The Wireless Data Crunch: Motivating Research in Wireless

The Wireless Data Crunch: Motivating Research in Wireless

Communications

Stephen Hanly

CSIRO-Macquarie University Chair in Wireless Communications

stephen.hanly@mq.edu.au

Wireless Growth Rate • Cooper’s law:

Number of “conversations” (voice or data) over a given area in

all of the useful radio spectrum has doubled every two-and-a-

half years for the past 104 years.

Wireless Growth Rate

1,000,000 x improvement in 45 years:

25 x ← due to more spectrum

5 x ← dividing spectrum into

narrower slices (FDM)

5 x ← advanced modulation and

coding techniques

1600 x ← frequency re-use

Wireless Data Crunch

In Feb. 2012, CISCO predicts:

• Mobile data demand will increase by 18 x between 2011-2016

By 2016:

• 2/3 mobile data traffic will be video

• Mobile connection speeds will need to increase by 9 x

• 1.4 mobile devices per capita

But how can networks adapt to this rate of change?!

This is the data crunch that is driving forward research in wireless communications

Claude Shannon and Capacity • In 1948, Claude Shannon wrote

down “Newtons laws of the

information age” in his classic

paper: “A Mathematical Theory

of Communication”

• This paper allows us to compute

the maximum possible bit rate, in

bits/sec, of any point-to-point

communication channel

• For the Gaussian noise channel:

bits/sec )1log( SNRWC

Claude Shannon and Capacity

• Formula above is for the additive white

Gaussian noise (AWGN) channel

• Sometimes you hear people claim to have

“broken the Shannon limit”

• The reality is they have only broken that

particular formula

• Just means channel model is NOT the

AWGN channel

Shannon gave the method to get the capacity of any point to

point channel

bits/sec )1log( SNRWC

Spectral Efficiency

• Since bandwidths can vary, its more useful to talk of

spectral efficiency, in bits/sec/Hz

• Let the received power be P mw and the Gaussian noise power

spectral density be s2 mw/Hz; so the

• Then the spectral efficiency limit is

2s

PSNR

zbits/sec/H 1log2

12

s

PC

• Later in the talk we’ll introduce an important extension:

area spectral efficiency, in bits/sec/Hz/m2

Claude Shannon and Capacity • Shannon didn’t just tells us a

capacity result

• He also indicated how to get

there: eg. use random Gaussian

codebooks in the AWGN channel.

• These ideas pointed the way to

Turbo codes and low density

parity check codes

• He also showed how to analyze

error probabilities

Random Code

• And design communication systems: eg. the source–channel

separation theorem tells us to design a layered system

x1

y ×

Beyond Shannon: Multiple Users • Shannon focused on the point-to-point

channel

• A cell in a mobile radio network has

multiple users

• The uplink (mobile to base station link) is

called a multiple access channel (MAC)

• Researchers that followed Shannon

(Ahlswede and Liao ’72) found the

capacity region of a MAC

• The downlink is called a broadcast channel

The Multiple Access Channel (MAC) • Lets look at the simplest two user

MAC: AWGN at the base station

• There is interference between the

two links which is reflected in a

tradeoff:

If R1 is the bit rate of user 1 and

R2 is the bit rate of user 2 then

there is a tradeoff between these

two rates

• There is a capacity region that

describes the set of achievable

rates

R1

R2

2-user Rate Region of AWGN MAC

R1

R2

x

x = interference as noise

= successive decoding

= FDMA curve

• The dominant line gives the best

pairs of rates – the optimal tradeoff

• FDMA curve touches the dominant

line at one point

• Treating interference as noise is

suboptimal

• Successive decoding is optimal

What About the Real World?

• In the real world we have cells

• Mobiles in one cell will interfere at another cell’s base station

2-user AWGN Interference Channel

• The rate region for this channel is

unknown

• An open problem for over 40 years!

• Recent work has characterized the rate region to within 1

bit/sec/Hz (Etkin, Tse, and Wang ‘08). Thus:

Rate region is known quite precisely when the noise is low

MACs in the Interference Channel

• User 1 and user 2 are heard at Rx1: MAC1

• User 1 and user 2 are heard at Rx2: MAC2

• The intersection of both rate regions is achievable

Message 1 Rx1 Message 1 Tx1

Message 2 Tx2 Message 2 Rx2

MACs in the Interference Channel

×

MAC2

MAC1

R1

R2

× = treat interference as noise

MACs in the Interference Channel

R1

R2 MAC2

MAC1

×

× = treat interference as noise

Han Kobayashi Scheme • Each user splits its data into two parts: private and common

• The private message is only decoded by the desired receiver (Rx1)

• The common message is decoded by both and cancelled

• The private data rate is too high to be decoded at Rx2 and is treated as Gaussian noise

Treat private message as noise at Rx2

Private message Common message Tx1 Rx1

Tx2 Rx2

Han Kobayashi Scheme

• Etkin, Tse and Wang ‘08 show this scheme can be optimized to

be within 1 bit/sec/Hz of capacity

• Still uses Gaussian random codes

• Will not work once we go to three or more users!

• Beyond Shannon: random codes are no longer any good

• New research shows we must look for structured codes

• A new concept called “interference alignment”

• Suppose users 1 and 2 use a random Gaussian codebook:

Gaussian Han-Kobayashi Not Optimal

Random Code

Sum of Two Random Codebooks Lattice Code for Users 1 and 2

User 0 Code Interference from users 1 and 2 fills the space: no

room for user 0.

Lattice codes can achieve constant gap

Gaussian Models

• It remains that the Gaussian model is a robust model

• It can be shown that the following formula gives achievable

rates for interference channels:

zbits/sec/H energy noise energy ceinterferen

energy signal1log

2

1

C

• Even this simple model is highly complex!

• Energy allocation can be optimized across bandwidth and time

• These optimization problems have been shown to be

mathematically intractable for large numbers of links

Symmetric Network Problem

N links:

• All links gains have the same value

• All cross-link gains have the same (other) value

e

e

1

1

N=2

Solution to Symmetric Network Problem

We might expect optimal solution to be:

• FDMA between the links when the cross-gain (e) is high

• Wideband (WB) frequency sharing when e is low

f 2

W

2

W

FDMA

one band for each link f 2

W

2

W

One band shared by all links

WB

Solution to Symmetric Network Problem

When the cross-gain (e) is low:

• FDMA between the links when the SNR is high

• Wideband (WB) frequency sharing when the SNR is low

• A mixture of these when the SNR is medium

f 2

W

2

W

FDMA

one band for each link f 2

W

2

W

One band shared by all links

WB

TV White Spaces and Unused Spectrum

• Is the spectrum really that congested?

• What about white spaces?

– eg TV bands that are not being used?

• Licensed (primary) users may not be active in some areas

• Room for secondary, unlicensed users?

• New spectrum opening up, and spectrum auctions

Fragmented Spectrum

Cognitive Radio

• Smart, agile radios that can sense and occupy un-utilized

spectrum

• Require flexible hardware, tuneable frequencies

• Overlay: search out unused bands – the FDM approach

• Underlay: UWB radios taking the WB approach

• Successive decoding can be used to strip off strong

interference

• Interference alignment strategies can reduce the spectral

footprint

Multiple Input, Multiple Output (MIMO)

• Multiple antennas can be used at the base

station

• Multiple users provide multiple antennas

• Similar in a lot of ways to a point to

point MIMO channel

• Beamforming:

– N antennas can create up to N non-

interfering beams

– Main challenge is channel

measurement

Coordinated Beamforming • If base stations share channel state information (CSI) over the

backhaul network coordinated beamforming

Cell 1 Cell 2

CSI

Distributed Antennas

• Throw away cells altogether (new architecture!)

• Backhaul transports both CSI and user data

• Now a giant broadcast channel with distributed antennas!

• Base stations must cooperate

CSI data

Small Cells • Spectral efficiency is really

measured in bits/sec/Hz/m2

• If we can decrease the cell sizes

then we increase spectral efficiency

by increasing frequency re-use

• Macrocells, microcells, picocells

• Tiny picocells can be used in

network hotspots

Femtocells • Femtocells are tiny cells formed

with cheap, off-the-shelf base

stations

• Femto base station are like WiFi

access point but use cellular

frequencies – plug and play

• They offload cellular traffic

onto owner’s broadband ISP

connection

• Networks are heterogeneous – a mix of short and long links,

mixture of planned and unplanned layouts

Multi-tier Heterogeneous Networks

Many research challenges:

• Closed versus open access

• Modelling and design (mixture of planned and unplanned)

• Interference avoidance (long versus short links)

• Power control

• Cell association

• Base station coordination

• Handovers (cell to cell and tier to tier)

Conclusions

• Exciting time to be in wireless research

• Great challenge is to drastically increase bits/sec/Hz/m2 to

match forecast demand

• Interference and congestion are the major challenges!

• Huge gap between theory and practice

– in MIMO, interference alignment, coordinating base stations

• New architectures and new algorithms

• Small cells, heterogeneous networks and cognitive radios offer

a prospect to meet the data crunch challenge!

Thankyou!