equal, packet is accepted and positive feedback is sent• Otherwise, negative feedback is provided to Tx


Position of Checksum Fields

• Often (especially for MAC- and link layer protocols) thepacket checksum is appended to the packet

• Reason: checksum computation can be done on the fly

(while transmitting) by the transmitter hardware, but for this

it must have seen all the bits of a packet, leaving appending

as the only option• Examples: Ethernet, Wireless LAN

• Some other protocols (e.g. TCP) store the checksum field

in the header


Parity Check

• Data stream is subdivided into small blocks, e.g. bytes

• Even parity: one additional bit is appended to a byte sothat total # of 1’s in data and parity bits is even, e.g.:

• 0100 1101 becomes 0100 1101 0

• 0101 1101 becomes 0101 1101 1

• Odd parity: similar, but total number of 1’s is odd

• Properties of parity check codes:

• All odd numbers of bit errors are reliably detected• All even numbers of bit errors are reliably not detected

• Parity check codes have traditionally been used on serial

interfaces, e.g. RS-232, where a parity bit has been

appended to each byte


Outline

Introduction


Introduction

Arithmetic in GF(2)

CRC Checksums


ARQ Methods


Introduction

• The following explains the arithmetic in GF(2), a simplealgebraic structure that:

• Is familiar (maybe in other guises) to computer scientists• Is the foundation for the computation of CRC checksums

• The presentation is mainly based on [13]


Field Axioms

DefinitionWe are given a non-empty set F and two operations

+ : F × F 7→ F

∗ : F × F 7→ F

for which the following holds (field axioms):

• Associativity: For all x ∈ F , y ∈ F and z ∈ F we have:

x + (y + z) = (x + y) + z

x ∗ (y ∗ z) = (x ∗ y) ∗ z

• Commutativity: For all x ∈ F and y ∈ F we have:

x + y = y + x

x ∗ y = y ∗ x


Field Axioms (2)

Definition• Additive identity: there exists an element 0 ∈ F , called identity, such that for each x ∈ F we have

x + 0 = 0 + x = x

• Multiplicative identity: there exists an element 1 ∈ F , called identity, such that for each x ∈ F we have

x ∗ 1 = 1 ∗ x = x

• Additive inverse: for every x ∈ F there exists an element y ∈ F such that

x + y = 0

This element is denoted as −x . This element is unique.

• Multiplicative inverse: for every x ∈ F with x 6= 0 there exists an element y ∈ F such that

x ∗ y = 1

This element is denoted as x−1. This element is unique.


Field Axioms (3)

Definition• Distributivity: For all x ∈ F , y ∈ F and z ∈ F we have

x ∗ (y + z) = (x ∗ y) + (x ∗ z)


Examples of Fields

• Infinite fields:

•

•

•

• There are also finite fields, called Galois Fields

• In honor of Evariste Galois, a french mathematician


GF(2) – Modulo-2 Arithmetic

• Set F = 0, 1

• Define the addition + as follows:

x y x + y

0 0 00 1 11 0 11 1 0

• This is precisely an XOR!!• You can check that 0 = 1 + 1 mod 2 holds

• Define the multiplication ∗ as follows:

x y x ∗ y

0 0 00 1 01 0 01 1 1

• Check that these operations satisfy the field axioms!!


GF(2) – Modulo-2 Arithmetic (2)

• The additive identity is 0

• The multiplicative identity is 1

• Additive inverses:

• 0 is the inverse of 0• 1 is the inverse of 1

The latter gives rise to the unfamiliar-looking (but true)

identity

1 + 1 = 0 = 1 − 1

and also to

0 − 1 = 1

This means: subtraction is the same as addition!


Polynomials over GF(2)

DefinitionBe n ∈. A polynomial f (X ) with one variable X and with

coefficients from GF(2) is of the form

f (X ) = fn · X n + fn−1 · X n−1 + . . .+ f1 · X + f0 =

n∑

k=0

fk · X k

where all fi are from GF(2). The polynomial is said to have

degree of n when fn = 1, otherwise it has a degree smaller

than n. The degree of f (X ) is denoted as deg f (X ).

• There are 2n polynomials over GF(2) with degree n


Addition of Polynomials over GF(2)

• Be f (X) =∑n

k=0 fk · X k and g(X) =∑m

k=0 gk · X k two polynomials

• For adding these polynomials, we add the coefficients of the same poweraccording to the rules of GF(2) aritmetic. For m ≤ n we get

f (X)+g(X) = (f0+g0)+(f1+g1)·X+. . .+(fm+gm)·X m+fm+1 ·Xm+1+. . .+fn ·X

n

• Example:

1 · X 5 + 0 · X 4 + 1 · X 3 + 1 · X 2 + 0 · X + 1

+ 1 · X 5 + 1 · X 4 + 0 · X 3 + 1 · X 2 + 0 · X + 1

= 0 · X 5 + 1 · X 4 + 1 · X 3 + 0 · X 2 + 0 · X + 0

or, using a shorthand notation (which ignores the formal variable X ):

101101

+ 110101

= 011000

• Under GF(2) addition and subtraction of polynomials are really the same!!


Multiplication of Polynomials over GF(2)

• The rule for multiplication can be obtained from expanding f (X) · g(X) andsumming up (according to the rules of GF(2)!!) the terms for the same power

• We have

f (X) · g(X) = c0 + c1 · X + c2 · X 2 + . . .+ cn+m · X n+m

with

c0 = f0 ∗ g0

c1 = f0 ∗ g1 + f1 ∗ g0

c2 = f0 ∗ g2 + f1 ∗ g1 + f2 ∗ g0

. . .

ck = f0 ∗ gk + f1 ∗ gk−1 + . . .+ fk ∗ g0

. . .

cn+m = fn ∗ gm

where the right-hand-sides are computed according to GF(2) rules


Multiplication of Polynomials over GF(2) – (2)

• We can multiply polynomials in a way similar as multiplying numbers

• Using the shorthand notation we have for example:

1 0 1 0 1 × 1 0 1 1

1 0 1 0 11 0 1 0 1

0 0 0 0 01 0 1 0 1

1 0 0 1 0 1 1 1


Rules for Polynomials over GF(2)

• Commutativity: for all polynomials f (X) and g(X):

f (X) + g(X) = g(X) + f (X)

f (X) · g(X) = g(X) · f (X)

• Associativity: for all polynomials f (X), g(X) and h(X):

f (X) + [g(X) + h(X)] = [f (X) + g(X)] + h(X)

f (X) · [g(X) · h(X)] = [f (X) · g(X)] · h(X)

• Distributivity: for all polynomials f (X), g(X) and h(X):

f (X) · [g(X) + h(X)] = [f (X) · g(X)] + [f (X) · h(X)]


Polynomial Division over GF(2)

• Be f (X) and g(X) two polynomials, deg g(X) > 0 and deg f (X) ≥ deg g(X)

• When f (X) is divided by g(X) then we get a unique pair of polynomials, say:q(X) and r(X) such that

f (X) = q(X) · g(X) + r(X)

where r(X) is called the remainder and q(X) is called the quotient, andfurthermore deg r(X) < deg g(X) holds

• Compare Euclids algorithm!!

• We can divide polynomials very similar to the long-division technique for dividinginteger numbers


Polynomial Division over GF(2) – (2)

• One example:

X6 +X5 +X4 +X +1 : X3 +X +1 = X3 + X2

−( X6 +X4 +X3 )

X5 +X3 +X +1

−( X5 +X3 +X2 )

+X2 +X +1

• And using the shorthand notation the same example becomes:

1 1 1 0 0 1 1 : 1 0 1 1 = 1100−( 1 0 1 1 0 0 0)

0 1 0 1 0 1 1−( 1 0 1 1 0 0)

1 1 1

• So the “quotient” is 1100 (or X3 + X2) and the remainder is 111 (or X2 + X + 1)


Outline

Introduction


Introduction

Arithmetic in GF(2)

CRC Checksums


ARQ Methods


CRC Computation and Transmission

• We are given k data bits, represented as polynomial k(X )with k coefficients (i.e. maximum degree is k − 1)

• We are furthermore given a so-called generator

polynomial g(X ), for example of degree 16 or 32

• We compute n − k = deg g(x) check bits as follows:

• compute new polynomial X n−k · k(X )

• amounts to shifting the bits of k(X ) to the left by n − k

places, filling up with zeros for the smallest coefficients

• Perform a polynomial division to give:

X n−k · k(X )

g(X )= q(X ) · g(X ) + r(X )

• Transmit the message m(X ) = X n−k · k(X ) + r(X )• Crucial property: m(X ) is evenly divisible by g(X )!!


CRC – Receiver Behaviour

• As a result of transmission, the channel possibly modifies

transmitted m(X ) to received m(X )

• The receiver:

• Divides received m(X ) by g(X )• If the remainder is zero, the message is accepted as correct• Otherwise, the message is rejected as erroneous

• Some comments:

• Errors are not detected when their bit pattern (taken as a

polynomial) is evenly divisible by g(X )• All odd numbers of errors can be detected when g(X ) can

be evenly divided by X + 1• All error bursts of length less than n − k are detected• As an approximation, error bursts of length r ≥ n − k are

not detected with probability 1/2r


A Justification

• The procedure rests on the fact that

m(X ) = X n−k · k(X ) + r(X ) is evenly divisible by g(X )

• This is true since:

m(X )

g(X )=

X n−k · k(X ) + r(X )

g(X )

= g(X ) · q(X ) + r(X ) +r(X )

g(X )

= g(X ) · q(X ) + r(X ) + r(X )

= g(X ) · q(X )

which is certainly divisible without remainder by g(X )• Can you justify the steps?


CRC – Some Standard Polynomials

• Quoted from [24, Sect. 6.3]

Name Specification

CRC-12 X12 + X11 + X3 + X2 + X + 1

CRC-16 X16 + X15 + X2 + 1

CRC-CCITT X16 + X12 + X5 + 1

CRC-32 X32 + X26 + X23 + X22 + X16 + X12 + X11 + X10 + X8 + X7 + X5 + X4 + X2 + X + 1


Summary and Important Points

• CRCs can be easily implemented in hardware

• They can also be implemented efficiently in software

(linear time in the number of message bits) [19], [21], [9]

Important Point

There is always a residual error rate, i.e. some probability that

errors are not detected by a CRC checksum!!

• Intuitive justification: when the number of checkbits is smaller than the number ofmessage bits, several messages are mapped to the same checksum. An errorpattern turning one of these messages into another one cannot be detected

• In many cases the residual error probability (especially for 32 bit CRCs) is prettysmall and errors can be neglected

• But: see [25]


Outline

Introduction



ARQ Methods


Outline

Introduction




Block Codes

Further Comments

ARQ Methods


Coding and Decoding – An Abstract View

• The channel is capable oftransmitting channel symbols froma finite set, e.g. 0, 1

• A coder maps source messages wto sequences xn of channel symbols

• The channel possibly introduceserrors and outputs a sequence yn ofchannel symbols

• The decoder applies a decodingfunction to yn and produces anestimate w of the transmittedmessage w

• When w 6= w we have a decodingerror


Code Rate

• Suppose that:

• the binary representation of source message requires k bits• the coder produces n > k channel bits

DefinitionThe ratio k

n is called the code rate. It expresses the amount of

redundancy added by code to the user data.

This is not the most general definition, see [6]


Goal of Coding Theory

Goal of Coding Theory

A key goal of coding theory is to find codes that achieve very

small probabilities of decoding errors while being reasonably

efficient, i.e. have a high code rate!

• Often, stronger codes (i.e. having lower probabilities of

decoding errors) tend to have lower code rates

Important Point

You have to pay for reliability with overhead!


A First Code: Repetition Coding

• Consider a channel with the following properties:

• The channel symbols are bits• Each bit is flipped with probility p < 1

2and transmitted

correctly with probability 1 − p• Bit errors are independent of each other

• The transmitter wishes to transmit a one-bit message w

• Coding scheme: the message bit is repeated n times, i.e.

0 7→ 000 . . . 00 ,n times 0

1 7→ 111 . . . 11 ,n times 1

where n is an odd number


A First Code: Repetition Coding (2)

• Suppose n = 5 and the receiver gets:

01011

• What is the most likely source message? 0 or 1?


A First Code: Repetition Coding (3)

• It can be shown that [17, Sect. 2.5]:• The decoding rule which minimizes the decoding error

probability is the majority voting rule

• Output the number that occurs most often in the received

sequence

• The probability for decoding error ǫ under the majority

voting rule is:

ǫ = 1 −

n−12

∑

i=0

(

n

i

)

pi(1 − p)n−i

• There is a problem: if we let our desirable ǫ shrink to zero,

the value n required to reach this ǫ grows without bound

• This means: arbitrarily reliable decoding not possible with

repetition coding with a strictly positive code rate


Shannons Noisy Channel Coding Theorem

• Can we do better than this? Can we transmit with arbitrarily

small decoding error probabilities at strictly positive rates?

• One of the initial [22] achievements of information theory is

the confirmation that this is indeed possible

• More precisely:

• For a given channel there exists a number C giving the

capacity of this channel• For any code rate R < C it is for given desired probability of

decoding error ǫ possible to construct a code having the

given rate R that has a decoding error probability less than ǫ


A Second Code: Parity Check Coding

• In this example code a block of 9user data bits is arranged in a 3 × 3matrix

• For each matrix row an even paritybit is computed (bits c1, c2, c3)

• For each matrix column an evenparity bit is computed (bits c4, c5, c6)

• In the lower right corner a parity bitfor the row and column parities iscomputed (bit c7)

• What happens after single-bit errors?

• What happens after two-bit errors?

• What happens after three-bit errors?


Outline

Introduction




Block Codes

Further Comments

ARQ Methods


Basic Operation

• In a block code:

• The user data stream is segmented into blocks of k bits• Each k -bit block is encoded independently of other

blocks to an n-bit codeword (n > k )

• The code rate is k/n

• The set of all possible source words has size 2k

• The set of all possible words in the code space has size 2n

• Out of these the code uses only 2k out of 2n elements

• These are called valid codewords• In many practical codes the set of codewords has some

algebraic structure, it could be a group or a vector space


Effect of Channel Errors

• Channel errors can turn:

• a valid codeword into a word from the set of 2n − 2k unused

codewords, then the decoder must guess which was the

transmitted codeword• a valid codeword into another valid codeword – this cannot

be detected by the code

• When facing an unused codeword y , decoders essentially

look for the valid codeword that is “closest” to y

• What can you do about errors that the decoder cannot

detect?

• How can we define “closest”? What notion of “distance” do

we use?

• Why do we choose the “closest”?


The Hamming Distance

DefinitionWe are given two words y1 and y2 of n bits length each. The

Hamming Distance of y1 and y2 is defined as the number of bit

positions in which y1 and y2 differ, it is denoted as d(y1, y2).The Hamming Distance of the Code, dH , is the minimum

Hamming Distance among all possible pairs of valid codewords.

• Note: d(y1, y2) can be computed by XOR’ing y1 and y2 and

counting the 1’s in the result

• Example:

y1 = 1011 0010 0101 1111

y2 = 1011 1010 1101 1011

y1 XOR y2 = 0000 1000 1000 0100

which implies that d(y1, y2) = 3


Decoding Rule

• Given the notion of Hamming distance we can formulate

the following decoding rule:

Decoding Rule

Let x1, x2, . . . , x2k denote the set of valid codewords. After

receiving a codeword y , the decoder outputs the xi that has the

smallest Hamming distance to y (ties are broken randomly), i.e.

we have:

x =xi∈x1,x2,...,x2kd(y , xi)

• The adoption of an algebraic structure (groups, vector

space) for the set of valid codewords often allows to

simplify the computation of as compared to an exhaustive

traversal of all valid codewords


Some Observations

• Suppose we have a code with Hamming distance dH

• Error patterns with no more than dH − 1 errors can bedetected reliably

• Why?

• Error patterns with no more than⌊

dH−12

⌋

errors can be

corrected reliably

• Why?


Practical Block Codes

• Block codes are still widely used, although better codes

exist

• Most popular classes:

• Reed-Solomon (RS) codes• Bose-Chaudhuri-Hocquenghem (BCH) codes


Outline

Introduction




Block Codes

Further Comments

ARQ Methods


Importance-based Coding

• The approach is quite simple: for many data types (e.g.

speech) not all source data bits have the same importance

=⇒ no need to encode all data bits in the same way

• Approach: Apply powerful coding (more redundancy) to

important data, less powerful coding to less important data!

• Some of the data types suitable for importance-basedcoding are:

• Speech and audio• Video

The (human) receiver of these data types does not

recognize small distortions in video frames / speech

samples


Importance-based Coding (2)

• Example: video coding with MPEG 2/4:

• Some bits describe global properties (“metadata”) of a

video stream, e.g. resolution, color depth, color palettes –

these bits are quite important• Other bits (DCT coefficients) describe picture contents –

the “low frequency” parts of a picture (large areas of the

same color) are more important than the “high frequency”

parts (sharp edges)

DCT = discrete cosine transform, has some similarities to discrete

Fourier transforms

• Importance-based coding is also applied in GSM speech

transmission


Outline

Introduction



ARQ Methods


ARQ Protocols

• ARQ = Automatic Repeat reQuest [3], [11], [15], [26]

• Basic idea:

• The transmitter augments each frame with a checksum• The receiver uses the checksum to check the integrity of a

packet and provides the transmitter with feedback• If needed, the transmitter retransmits the packet

Important Point

ARQ schemes are feedback-based or closed loop schemes.

They provide redundancy (retransmissions) only upon negative

feedback, on excellent channels they do not have significant

overhead


ARQ Protocols (2)

• Some immediate consequences:

• Transmitter must buffer packets for possible retransmission• Feedback channel needs bandwidth as well• Even for very few bit errors whole packet is retransmitted

• ARQ protocols differ:

• in the number of allowed outstanding frames /

unacknowledged frames• in the buffering requirements at receiver / transmitter• in the way feedback is provided (positive / negative

acknowledgement frames, timers)• in their maximum throughput under error conditions

• We discuss basic ARQ schemes, lots of variations!!


Assumptions

• In the following we assume one transmitter, one receiver

and a channel in between

• Data flows from transmitter to receiver

• Transmitter always has a new message available

• Packets are required to be delivered:

• Reliably (i.e. at least once)• Without duplication (at most once)• In-sequence

to the higher layers of the receiver

• Nomenclature:

• Message: refers to the higher-layer data• Packet: refers to the packet sent by error-control protocol,

includes the message data and header information


Assumptions (2)

• Packet header includes

• Source address• Destination address• Error-control information (depends on protocol)

• Packet also contains a checksum, which is perfect

• Upon receiving a packet the receiver:

• verifies checksum and drops it silently in case of failure• checks destination address, drops it when not destined to

him (address filtering)

• ARQ protocols also use acknowledgement packets,called ACKs, which include:

• Source address• Destination address• Error-control information (depends on protocol)


Reliable and Semi-Reliable Delivery

• Truly reliable delivery requires a potentially unbounded

number of retransmissions

• Protocols restricting the number of retransmissions are

called semi-reliable

• Almost all practical link- or transport-layer protocols are

semi-reliable


Outline

Introduction



ARQ Methods


Goback-N

Selective Repeat


Alternating-Bit Protocol

• The alternating bit protocol (ABP) [2] is the simplest of the

“serious” ARQ protocols

• It is also often referred to as send-and-wait

• Properties:

• guarantees in-sequence delivery if round-trip time is

bounded and timeout is chosen appropriately• simple to implement• requires one buffer at transmitter and one buffer at receiver• is reasonably efficient over links with propagation delays

smaller than the packet transmission time

• The packet header contains one extra field:

• seqno, having a width of one bit


ABP Transmitter


ABP Transmitter – Comments

• next is transmitters view on current sequence number

• tx-request is issued by higher layers to requestmessage transmission

• tx-request is accepted when transmitter is idle• tx-request is declined when transmitter is busy

• confirm primitives are sent to higher layer protocols

• Timeout value must be large enough to accommodate:

• propagation delay (twice)• transmission time for ACK packet• processing times

• Finding good timeouts is easy for a single hop and hard fora multi-hop network

• Why?


ABP Receiver


ABP – First Example


ABP – Second Example


ABP Properties

• Consider a packet of 1000 bits length, sent without gaps

• Medium bitrate is 100 Mbps

• Speed of light is 200,000 km/s in a cable

• Conclusion: packet has a geographical length of 2000 m

• Assume a link between Berlin/Germany and Christchurch

has ≈ 22, 000 km length

• Conclusion: 11,000 packets could be transmitted back to

back before the first one arrives in Berlin, 22,000 before

the first feedback returns to Christchurch

• But: ABP allows only one packet at a time in transit

Important Point

ABP is inefficient over “long fat pipes”, i.e. links with a large

bandwidth-delay product!!


ABP Properties (2)

• If the propagation delay is bounded and the timeout valueis appropriately chosen, then:

• ABP guarantees delivery• ABP delivers in-sequence• ABP avoids duplicate data delivery at receiver

• ABP does not work correctly if the network betweentransmitter and receiver may store packets (including acks)for arbitrary long times

• Question: Can you give example scenarios?


Outline

Introduction



ARQ Methods


Goback-N

Selective Repeat


Goback-N protocol

• Goback-N combats the inefficiency of ABP by allowing Noutstanding frames

• N is also called window size• outstanding = not yet acknowledged

• GB-N can be regarded as a generalization of ABP – stated

differently: ABP is actually a GB-1

• TCP originally used a variant of GB-N, recent versions

have selective acknowledgements

• Each packet is equipped with a sequence number, takenfrom a finite sequence number space

• Sequence numbers go from 0 to max_seqno (included)


GBN Operation

• The transmitter has a buffer for N packets – as long as this

buffer is not full, it accepts packets from its higher layers,

appends them to the buffer and transmits them asap

• Transmitter maintains variable next_seqno, initialized to

zero

• Receiver maintains variable expected_seqno, initialized

to zero

• If the receiver receives a packet with a seqno equal toexpected_seqno:

• expected_seqno is incremented modulo max_seqno + 1• The packet is delivered to higher layers• An acknowledgement is sent with expected_seqno


GBN Operation (2)

• If the receiver receives a packet with a different seqno:

• The packet is discarded• An acknowledgement is sent with expected_seqno, or no

ack is sent at all (to let the transmitter time out)

• If the transmitter gets a new packet from the applicationand the window is not full:

• The packet is equipped with the sequence number

next_seqno

• next_seqno is incremented modulo max_seqno + 1• The packet is transmitted asap (after any other not yet

transmitted packet that is)• A timer is started for this packet after its transmission


GBN Operation (3)

• If the transmitter receives an ack packet with sequencenumber ack_seqno:

• All packets in the window with sequence number smaller

than ack_seqno are regarded as acknowledged and are

removed from the window, opening it for further packets• the corresponding timers are canceled

• If transmitter gets timeout for oldest packet in the window:

• All pending timers are canceled• This packet and all subsequent packets are retransmitted• For each retransmitted packet a new timer is started

immediately after its transmission


GBN Example


Size of Sequence Number Space

• The sequence number space must be larger than the

window size N

• To see this, we use a counterexample where the size ofsequence number space is just N:

• Assume receivers seqno is expected_seqno=0• Assume the transmitter sends N packets with seqno’s from

0 to max_seqno

• The receiver receives all of them and sends

acknowledgements, at the end of this its seqno is again

expected_seqno=0• All acks are lost• The sender times out and sends all packets again• The receiver accepts all packets as new packets, since they

have the expected sequence numbers =⇒ duplicates!


GBN Properties

• GB-N can fill up a long fat pipe if N is chosen large enough

• The transmitter needs N buffers, the receiver only a single

buffer to accept an incoming packet

• If the network provides an upper bound on packet delays,the protocol ensures:

• in-sequence delivery• reliable delivery• no duplicates at the receiver

• If a packet fails, this packet and all subsequent packets are

retransmitted, even if the latter were correctly received

Important Point

If the packet error rate (PER) is small, Goback-N is reasonably

efficient, but for higher PERs the protocol retransmits many

correctly received packets and becomes inefficient


GBN Properties (2)

• With no bound on packet delays, correctness is harmed by:

• old data packets when their seqno equals receivers

expected_seqno

• old ack packets when their seqno fit into transmitters

current window

RemarkProtocols running over a large network use a large sequence

number space – TCP has 32-bit seqnos and IP has additionally

a mechanism to kill too old packets (TTL field)


Outline

Introduction



ARQ Methods


Goback-N

Selective Repeat


Selective Repeat Protocol

• Selective-Repeat (SR) works similar to Goback-N, butthere are differences:

• Receiver also has N buffers and maintains a “sliding

window”, giving the range of currently accepted sequence

numbers• The receiver buffers an out-of-order packet, given that its

sequence number fits into the current window, but does not

yet deliver it to higher layers• A packet arriving at the lower end of the sliding window is

directly delivered to the higher layers along with all buffered

packets immediately following the received packet – after

this the receiver advances its window accordingly, shifting

out the packets just delivered and creating space for new

packets


Selective Repeat Protocol (2)

• Some acknowledgement options:

• Receiver sends individual positive acknowledgements• Receiver sends negative acknowledgements (NACK) only

indicating missing frames• Receiver regularly sends a bitmap indicating the reception

status of its window elements

• Upon timer expiration or reception of a NACK the

transmitter retransmits only the indicated frame

• The sequence number space must have a size of at least

2N


Acknowledgement Variations

• Positive acknowledgements:• Receiver receives packet with seqno s

• Receiver sends ACK packet with seqno s

• Semantics: “i have successfully received this packet s”• Modified semantics for cumulative positive

acknowledgments: “i have successfully received this packet

s and all previous packets”

• Negative acknowledgement (Example):• Receiver receives packet with seqno s, previous received

packet had seqno s − 2• Receiver sends NACK packet with seqno s − 1• Semantics: “i have not received packet s − 1”• More generally, a NACK is issued when receiver notices

that some packet has not been successfully received• Can you imagine other methods to detect a failure when

there is no gap in the sequence numbers?

• NACKs can also be coupled with cumulative ACKs


Discussion: FEC and ARQ

• FEC does not require feedback channel, ARQ does

• Open-loop FEC has:

• constant overhead even without errors on the channel• constant throughput (if all errors can be corrected)• constant delay• variable residual error rate

• ARQ has:

• variable overhead: retransmissions occur only in case of

errors• variable delays (due to retransmissions)• very low residual error rate (determined by CRC quality and

allowed number of retransmissions)


Haowei Bai and Mohammed Atiquzzaman.Error modeling schemes for fading channels in wireless communications: Asurvey.IEEE Communications Surveys and Tutorials, 5(2):2 – 9, 2003.http://www.comsoc.org/livepubs/surveys.

K.A. Bartlett, R.A. Scantlebury, and P.T. Wilkinson.A note on reliable full-duplex transmission over half duplex lines.Communications of the ACM, 12(5):260ff, 1969.

D. Bertsekas and R. Gallager.Data Networks.Prentice Hall, Englewood Cliffs, New Jersey, 1987.

Ezio Biglieri.Coding for Wireless Channels.Springer, New York, 2005.

Leon Cohen.The history of noise – on the 100th anniversary of its birth.IEEE Signal Processing Magazine, 22(11):20–26, November 2005.

Thomas M. Cover and Joy A. Thomas.Elements of Information Theory.John Wiley & Sons, New York, second edition, 2006.

Victor DeBrunner, Linda DeBrunner, Longji Wang, and Sridhar Radhakrishnan.Error control and concealment for image transmission.


IEEE Communications Surveys and Tutorials, 3(1), 2000.http://www.comsoc.org/livepubs/surveys.

E. O. Elliot.Estimates of error rates for codes on burst-noise channels.Bell Systems Technical Journal, 42:1977–1997, September 1963.

David C. Feldmeier.Fast Software Implementation of Error Detection Codes.IEEE/ACM Transactions on Networking, 6(6):640–651, December 1995.

E. N. Gilbert.Capacity of a burst-noise channel.Bell Systems Technical Journal, 39:1253–1265, September 1960.

David Haccoun and Samuel Pierre.Automatic repeat request.In Jerry D. Gibson, editor, The Communications Handbook, pages 181–198. CRCPress / IEEE Press, Boca Raton, Florida, 1996.

Samir Kallel.Efficient hybrid arq protocols with adaptive forward error correction.IEEE Transactions on Communications, 42(2):281–289, February 1994.

Shu Lin and Daniel J. Costello.Error Control Coding.Prentice-Hall, Englewood Cliffs, New Jersey, second edition, 2004.


Shu Lin, Daniel J. Costello, and Michael J. Miller.Automatic-Repeat-Request Error-Control Schemes.IEEE Communications Magazine, 22(12):5–17, December 1984.

Hang Liu, Hairuo Ma, Magda El Zarki, and Sanjay Gupta.Error control schemes for networks: An overview.MONET – Mobile Networks and Applications, 2(2):167–182, 1997.

David J. C. MacKay.Information Theory, Inference, and Learning Algorithms.Cambridge University Press, Cambridge, UK, 2003.

Arnold M. Michelson and Allen H. Levesque.Error-Control Techniques for Digital Communication.John Wiley and Sons, New York, 1985.

Robert H. Morelos-Zaragoza.The Art of Error Correcting Coding.John Wiley & Sons, Chichester, UK, second edition, 2004.

Tenkasi V. Ramabadran and Sunil S. Gaitonde.A Tutorial on CRC Computations.IEEE Micro, 8(4):62–75, August 1988.

Tom Richardson and Ruediger Urbanke.Modern Coding Theory.Cambridge University Press, Cambridge, Massachusetts, 2008.


Dilip V. Sarwate.Computation of Cyclic Redundandy Checks via Table Look-Up.Communications of the ACM, 31(8):1008–1013, August 1988.

Claude E. Shannon.A mathematical theory of communication.Bell Systems Technical Journal, 27:379–423, 623–656, July, October 1948.

Bernard Sklar.A primer on turbo code concepts.IEEE Communications Magazine, 35(12):94–102, December 1997.


Jonathan Stone, Michael Greenwald, Craig Partridge, and James Hughes.Performance of checksums and crc’s over real data.IEEE/ACM Transactions on Networking, 6(5):529–543, 1998.

Andrew S. Tanenbaum.Computer Networks.Prentice-Hall, Englewood Cliffs, New Jersey, third edition, 1997.

H.S. Wang and N. Moayeri.Finite State Markov Channel - A Useful Model for Radio CommunicationChannels.IEEE Transactions on Vehicular Technology, 44(1):163–171, February 1995.


Yao Wang and Qin-Fan Zhu.Error Control and Concealment for Video Communication: A Review.Proceedings of the IEEE, 86(5):974–997, May 1998.

Introduction Shortest-Path Algorithms Distance-Vector Protocols Link-State Protocols Discussion and Further Topics Bibliograph


COSC 264Introduction to Routing

Dr. Andreas Willig1




UoC, 2014


Outline

Introduction

Fundamentals

Some “Unusual” Routing Protocols

Shortest-Path Algorithms

Bellman-Ford Algorithm

Dijkstra Algorithm

Distance-Vector Protocols

Protocol Operation

Problems

Link-State Protocols

Protocol Operation

Discussion and Further Topics

Discussion of DV and LS Protocols

Further Topics


On this Module

• Goals:

• Understand distinction between routing algorithms and

protocols• Understand distance-vector (DV) and link-state (LS) routing

protocols and their issues

• We explain the basic concepts of DV and LS routing, not

any particular routing protocol built on these paradigms

• This module is mainly based on [6], some parts also on [7,

Chap. 12] and [5]


Outline

Introduction






Outline

Introduction

Fundamentals







An Example Network

• We model a network as a graph

G = (V ,E)

• For simplicity, we define

V = 1, 2, . . . ,N

• Vertices v ∈ V correspond to

stations / routers

• E ⊂ V × V

• Edges e ∈ E correspond to

direct links, we write (i, j) or i − j

for a link between i ∈ V and

j ∈ V

• Edges are labeled with

non-negative numbers


An Example Network (2)

• Edge labels (called metrics) can be interpreted differently

• Interpretation as costs, e.g.:

• Delay• Monetary transmission costs• Geographical distance

Please note that cost is a more generic term than any of

the mentioned examples

• Interpretation as available resources, e.g.:

• Number of available phone trunks• Currently available capacity, given the set of flows that

already use this link


Fundamental task of routing

• Given a network G = (V ,E) , and given a fixed source

s ∈ V and destination node d ∈ V

• An m-hop path between source and destination is asequence of edges (i0, i1), (i1, i2), . . . , (im−1, im), so thati0 = s and im = d

• Alternative notation: i0 − i1 − . . .− im

Major task

For each source-destination pair in a network identify one or

more paths that are optimal (or at least of reasonable quality) in

some pre-defined sense.


Shortest-Path Routing

GoalThe goal of shortest-path routing algorithms is to select the

path that has the smallest total cost. To compute the total cost

of a path the costs of all its links are added.

• Routing in the Internet uses shortest-path routing

• But this is more complicated, though . . .

• Special case: minimum-hop routing is obtained when all

link costs are the same (e.g. = 1)

• Shortest-path routing is also known as least-cost routing


Widest-Path Routing

• Some networks judge the cost or value of a path based on

a non-additive property

• Example: dynamic call routing in POTS:

• Assume that for each network link we maintain the residualcapacity available on this link

• The residual capacity of a path is defined as the minimum

residual capacity of all the links on it• The goal is then to find the path with the maximum residual

capacity, since routing a new phone call on this path does

least likely create a bottleneck for future calls


Judgement Criteria

• Effectiveness: when a route between two nodes exists,

the routing algorithm / protocol should be able to find it

• Correctness: computed routes should be valid paths that

contain no circles (i.e. are loop-free)

• Simplicity: routing algorithms / protocols should be

computationally simple and require only little information

exchange among routers

• Robustness: a routing protocol must be able to cope with:

• link or station failures• newly established links or stations• changes in link metrics• congestion situations

by establishing new routes when old ones become

infeasible or are no longer optimal


Judgement Criteria (2)

• Stability: a routing protocol should not recomputeeverything upon minor changes in the network

• There is a tension between stability and robustness!

• Fairness: all users should be treated in the same way

• Optimality: different criteria, depending on perspective:

• Provider perspective: the network should carry as many

connections / packet flows as possible (maximizes revenue)• User perspective: generated routes should be “short”, “fast”

and offer good throughput


Classification Criteria for Routing Protocols

• Decision time (i.e. when are routing decisions made)

• For each packet (timescale of ms and below)• For each session / call (timescale of seconds to minutes)• At network configuration time (timescale of months / years)

• Decision place:

• Originating node / source (Source routing)• Central node (centralized routing)• Each node (distributed routing)

• Information used in decision:

• None• Local information (to a station)• Information from adjacent nodes (and local information)• Information from all nodes along a route• Information from all nodes in the network


Classification Criteria for Routing Protocols (2)

• Information update frequency:• Periodic• Upon major load change or topology change• Both

• Information update initiation:• Push: a node transmits information updates to its neighbors

on its own initiative• Pull: a node asks his neighbors (using request-response

exchange) for new information updates

• Information exchange channel:• In-band: routing information is exchanged on the same

network / channel as user data• Example: In the Internet routing protocols run on top of IP

• Out-of-band: routing information is transmitted on separatenetwork, no resource sharing at all with user data

• Example: in the POTS routing messages are exchanged

over separate network using the SS7 protocol


Routing Algorithms and Routing Protocols

• A routing algorithm solves routing problem in centralized

fashion, assuming full network information is available

• A routing protocol embeds a routing algorithm into a realnetworking context:

• It operates in a distributed environment• It incorporates explicit information exchange among nodes• Information exchange takes time and might fail, the protocol

must consider these possibilities

• A routing protocol mainly specifies which information is

exchanged between stations (and when), it is not

necessarily tied to any specific routing algorithm


Forwarding Table

• A forwarding table within a router maps to eachdestination address either:

• an outgoing interface (next-hop routing)• a full route to destination, which is then added to a packet

(source routing) and obeyed by all nodes on the path

Routing in the Internet uses next-hop routing

• The forwarding table:

• results from the execution of the routing protocol (dynamicrouting), or can be static / preconfigured (static routing)

• is changed on relatively large timescales, e.g. upon

topology changes, load changes or changes in metrics• is consulted for every packet


Outline

Introduction

Fundamentals







Hot-Potato Routing

• Rule: a router transmits a packet on the output link

currently having shortest queue / highest capacity / . . .

• Benefits:

• computationally simple• uses local information• no exchange of routing information required

• Drawbacks:

• No guarantees at all can be given


Randomized Routing

• Rule: a router transmits a packet on a randomly chosen

outgoing interface

• Benefits:

• Computationally simple• no exchange of routing information required• In a finite connected network this algorithm is guaranteed to

hit the destination node with probability one in absence of

link errors, congestion, etc.

• Drawbacks:

• Actual paths taken can be very long• Delays can be very long• Many hops give many opportunities for loosing packet


Flooding

• Rule: a router transmits a packet on all interfaces, exceptthe one it was received on

• A packet can arrive multiple times at destination or router• Router uses unique packet identifier (e.g. source address

and seqno) to avoid delivering duplicate data• Routers need these identifiers to avoid forwarding a packet

more than once (i.e. avoiding self-amplifying “explosion”)

• Benefits:

• simplicity, robustness, stability, . . .• requires no routing computations, no information exchange

• Drawbacks

• Extreme waste of resources• Security issue: all stations in the network get the packet

• Flooding can be an interesting option in networks where

data is transmitted only very rarely


Outline

Introduction






Introduction

• We discuss two shortest-path routing algorithms:

• Bellman-Ford• Dijkstra

There are other algorithms available, e.g. Floyd-Warshall

• See also [4]

• Both algorithms are centralized, i.e. they require that the

full information about network topology (nodes, edges and

metrics) is available to the algorithm

• Both algorithms play a prominent role in Internet routing


Some Notations

• Given a network G = (V ,E) with N = |V | stations

• i ∈ V and j ∈ V refer to some generic nodes / stations in

the network

• di,j is the direct link cost / metric between i and j , with:

• 0 ≤ di,j < ∞ when i and j are adjacent nodes• di,j = ∞ when i and j are non-adjacent nodes

• Di,j represents the total cost of the minimum cost path from

i to j in the Bellman-Ford algorithm, over one or multiple

hops, according to i ’s current knowledge

• Di,j represents the same thing for Dijkstra’s algorithm

• Ni represents the set of nodes adjacent to node i , i.e.

Ni = k ∈ V (i , k) ∈ E


Some Notations (2)

• In this example we have

d4,6 = 15, but D4,6 = D4,6 = 2

(by choosing the path 4 − 3 − 6)

• Furthermore, d1,6 = ∞, but

D1,6 = 3 (by choosing the path

1 − 4 − 3 − 6)

• We have N5 = 3, 4, 6


Outline

Introduction



Dijkstra Algorithm





Basic Idea

• B-F follows the dynamic programming principle [1], [2]

• We want to find a shortest route from s ∈ V to d ∈ V

• The following equations must be satisfied:

¯Ds,s = 0

¯Ds,d = mink∈Nd

Ds,k + dk ,d , for s 6= d

• Explanation: suppose node s already knows its least costs

Ds,k to the neighbors k ∈ Nd of d , then s’s least cost to d

is the minimum over all neighbors k of the costs Ds,k plus

the direct costs dk ,d

• But: how is Ds,k known and how to turn this into an

algorithm?


Approach

• In the Bellman-Ford algorithm the costs between a source

node s and all destination nodes d are computed iteratively

• The iterations are over the number of hops that packets

can take

• Change in notation: D(h)s,d gives the smallest cost among all

pathes between s and d that have at most h hops


A Simplified Bellman-Ford Algorithm

/ / Computes f o r a f i x e d node s the d is tances and the r o u t i n g/ / t r ee to a l l o ther nodes . Graph can be d i r ec ted or und i rec ted

/ / i n i t i a l i z a t i o n

D(0)s,s = 0 ; pred [ s ] = s ;

f o r a l l d wi th d 6= s do

D(0)s,d

= ∞ ; pred [ d ] = NULL ;

/∗ loop over a l l numbers o f hops ∗ /for h = 1 . . .N do :

D(h)s,s = 0

f o r a l l d 6= s do :

D(h)s,d

= mink∈NdD

(h−1)s,k

+ dk,d

pred [ d ] = k∈NdD

(h−1)s,k

+ dk,d


A Second Version

/ / Computes f o r a f i x e d node s the d is tances and the r o u t i n g/ / t r ee to a l l o ther nodes . Graph i s assumed to be d i r ec ted

/ / i n i t i a l i z a t i o n

Ds,s = 0 ; pred [ s ] = s ;f o r a l l d wi th d 6= s do

Ds,d = ∞ ; pred [ d ] = NULL ;

/∗ loop over a l l numbers o f hops ∗ /for h = 1 to N − 1 do :

foreach (v ,w) ∈ E do

when Ds,v + dv,w < Ds,w then

Ds,w = Ds,v + dv,w ;pred [ w ] = v


Comments

• The runtime of the algorithm for a single source is

O(|V | · |E |), for a whole network it becomes O(|V |2 · |E |)

• The algorithm shown is a simplified version of

Bellman-Ford that can handle only non-negative weights

• The algorithm can be extended to handle negative weights

as well, as long as no negative cycles are contained

• It is a centralized algorithm, complete network information

(di,j ) must be available at execution time

QuestionWhat is pred good for? What can you read off from it?


Outline

Introduction



Dijkstra Algorithm





Basic Idea

• Dijkstra’s algorithm is restricted to graphs with

non-negative weights

• It is greedy: in every situation it makes the choice that is

currently the best, without regard to future situations

• Here:

• The algorithm maintains a list S ′ of nodes that have not yet

been considered• In each step it removes k ∈ S ′ to which the source s has

the smallest known distance Ds,k

• For each neighbor x of k it is then checked if a path through

k to x is shorter than best so-far known path to x

• Remember that nodes are numbered from 1 to N, i.e.

V = 1, . . . ,N


The Dijkstra Algorithm


/ / i n i t i a l i z a t i o nS′ = V \ sDs,s = 0 ; pred [ s ] = s

f o r a l l d ∈ S′ doDs,d = ds,d ; pred [ d ] = NULL ;

when ds,d < ∞pred [ d ] = s

/ / main loopwhile S′ 6= ∅ do

k =m∈S′ Ds,m

S′ = S′ \ kfor j ∈ Nk do

when Ds,k + dk,j < Ds,j

Ds,j = Ds,k + dk,j

pred [ j ] = k


Comments

• The worst-case runtime of Dijkstra’s algorithm for a single

node is O(N2), but can be better for sparse graphs

• Dijkstra cannot handle negative metrics


Outline

Introduction






Outline

Introduction



Protocol Operation

Problems




Introduction

• The presentation here follows [6, Chap. 3]

• We want to turn the centralized Bellman-Ford algorithm

into a distributed protocol where nodes communicate only

with adjacent nodes

• For such a protocol we have to clarify:

• What kind of messages do nodes send to their neighbors,

what information is carried in them?• How often / when should this information be sent?


Basic Approach

• Previously we stated that the costs for the least cost routes

between source s and destination d satisfy:

¯Ds,s = 0

¯Ds,d = mink∈Nd

Ds,k + dk ,d , for s 6= d

• Observe here that Ds,k refers to a neighbor k of the

destination d

• A similar relationship must also hold for neighbors of s:

¯Ds,s = 0

¯Ds,d = mink∈Ns

ds,k + Dk ,d , for s 6= d

• What have we gained here?


Basic Approach (2)

• In the second form node s just uses information that:

• it has itself (s knows Ns and ds,k for direct neighbors)• it can obtain from direct neighbors (Dk,d , for k ∈ Ns)

• This gives hint: each node i transmits a vector, containingall values Di,d(t) it knows at time t , to its neighbors

• By this rule, node s receives Dk,d (t) from all its neighbors k• Upon receiving a value Dk,d (t) from a neighbor k , node s

can re-compute its own least-cost path to d by checking

whether Dk,d (t) + ds,k is smaller than currently known

least-cost – if so, node s stores k as next-hop for d• Some time later node s in turn transmits own vector with all

Ds,d (t), propagating changes further to downstream nodes• A node sends an updated version of its vector each time it

receives new information, or periodically (or both)• Over time, a node s receives sufficient vectors to have a

view on the whole network


Basic Approach (3)

• Note that initially a node s does not need to know the

whole network, it only needs Ns and ds,k for all k ∈ Ns

• Node addressing is handled outside the protocol

• For each destination address d , node s stores the next hop

and the least cost / distance

• Since nodes transmit a vector containing their distances

Ds,d to all known destinations d , this approach is called

distance-vector protocol

• Some of the Internets routing protocols are based on thisapproach, for example:

• RIP and RIP-2• BGP (uses variant of DV, named path vector routing)


DV Protocol Details

• In the following, let Dsk ,d(t) be the minimum cost from node

k to node d , as it is available to s at time t• Rationale for time dependency: it may take time for node k

to inform node s about changes in its least-cost values• Possible reasons:

• Processing delays at k , transmission delays• Node k transmits new cost vectors to some other node v

before transmitting it to s

• A protocol message sent by node i at time t has format:

Here:• Id= i indicates that i is the sender• a record Dst=d,Cost=c indicates that node i ’s current

least-cost path to node d has total costs of c• Node i includes such records for all destinations d it is

currently aware of


DV Protocol Details (2)

/ / I n i t i a l i z a t i o n/ / Node i s conf igured wi th unique node id , e . g . s

Ds,s = 0foreach k ∈ Ns do

Ds,k = ds,k

nexthop [ k ]=k

/ / p e r i o d i c t ransmiss ion o f own tab l e to neighbors (DV messages )on r e c e i v i n g t ransmi t−t imer do

t r ansm i t towards each k ∈ Ns the l i s t o f c u r r e n t l y \known d e s t i n a t i o n s and t h e i r most recent l e a s t costs

r e s t a r t t ransmi t−t imer



/ / recep t ion o f DV update from neighborson r e c e i v i n g DV message from neighbor k at t ime t do

when message conta ins new d e s t i n a t i o n d then

Ds,d = ds,k + Dsk,d

(t)

nexthop [ d ]=kforeach d e s t i n a t i o n d mentioned i n k ’ s DV message do

s to re Dsk,d

(t)

foreach d e s t i n a t i o n d known to s dowhen nexthop [ d ]=k then

Ds,d = ds,k + Dsk,d

(t)

/ / rou te computat ionforeach d e s t i n a t i o n d known to s do

foreach neighbor m do

when ds,m + Dsm,d

(t) < Ds,d then

Ds,d = ds,m + Dsm,d

(t)

nexthop [ d ]=m



/ / spec ia l cases

/ / a l i n k to a neighbor goes down / changes coston l i n k cost change towards neighbor k do :

ds,k = ∞ / / l i n k f a i l u r e , cost change also poss ib leforeach d e s t i n a t i o n d do

when nexthop [ d ]= k then

Ds,d = ∞ / / a l t e r n a t i v e l y : f resh route computat iont r ansm i t towards each m ∈ Ns the l i s t o f c u r r e n t l y \

known d e s t i n a t i o n s and t h e i r most recent l e a s t costs

/ / a l i n k to a neighbor comes upon l i n k c rea t i on towards neighbor k do :

update ds,k

t r ansm i t towards each m ∈ Ns the l i s t o f c u r r e n t l y \known d e s t i n a t i o n s and t h e i r most recent l e a s t costs



• How to detect a link failure?

• First option: check for sustained lack of periodic DV

messages• Second option: send separate hello messages frequently,

check for lack of answers

• How to detect a cost change for a link?

• Depends on the precise link metric• Example: delays depend on lengths of output queues


Comments

• The distances to destinations d as seen from source s

evolve over time, they change with message reception and

link failures or link establishments

• The period for periodic distance vector transmissions has

influence on convergence time

• A delay between detecting a link failure and transmitting an

updated distance vector can have significant influence on

protocol operation


Outline

Introduction



Protocol Operation

Problems




Convergence Time

• Consider the above example network

• Assume that all nodes are switched on at the same time t = 0

• Immediately after being switched on, each node informs its neighbors about itspresence

• Each node transmits its distance vector message every 60 seconds

• After receiving the distance vector messages the shortest path computationstakes one second


Convergence Time – Example Timeline

Time t = 0 s: all nodes are activated and send their initial DV messageTime t = 1 s: nodes finished their computations, result:

Nd 1

Dst Cost Outg

1 0 –2 1 1-2

Nd 2

Dst Cost Outg

1 1 2-12 0 –3 2 2-3

Nd 3

Dst Cost Outg

2 2 3-23 0 –6 1 3-6

Nd 6

Dst Cost Outg

3 1 6-36 0 –

Time t = 60 s: all nodes broadcast their DV messagesTime t = 61 s: nodes finished their computations, result:

Nd 1

Dst Cost Outg

1 0 –2 1 1-23 3 1-2

Nd 2

Dst Cost Outg

1 1 2-12 0 –3 2 2-36 3 2-3

Nd 3

Dst Cost Outg

1 3 3-22 2 3-23 0 –6 1 3-6

Nd 6

Dst Cost Outg

2 3 6-33 1 6-36 0 –

Time t = 120 s: all nodes broadcast their DV messagesTime t = 121 s: nodes finished their computations, result:

Nd 1

Dst Cost Outg

1 0 –2 1 1-23 3 1-26 4 1-2

Nd 2

Dst Cost Outg

1 1 2-12 0 –3 2 2-36 3 2-3

Nd 3

Dst Cost Outg

1 3 3-22 2 3-23 0 –6 1 3-6

Nd 6

Dst Cost Outg

1 4 6-32 3 6-33 1 6-36 0 –


Convergence Time – Summary

• In this example it took 121 s to let node 1 learn the full

network!!

• If we take any DV transmission operation as the beginning

of a round, it took three rounds to converge

• If the network diameter is K hops, it would take K rounds

to converge

• Cure:

• You cannot avoid to have K rounds• You can shorten a round by shortening transmission period• Problem: more DV messages, more overhead


Routing Loops – Example

• Suppose at time t0 routing has

converged, all nodes have

correct routing tables

• At time t1 link 3-6 fails

• At time t2 node 3 updates its

routing table entry D3,6 = ∞

• At time t3 node 2 sends a DV

message to node 3, including

D2,6 = 3

• At time t4 both nodes 2 and 3

perform a routing computation

for all known destinations and

update their routing tables


Routing Loops – Example (2)

Time t0: routing tables after convergence (only routes to node 6):Node 2

Dst Cost Outg

6 3 2-3

Node 3

Dst Cost Outg

6 1 3-6

Time t2: routing tables after node 3 recognized link failure to node 6:Node 2

Dst Cost Outg

6 3 2-3

Node 3

Dst Cost Outg

6 ∞ 3-6

Time t3: node 3 receives the following DV message from node 2

Time t4: routing tables after both node 2 and 3 performed routing computation:Node 2

Dst Cost Outg

6 3 2-3

Node 3

Dst Cost Outg

6 5 3-2

And here we have a loop!!


Routing Loops – Summary

• Routing loops can occur in DV protocols e.g. after link

failure or major increase in a link metric

• In this example, the loop would not have occured if node 3

would have updated its table and transmitted an updated

vector immediately after t1 and before time t3

• However, in a distributed environment such race conditions

cannot be entirely removed

• A more rigorous solution (called Diffusing UpdateAlgorithm) has been incorporated into the EIGRP protocol(see [6, Chap. 3], [3])

• EIGRP = Enhanced Interior Gateway Routing Protocol,

used in the Internet


Routing Loops – Time-To-Live Approach

• In networks where routing loops can occur, packets must

somehow be prevented from circulating forever

• A simple mechanism: time-to-live fields• A packet source adds to its packets a specific header field,

e.g. called time-to-live (TTL)• The TTL parameter either indicates a physical time that the

packet is allowed to circulate in the network, or it indicates

the maximum number of hops that a packet may take• Here we assume that TTL indicates max number of hops• A router behaves as follows:

• It reads the TTL field off from an incoming packet• If the TTL field is one and the router cannot directly reach

the final destination, the packet is dropped• Otherwise, the TTL field is decremented, written back into

the packet (with additional checksum re-calculation, if

necessary) and the packet is forwarded further

• This approach has been adopted by the IP protocol


Count-to-Infinity – Example

(see [5, Sec. 4.5])

• Time t0: network has converged

• Time t1: cost on link A-B increasesfrom 4 to 100, node B detects this


Count-to-Infinity – Example (2)

Time t0: routing tables after convergence:Node A

Dst Cost Outg

A 0 –B 4 A-BC 5 A-B

Node B

Dst Cost Outg

A 4 B-AB 0 –C 1 B-C

Node C

Dst Cost Outg

A 5 C-BB 1 C-BC 0 –

Time t1: node B detects link cost change, and computes new path to A (taking intoaccount his knowledge that C can offer a path of length 5 to A), giving:

Node A

Dst Cost Outg

A 0 –B 4 A-BC 5 A-B

Node B

Dst Cost Outg

A 6 B-CB 0 –C 1 B-C

Node C

Dst Cost Outg

A 5 C-BB 1 C-BC 0 –

We have a routing loop now!!

Time t2: node B informs node C via DV message that its new costs to A is 6, node Cre-calculates costs and route to A (which is via B), as:

Node A

Dst Cost Outg

A 0 –B 4 A-BC 5 A-B

Node B

Dst Cost Outg

A 6 B-CB 0 –C 1 B-C

Node C

Dst Cost Outg

A 7 C-BB 1 C-BC 0 –


Count-to-Infinity – Example (3)

• Continuation:

• Node C informs node B about its new cost (which is now 7)

and subsequently node B re-calculates its cost to 8• Node B informs node C about its new cost (which is now 8)

and subsequently node C re-calculates its cost to 9• and so on, and so on• The procedure stops when node B announces costs of 50,

then leading C to adopt the direct link C-A to C

• This behaviour is known as the count-to-infinity problem

• One cure:

• Split-horizon approach: when transmitting a DV message

on a link, include updates only for nodes for which the link

is not the next-hop link!• Convince yourself that this solves the present example• But: it does not solve the problem in general!


Another Problem with DV Protocols

• Suppose you are an evil or incompetent person and have

root access to a router running a DV protocol

• Can you imagine a way in which, by sending well-formed

DV messages, you can corrupt routing in parts of the

network?


Outline

Introduction






Outline

Introduction




Protocol Operation



The Dijkstra Algorithm


/ / i n i t i a l i z a t i o nS′ = V \ sDs,s = 0 ; pred [ s ] = s

f o r a l l d ∈ S′ doDs,d = ds,d ; pred [ d ] = NULL ;

when ds,d < ∞pred [ d ] = s

/ / main loopwhile S′ 6= ∅ do

k =m∈S′ Ds,m

S′ = S′ \ kfor j ∈ Nk do

when Ds,k + dk,j < Ds,j

Ds,j = Ds,k + dk,j

pred [ j ] = k


Basic Idea

• The Dijkstra algorithm needs the di,j as its main input

• Stated differently – the Dijkstra algorithm must know:

• the links (i.e. their start and end node)• and their costs / their state (up, down)

which is equivalent to knowing the whole topology

• In a distributed environment, this is not known a priori to a

node, but has to be discovered by use of a protocol

• Since such a protocol exchanges link information or link

states, they are known as link-state protocols

• Individual nodes then maintain a link-state database,

collecting all the di,j values


Some Assumptions

• Dissemination of link states is based on flooding, which isassumed to be reliable

• Means: each node in the network eventually gets the

information• But no guarantees are given as to when this happens• (Efficient) Implementation of reliable flooding is challenging

• Upon receiving new link-state information, a node performs

the shortest-path computation (e.g. using Dijkstra) locally,

based on its current link-state database

• A node has a mechanism to detect the costs and cost

changes of its outgoing links


Link State Advertisement (LSA) Messages

• SrcNode identifies the source of a link (its one end)

• DstNode identifies the destination of a link (its other end)

• Alternatively, we give a LinkID of the kind 1 7→ 2 to specify link

• Cost identifies the current link cost di,j

• Seqno is a sequence number

• TTL indicates the remaining time the LSA is valid (different meaning as in TTLmechanism discussed before)

• LSA messages are generated by each node for each of its outgoing links, theseare flooded reliably

• An example:


LSA Generation and Database Maintenance

• A node generates and floods LSA messages both:• Periodically, and• Upon topology / cost changes (triggered update)

• Periodic updates:• Period should be ≤ than the initial value of TTL field• A node receiving an LSA stores received data in link-state

database, recomputes routes and initializes a timer with the

value of the TTL field for this link• When this timer reaches 0:

• the link-state information is purged from the database• routes are re-computed without the purged link• neighbors are informed (flooding!) with special LSA with TTL

set to 0, forcing them to purge link-state information as well

• The chosen period has significant impact on overhead!!

• Updates upon topology / cost changes:• Whenever a node detects a change in cost / a link failure / a

new link, it immediately floods an LSA


The Need for Sequence Numbers

(see [6, Fig. 3.10])

• Suppose that LSA’s consist only ofSrcNode, DstNode, Cost

• Time t0: node 1 generates LSA[1, 1 7→ 2, 1] and sends this to nodes2 and 4

• Time t1: link 1 7→ 2 fails, node 1generates new LSA [1, 1 7→ 2,∞]and sends it to node 4

• Time t2: node 2 forwards LSA[1, 1 7→ 2, 1] to node 4 (as part of theflooding process)

ProblemHow can node 4 tell which LSA is correct?


The Need for Sequence Numbers (2)

• To circumvent this problem, sequence numbers are used

• More precisely:

• A node i maintains for each outgoing link o a separate

sequence number so

• Whenever i notices that the cost of link o has changed, it

sends a new LSA, includes the current value of so in it and

increments so afterwards (for periodic updates so does not

necessarily need to be updated)• Any other node k receiving a LSA for a link i 7→ j stores the

received sequence numbers• When node k receives a LSA from i for i 7→ j , it checks

whether the contained seqno is larger than the stored one

• If so, the new cost is extracted from the LSA and stored• Otherwise, the LSA is dropped

• Note: there is no coupling between sequence numbers at

different nodes (not even when sharing the same link)


The Sequence Number Space should be LARGE!!

• A SeqNo field in a frame is of finite width, say b bits

• Sequence number space is S = 0, 1, 2, . . . , 2b − 1

• At the end of S the numbers wrap around

• For b = 3 we have S = 0, 1, 2, . . . , 7

• Suppose a node receives two LSAs for a link, the first with

SeqNo=7, the other with SeqNo=2• Not clear whether SeqNo=2 was generated before or after

SeqNo=7!!

• To (almost) resolve this ambiguity:

• Sequence number space should be large, e.g. b = 32 or

b = 64, essentially removing need for wraparound• The TTL mechanism purges old link-state information


Sequence Numbers and Node Crashes

• Nodes increment seqno after each LSA transmission

• Nodes can crash

• How should a crashed node choose an initial sequence

number that is “larger” than what is present in the network?

• First approach:• Initialize sequence number to 0• Wait for a time of at least maximum TTL to ensure that all

old information has been purged• Then send a LSA with sequence number 0

• Second approach (Resynchronization):• Ask neighbor for his most recent copy of own LSA message• Extract seqno from this and increment it• Send an updated LSA message with new seqno• A recovered / new node also asks the neighbor for a copy

of its link-state database to be able to compute routes

quickly (i.e. much quicker than maximum TTL parameter)


The Hello Protocol

• One mechanism for checking state of link are periodicpacket exchanges with neighbor

• Hello packets• Link / neighbor is considered dead after subsequent failed

hello exchanges

• Period can be significantly smaller than maximum TTL

• The initial query of a neighbors link-state database is also

considered as part of the hello protocol


Discussion / Comments

• Each node performs route computations independently

• Because of finite propagation speed of link state changes,inconsistent views and routing loops can occur as well

• But: the flooding adopted by LS tends to resolve

inconsistencies much faster than in DV protocols

• LS protocol requires periodic flooding of LSA messages,one for each link

• Significant overhead• Reduce overhead by making period large, so that triggered

updates have major share of all updates


Outline

Introduction






Outline

Introduction






Further Topics


Differences between DV and LS Protocols

• Information maintained by nodes:

• DV: nodes know all other nodes and costs to reach them• LS: nodes know all links and their costs

• Protocol information messages:

• DV: carry per-node information• LS: carry per-link information

• Communication partners:

• DV: nodes talk only to their neighbors• LS: nodes inform the whole network

• Therefore:

• LS has much more overhead than DV!!• LS can propagate new information much quicker, shortens

the time for which inconsistencies can exist!


DV and LS Protocols – Scalability

• Both protocols have scalability issues:

• Nodes have to know all nodes / links in a network• This does not scale well to large networks like the Internet!

• Therefore, realistic protocols like OSPF are hierarchical:

• Network is partitioned into areas• A router in one area needs to know:

• How to reach each node in its own area• How to reach other areas (but not individual nodes in them!)

• Flooding of link states is restricted to own area• The hierarchy can have more levels

• Furthermore: routing is not done for the entire Internet,but only within smaller pieces (autonomous systems)

• Internet is partitioned into autonomous system (AS)• One AS is usually owned by one entity, which decides

about routing within the AS• Routing across different AS handled by BGP


DV and LS Protocols – Reliable Information Exchange

• We have seen that both protocols allow for periods withinconsistent views on the network

• One consequence: routing loops!

• Allowing loss of distance vector messages or LSA

messages would exacerbate consistency problems

• Therefore, many practical DV / LS protocols assume or

define a mechanism for reliable delivery of protocol control

messages (including DV messages and LSA messages)


Outline

Introduction






Further Topics


k Shortest-Path Algorithms

• It is sometimes useful to have not only a single (the best!)connection between source and destination, but several,for example k paths

• One usage: backup paths for fault tolerance• Another usage: load balancing over paths

• You typically do not want some k paths, but the k best ones

• One heuristic approach for finding link-disjoint paths:

• Start w/ full network G0, identify best path P1 (e.g. Dijkstra)• Compute network G1 by removing links of P1 from G0

• Identify best path P2 in G1, and so on . . .

• Variations:

• Remove one link of P1 at a time from G0

• Remove all intermediate nodes of P1 (and their direct links)

from G0, this creates node-disjoint paths


Max-Flow Problems

• Given a source node S and a destination node T , link labels denote capacity

• S wishes to transmit a continuous flow of constant rate to T

• Assume that source can use all links it wishes

• Assume that flows on a link can be split arbitrarily, e.g. a router can split a flow ofrate x to two outputs with rate x/2 each

• What is the maximum rate at which S can transmit data?

• Read about the Ford-Fulkerson algorithm, implement it and use it to solve thisproblem!!


Dimitri P. Bertsekas.Dynamic Programming and Optimal Control – Volume 1.Athena Scientific, Belmont, Massachusetts, 3rd edition, 2005.

Dimitri P. Bertsekas.Dynamic Programming and Optimal Control – Volume 2.Athena Scientific, Belmont, Massachusetts, 3rd edition, 2007.

J. J. Garcia-Luna-Aceves.Loop-free routing using diffusing computation.IEEE/ACM Transactions on Networking, 1(1):130–141, February 1993.

Bernhard Korte and Jens Vygen.Combinatorial Optimization – Theory and Algorithms.Springer, Berlin, third edition, 1005.

James F. Kurose and Keith W. Ross.Computer Networking – A Top-Down Approach Featuring the Internet.Addison-Wesley, Boston, fourth edition, 2001.

Deepankar Medhi and Karthikeyan Ramasamy.Network Routing – Algorithms, Protocols, and Architectures.Morgan Kaufmann, San Francicso, California, 2007.


The Internet IPv4 IP Helper Protocols Bibliography


COSC 264IP and Related Protocols

Dr. Andreas Willig1




UoC, 2014


Outline

The Internet

IPv4

Packet Format

IP Addressing

IP Forwarding and Routing


IP Helper Protocols

ARP

ICMP

DNS


About This Module

• Goals of this Module:

• Get a first idea of the Internet• Get to know the IP protocol and important support protocols

• Useful references:

• The “bible” on TCP/IP: [12] (old, but still great!)• Other references: [4], [11, Part V]• Internet protocols are published as requests-for-comment

(RFC) by the Internet Engineering Task Force (IETF), you

can access them via: http://www.ietf.org/rfc.html

• Most of these slides are based on [12]

http://www.ietf.org/rfc.html


Outline

The Internet

IPv4

IP Helper Protocols


The Internet

• The Internet is a packet-switched network

• It is a network of networks:

• It consists of many different networks, connected by routers• The networks or links can be of any technology:

• Ethernet• Optical point-to-point links• Wireless LAN• . . .• Carrier pigeons (RFC 1149)

• It is really large:

• The Internet Systems Consortium estimates about 900

million stations (called hosts) as of July 2012• See http://www.isc.org/solutions/survey

• It has a fairly complex topology [1]

http://www.isc.org/solutions/survey


The Internet (2)

• The end-to-end principle [10]:

• Most intelligent functions should be performed in end

points, not in the network• For example:

• The network knows how to deliver packets (using routers)• All functions making this delivery reliable are performed in

the end host, e.g. by the TCP protocol• There is no network-layer mechanism for reliable delivery• Keep the routers simple!

• The Internet is standardized by the IETF, the standards arecalled RFCs

• IETF = Internet Engineering Task Force (www.ietf.org)• RFC = Request For Comment

• For the design philosophy see [3]

www.ietf.org


The Hourglass Model for the Internet Protocol Stack

• “Everything over IP, IP over everything”


Outline

The Internet

IPv4

IP Helper Protocols


Introduction

• IP is specified in RFC 791 and many followup RFCs

• It is the network layer protocol of the Internet

• Some terminology:

• IP packets are called datagrams• End stations are called hosts• IP routers are called routers

• IP addresses are assigned to network interfaces:

• When a host has three Ethernet adapters, it has three IP

addresses, one for each adapter• Since most hosts have only one adapter, we speak of the IP

address of that host


IP Service – Best Effort

• Basic IP service is datagram delivery

• This service is:

• Connectionless: no connection or shared state is set up

before datagram delivery starts• Unacknowledged: IP does not use acknowledgements• Unreliable: on IP level no retransmissions are carried out• Unordered: IP does not guarantee in-sequence delivery [2]

• This kind of guarantee-nothing service is called best effort


Outline

The Internet

IPv4

Packet Format

IP Addressing



IP Helper Protocols


Packet Format


Packet Format (2)

• Where applicable (e.g. addresses), header is using big

endian byte ordering (also called network byte order)

• The HdrLen field:

• specifies the length of IP header as number of 32-bit words• If the Options field does not use a multiple of 32 bits, a

Padding field is used to fill up to 32 bits• When HdrLen > 5, then an Options field is present

• The TOS/DSCP field:

• TOS = Type Of Service, DSCP = DiffServ Code Point• Allows to mark packets for differentiated treatment to

achieve Quality-Of-Service (QoS), e.g. express priorities• DiffServ [6] is a framework for Internet QoS, another one is

IntServ [13]• Most routers ignore the TOS/DSCP field


Packet Format (3)

• The TotalLength field:

• Gives the total length of datagram in bytes (i.e. up to 65535)• Can be modified during fragmentation and reassembly• The TotalLength field is part of IP header, since some

technologies (Ethernet!) pad up frames to achieve

minimum frame size and do not reverse this

• The Identification field:

• Uniquely identifies each datagram sent by host / interface• Incremented by source host before sending new datagram• Routers do not touch this field

• The Flags field:

• Contains two flags relevant for fragmentation and

reassembly (DF, Don’t Fragment, and MF, More Fragments)


Packet Format (4)

• The FragmentOffset field:

• is used for fragmentation and reassembly• gives the offset of the current fragment within entire

datagram, in multiples of eight bytes

• The HeaderChecksum field:

• Is calculated over IP header only, not the data (TCP, UDP

etc. have their own checksums to cover their data)

• The Time-To-Live field:

• gives upper limit to number of routers a packet can traverse• decremented by each router, forces re-computation of

header checksum• when TTL reaches one and packet cannot be directly

delivered to destination, datagram is discarded and sender

is notified with ICMP message• Typical initial values: 32 or 64


Packet Format (5)

Protocol field Protocol

0x01 ICMP

0x02 IGMP

0x04 IP-in-IP Encapsulation

0x06 TCP

0x11 UDP

• The Protocol field indicates thehigher-layer protocol that generatedthe payload

• This field provides protocolmultiplexing

• Some values shown in table


Packet Format (6)

• The SourceAddress/DestinationAddress fields:

• SrcAddr indicates the initial sender of datagram• DstAddr indicates intended final receiver of datagram• Are of 32 bits width

• The Options field:

• Contains header field for optional IP features• One example option: source routing• Options are rarely used, we will not consider this anymore


Outline

The Internet

IPv4

Packet Format

IP Addressing



IP Helper Protocols


IP Address Representation

• IP addresses have a width of 32 bits

• They are supposed to be worldwide unique

• This is not really true anymore with NAT . . .

• IP addresses are written in dotted-decimal notation, e.g.:

130.149.49.77

where decimal (!) numbers are separated by dots

• They have an internal structure:

<network-id> <host-id>

where:

• <network-id> denotes a network (e.g. an Ethernet)• <host-id> refers to a host within this network

• The <host-id> must only be unique w.r.t. its network


Important Points

Important Point

A host address is tied to its location in the network, i.e. it is

coupled to network topology. When a host switches to another

network, it obtains another address and ongoing connections

(TCP!) are disrupted – IP therefore has no direct support for

mobility!!

Important Point

IP Routing is mostly concerned with networks, i.e. forwarding

tables in routers mostly store <network-id>’s – it is the

responsibility of the last router on a path to deliver an IP

datagram to a directly connected host.


Classful Addressing

• Initially so-called classful addressing has been used

• IP-addresses are subdivided into four classes:

• Class-A addresses: 7 bit network-id, 24 bits host-id, i.e. 128

class-A networks with maximum 224− 2 ≈ 16.7 million

hosts in each network

• Class-B addresses: 14 bit network-id, 16 bit host-id

• Class-C addresses: 21 bit network-id, 8 bit host-id

• Class-D addresses: 28 bits multicast group address


Classful Addressing (2)

• For each network-id there are two special host-ids:

• Host-id with all zeros refers to the network as such• Host-id with all ones is the broadcast address of this

network

• Example:

• 130.149.0.0 refers to a class-B network• 130.149.255.255 is broadcast address of this network• 130.149.49.123 refers to a particular host in that network


Classful Addressing – Discussion

• These three classes support networks of few distinct sizes

• Problems:

• With the growth of the Internet class-B addresses were

quickly exhausted, but many of the requesting

organizations do not really have 65534 hosts, these often

were poorly utilized• With the allocation of class-C addresses the routing tables

in Internet core routers quickly became large, which slows

down packet processing!


Subnetting

• Suppose that an organization:

• has a class-B network address, say 130.149.0.0• has its network internally divided into several LANs,• and wants to couple these by routers

• First option: allocate a class C address for each LAN

• requires additional addresses• increases size of routing tables in core routers

• Second option: use class-B address externally, subdividethis network internally

• The whole class-B network is seen by all external networks

only through one border router and one IP address• All internal networks are allocated addresses like

130.149.x.0 with an eight-bit host part• All internal routers and the border router know the internal

network structure and the networks


Classless Inter-Domain Routing

• CIDR = Classless Inter-Domain Routing

• Introduced 1993, specified in RFCs 1518, 1519, 4632

• Goal was to address the problems of classful addressing

by giving more fine-grained network sizes

• CIDR runs in conjunction with more modern routing

protocols like OSPF, RIP-2 or BGP

• In CIDR a network is specified by two values:

• A 32 bit network address• A 32 bit network mask (netmask)


CIDR – Netmask

• For a given 32-bit IP address the netmask specifies which

bits belong to the network-id and which bits belong to the

host-id

• The netmask consists of 32 bits, the left k bits are ones,

the remaining bits are zeros

• Examples:

Netmask Shorthand

11111111.11110000.00000000.00000000 /12

11111111.11111111.00000000.00000000 /16

11111111.11111111.11100000.00000000 /19

11111111.11111111.11111110.00000000 /23


CIDR – Netmask (2)

• Example: we are given the host address 192.168.40.3

and the netmask /24, then the hosts network address can

be computed as:

11000000.10101000.00101000.00000011 192.168.40.3

AND 11111111.11111111.11111111.00000000 /24

11000000.10101000.00101000.00000000 192.168.40.0

• The same example, now with netmask /21:

11000000.10101000.00101000.00000011 192.168.40.3

AND 11111111.11111111.11111000.00000000 /21

11000000.10101000.00101000.00000000 192.168.40.0

• In both examples the network addresses are the same, but

the networks are of different size

• To fully specify a network, one gives both network address

and netmask, e.g.:

192.168.40.0/21


CIDR – Netmask (3)

• In the network 192.168.40.64/28 there are 14addresses available:

• The netmask leaves four bits for the host-id, i.e. 16 values• The value 0000 is part of the network-id• The value 1111 is the broadcast address for this network


Supernetting

• Suppose an organization has allocated 16 networks of size/24 with contiguous network addresses, e.g.:

• 130.149.64.0/24

• 130.149.65.0/24

• . . .• 130.149.79.0/24

• With supernetting:

• these networks are summarized under the network address

130.149.64.0/20

• Routers outside any of these networks only have an entry

for 130.149.64.0/20 instead of 16 entries

• Can you figure out the formal conditions under which

supernetting is allowed?


Reserved IP address blocks

Address Block Current Usage

10.0.0.0/8 Private-use IP networks127.0.0.0/8 Host loopback network169.254.0.0/16 Link-local for point-to-point links (e.g. dialup)172.16.0.0/12 Private-use IP networks192.168.0.0/16 Private-use IP networks

(from: [8])

• Private-use IP addresses are often used for broadband

clients or by NAT boxes

• The “traditional” loopback address of a host is

127.0.0.1, but any address from the 127.0.0.0/8

network serves the same purpose

• Packets with private addresses are not routed in the public

internet, only within the provider network


Outline

The Internet

IPv4

Packet Format

IP Addressing



IP Helper Protocols


Simplified Packet Processing


Simplified Packet Processing (2)

• Packet processing chain is followed in routers and hosts• Incoming packets are checked for correctness and stored

in IP input queue – correctness includes:• right value in IP version field• correct IP header checksum

• Next, packet options are processed:• Options are rarely used• Special case: source routing option, then packet is

delivered to IP output stage

• Next, it is checked if packet is destined to this host / router

or to broadcast address• If so, protocol demultiplexing is carried out

• The protocol field in IP header is checked for its value• Packet payload is delivered to the software entity

implementing the indicated higher-layer protocol• Packet is not processed any further!


Simplified Packet Processing (3)

• If the packet is not destined to this host/router:• If packet forwarding is not enabled, the packet is dropped• Otherwise:

• Check if packet is destined to a directly reachable station

(e.g. on same Ethernet) – if so, deliver packet directly• If packet is not destined to directly reachable station, consult

forwarding table to determine next hop / outgoing interface• Decrement TTL value, drop packet when it reaches zero• Recompute packet header checksum (why?)• Hand packet over to outgoing interface

• Forwarding table is maintained by a routing daemon, i.e.

a process executing a routing protocol

• Note that datagrams to be routed can come from local

applications or from other hosts via IP input queue• Linux commands to inspect / modify forwarding table:

• netstat

• route


Forwarding Table Contents

• Each entry in the forwarding table contains:• Destination IP address, which can be either:

• a full host address (i.e. non-zero host-id)• a network address, with netmask

depending on the value of a flag• Information about next hop, either:

• IP address of next-hop router (must be directly reachable)• IP address of directly-connected network (network address)

• Flags:

• A flag telling whether destination IP is host or network• A flag telling whether next hop is a router or directly attached

network

• Specification of outgoing interface


Forwarding

• From forwarding table structure it is clear that a host /

router does not know the full path, but only next hop

• Forwarding table lookup for a packet with destination IPaddress dst proceeds in three stages:

• First look for an entry that is a full-host address matchingdst – if found, send packet to indicated next hop / outgoinginterface and stop processing

• This is not used very often

• Next look for an entry that is a network address matching

dst – if found, send packet to indicated next hop / outgoing

interface and stop processing• Finally look for special default entry – if found, send packet

to indicated next hop (the default router) and stop

processing• Otherwise drop packet, send ICMP message back to

original sender of datagram


Forwarding Tables in Hosts

• Most end hosts leverage the default route mechanism:• An end host can differentiate between packets to local

destinations and to all other destinations

• Question: suppose an end host has address

130.149.49.77 and is part of a /24 network – how does it

check whether a destination address dst belongs to another

host in the same network?

• Packets to local destinations are delivered directly• Packets to all other destinations are sent to default router

• Therefore, forwarding tables in end hosts can be made outof just two entries:

• One entry for the local network• The default route

• The default route must be configured


Forwarding Tables in Routers

• Most routers at the “border” of the Internet only have

routing tables for a subset of all networks attached to the

Internet, for all other networks they rely on default routers

• Some routers in the core:

• do not have a default router• are the default routers of other routers• must know (almost) all the Internet networks


Routing – Interior and Exterior Gateway Protocols

• Applying any distance-vector or link-state routing protocolto the whole Internet is hardly feasible

• Routing tables in routers would become too large• Internet dynamics would keep routing protocols busy all the

time (triggered updates)

• The Internet therefore is subdivided into autonomoussystems (AS):

• An AS is administered by one authority• An AS has a unique 16-bit identifier• Examples: a University campus, a corporation

• Routing protocols that route . . .

• within an AS are called interior gateway protocols• across AS are called exterior gateway protocols


Routing – Interior Gateway Protocols

• An AS can choose any of the interior gateway protocols to

determine routes between routers in the same AS

• Some interior gateway protocols:• RIP and RIP-2 (Routing Information Protocol):

• Defined in RFCs 1058 and RFC 1388• Both are distance-vector protocols, metric is hop-count• RIP-2 contains improvements to address DV problems

• EIGRP (Extended Interior Gateway Routing Protocol):

• Defined in [5], used in CISCO routers• Loop-free distance-vector protocol

• OSPF (Open Shortest-Path First):

• Defined in RFC 1247• It is a link-state protocol

• Note that all these protocols use shortest-path algorithms


Routing – Exterior Gateway Protocols

• Exterior gateway protocols (EGP) are used between

routers belonging to different AS

• Major EGP: BGP (Border Gateway Protocol)

• Defined in RFCs 1267 and 1268• BGP-V4: RFC 4271

• A BGP router stores for networks in foreign AS a numberof pathes towards this network:

• Such a path lists all intermediate AS, not individual routers• These paths are not determined based on costs only, but

also based on a policy• Policies are specified based on political or economical

considerations


Outline

The Internet

IPv4

Packet Format

IP Addressing



IP Helper Protocols


On the Choice of Packet Size

• The link-layer technologies underlying IP offer manydifferent maximally allowed packet sizes, e.g.:

• Ethernet: 1500 bytes• Gigabit Ethernet: 9000 bytes• IEEE 802.11 WLAN: 2312 bytes• ISDN: 576 bytes

• This maximum size is also known as maximum

transmission unit (MTU)

• Higher-layer protocols (TCP, UDP) and applications shouldnot be required to know these maximal sizes:

• One reason: “software hygiene”, separation of concerns• Another reason: it is not well defined:

• Different packets of the same flow can take different routes• A packet can use different technologies while on travel• Even if all packets go the same route, this route can change

due to link failures / restores



• IP hides this from upper layers, offers own maximummessage length of 65515 bytes to higher layers

• 65515 = 65535 - 20, 20 bytes is minimum size of IP header

• To cope with smaller MTUs:

• Sender IP instance partitions message into fragments• Fragment size is chosen as MTU of outgoing link• Each fragment is transmitted individually as a full IP packet,

with header information specifying that this is a fragment

and giving the position of fragment in whole message• IP instance at destination buffers received fragments,

re-assembles message and delivers to higher layers

QuestionWould it be useful to have intermediate IP routers perform

reassembly?


Fragmentation and Reassembly (2)

• In addition, every intermediate router can:

• fragment a full message• further fragment a fragment

when necessary for transmission on next hop

• When the destination receives the first fragment, it:

• Allocates buffer large enough for whole message• Starts a timer

• When all fragments arrive before timer expiration:

• Timer is canceled• Re-assembled packet is handed over to higher layers• Buffer is de-allocated

• When timer expires before all fragments have arrived:

• The already received fragments are dropped, buffer is freed• ICMP message (type 11, code 1) is sent to source host


Some Details

• Every message handed over to IP from higher layers hasits own identifier

• See identification field in IP header

• All fragment datagrams belonging to same message have:

• A full IP header• The same identification field• A TotalLength field reflecting the fragment size• Different values for FragmentOffset field (reflecting the

start of the present fragment within the whole message):

• FragmentOffset specifies offset in multiples of 64 bits

• The MF (more-fragments) bit set, except for the last

fragment, which has non-zero FragmentOffset


Some Details: The DF bit

• By setting the DF (don’t fragment) bit in the IP header a

source node forbids fragmentation by intermediate routers

• When a router receives a datagram with DF set, it:

• Checks whether outgoing link for this packet has an MTU

large enough to transmit the packet• If so, the packet is transmitted onto next hop• If not, the router drops the datagram and returns an ICMP

datagram to original IP source

• ICMP with type 2 (“destination unreachable”) and code 4

(“fragmentation required, but DF set”)


Some Details: The DF bit (2)

QuestionHow could you use this for the sender to determine the path

MTU, defined as the smallest MTU of all links along a path

between source and destination?


Fragmentation and Reassembly – Discussion

• Fragmentation/Reassembly creates significant overhead:

• Several datagrams transmitted per message, each one

having full IP header• Reassembly adds significant complexity to receiver• Upon loss of single fragment the whole message is possibly

re-transmitted by higher layers (TCP!)

• Fragmentation and reassembly complicates operation ofapplication-level firewalls, since these also must implementreassembly logic

• Application-level firewalls look at user data of packets• When user data is spread over several fragments, it must

collect them all

• Exception: the part of user data that is of interest is known to

fit in the first fragment


Outline

The Internet

IPv4

IP Helper Protocols


Outline

The Internet

IPv4

IP Helper Protocols

ARP

ICMP

DNS


Address Resolution Protocol – ARP

• IP addresses only have a meaning to IP and higher layers

• In an Ethernet, stations have own 48-bit MAC addresses

• An Ethernet station picks up a packet only if the destination

MAC address matches its own MAC address (ignoring

broadcast/multicast), IP addresses and other packet

contents are ignored

• An IP address is assigned to an Ethernet adapter

Important Question

How do other stations know to which MAC address a given IP

address refers, i.e. to which station an IP packet must be sent

(encapsulated in Ethernet packet)?


Address Resolution Protocol – ARP (2)

• ARP provides a binding service: it determines MAC

address for given IP address

• ARP is specified in RFC 826

• ARP is not restricted to Ethernet MACs, but in general is

geared towards LANs with broadcast capabilities

• ARP is dynamic:• The MAC address for a given IP address does not need to

be statically configured, but the protocol provides a

mechanism to determine this on-the-fly• Advantage: nodes can be moved or equipped with new

MAC adapters without any re-configuration• Disadvantage: a separate protocol is needed, bringing

additional complexity and requiring some bandwidth

• There is also a protocol that lets stations find an IP address

for given MAC address, this is called RARP (Reverse ARP)


Basic Operation of ARP

• Suppose that:

• We have two stations A and B attached to the same

Ethernet, having the following addresses:

Station A Station B

MAC 11:11:11:11:11:11 22:22:22:22:22:22

IP 130.149.49.11 130.149.49.22

• Both A and B are in the same IP network

130.149.49.00/24, which is an Ethernet network• Station A wishes to send an IP packet to address

130.149.49.22 and does not yet have any information

about the corresponding MAC address

• Each station maintains an ARP Cache, which stores the

mappings from IP to MAC addresses that the station

currently knows about


Basic Operation of ARP (2)

• Station A broadcasts an ARP-request message(displayed in wireshark as arp who-has), indicating:

• A’s own IP and MAC address• B’s IP address

Broadcasting means: packet is sent to Ethernet

broadcast address!!

• Any host C having an IP address other than

130.149.49.22 simply drops the ARP-request packet• Upon receiving the ARP request, host B (with IP address130.149.49.22) performs the following actions:

• It stores a binding between between A’s IP and MAC

address in its own ARP cache• It responds with an ARP-reply packet that includes:

• B’s MAC and IP address• A’s MAC and IP address

ARP reply is unicast to A’s MAC addr. (Why no broadcast?)


Basic Operation of ARP (3)

• Upon receiving ARP response from B, station A stores a

binding between B’s IP and MAC address in its ARP cache

• This procedure is called address resolution

• ARP does not make any retransmissions in case the ARP

request is not answered, this is left to higher layers

• If a station wants to send an IP packet to a localdestination with address xx.xx.xx.xx, it:

• first checks the ARP cache whether a binding for

xx.xx.xx.xx can be found• If so, the packet is encapsulated in an Ethernet frame and

directed to the MAC address found in the ARP cache• Otherwise, the address resolution procedure is started and

the packet is sent when the result is available


The ARP Cache

• The entries in an ARP cache are soft-state, entries aretypically removed 20 minutes after their creation

• Why?• Some implementations restart the timer after each

reference to an ARP cache entry

• Under Linux you can inspect your ARP cache with the

command:

/usr/sbin/arp -a

The path to the arp command can vary between systems


The ARP Frame Format

(See [12, Sect. 4])

• HardType determines the type of MAC addresses used, 0x0001 for Ethernet48-bit addresses

• ProtType determines the higher-layer protocol for which address resolutionneeds to be done, value 0x0800 for IP

• HardSize and ProtSize specify the size (in bytes) of the hardware and andprotocol addresses – they are 6 and 4 for Ethernet and IP

• Op distinguishes between ARP-request and ARP-reply, and some other types(RARP is covered as well)

• The remaining four fields are the mentioned address fields


Outline

The Internet

IPv4

IP Helper Protocols

ARP

ICMP

DNS


Introduction

• ICMP = Internet Control Message Protocol

• Specified in RFC 792

• This protocol:• Accompanies IP by allowing routers or destination hosts to

inform sender about “unusual” situations, including:

• There is no route to the destination• Destination host exists, but is not reachable• Fragmentation required but DF set

• Operates “on top” of IP, i.e. ICMP messages are

encapsulated in regular IP datagrams• Does not add any additional mechanisms (like error control)

to the IP service• IP sending host must not rely on ICMP messages


Message Format

• type and code specify actual ICMP message type and sub-type

• checksum covers ICMP header and data, with checksum assumed as zero


Some type/code Combinations

type code Meaning

0 0 Echo reply

3 0 Destination network unreachable3 1 Destination host unreachable3 2 Destination protocol unreachable3 3 Destination port unreachable3 4 Fragmentation required, but DF bit set3 6 Destination network unknown3 7 Destination host unknown

4 0 Source quench (Congestion control)

8 0 Echo request

11 0 TTL expired in transit11 1 Fragment reassembly time exceeded

• There are many more, e.g. for router advertisements, information aboutmalformed IP packets, etc.

• It is implementation-dependent, which ICMP messages are generated

• ICMP messages are often suppressed by firewalls, otherwise too muchinformation about internal network structures could be revealed


Some type/code Combinations (2)

• Source-quench (type=4, code=0):

• generated by an IP router when it has to drop a packet

because of congestion• Intention is to let source host throttle its rate

• TTL expiration (type=11, code=0):

• generated by an IP router when it drops a packet because

its TTL value reached zero

• Fragment reassembly timeout (type=11, code=1):

• Generated by destination when not all fragments of a

message have been received within timeout• Used to invite higher-layer protocol at sending host to

re-transmit message• IP itself does not perform any retransmission!


Some type/code Combinations (3)

• The “destination-unreachable” messages (type=3):• code=0 (destination network unreachable) and code=1

(destination host unreachable): generated when:• router finds that the cost to reach a non-directly connected

host are infinity (e.g. are link failure)• router could not deliver datagram to directly connected host

• code=2 (protocol unreachable): IP datagram refers to

non-existent higher-layer protocol in destination (compare

protocol-type field)• code=3 (port unreachable): used with TCP / UDP• code=6 (destination network unknown) and code=7

(destination host unknown): generated when:• a router could not determine a next-hop to a non-directly

connected host or network

• In these messages first 32 bits of the variable ICMP

message part are 0, following bytes contain IP header and

first few bytes of offending IP datagram


Outline

The Internet

IPv4

IP Helper Protocols

ARP

ICMP

DNS


Names vs. Addresses

• Names denote / refer to “things”

• In general: persons, cats, ships, . . .• In networks: nodes, networks, data, transactions, . . .• Often, but not always: names are unique

• Addresses: information needed to find these things

• Street address, IP address, MAC address• Often, but not always, unique• Addresses often have hierarchical structure to support their

intended use, e.g. in routing protocols

• Binding services: map names to addresses or vice versa(also called name resolution)

• Example: DNS maps www.canterbury.ac.nz to

132.181.2.23

• See [9, 7] for more about names and addresses

www.canterbury.ac.nz


The Domain Name Service (DNS)

• Initial specifications in RFC 1034 and RFC 1035

• DNS is responsible for mapping human-readable names to

addresses, it is a binding service

• DNS is used solely by applications, it has no role in the

TCP/UDP/IP protocols themselves• It has a distributed implementation:

• It consists of several name servers, which assist end hosts

in mapping a name to an address• No name server has the full knowledge of all bindings that

exist in the Internet• Besides mapping names to IP addresses it has additional

functions:• It allows to return an email server address for a given host• It allows to manage alias names for hosts

• It is also possible to perform reverse lookup, i.e. mapping

IP addresses to names


The DNS Name Space



The DNS Name Space (2)

• The name space is hierarchical:

• Arranged as a tree made of nodes• Each node has label of up to 63 characters• The domain name of any node is the (unique) list of all

labels that connect it with the unnamed root• All immediate children of a node must have distinct names

• In the written representation a full host name is

represented by its name, followed by its domain, all labels

are separated by “.”

• Example:

www.canterbury.ac.nz

Here:

• www is the host name• canterbury.ac.nz is its domain name


DNS – Zones

• A zone is a sub-tree of the namespace that isadministered separately

• Example: ac.nz

• A zone can be sub-divided into further zones, e.g. therecould be zones:

• canterbury.ac.nz

• massey.ac.nz

• For each zone multiple nameservers must be provided by

the administrative owner of that zone


DNS – Nameservers

• A nameserver keeps a table of all name 7→ IP-addressmappings in a zone

• When new host is added, administrator allocates name and

IP address and enters them into table• When host is removed, table entry is deleted as well

• There are primary nameservers and secondarynameservers:

• These are independent and redundant servers• Reason: fault tolerance• A primary nameserver reads the mapping table from a file• A secondary nameserver reads mappings from primary

nameserver (zone transfer)• Secondary nameservers update their tables regularly

against a primary nameserver

• A nameserver can handle several zones


DNS – Client Side

• Applications in hosts are DNS clients

• A DNS resolver library is linked to an application

• Under UNIX:

• see man page for gethostbyname for a C binding to the

resolver• nslookup is a command-line version of the resolver• The resolver reads a configuration file (often found under

/etc/resolv.conf, which contains a line like

nameserver 130.149.14.12

• The resolver uses the nameserver(s) specified in

/etc/resolv.conf to perform name resolution


DNS – Query Process

• The host hands over a name to its local resolver

• Example: www.canterbury.ac.nz

• The resolver library sends a request to its nameserver

• The nameserver:

• Checks if the requested name is in its zone table• If so, it returns a response to the resolver, which includes

the name 7→ IP-address binding• Otherwise, it contacts a root name server

• Currently there are 13 known root servers• The nameserver must know IP addresses of all of them as

part of its configuration

• The root server returns name and address of a nameserverresponsible for the top-level domain of the request

• Here: nz


DNS – Query Process (2)

• Continuation:

• It next connects to the nameserver for zone nz, which

returns name and address of the nameserver for zone

ac.nz

• It next connects to the nameserver for zone ac.nz which

returns name and address of the nameserver for zone

canterbury.ac.nz

• It next connects to the nameserver for zone

canterbury.ac.nz which then returns the IP address for

host www.canterbury.ac.nz


DNS – Caching

• A nameserver is required to store a name 7→ IP-addressmapping for a time that is indicated in the response of thefinal nameserver

• Caching

• Reason: when same name needs to be resolved shorttime later, it is not necessary to again involve allnameservers, the query can be handled from cache

• Load reduction on root name servers


David Alderson, Lun Li, Walter Willinger, and John C. Doyle.Understanding Internet Topology: Principles, Models and Validation.IEEE/ACM Transactions on Networking, 13(6):1205–1218, December 2005.

Jon C. R. Bennett, Craig Partridge, and Nicholas Shectman.Packet Reordering is Not Pathological Network Behaviour.IEEE/ACM Transactions on Networking, 7(6):789–798, December 1999.

David D. Clark.The design philosophy of the darpa internet protocols.ACM Computer Communication Review, 18(4):106–114, August 1988.

Douglas E. Comer.Internetworking with TCP/IP – Principles, Protocols and Architecture, volume 1.Prentice Hall, Englewood Cliffs, New Jersey, third edition, 1995.

J. J. Garcia-Luna-Aceves.Loop-free routing using diffusing computation.IEEE/ACM Transactions on Networking, 1(1):130–141, February 1993.

Kalevi Kilkki.Differentiated Services for the Internet.Macmillan Technical Publishing, Indianapolis, 1999.

James F. Kurose and Keith W. Ross.Computer Networking – A Top-Down Approach Featuring the Internet.Addison-Wesley, Boston, third edition, 2001.


Deepankar Medhi and Karthikeyan Ramasamy.Network Routing – Algorithms, Protocols, and Architectures.Morgan Kaufmann, San Francicso, California, 2007.

J. H. Saltzer.Naming and binding of objects.In R. Bayer, R. M. Graham, and G. Seegmüller, editors, Operating System – AnAdvanced Course, Lecture Notes in Computer Science, pages 99–208. Springer,1978.

Jerome H. Saltzer, David P. Reed, and David D. Clark.End-to-end arguments in system design.ACM Transactions on Computer Systems, 2(4):277–288, November 1984.


W. Richard Stevens.TCP/IP Illustrated Volume 1 - The Protocols.Addison-Wesley, Boston, Massachusetts, 1995.

Paul P. White and Jon Crowcroft.The integrated services in the internet: State of the art.Proceedings of the IEEE, 85(12):1934–1946, December 1997.

Computer Communication Networks

Documents

switching element advantage

network design

n stations

excursion bibliographyabout

number n of stations

network providers networks

links toenable communication

meshed network