rearchitecting wireless networks with phy layer components
DESCRIPTION
Rearchitecting Wireless Networks with PHY Layer Components. Romit Roy Choudhury Assistant Professor. A little bit about ourselves. Webpage. http://synrg.ee.duke.edu. Our Research. Mobile Computing (top down). Collaborative Sensing. Interfaces. Localization. Application. - PowerPoint PPT PresentationTRANSCRIPT
Rearchitecting Wireless Networkswith PHY Layer Components
Romit Roy ChoudhuryAssistant Professor
1
A little bit about ourselves
2
Webpage http://synrg.ee.duke.edu
3
Our Research
PHY
MAC / Link
Network
Transport
Security
ApplicationLocalization
Software Radios
Home networksMobility
EnergyLocation Privacy
Interfaces
Collaborative Sensing
Interference Mgmt.
Wireless Networking(bottom up)
Mobile Computing(top down)
Rate Control
Smart Antennas
Mobile Computing
Location SensingPhysical and Logical Localization
Information TelescopeMobile Phones for Collaborative Sensing
MobiSys 08 MobiCom 09Infocom 09, 10MobiCom 10
Smart ContentContext-aware content and compression
MobiSys 10Hotmobile 11
Micro-mobilityGesture and activity recognition
MobiHeld 09
5
Wireless Networking
Wired + WirelessInfrastructure Assisted Wireless
MobiCom 09Hotnets 08
Cross-LayerPHY Informed Protocol Design
MobiCom 09, 10Hotnets 09, 10
NSDI 10
SleepWellWiFi Energy Management
In Submission
Out of BandSensor Assisted Wireless Networking
LANMAN 10
6
Today’s Talk
1. Time to Frequency2. AccuRate3. CSMA/CN
Context
Cross-Layer Systems
Mobile Computing
Closing Thoughts
1. Virtual Telescope2. Location3. PhonePen
Cross-LayerPHY Informed Protocol Design
MobiCom 09, 10Hotnets 09, 10
NSDI 10
7
Context
8
Wireless Everywhere
9
Wireless usage increased by 25x in last 5 years
Cisco predicts 40x increase by 2013
Network outages a reality Major carriers forcing
customers to pay-per-byte
Skyrocketing Demands
FCC looking for 500 MHz spectrum by 2020 …But also calling for much better use of available spectrum
10
Problem is not of spectrum alone Under utilization of available spectrum a major problem
Significant leaps in achievable PHY capacity MIMO, OFDM, Coding, Beamforming …
Yet, this PHY capacity not visible to higher layers Inefficiencies in network design …
protocols … architecture
Capacity vs. Goodput
PHY Bitrate
LinkThroughput
11
The capacity-throughput gap is not new Researchers recognized need to share information across layers Cross layer approaches became popular
Cross layer optimization Several creative ideas … many analyzed and simulated
Layering too Restrictive?
Capacity
Throughput
12
The capacity-throughput gap is not new Researchers recognized need to share information across layers Cross layer approaches became popular
Cross layer optimization Several creative ideas … many analyzed and simulated
However, 2 deficiencies
Layering too Restrictive?
1. Lack of experimentation platform difficult to build practical working systems
2. Protocol designers untrained in communications cross layer ideas variants of originals uses some PHY layer info. Capacity
Throughput
13
Software Radios Software defined radios
Changing landscape of wireless systems
Protocol designers understanding PHY concepts, using them PHY community receiving feedback from practical systems
Full view of PHY layer enabling experimentation with holistic, unconventional ideas …
We intend to contribute here14
Our Goal:
Rearchitect wireless networks with full access to PHY layer capabilities
15
We instantiate our ideas through WiFi
However, the core ideas not specific to WiFi …choice of WiFi mainly from platform considerations
16
WiFiProtocol Structure
17
WiFi Structure
AP1 AP2
R1 R2
Packet for R1 Packet for R2
18
WiFi Structure
AP1 AP2
R1 R2
Random Backoff = 10
Random Backoff = 18
19
WiFi Structure
Time
AP1 AP2
R1 R2
Random Backoff = 10
Random Backoff = 18
AP1 = 10
AP2 = 1820
WiFi Structure
Time
AP1 AP2
R1 R2
RemainingBackoff = 0
RemainingBackoff = 8
AP1 = 0
AP2 = 821
WiFi Structure
Time
AP1 AP2
R1 R2
Transmit @ rate = r1
Channel Busy
AP1 = 0
AP2 = 8
Data
AP2 Waits
ACK
22
WiFi Structure
AP1 AP2
R1 R2
AP1 = 15
AP2 = 8
Data
AP2 Waits
ACK
NewBackoff = 15
RemainingBackoff = 8
23
WiFi Structure
AP1 AP2
R1 R2
AP1 = 7
AP2 = 0
Data
AP2 Waits
ACK
RemainingBackoff = 7
RemainingBackoff = 0
24
WiFi Structure
AP1 AP2
R1 R2
AP1 = 7
AP2 = 0
Data
AP2 Waits
ACK
Channel Busy Transmit @ rate r2
AP1 Waits
Data ACK
25
WiFi Structure
AP1 AP2
R1 R2
Data
AP2 Waits
ACK
Transmit Channel Busy
AP1 Waits
Data ACK
Data
26
WiFi Structure
AP1 AP2
R1 R2
Data
AP2 Waits
ACK
Channel Busy ACK NotReceived
AP1 Waits
Data✘Interference
27
WiFi Structure
AP1 AP2
R1 R2
Data
AP2 Waits
ACK
Channel Busy Adjust Rate & Retransmit
AP1 Waits
DataInterference
✘
Data
28
WiFi Structure
AP1 AP2
R1 R2
Data
AP2 Waits
ACK
Channel Busy Adjust Rate & Retransmit
AP1 Waits
DataInterference
✘
Data
29
WiFi Structure
AP1 AP2
R1 R2
Data
AP2 Waits
ACK
Channel Busy Adjust Rate & Retransmit
AP1 Waits
DataInterference
✘
Data
Channel Wastage
30
WiFi Structure
AP1 AP2
R1 R2
Data
AP2 Waits
ACK
Channel Busy Adjust Rate & Retransmit
AP1 Waits
DataInterference
✘
Data
Channel Wastage
Collision or Fading 31
WiFi Structure
AP1 AP2
R1 R2
Data
AP2 Waits
ACK
Channel Busy Adjust Rate & Retransmit
AP1 Waits
DataInterference
✘
Data
Heuristic Rate SelectionChannel Wastage
Collision or Fading 32
WiFi Structure
AP1 AP2
R1 R2
Data
AP2 Waits
ACK
Channel Busy Adjust Rate & Retransmit
AP1 Waits
DataInterference
✘
Data
RedundancyHeuristic Rate SelectionChannel Wastage
Collision or Fading 33
RedundancyHeuristic Rate SelectionChannel Wastage
34
PHY Layer Information
RedundancyHeuristic Rate SelectionChannel Wastage
(OFDM, Constellation, Interference Cancellation, Correlation …)
35
PHY Layer Information
RedundancyHeuristic Rate SelectionChannel Wastage
(OFDM, Constellation, Interference Cancellation, Correlation …)
Software Radios (USRP, WARP, SoRa)
36
PHY Layer Information
RedundancyHeuristic Rate SelectionChannel Wastage
(OFDM, Constellation, Interference Cancellation, Correlation …)
Software Radios (USRP, WARP, SoRa)
Cross-Layered Network Systems37
1. Channel Wastage due to Randomized Backing off
38
Backoff
Data
AP2 Waits
ACK AP1 Waits
Data ACK
Data
Per packet backoff 35% of channel wastage at 54 Mbps.Worse at higher data rates.
39
Fundamentally,backoff is not a time domain operation …
its implementation has been in the time domain
40
Fundamentally,backoff is not a time domain operation …
its implementation has been in the time domain
We intend to break away,and implement backoff on the frequency domain
41
Frequency Domain 802.11a/g PHY adopts OFDM
Wideband channel divided into 48 narrow sub-carriers Copes better with fast, frequency selective fading Purely a PHY layer motivation
MAC Opportunity Pretend OFDM subcarriers are integers Emulate randomized backoff
Frequency
Subcarriers: 1 2 3 4 … 48
42
T2F: Main Idea Pick random backoff, say 6 Submit signal on 6th subcarrier
0 47
6
0 47
18
AP1 Backoff = 6 AP2 Backoff = 18
43
0 47
6
0 47
18
Listen Antenna
Listen Antenna
6 18 6 18
Pick random backoff, say 6 Submit signal on 6th subcarrier
T2F: Main Idea
AP1 Backoff = 6 AP2 Backoff = 18
AP2 learns some other AP is winner.AP1 learns AP1 is the winner … hence, AP1 transmits
44
Subcarrier0 1 2 3 4 5
Second Round
What if Collision?
Introduce a second round of contention Winners of first go to second
Subcarrier0 1 2 3 4 5
First RoundWinner
45
Why beneficial?
Avg. temporal backoff = 16 slots = 144 micro sec.
Frequency backoff = 1 OFDM symbol = 4 micro sec
2 rounds of backoff = 8 micro sec.
Possible to do better …
46
Subcarrier0 1 2 3 4 5
Second Round
Creating a Queue
Subcarrier0 1 2 3 4 5
First Round
Winner Rank 2
47
Subcarrier0 1 2 3 4 5
Second Round
Creating a Queue
Subcarrier0 1 2 3 4 5
First Round
0 2 40 2 4 0 2 4Rank 1 Rank 2 Rank 3 EnablingTDMA
48
Improved Channel Utilization
Data DataData Data
Data DataData Data
WiFi: Contention per packet
T2F: OFDM contention per TDMA schedule
TDMA
49
Multiple Collision Domains What happens with real-world scatterred APs
50
Multiple Collision Domains What happens with real-world scatterred APs
B waits for O, but O waits for G
G 1
O 2 W 5B 3
51
Multiple Collision Domains What happens with real-world scatterred APs
B waits for O, but O waits for G
Protocol: Each node continuously carrier senses If node sees channel idle for DIFS, redo backoff
• New backoff = Old backoff – Backoff of past winner Else, if channel idle for PIFS duration
• Count ++• If Count = Rank, transmit
G 1
O 2 W 5B 3
52
Multiple Collision Domains What happens with real-world scatterred APs
B waits for O, but O waits for G
T2F precisely emulates 802.11 Only every backoff is nearly instantaneous
G 1
O 2 W 5B 3
G 1 W 5 4 O 2
B 3 1
53
Performance 10 USRP Testbed
Deployed in Duke
Quantify Reliable subcarrier detection Collision probability Net throughput gain over WiFi
54
Subcarrier DetectionSN
R in
dB
FFT NumberReliable subcarrier detection at 14dB
55
Collision Probability
Small collision probability in dense networks
Benefit of second round
56
Throughput Gain
Throughput Gain of 27% at 36Mbps with 15 APs57
Closing Thoughts
Contention is not fundamentally a time domain operation T2F shows feasibility in frequency domain Long standing overheads of backoff alleviated
T2F is complimentary to MIMO Additional antenna amenable to T2F
T2F not specific to WiFi Powerline Ethernet runs OFDM and performs backoff T2F applicable in such scenarios too …
58
2. Sub-Optimal Rate Selection
59
Data @ rate R ACK
What is the optimal rate R?
After backing off, AP transmits packet at rate R
60
Estimate from the optimal rate of the previous packet
After backing off, AP transmits packet at rate R
Data @ rate R ACK
What is the optimal rate R?
61
Estimate from the optimal rate of the previous packet
After backing off, AP transmits packet at rate R
Data @ rate R ACK
What is the optimal rate of the previous packet?
What is the optimal rate R?
62
After backing off, AP transmits packet at rate R
Data @ rate R ACK
What is the optimal rate of the previous packet?
What is the optimal rate R?
More generallyGiven any transmission at some rate R, what would have been the max rate R*,
at which that transmission would have been successful
Estimate from the optimal rate of the previous packet
63
✦ Recently PHY-based:✦ SoftRate [SIGCOMM ’09]
• Uses a BER heuristic to estimate bit rate• BER accurately identifies when to increase/decrease rate• However, may not be able to jump to optimal rate
Current Wireless Rate Selection
Data
ACK
HistoryInfo. Data
SNR
Frame Based SNR Based
SampleRate, RRAA RBAR, CHARM
We dive deeper into PHY … jump to the optimal rate64
Quick Background:Symbols, Modulation, Bit-rate
65
2 bits together
01111001 ....
11
00 10
01
Tx 4QAM Symbol
Data =
PHY Layer Symbols
66
2 bits together
01111001 ....
Dispersion11
00 10
01
Tx 4QAM Symbol
Data =
11
00 10
01
Rx 4QAM Symbol
11
01
11
01
Channel
Decoding Symbols
67
Dispersion 11
00 10
01
Rx 4QAM Symbol
2 bits together
01111001 ....Data =
11
00 10
01
Tx 4QAM Symbol
Channel
Decoding Symbols
68
Tx 16QAM Symbol
4 bits together
01111001 ....Data =
0111
Channel
Rx 16QAM Symbol
Dispersion
2 bits together
01111001 ....Data =
11
00 10
01
Tx 4QAM Symbol
Channel
Dispersion 11
00 10
01
Rx 4QAM Symbol
Different Modulations in 802.11
69
6 bits together
01111001 ....
Tx 64QAM Symbol
011110
Data =
Rx 64QAM Symbol
Channel
Dispersion
2 bits together
01111001 ....Data =
11
00 10
01
Tx 4QAM Symbol
Channel
Dispersion 11
00 10
01
Rx 4QAM Symbol
Different Modulations in 802.11
70
High Dispersion
Data 01111001 ....=
Tx 16QAM Symbol
WeakChannel
Wrongly demodulated symbol
Weak Channel Induces Error
Picking the right modulation is a function of channel quality Excess dispersion induces decoding error
71
01
0
StrongChannel
ModerateChannel
WeakChannel
0111
In General
72
01
0
StrongChannel
ModerateChannel
WeakChannel
0111
6 Mbps
24 Mbps
36 Mbps Smaller dispersion permits higher rate
In General
73
AccuRate Hypothesis: Symbol dispersion is independent of modulation
Observation: Dispersion indicates the optimal rate that should have been used for that packet
74
Hypothesis Verification
11
00 10
01
Tx 4QAM
Tx 16QAM
Channel
11
00 10
01
Rx QPSK
Rx16QAM 75
McKinley et. al., 2004, “EVM calculation for broadband modulated signals”
Hypothesis Verification
76
Observe symbol dispersion and select optimal modulation
Hypothesis: Symbol dispersion is independent of modulation
77
DataBPSK
4QAM
16QAM
78
DataBPSK
4QAM
16QAM
79
BPSK
4QAM
16QAM
Data
We call this Virtual Channel Replay
80
Channel Replay Vector
d1 Vector V = {d1, d2, ...., dn}d2
81
Receiver
Demodulator
Packet
Best Rate
BPSK Channel Replay
Demodulator CRCCheck
4QAM Channel Replay
Demodulator CRCCheck
16QAM Channel Replay
Demodulator CRCCheck
82
Optimal modulation ≠ Optimal rateBit-rate is a function of
both modulation and coding
Need to find the optimal <modulation,coding> for a received packet?
83
Receiver
Demodulator
Data BPSK Channel Replay1/2 Demodulator
CRC CheckDecoder
BPSK Channel Replay3/4 Demodulator
CRC CheckDecoder
QAM4 Channel Replay1/2 Demodulator
CRC CheckDecoder
QAM4 Channel Replay3/4 Demodulator
CRC CheckDecoder
18 Mbps
Decoder
6 Mbps
9 Mbps
12 Mbps
QAM64 Channel Replay3/4 Demodulator
CRC CheckDecoder
Best Rate
54 Mbps
84
Packet Failure
Packet failure implies incorrect dispersion vector
All packets have globally known preamble/postamble Used for detection and synchronization to new signals
Compute the replay vector for the preamble alone Precise dispersions known for this vector Replicate this vector to model the complete dispersion vector
Vector V’ = {d1, d2, ...., dn}d1
d2
e2 e2
Cannot replay this vector, V’
85
AccuRate needs to discriminate interference from collision
Rate selection needs to be independent of interference
86
How to Detect Interference?
Interference causes substantial symbol dispersion Reliable indicator
With InterferenceWithout Interference
87
How to Detect Interference?
Interference starts first: Preamble with high dispersion
Interference starts second: Postamble with high dispersion
Compare preamble with postamble dispersion
Data
DataInterference
Interference
88
Performance Evaluation
Used 802.11 like Tx and Rx design on USRP/GnuRadio Modulation: BPSK, QPSK, 16QAM, 64QAM Coding: Convolution coding with puncturing with rate 1/2, 3/4 Compare with Softrate, SNR-based
Testbed 10 traces at walking speed Trace based evaluation
Simulation Characterize AccuRate’s performance under high mobility Raleigh fading channel simulator ported to GnuRadio
RealChannel
Simulator
89
What is the True Optimal Rate?
Testbed Using train of packets (virtual packet) Virtual packet composed of train of small packets
• Each short packet at increasing rate
Virtual Packet 6Mbps 9Mbps 12Mbps 18Mbps 24Mbps 36Mbps 54Mbps
OptimalOptimal -1 Optimal+1
90
Rate Estimation Accuracy
AccuRate
Incorrect rate selection ~ 4%
For correctly received packets,
100% in Simulation,95% in Testbed
91
Interference Detection Accuracy
Detection Accuracy better at higher rates (91%)
Testbed
92
Overall Rate Selection Accuracy
93% accuracy in optimal rate selection
Correct
Overselect
Underselect
Testbed
93
AccuRate achieves 87% of the optimal throughput
Testbed
Throughput at Walking Speeds
94
3. Collision Detection in Wireless Networks
95
Data ✘
Transmitter infers a collision from the absence of an ACK.Prepares to retransmit the packet.
Interference Data ACK
96
Data Data✘
Bits transmitted unnecessarily
ACK
Interference
Bits transmitted redundantly
Undesirable Channel Wastage
97
Ethernet is Efficient
Called Collision Detection (CSMA/CD)
Collision
Transmitter detects a collision,and immediately aborts transmission.
Unfortunately, CSMA/CD not feasible in wireless networks …98
We ask: Can we emulate CSMA/CD in wireless networks
i.e., abort collisions right when they occur
99
MAC
PHY
MAC
PHY
Data Transmission (S1)
S=S1
Tx
Rx
Cros
s La
yer
Cros
s La
yer
CSMA/CN: Basic Idea
100
MAC
PHY
MAC
PHY
Data Transmission (S1)
S=S1
Tx
Rx
CollisionDetected
Cros
s La
yer
Cros
s La
yer
CSMA/CN: Basic Idea
101
MAC
PHY
MAC
PHY
Data Transmission (S1)
S=S1
Tx
Rx
Abort signaldetected
Abort Signal (S2)S=S1+S2
Cros
s La
yer
Cros
s La
yer
Collision.Send Abort
CSMA/CN: Basic Idea
102
MAC
PHY
MAC
PHY
Data Transmission (S1)
S=S1
Tx
Rx
ABORT
Abort Signal (S2)S=S1+S2
Cros
s La
yer
Cros
s La
yer
Collision.Send Abort
CSMA/CN: Basic Idea
103
3 Challenges
1. Detect collision in real time at the receiver2. Detect notification at the transmitter3. Design protocol to ensure correct transmissions aborted
MAC
PHY
MAC
PHY
Data Transmission (S1)
S=S1
Tx
Rx
ABORT
Abort Signal (S2)S=S1+S2
Cros
s La
yer
Cros
s La
yer
Collision.Send Abort
104
Wired
Wireless
Interference Cancellation Need to detect very weak notification signals
Opportunity Pass the Tx signal over wire Listen antenna has 2 copies
of the Tx signal
Both copies have same filter and frequency offset effects
Align the two signals using sampling offset information Subtract the wired signal from wireless
Correlate residue with collision notification
105
Collision Detection at Rx Receiver detects collision within 20 bytes Total turnaround time for CN signature 18us
Quicker turnaround Faster Tx abortion
Throughput gain over PPR
MAC
PHY
Median gain = 25%
106
In Closing …
Its not adequate to only use PHY layer information for higher layer protocols
Need to rethink protocol design holistically(with PHY layer capabilities folded into it)
We are trying to design such holistic systems, while respecting the lessons learnt from layering
107
Our Research
PHY
MAC / Link
Network
Transport
Security
ApplicationLocalization
Software Radios
Home networksMobility
EnergyLocation Privacy
Interfaces
Collaborative Sensing
Interference Mgmt.
Wireless Networking(bottom up)
Mobile Computing(top down)
Rate Control
Smart Antennas
Mobile Computing
Location SensingPhysical and Logical Localization
Information TelescopeMobile Phones for Collaborative Sensing
MobiSys 08 MobiCom 09Infocom 09, 10MobiCom 10
Smart ContentContext-aware content and compression
MobiSys 10Hotmobile 11
Micro-mobilityGesture and activity recognition
MobiHeld 09
109
Mobile Computing
Location SensingPhysical and Logical Localization
MobiCom 09Infocom 09, 10MobiCom 10
Smart ContentContext-aware content and compression
MobiSys 10Hotmobile 11
Micro-mobilityGesture and activity recognition
MobiHeld 09
Information TelescopeMobile Phones for Collaborative Sensing
MobiSys 08
110
Next generation mobile phones will havelarge number of sensors
Each phone may be viewed as a micro lensExposing a micro view of the physical world to the Internet
111
With 4.1 billion active phones in the world today
(the fastest growing computing platform …)
Our Vision is …
112
Internet
A Virtual Information Telescope
113
Free WiFi?
114
People
Virtual Information Telescope
Location Energy Privacy Interfaces . . .. . .
Apps
Research
115
Mobile Computing
Information TelescopeMobile Phones for Collaborative Sensing
MobiSys 08
Smart ContentContext-aware content and compression
MobiSys 10Hotmobile 11
Micro-mobilityGesture and activity recognition
MobiHeld 09
Location SensingPhysical and Logical Localization
MobiCom 09Infocom 09, 10MobiCom 10
116
RadioShackStarbucks
Physical LocationError
The dividing-wall problem117
Its possible to localize phones by sensing the ambience
such as sound, light, color, movement, orientation…
Hypothesis
118
Mobile Computing
Location SensingPhysical and Logical Localization
Information TelescopeMobile Phones for Collaborative Sensing
MobiSys 08 MobiCom 09Infocom 09, 10MobiCom 10
Micro-mobilityGesture and activity recognition
MobiHeld 09
Smart ContentContext-aware content and compression
MobiSys 10Hotmobile 11
119
Can phones automatically create video highlights of a social gathering?
120
Goal
Envisioning the end product Imagine a social party of the future Assume phones are wearable
GOAL: Create automatic video highlights of the occasion
The Idea: Mobile phones sense ambience Collaboratively infer an “interesting event” Select phone with good view of the event Stitch the recorded clips to form the highlights
Nokia Morph
MicrosoftSenseCam
AppleIPod Nano
121
Mobile Computing
Location SensingPhysical and Logical Localization
Information TelescopeMobile Phones for Collaborative Sensing
MobiSys 08 MobiCom 09Infocom 09, 10MobiCom 10
Smart ContentContext-aware content and compression
MobiSys 10Hotmobile 11
Micro-mobilityGesture and activity recognition
MobiHeld 09
122
PhonePoint Pen Using phone accelerometers
To write short messages in the air
123
PhonePen words
124
See PhonePoint Pen YouTube Video at
http://synrg.ee.duke.edu/media.htm
125
Thank You
Visit our Systems Networking Research Group (SyNRG)
Google “SyNRG”
126
127
Ordered Transmissions Each node has global view of backoff values
Computes its rank among contending APs Enables TDMA
470 12
Rank in TDMA: 3AP1
Self Backoff
Other’s Backoff
128
App Store 2020 OmniSearch
Is there live music at the plaza now? Do people around 9th street feel safe?
LiveLearn Learn about Mt. Everest via videos from base camp, interact with Sherpas Zoom into a location, and replay the history from any time in the past
VirtualWindShield Content overlayed on vehicle windshield … show street view … speed limit
BlindSight Sensors “see” the surrounding and speaks into the ears of blind people
MyLife’sImportantBits Automatic video highlights of your everyday life
AirWrite, MindWrite Write SMS in the air … send SMS by thinking
129
Beyond Location, Energy, Privacy …
Humans part of the end device 3 eyes (2 human and 1 camera) … 2 CPUs (1 Intel and 1 brain)
Activity, emotion, and intent recognition Feasible from the ability to sense and mine multi-dimensional data
streams Extend to vision / cognition
Information distillation Pulling out the “signal” from the noise
Agile networking Sensor assisted wireless communications
User interface Intuitive, implantable, cognitive
130
Caveat:
We are not innovating in communications.
We are adopting well known techniques to rearchitect wireless networks
131
We Ask:
If network protocols were designed with joint PHY and higher layer information,
would they be the same as what we have today?
The answer is: No
132
We Ask:
If network protocols were designed with joint PHY and higher layer information,
would they be the same as what we have today?
The answer is: No
This talk explains why not, and carves out opportunities of improvement.
133
134
Problem is not of spectrum alone Significant leaps in achievable PHY capacity MIMO, OFDM, Coding, Beamforming …
This PHY capacity not visible to higher layers Inefficiencies in network design … architecture
Capacity vs. Goodput
WiFi Structure
Time
AP1 AP2
R1 R2
Random Backoff = 10
Random Backoff = 18
AP1 = 10
AP2 = 18135
Today’s Talk
Cross-LayerPHY Informed Protocol Design
Context
Cross-Layer Systems
Mobile Computing
Closing Thoughts
1. Time to Frequency2. AccuRate3. CSMA/CN
1. Virtual Telescope2. Location3. PhonePen
MobiCom 09, 10Hotnets 09, 10
NSDI 10
136
Backoff
Data
AP2 Waits
ACK AP1 Waits
Data ACK
Data
137
138
Overwhelming Adjacent Subcarrier
256 Point FFT
139
Solution: Use a Higher Point FFT
Subcarrier Number
SN
R (d
B)
SN
R (d
B)
-10 -5 0 5 10 Subcarrier Number
70
60
50
40
30
20
10
0
512 point FFT
1024 point FFT
Reliable subcarrier detection at 15dB
Data
140
Subcarrier Detection Accuracy
141
Dete
ctio
n Ac
cura
cy
Distance in Subcarriers
Reliable subcarrier detection at 14dB
Simulation Testbed
Rate Estimation Accuracy
For correctly received packets,
100% in Simulation,95% in Testbed 142
143
Practical Requirements?
Collision
Transmitter cannot detect collision Receiver needs to detect it
Receiver needs to convey Collision notification to the transmitter
Transmitter needs an additional antenna To receive the notification
Collision Notification
144
CSMA/CN Summary CSMA/CN imitates CSMA/CD in wireless
Aborts collision upfront, instead of recovering from it (PPR)
Correlation is the key Enough for 1 bit feedback
2 important future directions How can protocols be redesigned with such few bit feedbacks Can we move from correlation to decoding (full duplex)
Prevention Better than Cure
145
Practical Challenge: High Self Signal
AccuRate performs well even under high mobility
Under Varying Mobility
146
AccuRate estimates the optimal rate for an already received packet
What is the performance if the next transmission uses this rate?
147
Distance
Signalpower
Collision
Thus, wireless networks must perform less efficient CSMA/CA148
MAC
PHY
MAC
PHY
Tx
Rx
Cros
s La
yer
Cros
s La
yer
We propose CSMA/CN
149
MAC
PHY
MAC
PHY
Data Transmission (S1)
S=S1
Tx
Rx
Check forCollision
Cros
s La
yer
Cros
s La
yer
CSMA/CN: Basic Idea
150
MAC
PHY
MAC
PHY Cros
s La
yer
Data Transmission (S1)
S=S1
Tx
Rx
Check forCollision
Search forAbort
Cros
s La
yer
CSMA/CN: Basic Idea
151
MAC
PHY
MAC
PHY
Data Transmission (S1)
S=S1
Tx
Rx
Search forAbort
Cros
s La
yer
Cros
s La
yer
Check forCollision
CSMA/CN: Basic Idea
152
MAC
PHY
MAC
PHY
Data Transmission (S1)
S=S1+S2
Tx
Rx
Collision.Send Abort
Search forAbort
Abort Signal (S2)
Cros
s La
yer
Cros
s La
yer
CSMA/CN: Basic Idea
153
MAC
PHY
MAC
PHY
Data Transmission (S1)
S=S1
Tx
Rx
Abort signaldetected
Abort Signal (S2)S=S1+S2
Cros
s La
yer
Cros
s La
yer
Collision.Send Abort
CSMA/CN: Basic Idea
154
MAC
PHY
MAC
PHY
Data Transmission (S1)
S=S1
Tx
Rx
ABORT
Abort Signal (S2)S=S1+S2
Cros
s La
yer
Cros
s La
yer
Collision.Send Abort
CSMA/CN: Basic Idea
155
156
3 Challenges1. Detect collision in real time at the receiver2. Detect notification at the transmitter3. Design protocol to ensure correct transmissions aborted
MAC
PHY
MAC
PHY
Data Transmission (S1)
S=S1
Tx
Rx
ABORT
Abort Signal (S2)S=S1+S2
Cros
s La
yer
Cros
s La
yer
Collision.Send Abort
Challenge 1Data Data
Correlate for Preamble + SoftPHY hints
Detect collision in real time
Main Idea Receiver correlates for preamble If preamble present
Continuously monitor symbol level confidence (SoftPHY)
Achieves 96% collision detection accuracy at the receiver157
Challenge 1Data Data
Correlate for Preamble + SoftPHY hints
Detect collision in real time
Main Idea Receiver correlates for preamble
If preamble present Continuously monitor symbol level confidence (SoftPHY)
Confidence (symbol) = Log likelihood ratio (symbol) BER = Avg. (Confidence (symbol) If BER_i > 3*BER_(i-1) Announce Collision
Achieves 96% collision detection accuracy at the receiver158
Challenge 2 Detect collision notification
In the same frequency channel
Main Idea No need to decode the abort signal correlate for signature
Data
Abort
159
Challenge 2 Detect collision notification
In the same frequency channel
Main Idea No need to decode the abort signal correlate for signature
Data
Abort
Cor
rela
tion
Sample Number
Works when notification is no weaker than 18dB of self-signalCorrelation spikes whenever notification arrives
160
Wired
Wireless
Challenge 2 Need to detect very weak notification signals
Opportunity Pass the Tx signal over wire Listen antenna has 2 copies
of the Tx signal
Both copies have same filter and frequency offset effects
Align the two signals using sampling offset information Subtract the wired signal from wireless
Correlate residue with collision notification
161
Reliable notification detection until 34dB below
(Self Signal) - (Notification Signature)(dB)
Fr
actio
n of
False
+/-
False positivesFalse negatives
Challenge 2 Performance with cancellation
162
163
Challenge 3: Protocol Design
Data Data
T1
R
Collision
T2
R2R1
Signal starts before interference
164
Data Data
R
Correlate (Sign(R1))
Sign(R1) Sign(R2)
Collision
T1T2
R2R1
Signal starts after interferenceChallenge 3: Protocol Design
Data Data
R
Sign(R1)
Corr (Sign(R1))
Notification!Stop Tx
Collision
T1T2
R2R1Correlate (Sign(R1))
Sign(R1) Sign(R2)
Signal starts after interference Notification aborts the correct transmission
Challenge 3: Protocol Design
165
166
Performance Evaluation 10 node USRP testbed BPSK, QPSK modulation Signature size: 20 bytes Topologies with three links doing CSMA/CN Compare with 802.11 and PPR
PPR retransmits only suspected bits of the packet
167
Collision Detection at Rx Receiver detects collision within 20 bytes Total turnaround time for CN signature 18us
Quicker turnaround Faster Tx abortion
Throughput gain over PPR
MAC
PHY
Median gain = 25%
168
CSMA/CN Summary
CSMA/CN imitates CSMA/CD in wirelessAborts collision upfront, instead of recovering from it (PPR)
Prevention Better than Cure
169
In Closing …
People
Virtual Information TelescopeFrontend
Backend
170
Location Sensing
171
Smart Content
172
User Interfaces
173
Some Other Projects
174
Wired + WirelessInfrastructure Assisted Wireless
MobiCom 09Hotnets 08
SleepWellWiFi Energy Management
In Submission
WiFi as sensorsSleepWellHome networksSAWCSmart antennas
Wide agreement on the need for cross layer systems Exploit information from other layers Optimize protocols/networks
Not a new idea Discussed for several decades Variety of creative thought experiments Several ideas developed and simulated …
Layering Too Restrictive?
175
Yet, cross-layer systems have not made itto the main streamThe PHY and MAC layer gap remains …
even widens as each area advances independently
The MAC layer operates at a coarser granularity than PHY PHY: Signals, bits MAC: packets, frames
Wireless networks have been designed on this principle Data and Control … both use packets
• Even 1 bit feedback, takes up a discrete resource unit
Cross-layer systems allow flexible granularity Control information can be at the granularity of bits/signals Data communication at the granularity of packets/bandwidth
Efficient, streamlined wireless systems
In Closing …
176