Geosphere: C Consistently T y Turning M MIMO C O Capaci city y - - PowerPoint PPT Presentation

geosphere c consistently t y turning m mimo c o capaci
SMART_READER_LITE
LIVE PREVIEW

Geosphere: C Consistently T y Turning M MIMO C O Capaci city y - - PowerPoint PPT Presentation

Geosphere: C Consistently T y Turning M MIMO C O Capaci city y into T Throughput Konstan'nos Nikitopoulos Juan Zhou, Ben Congdon , Kyle Jamieson 5G Innova)on Centre


slide-1
SLIDE 1

Geosphere: C Consistently T y Turning M MIMO C O Capaci city y into T Throughput

Konstan'nos ¡Nikitopoulos ¡ 5G ¡Innova)on ¡Centre ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ University ¡of ¡Surrey ¡ Juan ¡Zhou, ¡Ben ¡Congdon ¡, ¡Kyle ¡Jamieson ¡ Department ¡of ¡Computer ¡Science ¡ University ¡College ¡London ¡

SIGCOMM, ¡Chicago ¡2014 ¡

1 ¡

slide-2
SLIDE 2

Need t to S Sca cale W Wireless C Capaci city… y…

Photo ¡credit: ¡www.cultofmac.com ¡ ¡ ¡ 2 ¡

How ¡can ¡we ¡scale ¡capacity? ¡ Time ¡Division ¡ Mul'ple ¡Access ¡ (uplink) ¡ Increasing ¡the ¡number ¡of ¡users ¡results ¡in ¡a ¡ reduc)on ¡in ¡per-­‑user ¡throughput ¡

slide-3
SLIDE 3

MIMO w O with S Spatial M Multiplexing

Photo ¡credit: ¡www.cultofmac.com ¡ ¡ ¡ 3 ¡

Ques'on: ¡How ¡can ¡we ¡most ¡efficiently ¡demul)plex ¡the ¡ mutually ¡interfering ¡informa)on ¡streams? ¡

slide-4
SLIDE 4

Mot Motivat ation

  • n

4 ¡

q Problem 1: Zero-forcing (e.g., [SAM, Mobicom ’09], [Bigstation, Sigcomm ’13]) suffers as APs get more antennas.

slide-5
SLIDE 5

5 ¡

SNR degradation (dB) CDF across MIMO channels

4x4: ¡More ¡than ¡10 ¡dB ¡degrada)on ¡ ¡ (x10 ¡noise ¡amplifica)on) ¡for ¡50% ¡of ¡the ¡links ¡ 2x2: ¡More ¡than ¡5 ¡dB ¡degrada)on ¡ ¡ (x3 ¡noise ¡amplifica)on) ¡for ¡30% ¡of ¡the ¡links ¡

Motivation: Z Zero-f

  • forci

cing s suffers

slide-6
SLIDE 6

Mot Motivat ation

  • n

6 ¡

q Problem 1: Zero-forcing (e.g., [SAM, Mobicom ’09], [Bigstation, Sigcomm ’13]) suffers as APs get more antennas. q Problem 2: Optimal solutions are very computationally complex and, therefore, cannot scale to high transmission rates. Geosphere: ¡Enables ¡op)mal ¡detec)on ¡at ¡a ¡reasonable ¡ complexity ¡by ¡employing ¡geometrical ¡reasoning. ¡

slide-7
SLIDE 7

Zero-F

  • Forci

cing A Amplifies N Noise

1

x

2

x

1

y

2

y

h21 h

11

h

12

h22

7 ¡

The ¡Noiseless ¡ Case: ¡

y1 y2 ! " # # $ % & & = h

11

h

12

h21 h22 ! " # # $ % & & x1 x2 ! " # # $ % & & ⇔ Y = HX X = x1 x2 ! " # # $ % & &= H−1Y

The ¡Zero-­‑Forcing ¡solu'on ¡is: ¡

slide-8
SLIDE 8

Zero-F

  • Forci

cing A Amplifies N Noise

ˆ X = H−1Y = H−1 HX + N

( )

ˆ X = X + H−1N

With ¡noise ¡Zero-­‑Forcing ¡gives ¡

8 ¡

Noise ¡amplifica'on ¡

1

x

2

x

1

y

2

y

h21 h

11

h

12

h22

y1 y2 ! " # # $ % & & = h

11

h

12

h21 h22 ! " # # $ % & & x1 x2 ! " # # $ % & & + n1 n2 ! " # # $ % & & ⇔ Y = HX + N

slide-9
SLIDE 9

2 possible

min arg ˆ Hx y x

x

− =

Maximum-L

  • Likelihood De

Detect ction a and S Sphere De Deco coding

  • Minimizes detection errors
  • Finding the ML solution by exhaustive search is impractical

9 ¡

2 1 1 11 1

' ' ⎥ ⎥ ⎥ ⎦ ⎤ ⎢ ⎢ ⎢ ⎣ ⎡ ⎥ ⎥ ⎥ ⎦ ⎤ ⎢ ⎢ ⎢ ⎣ ⎡ − ⎥ ⎥ ⎥ ⎦ ⎤ ⎢ ⎢ ⎢ ⎣ ⎡

N NN N N

x x R R R y y

  • Sphere Decoder uses QR decomposition to transform the problem into

ˆ x = arg min

possible x y'− Rx 2

d xN

( )

d xN, xN−1

( )

d xN,..., x1

( )

slide-10
SLIDE 10

b c d a a b c d a b c d a b c d a b c d

:

N

x :

1 − N

x

Maximum-L

  • Likelihood De

Detect ction a and S Sphere De Deco coding

{ } (

) ( ) { }

… + + =

− ∈ 1 ,.... ,

, min ˆ

N N N b a x

x x d x d

i

x

Therefore, ¡the ¡ML ¡problem ¡transforms ¡to: ¡

( )

a d

( )

d a d ,

( ) ( )

d a d a d , +

Node’s ¡Par'al ¡Distance ¡(PD) ¡

a b c d

xi ∈

if ¡

Client ¡N-­‑1: ¡ Client ¡N: ¡

Sphere ¡Decoding ¡Problem: ¡Find ¡the ¡ leaf ¡with ¡the ¡minimum ¡PD ¡

10 ¡

slide-11
SLIDE 11

b c d a a b c d a b c d a b c d a b c d a b c d

Maximum-L

  • Likelihood De

Detect ction a and S Sphere De Deco coding

2

) ( r b PD >

11 ¡

  • To avoid exhaustive search, original SDs search just a subset
  • f tree nodes (with PD<r2).
  • Such approaches cannot guarantee the ML solution.

How ¡can ¡we ¡minimize ¡the ¡number ¡of ¡visited ¡ nodes ¡without ¡compromising ¡op)mality? ¡

slide-12
SLIDE 12

Geosphere’s ( (and ET ETH-S

  • SD’

D’s(1

(1)) t

tree t traversal a and p pruning

Example: ¡3x3 ¡system ¡with ¡four ¡element ¡constella)on ¡(= ¡64 ¡tree ¡nodes) ¡

12 ¡

b c d a a b c d a b c d a b c d a b c d a b c d

(1) ¡Burg, ¡Andreas, ¡et ¡al. ¡"VLSI ¡implementa)on ¡of ¡MIMO ¡detec)on ¡

using ¡the ¡sphere ¡decoding ¡algorithm." ¡IEEE ¡Journal ¡of ¡Solid-­‑State ¡ Circuits, ¡40.7 ¡(2005): ¡1566-­‑1577. ¡

slide-13
SLIDE 13

b c d a a b c d a b c d a b c d a b c d a b c d

Geosphere’s ( (and ET ETH-S

  • SD’

D’s) t tree t traversal a and p pruning

13 ¡

+∞ =

2

r

Sort ¡by ¡PD ¡

slide-14
SLIDE 14

b c d a a b c d a b c d a b c d a b c d a b c d

Geosphere’s ( (and ET ETH-S

  • SD’

D’s) t tree t traversal a and p pruning

+∞ =

2

r

14 ¡

Sort ¡by ¡PD ¡

¡

slide-15
SLIDE 15

b c d a a b c d a b c d a b c d a b c d a b c d

Geosphere’s ( (and ET ETH-S

  • SD’

D’s) t tree t traversal a and p pruning

+∞ =

2

r

15 ¡

Sort ¡by ¡PD ¡

¡

slide-16
SLIDE 16

b c d a a b c d a b c d a b c d a b c d a b c d

+∞ =

2

r ) , , (

2

c c b PD r =

16 ¡

Sort ¡by ¡PD ¡

¡

Geosphere’s ( (and ET ETH-S

  • SD’

D’s) t tree t traversal a and p pruning

Radius ¡update ¡

slide-17
SLIDE 17

b c d a a b c d a b c d a b c d a b c d a b c d

) , , (

2

c c b PD r =

2

) , ( r d b PD >

17 ¡

Sort ¡by ¡PD ¡

Geosphere’s ( (and ET ETH-S

  • SD’

D’s) t tree t traversal a and p pruning

slide-18
SLIDE 18

b c d a a b c d a b c d a b c d a b c d a b c d

) , , (

2

c c b PD r =

2

) ( r c PD >

18 ¡

Sort ¡by ¡PD ¡

Geosphere’s ( (and ET ETH-S

  • SD’

D’s) t tree t traversal a and p pruning

slide-19
SLIDE 19

Geosphere’s ( (and ET ETH-S

  • SD’

D’s) t tree t traversal a and p pruning

) , , (

2

c c b PD r =

We ¡can ¡find ¡the ¡ML ¡solu)on ¡by ¡visi)ng ¡only ¡5 ¡nodes ¡ ¡

19 ¡

How ¡can ¡we ¡minimize ¡sor)ng ¡complexity? ¡

b c d a a b c d a b c d a b c d a b c d a b c d

slide-20
SLIDE 20

Traditional S Sorting a and P PD ca D calcu culations

Single ¡Dimensional ¡Constella'ons ¡ Dense ¡two-­‑dimensional ¡symmetric ¡constella'ons ¡

20 ¡

1 ¡ 2 ¡ 3 ¡ 4 ¡ ¡1 ¡ ¡2 ¡ ¡3 ¡

Visi)ng ¡3 ¡nodes ¡requires ¡3 ¡PD ¡calcula)ons ¡ ¡ ¡ ¡ Visi)ng ¡3 ¡nodes ¡requires ¡(constella'on ¡size)1/2+2 ¡PD ¡calcula)ons ¡ ¡ Half ¡distance ¡ between ¡ symbols ¡ Selected ¡ ¡ Symbol ¡ Received ¡ Signal ¡ Transmiged ¡ Symbol ¡

𝑏 ¡ 𝑐 ¡ 𝑑 ¡ 𝑒 ¡ 𝑓 ¡ 𝑔 ¡ 𝑕 ¡

slide-21
SLIDE 21

Geosphere’s 2 2D z D zig-za zag

b

c e g

d f

21 ¡

1 ¡ 2 ¡ 3 ¡

Visi)ng ¡3 ¡nodes ¡requires ¡4 ¡PD ¡calcula)ons ¡ ¡ Half ¡distance ¡ between ¡ symbols ¡ Selected ¡ ¡ Symbol ¡ Received ¡ Signal ¡ Transmiged ¡ Symbol ¡

slide-22
SLIDE 22

Geosphere’s Ea Early P y Pruning

22 ¡

Dmin can ¡be ¡pre-­‑calculated ¡for ¡all ¡constella)on ¡symbols ¡ (func)on ¡of ¡QAM ¡geometry) ¡

a

D(a)

¡

Actual ¡distance ¡

¡

Dmin(a) ≤ D(a)

¡

Lower ¡limit ¡of ¡the ¡ distance ¡

¡

We ¡can ¡avoid ¡PD ¡calcula)ons ¡by ¡first ¡ checking ¡Dmin ¡meets ¡ ¡the ¡pruning ¡criterion. ¡

Half ¡distance ¡ between ¡ symbols ¡ Received ¡ Signal ¡ Transmiged ¡ Symbol ¡

slide-23
SLIDE 23

Ev Evaluation

23 ¡

q Both by using: Ø WARP-based testbed in indoor (office) environment (5GHz band, 20MHz bandwidth, 64-OFDM) Ø Simulations (using Rayleigh and empirically measured channels) q We compare Geosphere: Ø With Zero-forcing for throughput Ø With ETH-SD for complexity

Up Up
  • client ¡

AP ¡

slide-24
SLIDE 24

Geosphere’s T Throughput G Gains f for 4 4 A AP A Antennas

24 ¡

¡x2 ¡Gain ¡

¡x2 ¡Gain ¡ 2 ¡clients ¡ 3 ¡clients ¡ 4 ¡clients ¡ Average ¡SNR ¡per ¡stream ¡(dB) ¡ Zero-­‑forcing ¡ Geosphere ¡ Net ¡throughput ¡(Mbps) ¡ ZF ¡is ¡less ¡subop)mal ¡when ¡ we ¡sacrifice ¡throughput ¡

slide-25
SLIDE 25

Computational C Complexity G y Gains f for 4 4 A AP A Antennas

25 ¡

Rayleigh ¡channels ¡ Complexity ¡ ¡ (in ¡PD ¡calcula)ons) ¡ ¡ ¡ ¡ ¡16-­‑QAM ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡64-­‑QAM ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡ ¡256-­‑QAM ¡ ¡ ¡ ¡ ¡ ¡ ¡

¡ ¡ ¡four ¡clients ¡and ¡four ¡AP ¡antennas ¡ ¡

¡ packet-­‑error-­‑rate ¡of ¡10% ¡ Empirically ¡measured ¡channels ¡ ETH-­‑SD ¡ Geosphere ¡

For ¡two ¡clients ¡the ¡complexity ¡is ¡ ¡ ~3 ¡PD ¡calcula)ons ¡across ¡all ¡QAM ¡constella)ons! ¡

slide-26
SLIDE 26

Related W Work

q The sphere decoding literature is very rich1. q However, already proposed approaches q Cannot efficiently support for very dense symbols constellations &, q Guarantee optimal performance &, q Efficiently adjust their complexity according to the MIMO channel utilization.

26 ¡

1“SD Sequence determination” IEEE SARNOFF ’09], [“K-Best SD” IEEE JSAC

’06/10, IEEE ISCAS ’08, IEEE TVLSI ’09],[“Fixed Complexity SD”, IEEE TCOM ’08], “Probabilistic Pruning” IEEE TSP ’08]…

slide-27
SLIDE 27

Concl clusions

q Low complexity detection become highly suboptimal when increasing the number of antennas. q ML detection allows scaling capacity in MIMO networks but it very complex. q Geosphere enables ML detection pragmatic systems with dense constellations q Future research will be focused in extending Geosphere to Ø Shannon capacity achieving soft-receiver processing Ø Large MIMO systems

27 ¡