joint in-band scheduling and interference mitigation in 5g hetnets
TRANSCRIPT
May 20, 2016 1 / 17
Joint In-Band Backhauling and InterferenceMitigation in 5G Heterogeneous Networks
Trung Kien Vu, Mehdi Bennis, Sumudu Samarakoon,Me’rouane Debbah†, and Matti Latva-aho
Centre for Wireless Communications, University of Oulu, Oulu,Finland, and †Mathematical and Algorithmic Sciences Lab, Huawei
France R&D, Paris, France.. Email: [email protected].
May 20, 2016
May 20, 2016 2 / 17
Outline
Introduction
System Model
Problem Formulation
Lyapunov Framework
Numerical Results
Conclusions
May 20, 2016 3 / 17
IntroductionI To meet the massive mobile data demand1:
I Advanced spectral-efficiency technique (Massive MIMO)I Dense deployment of small cellsI High frequency bands
Figure: Cisco Forecasts 30.6 Exabytes per Month of Mobile Data Traffic by 2020
12020: Beyond 4g radio evolution for the gigabit experience,” White Paper, Noikia SiementsNetworks, 2011.
May 20, 2016 3 / 17
IntroductionI To meet the massive mobile data demand1:
I Advanced spectral-efficiency technique (Massive MIMO)I Dense deployment of small cellsI High frequency bands
Figure: Cisco Forecasts 30.6 Exabytes per Month of Mobile Data Traffic by 2020
12020: Beyond 4g radio evolution for the gigabit experience,” White Paper, Noikia SiementsNetworks, 2011.
May 20, 2016 3 / 17
IntroductionI To meet the massive mobile data demand1:
I Advanced spectral-efficiency technique (Massive MIMO)I Dense deployment of small cellsI High frequency bands
Figure: Cisco Forecasts 30.6 Exabytes per Month of Mobile Data Traffic by 2020
12020: Beyond 4g radio evolution for the gigabit experience,” White Paper, Noikia SiementsNetworks, 2011.
May 20, 2016 3 / 17
IntroductionI To meet the massive mobile data demand1:
I Advanced spectral-efficiency technique (Massive MIMO)I Dense deployment of small cellsI High frequency bands
Figure: Cisco Forecasts 30.6 Exabytes per Month of Mobile Data Traffic by 2020
12020: Beyond 4g radio evolution for the gigabit experience,” White Paper, Noikia SiementsNetworks, 2011.
May 20, 2016 4 / 17
Solutions
I The interplay between Massive MIMO and a densedeployment of self-backhaul small cells(SCs)
I The problem of joint scheduling, interference mitigation,and in-band wireless backhauling
I A network utility maximization problem subject todynamically varying wireless backhaul and network stability
May 20, 2016 4 / 17
Solutions
I The interplay between Massive MIMO and a densedeployment of self-backhaul small cells(SCs)
I The problem of joint scheduling, interference mitigation,and in-band wireless backhauling
I A network utility maximization problem subject todynamically varying wireless backhaul and network stability
May 20, 2016 4 / 17
Solutions
I The interplay between Massive MIMO and a densedeployment of self-backhaul small cells(SCs)
I The problem of joint scheduling, interference mitigation,and in-band wireless backhauling
I A network utility maximization problem subject todynamically varying wireless backhaul and network stability
May 20, 2016 5 / 17
ToolsI Random Matrix Theory2
I Large number of antennas N , number of UEs K
I Stochastics optimization3
I Large number of variables and constraints, and dynamicload.
I Success approximation convex method4
I Convert the non-convex program by its solvable convexupper bound.
2S. Wagner, R. Couillet, M. Debbah, and D. Slock, “Large system analysis of linearprecoding in correlated MISO broadcast channels under limited feedback,” IEEE Transactionson Information Theory, vol. 58, no. 7, pp. 4509–4537, 2012.
3 M. J. Neely, “Stochastic network optimization with application to communication andqueueing systems,” Synthesis Lectures on Communication Networks, vol. 3, no. 1, pp. 1–211,2010.
4A. Beck, A. Ben-Tal, and L. Tetruashvili, “A sequential parametric convex approximationmethod with applications to nonconvex truss topology design problems,” Journal of GlobalOptimization, vol. 47, no. 1, pp. 29–51, 2010.
May 20, 2016 5 / 17
ToolsI Random Matrix Theory2
I Large number of antennas N , number of UEs K
I Stochastics optimization3
I Large number of variables and constraints, and dynamicload.
I Success approximation convex method4
I Convert the non-convex program by its solvable convexupper bound.
2S. Wagner, R. Couillet, M. Debbah, and D. Slock, “Large system analysis of linearprecoding in correlated MISO broadcast channels under limited feedback,” IEEE Transactionson Information Theory, vol. 58, no. 7, pp. 4509–4537, 2012.
3 M. J. Neely, “Stochastic network optimization with application to communication andqueueing systems,” Synthesis Lectures on Communication Networks, vol. 3, no. 1, pp. 1–211,2010.
4A. Beck, A. Ben-Tal, and L. Tetruashvili, “A sequential parametric convex approximationmethod with applications to nonconvex truss topology design problems,” Journal of GlobalOptimization, vol. 47, no. 1, pp. 29–51, 2010.
May 20, 2016 5 / 17
ToolsI Random Matrix Theory2
I Large number of antennas N , number of UEs K
I Stochastics optimization3
I Large number of variables and constraints, and dynamicload.
I Success approximation convex method4
I Convert the non-convex program by its solvable convexupper bound.
2S. Wagner, R. Couillet, M. Debbah, and D. Slock, “Large system analysis of linearprecoding in correlated MISO broadcast channels under limited feedback,” IEEE Transactionson Information Theory, vol. 58, no. 7, pp. 4509–4537, 2012.
3 M. J. Neely, “Stochastic network optimization with application to communication andqueueing systems,” Synthesis Lectures on Communication Networks, vol. 3, no. 1, pp. 1–211,2010.
4A. Beck, A. Ben-Tal, and L. Tetruashvili, “A sequential parametric convex approximationmethod with applications to nonconvex truss topology design problems,” Journal of GlobalOptimization, vol. 47, no. 1, pp. 29–51, 2010.
May 20, 2016 5 / 17
ToolsI Random Matrix Theory2
I Large number of antennas N , number of UEs K
I Stochastics optimization3
I Large number of variables and constraints, and dynamicload.
I Success approximation convex method4
I Convert the non-convex program by its solvable convexupper bound.
2S. Wagner, R. Couillet, M. Debbah, and D. Slock, “Large system analysis of linearprecoding in correlated MISO broadcast channels under limited feedback,” IEEE Transactionson Information Theory, vol. 58, no. 7, pp. 4509–4537, 2012.
3 M. J. Neely, “Stochastic network optimization with application to communication andqueueing systems,” Synthesis Lectures on Communication Networks, vol. 3, no. 1, pp. 1–211,2010.
4A. Beck, A. Ben-Tal, and L. Tetruashvili, “A sequential parametric convex approximationmethod with applications to nonconvex truss topology design problems,” Journal of GlobalOptimization, vol. 47, no. 1, pp. 29–51, 2010.
May 20, 2016 6 / 17
System Model
MBSFD-SC
MUE
Massive MIMO Antennas
beamforming
D: Queue buffer
Dataflow
ISD: 250, 125, 100, ..., 35 m
Q: Network Queue
Data
FD-SC
SUEwir
eless b
ackhau
l
data access
MUE served by MBS
and interfered by SCs
SUE served by SC only
and interfered by MBS
Figure: Network Scenario.
May 20, 2016 7 / 17
Network Assumptions
I N number of antennas, M number of macro users (MUEs),S number of SCs
I N ≥ (K = M + S) ≥ 1
I Dense deployment of SCsI Full-duplex capacityI Two antennas, small cell user (SUE) per each
I Co-channel time-division duplexing (TDD) protocol
I Imperfect channel state information (CSI)
May 20, 2016 8 / 17
Queue Buffer at SC
I rs(t) and rm(t) be the data rates from MBS to SC andMUE, respectively.
I rfws (t) be the data rate of from SC to SUE.
I D = (D1(t), D2(t), ...) as a data queue at SCs,
Ds(t + 1) = max[Ds(t) + rs(t)− rfws (t), 0] (1)
May 20, 2016 9 / 17
Network Utility MaximizationNetwork utility maximization problem:
max f(r̄) (2)subject to r̄ ∈ R (3)
D̄ <∞ (4)network stability (5)
I The network utility function f =∑k f(r̄k), f(r̄k) is assumed to be twice differentiable,
concave.I The rate region R
I r̄ , limt→∞
1t
t−1∑τ=0
E[r(τ)] denotes the time average expectation of the Ergodic data rate*.
I D̄ , limt→∞ 1t
∑t−1τ=0 E
[|D(τ)|
]
May 20, 2016 10 / 17
How to apply the Lyapunov framework
Convert to a queuing network
Convert all constraints by introducing new variables
Replace them by virtual queue
Write the Lyapunov function including
all network queues and virtual queue
Use the Lyapunov plus penalty technique to solve the original problem
by minimizing the Lyapunov function and the objective function
Decouple into several subproblems which are solvable and independence.
Figure: Lyapunov Framework
May 20, 2016 11 / 17
Flowchart of Proposed Algorithm
Algorithm 1: Inband Scheduling
& FD Mode Control
Beamforming Design
Power Allocation
Network Queue Update
Tim
eIn
dic
es
Alg
ori
thm
1
Qu
eu
eU
pd
ate
Pow
erA
lloca
tion
DL
Tra
nsm
issi
on
DL Transmission
Network Queue
& Channel Initialization
CSI ReportUniform Power Assign
May 20, 2016 12 / 17
Simulation Setup
I Different frequency bands: 2.4, 5, and 28 GHz.
I HomNet: where the MBS with Massive MIMO will serve all UEs.
I Number of SCs is increased by reducing the inter-site-distance.
I Number of antennas is twice as number of users, which ensuringthe closed-form expression of user rate is hold.
May 20, 2016 13 / 17
Average UE ThroughputI 28 GHz achieves 56× and 62× gain as compared to 2.4 GHz due to 50× larger, when the
ISD is 250 m and 80 m, respectively.I In ultra-dense deployment, the UE throughput at 2.4 GHz is below 10 Mbps, whereas
473 Mbps per UE is achieved by using 28 GHz for 33 m of ISD.
Number of Small Cells per km216 64 144 400 900A
chie
vab
le A
ver
age
UE
Th
rou
gh
pu
t [G
bp
s][B
lue]
0
1
2
3
4
28 GHz
HetNet-HybridHomNetArrival Mean Rate
0
20
40
Aver
age
Queu
e L
ength
[G
b][
Red
]
0
10
20
30
40
16 64 144 4000
0.2
0.410 GHz
0
2
4
16 64 100 1440
0.12.4 GHz
0
0.2
0.4
ISD=80mISD=100m
ISD=125m
ISD=250m
ISD=50m
ISD=33m
Figure: Achievable Average UE throughput and Network Queue length versus number of
Small Cells at 28 GHz, 10 GHz, and 2.4 GHz.
May 20, 2016 14 / 17
Cell-Edge UE ThroughputI 1 Gbps and 0.73 Gbps in case of HetNet-Hybrid and HomNet, respectively.
Number of Small Cells per km2
16 64 144 400 900
Ach
iev
able
Cel
l-E
dg
e U
E T
hro
ug
hp
ut
[Gb
ps]
0
1
2
3
28 GHz
HetNet-HybridHomNetArrival Mean Rate
16 64 144 4000
0.2
0.410 GHz
16 64 100 144
0.03
0.062.4 GHz
ISD=33mISD=50m
ISD=80m
ISD=125m
ISD=100m
ISD=250m
Figure: Achievable 5th% UE throughput versus number of Small Cells at 28 GHz, 10 GHz,
and 2.4 GHz
May 20, 2016 15 / 17
Utility-Queue length tradeoffI There exists an [O(1/ν),O(ν)] utilityqueue backlog tradeoff, which leads to an
utility-delay tradeoff.
ν 10 5×105 10
6 1.5×10
6 2×10
6 2.5×10
6
Aver
age
Net
work
Uti
lity
[G
bps]
[B
lue]
6
8
10
12
14
16
18HetNet-Hybrid
HomNet
Aver
age
Queu
e B
acklo
g [
Gb][
Red
]
6
8
10
12
14
16
18
Figure: Impact of control parameter ν on the Utility and Network Backlogs at 28 GHz when
K = 16, N = 64.
May 20, 2016 16 / 17
Conclusions
I Joint inband scheduling and interference mitigationoptimization.
I Lyapunov framework solution to decouple the problem.
I Performance gains of three combined techniques in 5GHetNets.
May 20, 2016 17 / 17
Thank You!