1 Order Matters: Interference-Aware Transmission Reordering in Wireless Networks.

1

Order Matters: Interference-Aware Transmission

Reordering in Wireless Networks

2

Wireless Networks Interference Limited

Packet decoded successfully When interference substantially lower Else, collision

CollisionCollision

IEEE 802.11

3

Phy Layer Capture

Concurrent transmissions may not necessarily cause collision Possible to decode the frame with higher SINR As long as receiver not “locked” onto interference Known as PHY layer capture

What is locking onto a signal? What is locking onto a signal?

4

Implications of Capture

When stronger signal is of interest, AND Arrives within PLCP window

Concurrency feasible

PLCP

5

Implications of Capture

When stronger signal is of interest, AND Arrives within PLCP window

Concurrency feasible

However, PLCP duration small Probability of precise timing also small

millisecond20 us

6

Capture and MIM

Message in Message (MIM) Strong frame arrives after preamble of interfering frame Receiver locked onto interference by then, and decoding However, continues searching for another preamble Strong message can be extracted while in another message

7

Caveats

Recognizing arrival of new preamble

requires new preamble to be have higher SINR

Only then correlation shows a high value

8

SINR for MIM a function of relative arrival order and timingSINR for MIM a function of relative arrival order and timing

SINR Requirements [Lucent NIC]

4 dB

4 dB

10 dB

10 dB

9

Order Matters

Some signal-arrival orders will permit concurrency Productive

But the reverse order may cause collision Unproductive

Example …

10

MIM Aware Scheduling

AP1 must start first, followed by staggered

transmission from AP2 -- concurrency feasible



10 dB5 dB

11






10 dB5 dB

In general, weaker transmission must start first,

stronger receiver suppresses it, and extracts own signal



12






10 dB5 dB




stronger receiver suppresses it, and extracts own signalObserve that 802.11 does not enforce this order,

hence will fail to exploit MIM capabilities

Observe that 802.11 does not enforce this order,

hence will fail to exploit MIM capabilities

13

Problem Definition:

Design an MIM-aware scheduling algorithm that reorders transmissions to augment concurrency

What is the bound on improvement?How to cope with time-vaying channel?How to sustain fairness and starvation?

14

Solution Space

Shuffle A centralized MIM-aware scheduling protocol for

Enterprise wireless LANs (EWLAN)

AP2AP2

ControllerController

AP1AP1 AP3AP3

15

Solution Space

Shuffle A centralized MIM-aware scheduling protocol for

Enterprise wireless LANs (EWLAN)

Why EWLAN?1. Becoming popular in single-admin environments

§ Offices, warehouses, libraries

2. Understand MIM for centralized systems, then goto distributed

§ Need to walk before running

AP2AP2

ControllerController

AP1AP1 AP3AP3

16

Shuffle: 3 Main Components

1. Measuring Interference Relation: Rehearsal Characterize interference map to identify MIM opportunity Cope with time-varying channel conditions

2. Packet Scheduler Use rehearsal outcome to schedule transmissions Scheduling = Reordering and staggering Protect from unfairness and starvation

3. Schedule Coordinator Execute MIM-aware schedule Cope with failures, retransmissions, and centralized bottleneck

17

Main Assumptions

Dominant download traffic Upload handled through periodic “upload windows”

Processing time and latencies Powerful controller, thin APs Wired backbone fast, but can become bottleneck

Additive Interference Total interference = sum of individual interferences

18

Feasibility First

What is the maximum improvement with Shuffle

in finite network scenarios?

Determine the optimal link selection, and their relative

order of initiation, to achieve this bound

Observe that graph coloring inapplicable

19

Analysis

Optimal MIM-aware link scheduling is NP-Hard Proof:

• Reduction from Independent Set selection

• MIM scheduling is special case

• Set SL (Signal Last SINR)=

• Relative order requires the optimal choice of links first.

• Hence, NP-Hard

20

Integer Linear Program

Use ILP to upper bound improvement For a large number of finite-sized topologies

21

CPLEX Results

MIM comparison with non MIM Substantial improvement feasible, worth researching

22

Protocol Design

Measuring Interference Relations:

Rehearsal

23

Rehearsal

Central controller needs link conflict information Graph coloring notion of conflict not applicable Conflicts also change with time-varying channel

Basic idea: Controller orchestrates a rehearsal of transmissions Clients and APs record RSSI values as instructed times Recorded RSSI correlated at controller Inteference graph generated

24

Rehearsal

At network initialization APs and clients informed about time of transmissions Each AP transmits sequence of probes at base rate Clients transmit probes, piggybacks recorded RSSI values

At the end, APs forward gathered values to controller Controller derives interference map

• using additive interference assumption

25

Interference Map

Pairwise interferences mapped Controller populates table

i

j

Interference

from i to j

Interferer

Sniffer

. . .. . .

26

Rehearsal

Opportunistic rehearsal Continuous rehearsal expensive Utilize regular transmissions to piggyback overheard RSSI

Coping with Fading Convergence may take long with opportunistic Handling loss will require immediate conflict information Perform self-corrective rehearsal using data packets

• Schedule packets conservatively to also serve the rehearsal purpose

27

RehearsalRehearsal

MIM Scheduler

(Optimal NP-Hard)

MIM Scheduler

(Optimal NP-Hard)

i

j

Interference

from i to j

28

MIM-Aware Scheduler

Scheduler operation: Choose non-conflicting packets from queue Determine their relative starting order + stagger durations Dispatch batch to AP

Scheduler goal: Maximize batch size Protect from starvation Ensure high fairness

29

Greedy Heuristics

Basic greedy Fix a queue lookahead size for scheduling (say L) Controller takes in-order packets from FIFO queue Packet j scheduled if no conflict with pkts already scheduled

• Conflict is a function of SL and SF thresholds

If conflict, packet j postponed for next batch

No starvation, Good Fairness Every batch, a packet progresses in queue Head of the batch always transmitted

O(n2 )O(n2 )

30

Greedy Heuristics

Randomized Greedy Perform basic greedy on randomized subsets of queue Probability of choosing packets biased

• Earlier in the queue have higher probability Choose largest batch among all solutions

Least-Conflict Greedy Compute packet score = # of pair-wise conflicts

• Score higher if pkt must start earlier, lower else Sort packets based on score Perform basic greedy on this sorted order Incorporate aging for fairness/starvation

O(n2 logn)O(n2 logn)

O(n2)O(n2)

31

Optimal Vs Greedy (variants)

32

RehearsalRehearsal

MIM Scheduler

(Optimal NP-Hard)

MIM Scheduler

(Optimal NP-Hard)

Schedule Coordinator

(ReTx, Prefetch, Predict)


(ReTx, Prefetch, Predict)

< Batch of packets, Schedule >

33


Packets dispatched to APs Time synchronized between APs and contollers Pipeline Controller to AP, and AP to Client transmissions APs transmit at specified time

34


ACK Transmission Controller embed ACK schedule in Data Packet header Clients follow schedule (MIM-aware) AP forwards ACKs to controller (ACKs may have RSSI) When no ACK, AP forwards NACK Lost packets scheduled with highest priority

35

Batch Selection

C1C4C3C1C2C1

Queue

Batch

AP1

AP2

AP3

C1

C2

C3

C4

36

Controller sends packets to APs

AP1

AP2

AP3

C1

C2

C3

C4

C1

C4C3

C1C2

C1

Queue

Batch

37

AP1

AP2

AP3

C1

C2

C3

C4

APs/Clients Stagger transmissions

C1

C3

C4

ACK

ACK

ACK

Data Staggering Order: AP2-AP3-AP1

ACK Staggering Order: C4-C3-C1

38

Coping with Fading Loss

Time varying channel Interference graph changes Subsequent MIM scheduling can cause further failures

Immediate corrective rehearsal Controller identifies links suspected of fading Schedules a packet batch only for these APs

• This is a partial rehearsal

• Packets are transmitted in serial order

APs and clients unaware, send Data and ACKs Controller updates interference map from ACK RSSIs

39

Pipelining Batches

Batch - ACK - Batch inefficient APs remain idle between batches (not negligible)

Controller sends 2 batches to AP AP sends batch 1 and receives ACKs Batch 2 started, ACKs forwarded to controller in parallel Controller processes ACKs and next batch in parallel Controller schedules batch 3, sends to APs AP finishes batch 2 Repeat …

40

Testbed Evaluation

Looks Familiar?

41

Linux Laptops + Soekris + WiFi Device Driver

Validating that order really matters

42

Evaluation

Shuffle performs better

43

Ordering

Not all ordering is same -- hence, random not enough

44

Future Work

MIM aware routing Choose paths such that MIM is maximally activated

Distributed Shuffle We presented for EWLAN What about residential WLANs, organic emergence?

Can postambles be useful? As opposed to Preamble (Hari Balkrishnan, NSDI 08)

45

Conclusion

Necessary to pay attention to PHY layer capabilities Interference cancellation (its first steps) one example

MIM is ability to extract frame of interest Even under ongoing interference Provided some (relative order, SINR) conditions hold

Facilitating these conditions can enable MIM Rich performance gains feasible

MIM-aware link layer scheduling necessary

46

Conclusion

Shuffle - MIM scheduling for EWLANs EWLANs proliferating, also foundation for distributed case

NP-Hard problem Bounds characterized through linear programming

Greedy scheduling heuristics perform well Performance close to optimal

Evaluation on Linux testbed + Soekris boxes Consistent improvement, even under fading and losses

47

Questions?

48

Interference Cancellation in Wireless LANs

49

Successive Interference Cancellation (SIC)

State of the art allows only one reception The stronger one

SIC enables a receiver to receive both signals Stronger signal decoded and subtracted Residual signal decoded from the residue

50

SIC based WLANs

Existing schemes require SINR > Game of out-shouting each other

SIC offers payoff if transmitter Either out-shouts or whispers Fundamental changes for protocol design

51

MIM + SIC

Ongoing work on GNU radios SIC + MIM implementation can enable protocols

52

Diving a Little More

Modulation 101 How do you transmit a bit sequence? Need to convert bits onto analog domain … convert back

Basic Idea:• Change the properties of a signal (amplitude, frequency, phase) to

reflect the bit that you are trying to convey

• Example: Shout loud for +1, whisper for 0

• Shouting loud is a symbol for +1 … whispering is a symbol for 0

53

Modulation 101

But humans can modulate their voice better So why not shout/whisper at different levels?

• Each level a symbol -- each symbol carries multiple bits

11

1001

00

54

Modulation 101

But humans can modulate their voice better So why not shout/whisper at different levels?

• Each level a symbol -- each symbol carries multiple bits

11

1001

00

So if recevier samples at the right time (during peak or trough) then it can get a value. Since it knows the ranges for each symbol, it knows what symbol was received … hence, what bit sequence.

So if recevier samples at the right time (during peak or trough) then it can get a value. Since it knows the ranges for each symbol, it knows what symbol was received … hence, what bit sequence.

55

With Actual Signals

More opportunity: Modulate a Sin(.) and a Cos(.) signal with different bits This is like 2D space

• Called constellation diagram

Send the sum of both as

a single symbol

Receiver gets sum, and can

extract both using a

coherent demodulator

56

Quadrature Amplitude Modulation (QAM)

Mathematically

57

16 QAM reception

Receiver gets a dot Computes nearest neighbor as the transmitted symbol

Hence, the bits are

now decoded

QuickTime™ and aTIFF (Uncompressed) decompressor

are needed to see this picture.

58

16 QAM reception

Receiver gets a dot Computes nearest neighbor as the transmitted symbol

Hence, the bits are

now decoded

QuickTime™ and aTIFF (Uncompressed) decompressor

are needed to see this picture.Do you see why higherData rate increases the

Probability of error?

Because, separation betweenSymbols become smaller

Do you see why higherData rate increases the

Probability of error?

Because, separation betweenSymbols become smaller

59

Easier Said Than Done

Lots of issues: Rx is assumed to know the phase of transmitted signal

• So that Rx can sample at the right time

But difficult because signals getting reflected Also, Rx’s frequency needs to be exactly same as Tx

Recall PLCP It helps in straightening these out

60

So Then … How do You Do SIC?

Basic Idea: Let received combined signal be S’ …

• Stronger received signal be S1, weaker received signal be S2

Synchronize with the stronger signal• By detecting PLCP

Demodulate by treating the weaker as interference• Get the bits out

Now, model the stronger signal based on the bits (S1’)• To see how it would look without the interference (S1’ != S1)

Now subtract: i.e., S2’ = S’ - S1’ Demodulate S2’ to get the bits out

How

61

Modeling a Symbol from Bits

00

01

10

11

62


00

01

10

11

S’

63


00

01

10

11

S’

S1’

64


00

01

10

11

S’

S1’

-S1’

65


00

01

10

11

S’

S1’

-S1’

S2’ = S’ + (-S1’)Thus weaker signal is bit 11

66

All Modeling, Subtracting in Software

USRP (Universal Software Radio Peripheral) Connected to laptop for doing processing

This paper demonstrated offline SIC• All signals received, and procesing done after that

67

ZigZag Decoding: Combating Hidden Terminals in Wireless Networks

Shyamnath Gollakota and Dina Katabi

MIT CSAIL

SIGCOMM 2009

68

Hidden Terminal Problem

Leads to low utilization of bandwidth and unfairness in channel access

RTS/CTS induced too much overhead Collided packets may still be decodable!

Alice BobAPX

69

Basic idea of ZigZag Decoding

Chunk 1 from user A from 1st copy of collided packet can be decoded successfully Subtract from 2nd copy to decoded the Chunk 1 of user B

• Subtract from 1st copy of collided packet to decode Chunk 2 from user A– Subtract from 2nd copy of collided packet to decode Chunk 2 from user B

70

Wait! What about Shannon Capacity?

Requires retransmissions if collision occurs No overhead if no collision

R1

R2

TDMA

71

Other alternatives

CDMA Incompatible with WLAN Low efficiency in high SNR

Successive interference cancellation (SIC) Chunk == packet Decode the strong signal first, subtract from the sum and

then decode the weak signal No need for retransmissions Both transmitters need to transmit at a lower rate

72

Patterns that ZigZag Applicable

Both backward and forward decoding can be used

73

Technical Barriers

How do I know packets collide Matching collision happened? (P1, P2) and (P1’, P2’) Frequency offset between transmitter and receiver Sampling offset Inter-symbol interference What if errors occur in chunks Acknowledgement? } subtraction is

non-trivial

74

Evaluation

14-node GNURadio testbed USRP with RFX2400 radio (2.4 GHz) BPSK Bit rate 500kbs 32-bit preamble 1500-byte payload, 32-bit CRC

Deficiency in GNURadio Cannot coordinate transmission and reception very closely CSMA, ACK

Transmitter Receiver

Software

75

Micro-benchmark

76

Alice & Bob

Bob’s location is fixed, Alice moves closer to the base-station

77

Impact of SNR on BER

Alice & Bob at fixed and equal location

Vary transmission power level

78

Testbed Results

Pick two senders randomly 10% hidden terminals, 10% partial, 80% perfect

79

Three hidden terminals

80

Conclusion

ZigZag improves fairness & throughput Further research

Combination of analog network coding

81

Questions?

82

Preliminary on communication

BPSK: 0 -> -1 1 -> 1

http://en.wikipedia.org/wiki/QPSK

driftfrequency todue ][][][

collision of presencein ][][][][

channel)invariant time(a ][][][

2 nwenHxny

nwnynyny

nwnHxny

fTj

BA

+=

++=+=

πδ

83

Collision Detection

Preamble Pseudo random number Correlation with moving

window thresholding

∑=

=ΔΓL

kB ksH

1

2|][|)('

84

Matching collision

Given (P1 + P2(Δ)) and (P1’, P2’(Δ’)), how to determine that P1 = P’ and P2 = P2’’ Determine offset first Correlation of P2(Δ) and P2’(Δ’)

85

Decode matching collision

Decode iteratively Re-encoding

Computing channel parameters• Channel gain estimated from

• Frequency offset and sampling error 1) coarse estimation from previously successful reception 2) iterative estimation

• Inter-symbol interference: take the inverse of linear filter (for removal of ISI)

∑=

=ΔΓL

kB ksH

1

2|][|)('

∑−=

+=L

LlISIl lixhix ][][

86

Decode matching collision (cont’d)

Re-encoding Account for sampling error

))((sin][][

][][ 2

inciyny

enxHny

AAAA

TfjAAA

A

−+=+

=

∑∞

∞−

μπμ

πδ

87

What about errors?

Will errors in decoding have a cascading effect? Error propagation dies out exponentially

• Error correction capability of modulation

Forward and backward decoding

88

Acknowledgement

Use as much synchronous acknowledgement as possible for backward compatibility

89

Duke EWLAN Topology

Client, AP placement traces used to feed Qualnet Fading models from Qualnet

Only 4 topologies shown in graph

90

Increasing AP Density

Shuffle throughput higher in denser conditions Greater scope to “squeeze in” transmissions in space

91

Latency Improves

Latency increases due to higher concurrency As well as from TDMA scheduling

92

Under Channel Fading

Corrective rehearsal effective to cope with fading We observed loss fraction of 12% under Ricean.

93

Ongoing Work

Integrating upload traffic Proposing upload windows Can be opportunistically used for download (ZMAC) Can be used to accommodate client joins, departure

Interference from external networks affect schedule Need to treat border APs separately

Interference cancellation may decode both signals More powerful than MIM, hence, new MAC necessary We are investigating possibilities through GNU radio

94

Today’s Menu

Spotlight

ShuffleShuffle

Micro-BlogMicro-Blog

Mingle

95

Micro-Blog:A Virtual Information Telescope using Mobile Phones and Social Participation

96

Mobile Phones = Powerful Sensors

Next Generation Mobile Phones Variety of embedded sensors

- Cameras, mic., accelerometer, health monitor, RFID reader

3 Billion active phones 2009 - phone sales will surpass computers Convergent device accepted technologically, socially

97

Vision

Envision each mobile phone as a virtual lens

Imagine an Information Telescope over 3 billion lenses

Enabling you to zoom in and perceive any part of the worldthrough the eyes and ears of these phones

And even querying them in real time, with automatic, social, or participatory replies

98

Micro-Blog

Virtual TelescopeVirtual Telescope

Internet, Cellular Networks

Internet, Cellular Networks Visualization ServiceVisualization Service

Web ServiceWeb Service

Sensors

Phones

People

Physical SpacePhysical Space

99

Prototype: From Japan

100

Prototype: From Sydney

101

Post-Its in the Air

Information superimposed on virtual space Google maps, Microsoft SensorMap, etc.

Feasible to superimpose on physical space As if sticky notes floating in the air Downloadable into mobile phones

… Prototype for Duke campus

102

Micro-Blog [mobisys2008]

Project Website http://microblog.ee.duke.edu

Project live at http://152.3.193.194/microblog/dev7/microblog.php

103

So, where exactly is the research here ???!!**

104

Many Challenges

Energy-Aware Localization [mobisys08_poster]

GPS offers 7 hours battery, but hi accuracy Alternates tradeoff accuracy for energy

105

Many Challenges

Location Privacy Users do not want to reveal location Partial location important for contextual info.

Incentives No reason for user to participate Designing incentive schemes

Too much information entering the system Information distillation critical

… many many more

106

Thanks a lotfor your time and patience

SyNRG Homepage: http://synrg.ee.duke.edu

107

Analogy

Imagine a graphic equalizer How do you know what setting will play the song best?

If each song had a “known tune” preceding it• You could set the graphic equalizer based on the tune

• Then listen to the song well

• Analogous to “locking” on to the song

108

Similarly …

Payload in data frame preceded with PLCP PLCP like pilot signal Receiver uses for synchronization/correction with Tx

During synchronization, Rx susceptible to distraction Once synchronized, following bits can be well decoded

• However, if strong interference, then collision

BackBack

109

Backup Slides

110

111

The Menu

Spotlight

Shuffle

Micro-Blog

Mingle

112

Optimal link schedule w, w/o MIM:shows potential gain with MIM-

awareness

Optimal link schedule w, w/o MIM:shows potential gain with MIM-

awareness

Integer Programming

113

Vision

Design a (software) information telescope to zoom into a any part of the world,

and view it through virtual lenses located there

Design a (software) information telescope to zoom into a any part of the world,

and view it through virtual lenses located there

Query the lenses in real timeQuery the lenses in real time

Incentivize participatory sensingEnable automatic sensing

Incentivize participatory sensingEnable automatic sensing

114

Our Research

PHY

MAC / Link

Network

Transport

Security

ApplicationIncentives

Channel fluctuations

Spatial Reuse

MobilityEnergy Savings

EavesdroppingLoss Discrimination

Privacy

Ubiquitous Services

Interference Mgmt.

What can be enabled(bottom up)

What can be enabled(bottom up)

What are the visions(top down)

What are the visions(top down)

115

Shuffle: 3 Main Components

1. Measuring interference relationship - Rehearsal- Controller orchestrates rehearsal- Each node measures interference map from all others- Result is a network-wide interference map

2. Scheduler determines links and order of transmission- From the interference map- Scheduling NP-Hard --> approximation algorithms- Packets scheduled in batches

3. Schedule manager executes schedule- Copes with failures, fading, mobility- Performs pre-fetching, speculation, prediction

116

Collection of wireless hosts Relay packets on behalf of each other Together form an arbitrary topology May be connected to wired infrastructure

2 reasons to prefer multihop Capacity and Power constraint

Wireless Multihop Networks

B

AC

D

117

MIM Scheduling

0/1 Solution to ILP satisfies

Hence, IP solution is the time-ordered schedule

118

2 Key Architectures

Single hop networks

Multi-hop networks

119

Wireless Single Hop Networks

Cellular Networks

Distributed WLANs

Centralized Enterprise WLANs

120

Collection of wireless hosts Relay packets on behalf of each other Together form an arbitrary topology May be connected to wired infrastructure

2 reasons to prefer multihop Capacity and Power constraint

Wireless Multihop Networks

121

The Context

The edge of the internet becoming wireless 167,000 hotspots by 2008 end [GartnerSurvey06]

75 million user base Mesh network extensions to rural regions

Many Motivations to get unplugged Unrestricted mobility Significantly lower deployment/maintenance cost Ease of use

122

Proliferating Applications and Technologies

When combined in synergy …

Mesh NetworksMesh Networks Sensor NetworksSensor Networks

Social CommunitiesSocial Communities

Mobile NetworksMobile Networks

Ad Hoc NetworksAd Hoc Networks

RFID TrackingRFID Tracking

Personal Area NetworksPersonal Area NetworksHybrid NetworksHybrid Networks

Mobile BloggingMobile Blogging

Location ServicesLocation Services

Smart ClothesSmart Clothes

Information MappingInformation MappingGamingGaming

123

The Key Intuition

Ability to Decode = Ability to Cancel

In other words,

knowing the structure of interference, helps in coping with it

In other words,

Stronger, decipherable interference better than weak, undecipherable ones

124

Theory Vs Practice

Theoretically, cancellation feasible

In practice, perfect cancellation difficult Suppression beneficial Along with some help from higher SINR requirements

Suppression + Higher SINR = Concurrency

1 Order Matters: Interference-Aware Transmission Reordering in Wireless Networks.

Documents

signal slide

db slide

message slide

staggered transmission

concurrency feasible

weaker transmission

stronger signal

mim aware scheduling