Breaking the Interference BarrierBreaking the Interference Barrier

David TseDavid Tse

Wireless FoundationsWireless Foundations

University of California at BerkeleyUniversity of California at Berkeley

Mobicom/Mobihoc Plenary TalkMobicom/Mobihoc Plenary Talk

September 13, 2007September 13, 2007

The Interference BarrierThe Interference Barrier

Lots of recent advances in physical layer wireless Lots of recent advances in physical layer wireless communication (multiple antennas MIMO, space-time codes, communication (multiple antennas MIMO, space-time codes, opportunistic scheduling, turbo codes, hybrid ARQ….)opportunistic scheduling, turbo codes, hybrid ARQ….)

From theory to practice in a decade.From theory to practice in a decade.

Gains pertain mainly to point-to-point or multiple access Gains pertain mainly to point-to-point or multiple access performance.performance.

But performance of many wireless systems ultimately limited But performance of many wireless systems ultimately limited

by by interference interference..

Breaking this interference barrier will be the next step.Breaking this interference barrier will be the next step.

Examples of Interference BarrierExamples of Interference Barrier

Cellular networksCellular networks: inter-cell interference: inter-cell interference

Ad hoc networksAd hoc networks: interference from simultaneous : interference from simultaneous transmissions transmissions

Wireless LANSWireless LANS: interference between adjacent networks: interference between adjacent networks

Cognitive networksCognitive networks: interference between primary and : interference between primary and secondary users and between multiple secondary systemssecondary users and between multiple secondary systems

Talk OutlineTalk Outline

We discuss several We discuss several speculativespeculative approaches to break the approaches to break the interference barrier:interference barrier:

cooperative distributed MIMOcooperative distributed MIMO

exploiting mobility to localize interference exploiting mobility to localize interference

interference alignmentinterference alignment

Key message: Key message:

Solving the interference problem requires a combination of Solving the interference problem requires a combination of physical layer and architectural ideas.physical layer and architectural ideas.

Traditional Interference Management Traditional Interference Management in Cellular Systemsin Cellular Systems

Narrowband (eg. GSM)Narrowband (eg. GSM)

Inter-cell interference made negligible at Inter-cell interference made negligible at the price of poor frequency reusethe price of poor frequency reuse

Wideband (eg. CDMA, OFDM)Wideband (eg. CDMA, OFDM)

Universal frequency reuse but system is Universal frequency reuse but system is interference-limited.interference-limited.

Example: WiMax is Interference-LimitedExample: WiMax is Interference-Limited

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1

F1 = 10MHz

SIR = 2 dB

SNR = 20 dB

Universal Reuse : 6 dominant interferers

SIR to SNR gap = 18dB

Source: Intel WiMax simulations

Fractional Reuse: A Partial SolutionFractional Reuse: A Partial Solution

Universal reuse for cell-interior users.Universal reuse for cell-interior users.

Orthogonal bands for cell-edge users.Orthogonal bands for cell-edge users.

But cell-edge users are still the bottleneck.But cell-edge users are still the bottleneck.

f2 f3f0

f0

Tale of Two Cell-Edge UsersTale of Two Cell-Edge Users

keep users on orthogonal bands: lose half the effective keep users on orthogonal bands: lose half the effective bandwidth but avoid interferencebandwidth but avoid interference

Best of both worlds?Best of both worlds?

Yes, base-stations can cooperate to form a Yes, base-stations can cooperate to form a distributed MIMOdistributed MIMO array.array.

MIMO in One SlideMIMO in One Slide

M by M MIMO system with a sufficiently random channel supports M simultaneous data streams.

Signal space at Rx array (M=2)

direction of signalfrom Tx antenna 1

Infrastructure CooperationInfrastructure Cooperation

Base stations cooperate to form a macro-array to jointly Base stations cooperate to form a macro-array to jointly decode in the uplink and transmit in the downlink.decode in the uplink and transmit in the downlink.

Turns harmful inter-cell interference into useful signalsTurns harmful inter-cell interference into useful signals

High-speed connectivity to a central processing unit. High-speed connectivity to a central processing unit.

Simulation in a Hexagonal Cellular System Simulation in a Hexagonal Cellular System

Rise-over-thermal = 6dB 2 Rx antennas per BS

cooperation

single-cell processing

(Alessandro et al 06)

Cooperation in Ad Hoc Networks Cooperation in Ad Hoc Networks

Capacity of ad hoc networks limited by mutual Capacity of ad hoc networks limited by mutual interferenceinterference between simultaneous between simultaneous transmissions.transmissions.

How can cooperation between mobiles How can cooperation between mobiles improve capacity?improve capacity?

Unlike infrastructure-based cellular systems, Unlike infrastructure-based cellular systems, such cooperation comes at an over-the-air such cooperation comes at an over-the-air transmission cost.transmission cost.

Will the overhead swamp the cooperation Will the overhead swamp the cooperation gain?gain?

Scaling Law Formulation Scaling Law Formulation (Gupta-Kumar 00)(Gupta-Kumar 00)

n nodes randomly located in a fixed area.n nodes randomly located in a fixed area.

n randomly assigned source-destination pairs.n randomly assigned source-destination pairs.

Each S-D pair demands the same data rate.Each S-D pair demands the same data rate.

How does the total throughput T(n) of the network How does the total throughput T(n) of the network scale with n?scale with n?

How much can Cooperation Help? How much can Cooperation Help?

?

Can we get linear scaling with more sophisticated cooperation?

Arbitrarily closely. (Ozgur,Leveque,T. 06)

T(n) = £(1) T(n) = £(pn)

Courtesy: David Reed

Gupta-Kumar Capacity is Gupta-Kumar Capacity is InterferenceInterference-Limited-Limited

Long-range transmission causes too much interference.Long-range transmission causes too much interference.

Multi-hop means each packet is transmitted many times.Multi-hop means each packet is transmitted many times.

To get linear scaling, must be able to do many To get linear scaling, must be able to do many simultaneous long-range transmissions.simultaneous long-range transmissions.

How to deal with interference?How to deal with interference?

A natural idea: distributed MIMO!A natural idea: distributed MIMO!

But But cooperation overheadcooperation overhead is bottleneck. is bottleneck.

What kind of cooperation architecture minimizes What kind of cooperation architecture minimizes overhead?overhead?

A 3-Phase SchemeA 3-Phase SchemeDivide the network into clusters of size M nodes.Divide the network into clusters of size M nodes.

Focus first on a specific S-D pair.Focus first on a specific S-D pair.

source s wants to send M bits to destination d.source s wants to send M bits to destination d.

Phase 1 :Setting up Tx cooperation:1 bit to each node in Tx cluster

Phase 2:Long-rangeMIMO between s and d clusters.

Phase 3:Each node in Rx clusterquantizes signal into k bitsand sends to destination d.

Parallelization across S-D PairsParallelization across S-D PairsPhase 1:Clusters work in parallel.Sources in each cluster taketurn distributing their bits.

Total time = M2

Phase 2:1 MIMO trans.at a time.

Total time = n

Phase 3:Clusters work in parallel.Destinations in each clustertake turn collecting their bits.

Total time = kM2

Back-of-the-Envelope Throughput CalculationBack-of-the-Envelope Throughput Calculation

total number of bits transferred = nMtotal number of bits transferred = nM

total time in all three phases = Mtotal time in all three phases = M22 + n + kM + n + kM22

Throughput: bits/secondThroughput: bits/second

Optimal cluster sizeOptimal cluster size

Best throughput: Best throughput:

M ¤ =pn

nMM 2+n+kM 2

pn

Further ParallelizationFurther Parallelization

In phase 1 and 3, MIn phase 1 and 3, M22 bits have to be exchanged within each bits have to be exchanged within each cluster, 1 bit per node pair. cluster, 1 bit per node pair.

Previous scheme exchanges these bits one at a time (TDMA), Previous scheme exchanges these bits one at a time (TDMA), takes time Mtakes time M22..

Can we increase the spatial reuse ?Can we increase the spatial reuse ?

Can break the problem into M sessions, each session involving Can break the problem into M sessions, each session involving M S-D pairs communicating 1 bit with each other:M S-D pairs communicating 1 bit with each other:

cooperation = communicationcooperation = communication

Any better scheme for the small network can build a better Any better scheme for the small network can build a better scheme for the original network.scheme for the original network.

RecursionRecursion

Lemma: A scheme with thruput Lemma: A scheme with thruput MMbb for the smaller network yields for the smaller network yields for the original network a thruput:for the original network a thruput:

n1

2¡ b

f (b) =1

2¡ b

MIMO + Hierarchical CooperationMIMO + Hierarchical Cooperation-> Linear Scaling-> Linear Scaling

..

Setting up Tx cooperation

Long-range MIMO

Cooperateto decode

By having many levels of hierarchy, we can get as close to linear scaling as we wish.

Linear Scaling with Less Work?Linear Scaling with Less Work?

Linear scaling means that the capacity of the network is not Linear scaling means that the capacity of the network is not significantly limited by significantly limited by interferenceinterference..

But the hierarchical scheme requires tracking of channel But the hierarchical scheme requires tracking of channel information as well as significant cooperation between nodes.information as well as significant cooperation between nodes.

Can one get linear scaling with less work?Can one get linear scaling with less work?

Yes, if nodes are mobile.Yes, if nodes are mobile.

Mobility Can Help!Mobility Can Help!

(Grossglauser and T. 01)(Grossglauser and T. 01)

Suppose nodes move randomly and independently.Suppose nodes move randomly and independently.

A linear throughput can be achievedA linear throughput can be achieved

if one is willing to wait.if one is willing to wait.

Throughput is averaged over the time-scale of mobility.Throughput is averaged over the time-scale of mobility.

Direct Communication Does Not WorkDirect Communication Does Not Work

The source and destination are nearest neighbors only O(1/n) The source and destination are nearest neighbors only O(1/n) of the time.of the time.

Detour: Multiuser Diversity in Cellular SystemsDetour: Multiuser Diversity in Cellular Systems

By opportunistically scheduling transmissions to users with instantaneously strong channels, multiuser diversity gain is achieved.

Multiuser Diversity via RelayingMultiuser Diversity via Relaying

Multiuser diversity created artificially using all other nodes as Multiuser diversity created artificially using all other nodes as relays.relays.

Phase I: Source to RelaysPhase I: Source to Relays

At each time slot, source At each time slot, source relays a packet to nearest relays a packet to nearest neighbor.neighbor.

Different packets are Different packets are distributed to different relay distributed to different relay nodes.nodes.

Phase 2: Relays to DestinationPhase 2: Relays to Destination

Steady state: all nodes have Steady state: all nodes have packets destined for D.packets destined for D.

Each relay node forwards Each relay node forwards packets to D only when it packets to D only when it gets close.gets close.

Phase I and II StaggeredPhase I and II Staggered

O(1) throughput from S to DO(1) throughput from S to D

Communication is confined to nearest neighbors, but each Communication is confined to nearest neighbors, but each packet goes through at most two hopspacket goes through at most two hops

Load is distributed evenly between all relay nodes, enabling Load is distributed evenly between all relay nodes, enabling every S-D pair to follow the same strategy.every S-D pair to follow the same strategy.

Linear Scaling without Cooperation?Linear Scaling without Cooperation?

The two approaches rely on some sort of cooperation to The two approaches rely on some sort of cooperation to mitigate interference.mitigate interference.

Is cooperation really necessary?Is cooperation really necessary?

Spectrum Sharing RevisitedSpectrum Sharing Revisited

Working assumption:Working assumption:

only one transmission on each only one transmission on each

time-frequency-space resource.time-frequency-space resource.

Implicit assumption: Implicit assumption:

spectrum is a common “ether” spectrum is a common “ether” shared by all.shared by all.

But is this metaphor correct?But is this metaphor correct?

Tx 1

Tx 2

Tx n

Rx 1

Rx 2

Rx n

Channel11

Channel21

Channel2n

Channelnn

::

Interference Alignment ExampleInterference Alignment Example

(Cadambe-Jafar 07) (Cadambe-Jafar 07)

All direct channels delay transmission All direct channels delay transmission by 1 symbol time.by 1 symbol time.

All cross channels delay by 2 symbol All cross channels delay by 2 symbol times.times.

Each user can transmit every other Each user can transmit every other symbol time, yet no interference!symbol time, yet no interference!

What matters is what happens at the What matters is what happens at the receiver, and each receiver sees a receiver, and each receiver sees a different picture.different picture.

So all the interference can be aligned So all the interference can be aligned onto one symbol time and yet the onto one symbol time and yet the signal is orthogonal to the signal is orthogonal to the interference. interference.

Tx 1

Tx 2

Tx n

Rx 1

Rx 2

Rx n

Channel11

Channel12

Channel2n

Channelnn

::

Interference Alignment: GeometryInterference Alignment: Geometry

Tx 1

Rx 3

Rx 2Tx 2

Tx 3

Rx 1H11

Recurring ThemeRecurring Theme

Channel diversity is a key resource for breaking the Channel diversity is a key resource for breaking the interference barrier.interference barrier.

The three approaches can be viewed as ways to exploit this The three approaches can be viewed as ways to exploit this diversity:diversity:

Hierarchical cooperation to exploit MIMO gain.Hierarchical cooperation to exploit MIMO gain.

Mobility and relaying to exploit multiuser diversity gain.Mobility and relaying to exploit multiuser diversity gain.

Interference alignment to exploit diversity between direct and Interference alignment to exploit diversity between direct and cross channels. cross channels.

ConclusionsConclusions

Breaking the interference barrier is the next step in the Breaking the interference barrier is the next step in the evolution of wireless systems.evolution of wireless systems.

We focus on speculative ideas in this talk.We focus on speculative ideas in this talk.

Hopefully they provide some food for thought for system Hopefully they provide some food for thought for system builders.builders.