11th March 2010 1

Physical layer network coding for next generation

wireless broadband

Alister Burr, University of Yorkagb1@ohm.york.ac.uk

Agisilaos Papadogiannis, Chalmers University

11th March 2010 2

Outline

The challenge of next generation wireless networks

Next generation network architectures MIMO and MIMO cellular Multi-user and Network MIMO Physical layer network coding Conclusions

11th March 2010 3

Wireless networks – some history

Began with Marconi in 1890’s … 1st generation

analogue, telephony ~1980 (Japan) 2nd generation

digital, some data ~1992 3rd generation

CDMA, flexible services, up to 384 kbit/s ~2002 4th generation

OFDM(A), full Internet access, up to 1 Gbit/s ???

11th March 2010 4

4G

The “next generation” has been discussed ever since 3G standards were finalised a decade ago however it was not initially clear what form the

“fourth generation” might take However starting from 2002 ITU-R has defined the

requirements for IMT-Advanced which has since then been generally accepted as

the definition of 4G Key requirement is 100 Mbit/s for high mobility and

1 Gbit/s for low mobility Standards currently under development:

Mobile WiMAX (IEEE 802.16e/m) 3GPP LTE-Advanced

11th March 2010 5

The dream

To provide full Internet connectivity to everyone, anywhere which means wirelessly

Next generation wireless research has usually focussed on a ‘headline’ maximum data rate but of course this will not be the rate most users

experience, and probably is not the most important figure In densely-populated cities a network for everyone

must provide extremely high capacity densities

11th March 2010 6

Required capacity density

Average population density in European cities ranges from 3400 - 5400/km2

however in commercial district in working hours it will be much higher

say 8000/km2

Suppose 10% subscribe, and 20% of those require access at busy hour

Expected data rate 5 Mbit/s8000/km2 10% 20% 5 Mbit/s = 800 Mbit/s/km2

11th March 2010 7

Capacity density of 1G

e.g. AMPS, U.S.A: ~400 channels in each direction ~15 km radius cells 700 km2

re-use ~1:10 0.06 channels/km2, equivalent to approx 1.8 kbit/s/km2

in approx 50 MHz

11th March 2010 8

Current and 4G systems

Currently one base station serves about 1km2

4G bandwidths proposed are ~ 40 MHz Best available bandwidth efficiency averages

about 2 bits/s/Hz across cell hence capacity density is 80 Mbit/s/km2

assumes 100% frequency re-use We need an order of magnitude more! 10 more bandwidth unlikely to be available

11th March 2010 9

Increased density

We can also increase number of cells

BUT need many

more cell sites interference

11th March 2010 10

Cell edge problem

We may be able to increase bandwidth efficiency (bits/s/Hz) use (e.g) advanced MIMO

techniques BUT mobile close to cell

edge suffers interference from adjacent cells

Conventionally we reduce frequency re-use but this reduces available

bandwidth by factor 3 or more

11th March 2010 11

BuNGee

University of York is part of a European project tackling these problems Beyond Next Generation Mobile Broadband

(BuNGee) Proposes:

new hierarchical network architecture based on wireless backhaul

Advanced MIMO techniques for high bandwidth efficiency

Self-organising network for optimal spectrum use

Goal: 1 Gbit/s/km2

11th March 2010 12

Outline

The challenge of next generation wireless networks

Next generation network architectures MIMO and MIMO cellular Multi-user and Network MIMO Physical layer network coding Conclusions

11th March 2010 13

Achieving capacity density

Will probably need a combination of the approaches mentioned: More spectrum Improved bandwidth efficiency

especially increased use of MIMO Increased frequency re-use (100%) Reduced cell size

requires low base station installation cost, and a cost-effective backhaul network

11th March 2010 14

Wireless backhaul

Simple comparison with 4G proposals suggests we may need ~10 BSs per km2!

We believe that the only cost-effective way to provide this is by wireless backhaul

However must allow for spectrum used by backhaul links

Hence must minimise backhaul load

11th March 2010 15

BuNGee architecture

11th March 2010 16

Some figures

Assume HBS serves 1 km2

Assume total 40 MHz available 20 MHz for MS-ABS (access links); 20 MHz ABS-HBS (backhaul)

Assume average 2 bits/s/Hz across cell Then capacity per ABS = 20 2 = 40 Mbit/s No. ABS per HBS = 1 Gbit/s / 40 Mbit/s = 25 Area served by ABS = 1 km2/25 = 40 000 m2, or

200m square

11th March 2010 17

Wireless mesh network

Since cells are very small, mobile (MS) may be

served by more than one ABS

MSs now served by optimum combination of available ABSs

Practically abolishes concept of cells!

Overall network looks more like a wireless mesh network

HBS

ABS

MS

Relay

MS

ABS

ABS

11th March 2010 18

Wireless relaying

A cell can be extended by adding fixed, or infrastructure relays very similar

architecture to wireless backhaul

with relays replacing ABSs

may allow direct connection of MSs to hub

RelayRelay

11th March 2010 19

Hierarchical wireless network

A generalised framework for network architectures involving wireless backhaul and/or relaying

we might allow: more than one layer of

relays direct connections between

nodes on the same level MSs to connect to different

relay levels Again, similar to mesh

network in structure

11th March 2010 20

Outline

The challenge of next generation wireless networks

Next generation network architectures MIMO and MIMO cellular Multi-user and Network MIMO Physical layer network coding Conclusions

11th March 2010 21

MIMO link model

MIMO = Multiple Input, Multiple Output i.e. multiple antennas at each end of a link

Input and output signal can be modelled as (1 nT) and (1 nR) vectors, s and r; noise (1 nR): n

channel modelled as a matrix H: element Hij gives propagation between transmit antenna j and receive antenna i

r = H s + n

ENCODER

DECODER

H11

H21

HnR1

s r

Tn

nR

11th March 2010 22

Eigendecomposition

s' UH

n

VHsr

r'

nTnR

n1

n

Multiply by matrices U and V at input and output of channel, where

columns of U,V are transmit and receive eigenvectors of the channel

Then U H VH =

a diagonal matrix with the square roots of the eigenvalues of HHH on the diagonal, and r’i = i s’i + n’i

i.e. we create a set of uncoupled channels, whose power gains are the eigenvalues

Each eigenvector can be treated as a steering vector for antenna array transmit/receive beam patterns

11th March 2010 23

Beamforming model

Another way of viewing MIMO: each input stream corresponds to a beam from the

antenna array towards a multipath signal can create as many such beams as there are antennas hence can transmit up to n = min(nT, nR) beams

provided there are enough multipaths

s' U VH r'

n

11th March 2010 24

MIMO capacity

Capacity and bandwidth efficiency approximately multiplied by no. of streams, n

Slope of curves proportional to n called multiplexing

gain Dramatic capacity

gain!

0 5 10 15 20 25 300

5

10

15

20

25

30

35Rayeigh capacity

SNR(dB)

Cap

acity

(bi

ts/s

/Hz)

nT=4 nR=4nT=2 nR=4

nT=1 nR=4

nT=4 nR=2

nT=2 nR=2

nT=4 nR=1nT=1 nR=1

11th March 2010 25

MIMO in interference

MIMO cellular system also subject to interference Beamformer applies linear weights to maximise

SNIR at output filters signals from different directions to maximise

signal to noise-plus-interference ratio

s

sint dint

d

r x

H

H int

W V

U

Uint

matched filter

n

d̂

Beamformer

11th March 2010 26

Capacity of MIMO cellular

4 antenna elements

MIMO system capacity around 3 SISO

and more than 1.5 SIMO (smart antenna)

Beamformer (“prewhitening”) very important

Interference limited

11th March 2010 27

Implications for 4G networks

MIMO can dramatically increase link capacity and significantly increase cellular capacity Note that capacity is mainly affected by

n = min(nT, nR)

Still severely limited by inter-cell interference

Can we reduce the effect of interference?

11th March 2010 28

Outline

The challenge of next generation wireless networks

Next generation network architectures MIMO and MIMO cellular Multi-user and Network MIMO Physical layer network coding Conclusions

11th March 2010 29

Multi-user MIMO systems

It has been known since the 1960s that the optimum means of sharing a channel between several users may be by simultaneous, mutually interfering transmission as opposed to time-division multiplexing, or

other orthogonal multiplexing Information-theoretic approach:

Multiple Access Channel - MAC (uplink) Broadcast Channel - BC (downlink)

11th March 2010 30

2-user SISO MAC

Rate region: set of achievable rates of the two sources

C1 and C2 are capacities of two channels without the other

“Corner points” achieved by successive interference cancellation

“Time sharing” sum rate limited to dashed line in general achievable sum

rate Csum exceeds this

1 1 1 2 2 1

2 2 2 1 2 2

1 2 1 2

2 1 2

; log 1

; log 1

, ;

log 1

sum

R C I X Y X P N

R C I X Y X P N

R R C I X X Y

P P N

R1

R2

C2

C1

Csum

time-sharingrate

11th March 2010 31

Multi-user MIMO MAC

In a multi-user MIMO multiple access channel (uplink), sum rate capacity limit is capacity of MIMO channel

formed by combining all Tx antennas of all users

For nU users with nT Tx antennas, average Tx power Si and Tx time Ti (each)

where:

denotes capacity of nTnR MIMO channel with SNR S

, ,1

, ,1

1 U

T R T R

UU T R U T R

n

TS i n n i n nii

nsum n n n i n n ni

TC T C S C S

T T

C C S C S

1 1; ;U Un ni ii iS S T T

,T Rn nC S

11th March 2010 32

Multi-user MIMO

Conventional TDMA/FDMA is equivalent to time-sharing divides “headline” rate by no. of channels

MU-MIMO allows several users to share same time slot/channel

Users/BS can act as a single nT nU nR MIMO system

Usually more BS than terminal antennas multiplexing gain no

longer limited by no. terminal antennas

BS

nR

nT

nU

11th March 2010 33

Symmetric and asymmetric

Modest advantage when nT = nR (symmetric links) Large advantage when nT nR (asymmetric links)

max. multiplexing gain becomes min(nR, nTnU)

0 5 10 15 20 25 300

10

20

30

40

50

60

70

80

SNR (dB)

Cap

acity

(bi

ts/s

/Hz)

Sum rate (black) and time-share capacity (red), 8 users, nT=2, nR=8

0 5 10 15 20 25 300

5

10

15

20

25

30

35

40

SNR (dB)

Cap

acity

(bits

/s/H

z)

Sum rate (black) and time-share capacity (red), 8 users, nT=4, nR=4

11th March 2010 34

Broadcast channel

Broadcast channel simply means one transmitter to many receivers obvious example is radio/TV broadcasting, where

same message is intended for all also applies to cellular downlink, where a different

message is intended for each receiver In latter case can define a

capacity region like that for multi-user MIMO

Again time sharing (TDMA/FDMA) is sub-optimal

R1

R2

C2

C1

Csum

time-sharingrate

11th March 2010 35

Dirty paper coding (DPC)

This is an information-theoretic result which applies to a channel subject to interference where the interference is known at the transmitter r = s + i + n where s is the information signal, i is the interference, and n is

noise Achievable capacity of this channel is

log2(1 + PS/Pn) i.e. the same capacity as if the interference were not present note this is not achieved by “pre-cancelling” the interference

On a broadcast downlink, the signal to one user is interference to another, and is known at the transmitter

11th March 2010 36

Precoding in practice

Dirty paper precoding in principle operates by selecting a codebook (set of transmitted codewords) depending on the interference

A more practical scheme following the same principle is Tomlinson-Harashima precoding (THP)

Use modulo function f(x), and transmit: where d is data, i is interference and

k is some integer At the receiver apply the modulo

operation again: interference is removed some degradation due to “folding” of

noise

A

f(x)

x

mod ,2 2

A Af x x A

s f d i d i kA

d f r f s i n

f d n kA

d n k A

11th March 2010 37

Linear beamforming

Or simpler still, we can simply form beams to each user ensuring also that we null interference to other

users this is a purely linear operation

11th March 2010 38

Network MIMO

So far we have assumed that signals from other cells must be treated as interference

However it is possible for several (in principle, all) base stations to cooperate to transmit to a given mobile

or to receive from that mobile Then there is in principle no CCI!

since all received signals are exploited as signals

The entire system then operates as a multi-user MIMO system with (on the uplink) nTnU nC transmit and nRnC receive antennas

where nC is the number of cooperating cells

in principle multiplexing gain approaches min(nRnC, nTnU nC)

11th March 2010 39

Practical limitations

There are of course practical limitations to this concept: Where should the processing be performed?

In a hierarchical network like BuNGee, at the hub Distributed methods also possible, with processing

at cooperating BSs Computational complexity

Synchronisation Is it feasible to keep cooperating base stations

phase synchronous? (especially for downlink) Backhaul capacity

11th March 2010 40

Backhaul capacity requirements

On the downlink, if two BSs cooperate to communicate with an MS, that MS’s data should be sent to both could double backhaul requirements

On the uplink, neither may be able to decode the MS without the signal from the other hence analogue signal may need to be

transmitted over the backhaul in high precision

may increase backhaul requirements by several times

Need to ensure backhaul links are efficiently used

r1r2

a1r1 + a2r2

s1 s2

b2s2b1s1

11th March 2010 41

Limitations of in-band backhauling

If we use wireless backhaul, we must account for bandwidth occupied

Can we re-use the same spectrum in backhaul and access segments? in-band backhauling

Duplexing restrictions of ABSs usually prevent same resources being used in the two segments in the same place

11th March 2010 42

Outline

The challenge of next generation wireless networks

Next generation network architectures MIMO and MIMO cellular Multi-user and Network MIMO Physical layer network coding Conclusions

11th March 2010 43

Network coding

A network node applies a joint coding function to two (or more) incoming data streams instead of simply switching between

them In this simple example (the “butterfly

network”) the central node applies the XOR function (modulo-2 addition) then both streams can be recovered

at both output nodes without network coding the central

link would have to have twice the capacity to achieve this

11th March 2010 44

Two-way relay channel (2WRC)

Allows a relay to support transmissions in two directions at once

Relay broadcasts XOR combination of two incoming streams

Each destination can then reconstruct data intended for it by XOR combination with the data it transmitted

SA

DB

RSB

DA

a b

a ba b

11th March 2010 45

Physical layer network coding (PLNC)

In a wireless network, we do not have discrete, non-interfering paths except by using TDMA or FDMA

Signals: are broadcast to all nodes within range combine additively in signal space

However it is still possible to extract a joint information stream equivalent to XOR combination

11th March 2010 46

PLNC for 2WRC

System operates in two phases Phase 1: sources transmit

simultaneously Phase 2: relay transmits

Assume both sources transmit BPSK {+1, -1} {1, 0}

SA

DB R

SB

DA

a b

a ba b

a + b

0+211

1001

1010

0-200

a ba+bba

SA, SB

SA, SB

time

Phase 1 Phase 2

11th March 2010 47

For comparison

Without network coding

Network layer network coding

Physical layer network coding

SA

DB

RSB

DA

a b

a ba b

time

SA

R(A)

SB

R(B)

time

SA SB

R(AB)

time

SA, SB

R(AB)

11th March 2010 48

Relay data compression

Cooperative diversity: relay provides extra diversity if S – D link fades

Phase 1: Source transmits to relay and destination Phase 2: Relay transmits to destination Note signal at relay is correlated with Phase 1 signal at

destination since both arise from the same data

This allows distributed compression using Slepian-Wolf coding reduces the data relay must transmit

S

R

D

11th March 2010 49

Slepian-Wolf coding

If two data sources are correlated, their joint information content is less than the sum of their separate content

Can exploit this to compress the data even though encoders

are separate White area on graph gives

rate region range of possible

compressed rates to allow reconstruction

S1

R1

R2

C1

D

S2 C2

S’1

S’2

R1

R2

11th March 2010 50

PLNC in network MIMO

Example: 2 terminals connected

to hub via 3 BS B2 can use PLNC, to share

its link with the hub betweentwo terminals

It can use distributed compression, since the data is correlated with that via B1 and B3

Reduces backhaul load

T1 T2

B1 B2 B3

Hub

11th March 2010 51

Two-way relaying in hierarchical wireless network

Duplexing constraints can be alleviated using 2-way relaying allows access and backhaul

resource to be shared at each ABS however neighbouring ABSs and

corresponding MSs may interfere We have analysed this

considering a simple scenario Assume 2 ABSs share resources

over 2 slots i.e. both ABSs and MSs transmit

simultaneously on same channel consider amplify-and-forward and

network coded decode-and-forward relaying

11th March 2010 52

Capacity results

Plot resulting capacity against strength of interfering links for small interference

DF is better for large interference

AF capacity increases again

both MSs can exploit both ABSs

exploits a form of network MIMO

Generally 2 slots are better than three or four

11th March 2010 53

Conclusions (1)

Next generation wireless networks will be required to handle much larger capacity densities

MIMO techniques can greatly increase link capacity but still seriously affected by interference in cellular

networks Multi-user MIMO can greatly increase capacities

especially for asymmetric systems Network MIMO can further increase capacity

and also largely eliminate interference however increases backhaul load

11th March 2010 54

Conclusions (2)

Physical Layer Network Coding can better exploit backhaul capacity by allowing backhaul links to be shared by exploiting correlation of signals travelling by

different routes by overcoming duplexing constraints on spectrum

sharing All these techniques will be essential elements of

next generation networks!

11th March 2010 55

Acknowledgements

With thanks to our collaborators : Prof. Jan Sykora, Czech Technical University,

Prague Prof. Tad Matsumoto, JAIST, Japan Prof. Meixia Tao, Shanghai Jiao Tong University

and funders: European Commission, FP7 project “BuNGee” RCUK “Sciences Bridges China” programme COST 2100 programme Czech Technical University