ParCast: Soft Video Delivery in MIMO-OFDM WLANs

Xiaolin Liu Wenjun HuQifan Pu Feng Wu Yongguang Zhang

Microsoft Research AsiaUniversity of Science and Technology of China

Wireless trends

Wireless trends

• Channel capability

Trend: MIMO-OFDM becoming primitives

Wireless trends

• Channel capability • Application demand

Abcdef…

Trend: MIMO-OFDM becoming primitives

Trend: Video traffic dominatesVideo over wireless to rule!

Traditionally…Divide into blocks and Transform

10000100101001110010001101001100101001001110

……

Quantize + entropy coding

1000010010100111001000

1101001100101001001110

……

Packetize

100001001011000011001110010000011100

110100110011101001010010011100101110

+ protection

Separate source and channel codesModulate

Challenge for source coding

Digital rates do not fall back gracefully– Cliff effects at some SNR– Glitches due to bit errors

Vide

o qu

ality

Channel quality

Vide

o qu

ality

Successfully received packet index

Ideally, want graceful degradation with the channel

Challenge for channel coding

Efficiency vs complexity tradeoff

– 100+ heterogeneous subchannels in practice• E.g., 52 OFDM subcarriers in 802.11n per 20MHz, 2-4

MIMO spatial paths per subcarrier – One code for all does not fully utilize capacity– Subchannel specific modulation impractical

Chan

nel

qua

lity

Subchannel indexUnicast over MIMO-OFDM resembles broadcast over narrowband SISO channels

Ideally, want one code to gracefully degrade on all subchannels

Our solution

• A single joint source-channel code to work gracefully with SNR on all subchannels– Take a leaf from SoftCast’s book – Apply similar principles to MIMO-OFDM unicast

Source characteristics

First-frame 8x8 block DCT coefficient energy distribution

Energy difference over several orders of magnitude

Source coding goals

First-frame 8x8 block DCT coefficient energy distribution

Essential to protect the most important bands well

Discard least important bands for compression

Channel characteristics

3x3 MIMO-OFDM subchannel gains on 20MHz channel

Gain difference over several orders of magnitude

Channel coding goals

3x3 MIMO-OFDM subchannel gains on 20MHz channel

Want to utilize the good subchannels

Want to mitigate effects from the bad subchannels

Source-channel similarities

Source: First-frame 8x8 block DCT coefficient energy distribution

Channel:3x3 MIMO-OFDM subchannel gains on 20MHz channel

Source-channel synergy

Source

Decompose

Allocate bits

Source-channel synergy

Source Channel Joint

Decompose

FreqPo

wer

Allocate power

ParCast: Parallel Video Unicast

ParCast: Parallel video unicast

Freq

Preprocessing

a c e

b d f

a + bi c + di e + fi

Analog m

odulation and transm

ission

Sorting and matching

Power allocation

Separating components

Decoding per subchannel

Pow

er

ParCast: Preprocessing

• Video source:

• Channel:

Whole-frame 3D-DCT

a+bi c+di

f+gi h+ki

s1 0

s2

Divide into chunks

Precode

s1s20

Real values with variances λ1 , λ2 , …

Decorrelate and separatesource/channel components

ParCast: Sorting and matching

• Video source: chunks sorted by importance

• Channel: subchannels sorted by gain

• Matching Fine-grained unequal error protection

ParCast: Power allocation

• More power → higher rate/lower distortion• Can formulate/solve optimization problem – Fixed total budget– Minimize ∑ distortion over components

• Source/channel component pair as unit – Power weight = f(λi/si

2)

Helpful when given similar energy spread for both source/channel

ParCast: Power allocation

• More power → higher rate/lower distortion• Can formulate/solve optimization problem – Fixed total budget– Minimize ∑ distortion over components

• Source/channel component pair as unit – Power weight = f(λi/si

2)

– Discard pair if λi/si2 too small

– Per-chunk whitening for hardware considerationCompression and protection in a single code

Further unequal error protection

ParCast: Analog modulation

OFDM symbol OFDM symbol OFDM symbol…Preamble

44 OFDM symbols per packet

…

Chunk i: 1 2 3 4 5 6 7 8 9 Chunk j: 1 2 3 4 5 6 7 8 9

IQ

Complex subchannel

Soft delivery per subchannel to leverage channel capacity

Nspatial path x Nsubcarrier complex symbols

ParCast: Decoding

• No standard MIMO decoding• LLSE to decode DCT coefficients per chunk• Inverse 3D-DCT

ParCast: Parallel video unicast

Freq

Preprocessing

a c e

b d f

a + bi c + di e + fi

Analog m

odulation and transm

ission

Sorting and matching

Power allocation

Separating components

Decoding per subchannel

Pow

er

ParCast: Parallel video unicast

Freq

a c e

b d f

a + bi c + di e + fiParallel, independent

encoding/transmission and reception/decoding

Pow

er

ParCast: Parallel video unicast

Freq

a c e

b d f

a + bi c + di e + fi

Joint source-channel codingLossless compression over

lossy communication

Pow

er

ParCast: Parallel video unicast

Freq

a c e

b d f

a + bi c + di e + fi

Linear codec per subchannel for unicast over MIMO-OFDM

Pow

er

Performance

Implementation

• Source codec in Matlab• Channel dependent modules implemented

on Sora • Channel trace

driven simulation [Halperin et al, SIGCOMM 2010]

Evaluation

• Microbenchmarks– Software radio based experiments– Effects of individual modules • Channel precoding, matching, joint power allocation

• Video quality (PSNR) Comparison– Variants of SoftCast adapted to MIMO– Omni-MPEG over 802.11n– Layered SVC over MIMO

Performance: ParCast vs SoftCast

• Must decorrelate both! – For best unequal protection

• SoftCast does not work well over MIMO– Mixing source chunks =

mixing subchannels

Separate source, mixed channelSoftCast (mixed source) ParCast (Separate source/channel)Mixed source, separate channel

Performance: Stationary links

~ 10 dB

Performance: Mobile scenariofootball

~ 5 dBFairly stable with

delayed CSI

Channel feedback delay period

Related work

• Unequal error protection over MIMO/OFDM– Layered video coding, one layer per subchannel

• Channel codes for graceful degradation– HM over STBC (broadcast), Apex (unicast), etc.

• Joint source-channel coding (single antenna)– Jointly optimizing separate codes– A single code for compression/protection• SoftCast (broadcast), FlexCast (unicast)

Conclusion

ParCast is simple yet effective• One code to work gracefully with SNR on all MIMO-

OFDM subchannels• Joint source-channel coding– Motivated by source-channel synergy– Non-uniform distribution at both provides inherent

unequal error protection• Parallel delivery per source chunk/subchannel– Treating unicast as broadcast– Graceful degradation via linear codec

Thank you!

Questions?