Outline Cooperative Communicationsthomaszemen.org/Cooperative/Lecture_10_handout.pdf · 2015. 2....
Transcript of Outline Cooperative Communicationsthomaszemen.org/Cooperative/Lecture_10_handout.pdf · 2015. 2....
Cooperative CommunicationsLecture 10
Thomas Zemen Paolo Castiglione
May 26, 2011
Thomas Zemen and Paolo Castiglione May 26, 2011 1 / 36
Outline
Today, Lecture 10Partner selection algorithms
Introduction to cross-layer design for cooperative systemsLink level outage performance analysisPartner selection for static cooperative multiple accessLifetime vs fairness
Thomas Zemen and Paolo Castiglione May 26, 2011 2 / 36
Cooperation: also between OSI layersmore point-to-point links are involved for one transmission:separation of OSI layers is no longer optimal
Thomas Zemen and Paolo Castiglione May 26, 2011 3 / 36
PHY+MAC: grouping & partner selection (I)
Layers are inter-dependent if a user (BER etc.) or a network (lifetimeetc.) performance metrics need to be optimized
PHY layer:cooperative scheme: DF, AF, cooperative beamforming, interferencealignment..estimation of channel parameters: SNR, channel gain, Dopplerand/or delay spread..
MAC layer:resource allocation: power, bandwidth, rate, delay..grouping and partner selection: which relays for which nodes? whichinterference alignment cluster?
Thomas Zemen and Paolo Castiglione May 26, 2011 4 / 36
PHY+MAC: grouping & partner selection (II)Recalling the cooperative region for coded cooperation over time-variantchannels (lecture 6):[collection of mobility (TBP) and channel (γ, p) settings for which codedcooperation is beneficial in terms of user BER]
0.5 1 1.5 2 2.5 310
-4
10-3
10-2
10-1
TBP
Inte
r-M
SB
lock
Err
or
Pro
ba
bili
typ
= 0 dB
= 5 dB
= 10 dB
= 15 dB
= 20 dB
Analytical boundaryEmpirical approximation
Figure source: [1, Fig. 11]
Partners must be chosen such thatinter-user outage probability p issmall (close partners) and uplinkSNR γ is critical.Velocity (Doppler-spread) indicatesif cooperation is advantageous.
Thomas Zemen and Paolo Castiglione May 26, 2011 5 / 36
PHY+MAC: grouping & partner selection (III)
Mobility asymmetry (β) affects theperformance (long-term statisticsunbalances are drawbacks forcooperative MICRO-diversity)partners chosen with similar SNR
and velocity values.
0.5 1 1.5 2 2.5 310
-4
10-3
10-2
10-1
TBP1= TBP2
Inte
r-M
SB
lock
Err
or
Pro
ba
bili
typ
= 10 dB Analytical boundary
= 3/4
= 1/2
= 1
Figure source: [1, Fig. 12]
Thomas Zemen and Paolo Castiglione May 26, 2011 6 / 36
Literature
First paper on the topic:A.Hunter, T.E.; Nosratinia, “Distributed protocols for usercooperation in multi-user wireless networks,” Proceedings IEEE
GLOBECOM ’04
Huge literature covering many different types of problems.Suggested papers:
Z. Lin, E. Erkip, A. Stefanov, “Cooperative regions and partnerchoice in coded cooperative systems,” IEEE Transactions on
Communications, 2006V. Mahinthan, Lin Cai, J.W. Mark, Shen Xuemin, “Maximizingcooperative diversity energy gain for wireless networks,” IEEE
Transactions on Wireless Communications, 2007
Thomas Zemen and Paolo Castiglione May 26, 2011 7 / 36
Partner selection for static cooperative multiple accessMeasurement network topology and system/channel modelAF outage analysis over Ricean fadingGoal and research questionEnergy consumption optimization (based on outage analysis)Optimal and suboptimal solutionsPerformance results, optimality degree and robustness
Thomas Zemen and Paolo Castiglione May 26, 2011 8 / 36
Network and Channel ModelN battery-powered static indoor nodescommunicate to an outdoor BS (j = 0).
50m
OfficeBase Station
ij
0,0, , jj LK 0,0,,
iiLK
,,,
iiLK
j j
Stanford 2008measurements map(Picture: © Google Map)
BS
45m
Network setup
Modeling from meauserements: Ricean fading |hi,j | on each link (i , j).K-factors and path loss (Ki,j , Li,j) modeled as log-normal random variates
with distance dependent means.
Thomas Zemen and Paolo Castiglione May 26, 2011 9 / 36
System Model
Simple cooperative multiple access scheme: TDMA and AF.
User j-thdata
User i-th amplified signal
ij
TDMA frame
User i-thdata
User j-th amplified signal
Beacon
slot
Node i transmits power ρi for the whole slot (also for relaying).γi,j = ρi |hi,j |2 /σ2: instantaneous SNR on the link (i , j).
Thomas Zemen and Paolo Castiglione May 26, 2011 10 / 36
Outage Analysis
Direct transmission on the link (i , j): outage probability tightlyapproximated1
Pr[γi,j < γdir
th ] ≈�γdir
th · σ2/ (ci,jρi)�di,j
.
Ricean fading: diversity di,j = 1 and coding gain ci,j
ci,j =exp (Ki,j)
Li,j (Ki,j + 1) .
1γdirth depends on the target spectral efficiency and the modulation/code.
Thomas Zemen and Paolo Castiglione May 26, 2011 11 / 36
Estimated Coding Gainsfrom I2O Measurements
0 5 10 15 20 25 30 35 40-40
-20
0
20
40
60
80
100
Path Loss [dB] Li,,j
Coding gain(Ricean fading)
Coding gain(Rayleigh fading)
Codin
gG
ain
[dB
]j,ic
Coding gain(Rayleigh fading)
0 5 10 15 20 25 30 35 40
j=0j=0
Coding gain(Ricean fading)
I2I I2O
Figure source: [3, Fig. 2]
Real coding gain valueshighlight the poorness of the Rayleigh fading assumption ci,j = L
−1i,j .
Thomas Zemen and Paolo Castiglione May 26, 2011 12 / 36
AF outage anlysisAF relaying2: The dynamic amplification factor at relay j sets theeffective SNR for source node i
γ(i,j),0 = γi,0 +
�1γi,j
+1γj,0
+1
γi,jγj,0
�−1.
Pr[γ(i,j),0 < γAF
th ] ≈�
γAF
th· σ2
cAF
(i,j),0 ·√ρiρj
�dAF(i,j),0
,
diversity gain dAF
(i,j),0 = 2; effective coding gain cAF
(i,j),0:
cAF
(i,j),0 =
�1
ci,0·�
1ci,j
+1
cj,0
��− 12
.
2γAFth depends on the target spectral efficiency and the modulation/code.
Thomas Zemen and Paolo Castiglione May 26, 2011 13 / 36
Energy Consumption
Set by the target outage probability, the spectral efficiency andthe modulation/coding format.Over-head and synchronization aspects translated into additionalenergy terms.
Thomas Zemen and Paolo Castiglione May 26, 2011 14 / 36
Goal and Research Question
GOAL: to maximize the network lifetime,i.e. to minimize the maximum energy consumption,by allowing for a variable number of cooperating pairs of nodes.
QUESTION: which are the optimaland the low-complexity suboptimal pairing strategies?
Thomas Zemen and Paolo Castiglione May 26, 2011 15 / 36
Energy Consumption Optimization
P is the set of candidate pairing sets ξ (disjoint pairs)[EXAMPLE: for N = 15 the number of candidate pairing sets |P| � 107]
!"#$%$&'#(!!) *+#,-$#$.$&/'#0!.+12$3#4'#5#+*./($#.!3$
!"#$%$&'#$3/$ *+#,-$#256*212#$.$&/'#0!.+12$3#4'#,7!#0!!)$&5,*%$#.!3$+#
8),*2*95,*!.#!.#!"!#$%&'()%)*+,-.//0,1"!!*1)*2,34567849,:('&;
:;<=>?
@ A
B
@ A
B
@ A
B
P: ={!1, !2, !3, !4}!1 : ={(1,2),3} !2 : ={(2,3),1} !3 : ={(1,3),2}
@ A
B
!4 : ={1,2,3}
Thomas Zemen and Paolo Castiglione May 26, 2011 16 / 36
Optimal Solution
Min-max problem: find the pairing set ξ for which the maximum energyconsumption Emax(ξ) is minimum.
ξ = arg minξ∈P
Emax(ξ).
Optimal solution to the combinatorial optimizations problem(iterated Gabow algorithm):
Computational complexity O(N5): too slow for large networks!BS (genie) would need to know all the energy consumptions for allpaired and single nodes: impracticable!
Thomas Zemen and Paolo Castiglione May 26, 2011 17 / 36
Suboptimal Solution
Existing Worst-Link-First algorithms: based on second order statistics(rx powers or signal strenghts, link path loss..)We call them “Path-Loss based”: WLF-PL
Novelty in our algorithm: the metric is the uplink coding gain3.
3Furthermore the novel algorithm considers odd N networks.
Thomas Zemen and Paolo Castiglione May 26, 2011 18 / 36
WLF-CG algorithm
Worst-Link-First Coding-Gain (ci,j) based algorithm, O(N2):
1) Node ( i ) selects candidate partners ( j ) : ci,j / ci,0 » 1
BS
i
j
j
j
BS worst-uplink i
best-uplink j
2) Iterate until list is empty:if best-uplink node is candidate partner of the worst-uplink node,pair and remove from list worst- and best-uplink nodes;otherwise leave single and remove from list the worst-uplink node.
3) The BS communicates the pairing solution
BS
i
j
Thomas Zemen and Paolo Castiglione May 26, 2011 19 / 36
Performance Results
Network topology similar to the one of the Stanford measurementcampaign.Results averaged over 5 × 104 scenarios.
Target outage probability 10−3;spectral efficiency = 1;Gaussian code-book.
For each scenario, (Ki,j , Li,j) generated according to the I2Ostochastic model.
Thomas Zemen and Paolo Castiglione May 26, 2011 20 / 36
Energy Gains over No-Cooperation
3 5 79 1113 15 25 35 45 55
N (odd)10 20 30 40 504 6 8 1214
0
5
10
15
20
25
30
35
N (even)
Energ
ygain
[dB
]
WLF-CG
random pairing
optimal solution
WLF-PL
Figure source: [3, Fig. 3]
Huge energy gains over direct transmission, random pairing andexisting algorithms!With high probability ci,j � ci,0: less sensitivity to inter-nodechannels qualities → WLF-CG almost optimal.
Thomas Zemen and Paolo Castiglione May 26, 2011 21 / 36
K-factor Estimation Errors
-10 -5 0 5 1020
22
24
26
28
30
-10 -5 0 5 1010
12.5
15
17.5
20
22.5
25
Ene
rgy
ga
in[d
B]
WLF-CG
WLF-PL
optimal solution (
N=4 N=12
K2 [dB]Mean square error of the K-factor estimation !
K2
! =0)
Figure source: [3, Fig. 4]
For small networks, WLF-CG robust to the K-factor estimationerrors.
Thomas Zemen and Paolo Castiglione May 26, 2011 22 / 36
Summary
Realistic coding gains, calculated from measurements→ the Rayleighfading assumption does not hold in practical I2O scenarios.Partner selection algorithm (WLF-CG) exploits the local knowledgeof the fading statistics.The WLF-CG algorithm gains up to 11dB compared to existinggreedy algorithms in static multiple access.What happens though if nodes do not fully tolerate to share
their resources to the partners? See next section
Thomas Zemen and Paolo Castiglione May 26, 2011 23 / 36
Lifetime vs fairnessUser energy gain and user-fairness
Goal and research question (revisited)Energy consumption optimization with user-fairness constraintOptimal and suboptimal solutions (revisited)Performance results and optimality degree
Thomas Zemen and Paolo Castiglione May 26, 2011 24 / 36
User energy gain
Energy consumption Ei,0 (direct transmission) and EAF
(i,j),0 (AF) setby the target outage probability and the spectral efficiency (andalso the modulation/coding format)The energy gain or loss gi,j achieved by user i cooperating withpartner j
gi,j =Ei,0
EAF
(i,j),0
Thomas Zemen and Paolo Castiglione May 26, 2011 25 / 36
User-fairness (I)
The energy gain or loss gi,j is constrained by the user-fairness g
gi,j = geff(R)× gµD × g
MD
i,j =Ei,0
EAF
(i,j),0≥ g
geff(R) =�2R − 1
�/�22R − 1
�< 1 is the rate penalty factor;
gµD = (2p)1/dAF(i,j),0 /p =
�2/p ≥ 1 is the micro-diversity gain
(small-scale);gMD
i,j = cAF
(i,j),0/ci,0 is the macro-diversity gain (large-scale).For ci,j � ci,0: gMD
i,j � gMDi,j =
�cj,0/ci,0
Thomas Zemen and Paolo Castiglione May 26, 2011 26 / 36
User-fairness (II)
User-fairness g has here a twofold interpretation:a maximum amount of energy loss, compared to the non-cooperativeoption, can be tolerated by the user (g < 0dB);a minimum energy gain for the user is introduced as incentive tocooperate (g ≥ 0dB).
Thomas Zemen and Paolo Castiglione May 26, 2011 27 / 36
No-coop 2x1 MISO
AF cooperation 2 x 1 MISO system
compared to
User-fairness no longer applies.
i.i.d. MISO links: only the micro-diversity gain is available (Alamouti)
gMISO =gµD
2 ,
where factor 1/2 is due to the power splitting between the antennas
Thomas Zemen and Paolo Castiglione May 26, 2011 28 / 36
Goal and research question
GOAL: to maximize the network lifetime (same as before),i.e. to minimize the maximum energy consumption (among theusers),by allowing for a variable number of cooperating pairs of users.
QUESTION: which are the optimaland the low-complexity suboptimal pairing strategies,that account for the user-fairness constraint?
Thomas Zemen and Paolo Castiglione May 26, 2011 29 / 36
Optimal solution
!"#$%$&'#(!!) *+#,-$#$.$&/'#0!.+12$3#4'#5#+*./($#.!3$
!"#$%$&'#$3/$ *+#,-$#256*212#$.$&/'#0!.+12$3#4'#,7!#0!!)$&5,*%$#.!3$+#
8),*2*95,*!.#!.#!"!#$%&'()%)*+,-.//0,1"!!*1)*2,34567849,:('&;
:;<=>?
@ A
B
@ A
B
@ A
B
P: ={!1, !2, !3, !4}!1 : ={(1,2),3} !2 : ={(2,3),1} !3 : ={(1,3),2}
@ A
B
!4 : ={1,2,3}
Min-max problem: find the pairing set ξ for which the maximum energyconsumption Emax(ξ) is minimum
ξ = arg minξ∈P
Emax(ξ)
s.t. gi,j ≥ g , ∀(i , j) ∈ ξ
Edges not meeting the user-fairness constraint are discarded:only a subset of P fulfills the constraint
Thomas Zemen and Paolo Castiglione May 26, 2011 30 / 36
Modified WLF-CG algorithm
Worst-Link-First Coding-Gain (ci,j) based algorithm, O(N2):
1) Node ( i ) selects candidate partners ( j ) : ci,j / ci,0 >> 1. [The inter-user channel ci,j will not impact the performance of ( i ) :
BS
i
j
j
j
BS worst-uplink i
best-uplink j
2) Iterate until list is empty:if best-uplink candidate partner ( j ) of the worst-uplink node ( i ) fulfilling:
pair and remove from list worst- and best-uplink nodes;otherwise leave single and remove from list the worst-uplink node.
3) The BS communicates the pairing solutionBS
i
j
( ) ( ))(for partner candidate also )(
fairness-userj icc
ggggRg
jij
ijji
0,,
MD,
MD,
Deff
&
~,~min
>>
!"" µ
]MD,
MD,
~ jiji gg #$%&
Thomas Zemen and Paolo Castiglione May 26, 2011 31 / 36
Performance results
Network topology similar to the one of the Stanford measurementcampaign.Results averaged over 105 scenarios.
Target outage probability 10−3;spectral efficiency = 1;Gaussian code-book;ci,j > ci,0 + 20dB in the WLF-CG.
For each scenario, (Ki,j , Li,j) generated according to the I2Ostochastic model.
Thomas Zemen and Paolo Castiglione May 26, 2011 32 / 36
Lifetime gains vs user-fairness
-80 -70 -60 -50 -40 -30 -20 -10 0 100
5
10
15
20
25
30
Network Lifetime Gain [dB]
optimal solutionWLF-CG
N = 10
N = 6
N = 50
g MISO |dB = 13.5 dB
User Fairness [dB]g dB 11.7 | g dBmax =
Figure source: [4]
For less stringent user-fairness requirements (g < 0dB),WLF-CG better than Alamouti! Large benefits of macro-diversity
WLF-CG suboptimal, e.g. up to 6dB loss for N = 10:due to the conservative rule ci,j > ci,0 + 20dB (necessary for theuser-fairness)
Thomas Zemen and Paolo Castiglione May 26, 2011 33 / 36
Is cooperation attractive?
-80 -70 -60 -50 -40 -30 -20 -10 0 100
10
20
30
40
50
60
70
80
90
User Fairness [dB]
% cooperative users
N = 50
N = 10
N = 6
g dB 11.7 | g dBmax =
optimal solutionWLF-CG
Figure source: [4]
More severe the user-fairness constraint (g ≥ 0dB): less likely tocooperate (especially for small N)Less stringent constraints (g < 0dB): cooperative option becomesmore attractive
Thomas Zemen and Paolo Castiglione May 26, 2011 34 / 36
Summary
User-fairness: twofold definition in AF cooperation1) max energy loss tolerated 2) min energy gain required by the userPairwise grouping (low-complexity algorithm):I2O network lifetime increase by 10 compared to Alamouti,by 200 compared to single-antenna transmission[50 indoor users, tolerance up to 10dB energy loss]Macro-diversity has a huge impact on performances
Thomas Zemen and Paolo Castiglione May 26, 2011 35 / 36
References I[1] P. Castiglione, M. Nicoli, S. Savazzi and T. Zemen,“Cooperativeregions for coded cooperation over time-varying fading channels,”Proc.
of Int. ITG/IEEE Workshop on Smart Antennas 2009
[2] Z. Wang, Member, G. B. Giannakis, “A simple and generalparameterization quantifying performance in fading channels,” IEEE
Trans. on Comm., 2003[3] P. Castiglione, S. Savazzi, M. Nicoli, T. Zemen, ”Impact of fadingstatistics on partner selection in indoor-to-outdoor cooperative networks”,Proc. of IEEE ICC 2010
[4] P. Castiglione, S. Savazzi, M. Nicoli, G. Matz, “Partner Selection inCooperative Networks: Efficiency vs Fairness in Ricean Fading Channels”,Proc. of IEEE ISCCSP 2010
[5] I. Stanojev, O. Simeone, U. Spagnolini, Y. Bar-Ness, R. L. Pickholtz,“Cooperative ARQ via Auction-Based Spectrum Leasing,” IEEE Trans.
on Comm., 2010
Thomas Zemen and Paolo Castiglione May 26, 2011 36 / 36