optimization and self-optimization for lte networks · pdf fileoptimization and...

PhD Thesis Defense on

Optimization and Self-optimization for LTE Networks

Author: Abdoulaye TALL

Supervisors: Zwi ALTMANEitan ALTMANRichard COMBES

December 17th, 2015

Introduction

Context

More and morecomplex networks

Highly competitivemarket withsteadily decreasingprices/revenues

Example: Orange France

Sites: 19000+ 2G, 19000+ 3G, 7000+ 4G

Frequency bands: 700 MHz, 800 MHz, 900 MHz,1.8 GHz, 2.1 GHz and 2.6 GHz

Indoor femto-cells

Challenge:Optimize resource utilization and Control the OPEX.

Enabler:Automation with Self-Organizing Network (SON) for

self-configuration, self-optimization and self-healing.

Abdoulaye Tall Optimization and Self-optimization for LTE December 17th, 2015 2 / 56

Introduction

Thesis objectives

Design self-optimizing algorithms for use cases of interest: hetnets and activeantenna systems.

RequirementsSimple

Non-Heuristic

Fast

Robust


Introduction

SON design methodology

Network/KPI modeling

Queuing theory

Probability theory

Mobile technology

Problem Formulation

Algorithm Design

Game theory

Convex Optimization

Stochastic Approximation

Proof-of-concept Simulator/Demonstrator


Overview

1 Introduction

2 Background

3 Heterogeneous NetworksLoad balancingInterference coordinationNumerical results

4 Active Antenna SystemsVertical SectorizationVirtual SectorizationMultilevel Beamforming

5 SON CoordinationConceptProblem formulation and SolutionUse case

6 Conclusion & PerspectivesConclusionPerspectives


Overview

1 Introduction

2 Background

3 Heterogeneous Networks

4 Active Antenna Systems

5 SON Coordination

6 Conclusion & Perspectives


Background

Flow level network model

Base station modeled as M/G/1 PS queue

User’s data rate R(r) = min(Rmax, η log2(1 + SINR(r)))

Key performance indicators:

Load: ρ = traffic demandservice rate

Mean user throughput: µ = R(1− ρ)

Others by simulation.


Overview

1 Introduction

2 Background



5 SON Coordination



Heterogeneous Networks

Introduction

Nodes with different transmit powers

Nodes with different propagation conditions

Nodes with low processing capabilities

Dense network with more interference problems


Overview

1 Introduction

2 Background



5 SON Coordination



Heterogeneous Networks Load balancing

Literature review

H. Kim et al., IEEE INFOCOM 2010:

s∗u = argmaxsRu,s(1− ρs)α (1)

with Ru,s (data rate from BS s to user u), ρ (load) and α ∈ R+.

=⇒ Minimizes∑

s(1−ρs )1−α

α−1 .

R. Combes et al., IEEE INFOCOM 2012:

Ps(k + 1) = Ps(k)(1 + εk(ρ1(k)− ρs(k))) (2)

with P (pilot power).

=⇒ Minimizes maxs ρs


Heterogeneous Networks Load balancing

End-to-end load model

Classical BS load definition:

ρ = min

(1,

∫A

λ(r)E(σ)

R(r)dr

), (3)

End-to-end load definition

ρg = min

(1,

∫A

λ(r)E(σ)

min(CBH ,R(r))dr

). (4)

where

CBH : backhaul capacity,

λ(r): arrival rate at location r ,

E(σ): mean file size,

R(r): peak rate at location r .


Overview

1 Introduction

2 Background



5 SON Coordination



Heterogeneous Networks Interference coordination

Almost Blank Subframe (ABS) mechanism

Time-domain interference mitigation in heterogeneous networks.

Used in conjunction with Cell Range Extension (CRE).

Illustration of Almost Blank Sub-Frames in a HetNet



ABS-based interference coordination (eICIC)

Opimization Problem: trade-off between small cells’ users SINR and macro cells’capacity.

θ: ABS ratio applied by the considered macro cells.

Ru = (1− θ)Ru,m: Average data rate of macro user.

Ru = (1− θ)Rno ABSu,p + θRABS

u,p : Average data rate of small cell user.

Performance criteria: α-fair utility of users’ throughput

Uα(θ) =

∑

all users

logRu if α = 1

∑all users

R1−αu

1−α otherwise

(5)



Lower-bound PF utility for low complexity algorithm

Exact Proportional Fair (α = 1) utility:

UPF exact(θ) =M∑

m=1

∑u∈m

log((1− θ)Ru,m) +∑u∈p

log((1− θ)Rno ABSu,p + θRABS

u,p ) (6)

Lower bound utility:

UPF approx(θ) =M∑

m=1

∑u∈m

log((1− θ)Ru,m) +∑u∈p

1

2log(2(1− θ)Rno ABS

u,p )

+∑u∈p

1

2log(2θRABS

u,p )

(7)

Optimal ABS ratio in closed-form

θ =Np

2(Np +∑M

m=1 Nm)(8)



Frequency splitting (orthogonal deployment)

Orthogonal frequency used between macro cells and small cells

Completely eliminates interference between macro cells and small cells at theprice of reduced frequency reuse.


Overview

1 Introduction

2 Background



5 SON Coordination



Heterogeneous Networks Numerical results

Numerical results - Scenario

Network parameters

Number of macro BSs 3

Number of small BSs 12

Number of interfering macros 6 × 3 sectors

Macro Cell layout hexagonal trisector

Small Cell layout omni

Intersite distance 500 m

Bandwidth 10MHz

Scheduling Round-Robin

Channel characteristics

Thermal noise -174 dBm/Hz

Macro Path loss (d in km) 128.1 + 37.6 log10(d) dB

Small cell Path loss (d in km) 140.7 + 36.7 log10(d) dB

Algorithms Parameters

SON update frequency every event

Step size of ABSrO 10−4

Step size of FSO 5.10−5

Network layout scenario



Numerical results - Scenario cont’d.

Traffic spatial distribution uniform

λ 14 users/s/km2

λh 6 users/s/km2

Service type FTP

Average file size 6 Mbits

NoSON: baseline.LBonly: load balancing only.AFUAonly: alpha-fair user association (afua) only.LB-CCD-approx: load balancing with approximate ABS-based eICIC.AFUA-CCD-approx: afua with approximate ABS-based eICIC.LB-OD-approx: load balancing with approximate frequency-based eICIC.AFUA-OD-approx: afua with approximate frequency-based eICIC.



Numerical results - Performance

0

2

4

6

8

10

−46%

65%

424%

98%

411%

108%

Geo

met

ric M

ean

Thr

ough

put (

Mbp

s)

NoSO

N

LBon

lyAFU

Aonly

LB_C

CD_app

rox

AFUA_C

CD_app

rox

LB_O

D_app

rox

AFUA_O

D_app

rox

α-fair utility (α = 1)

0

20

40

60

80

100

Res

ourc

e U

tiliz

atio

n (%

)

NoSO

N

LBon

lyAFU

Aonly

LB_C

CD_app

rox

AFUA_C

CD_app

rox

LB_O

D_app

rox

AFUA_O

D_app

rox

At Small CellsAt macro Cells

Loads

0

5

10

15

60%31%

136%

65%

152%

27%

Mea

n U

ser

Thr

ough

put (

Mbp

s)

NoSO

N

LBon

lyAFU

Aonly

LB_C

CD_app

rox

AFUA_C

CD_app

rox

LB_O

D_app

rox

AFUA_O

D_app

rox

Mean User Throughput

0

1

2

3

−45%

60%

368%

118%

307%

74%

Cel

l Edg

e U

ser

Thr

ough

put (

Mbp

s)

NoSO

N

LBon

lyAFU

Aonly

LB_C

CD_app

rox

AFUA_C

CD_app

rox

LB_O

D_app

rox

AFUA_O

D_app

rox

Cell Edge throughput



Exact vs Approximate algorithms

0 10 20 30 40 500

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

User Throughput (Mbps)

CD

F (

%)

Exact utility ABSrO with LBLower Bound utility ABSrO with LBExact utility ABSrO with AFUALower Bound utility ABSrO with AFUA


Overview

1 Introduction

2 Background



5 SON Coordination



Active Antenna Systems

Introduction

Base Station architecture evolution


Overview

1 Introduction

2 Background



5 SON Coordination



Active Antenna Systems Vertical Sectorization

Vertical Sectorization (VeSn) description

Vertical separation of the two beams in the same sector

Sector divided in two cells: inner and outer with resp. vertical tilts θinner andθouter with θinner > θouter

I Transmit powers:Pi and Po for inner andouter cells resp. withPi + Po = Ptotal.

I Possible implementations:bandwidth sharing or fullreuse.



VeSn with frequency reuse one

Description

Inner and outer sectors reuse the whole available bandwidth.

Power budget split equally between inner and outer sectors.

Advantages

Increased capacity.

Increased antenna gain for inner cell users.

Drawbacks

Reduced transmit power.

More interference.

Requirement: SON controller for VeSn feature activation.



VeSn feature activation problem

Activation rule

Action =

{VeSn ON if µON(ρi , ρo) > µOFF(ρi , ρo)VeSn OFF otherwise

(9)

Decision Boundary

µON(ρi , ρo) = µOFF(ρi , ρo) (10)

(VSOFF) : a1ρ2i + b1ρiρo + c1ρ

2o + d1ρi + e1ρo = 0 (11)

(VSON) : a2ρ2i + b2ρiρo + c2ρ

2o + d2ρi + e2ρo = 0 (12)

with a1, a2, b1, b2, c1, c2, d1, d2, e1, e2: parameters depending on traffic and datarate distributions in the sector.



VeSn activation controller calibration

Data from realistic network simulator

Activation

0 0.2 0.4 0.6 0.8 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

ρinner,off

ρ oute

r,of

f

Better activateStay deactivated

Deactivation

0 0.2 0.4 0.6 0.8 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

ρinner,on

ρ oute

r,on

Stay activatedBetter deactivate

Data used to estimate parameters of the decision boundaries: classificationproblem.



VeSn activation controller performance

ρinner

ρ oute

r

0 0.2 0.4 0.6 0.8 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Decision boundary for VS OFFDecision boundary for VS ON

Activation controller

0 1000 2000 3000 4000 5000 6000 7000 80000

2

4

6

8

10

12

Tra

ffic

dem

and

(Mbp

s)

Time (s)

Inner cellOuter Cell

Traffic scenario

0 1000 2000 3000 4000 5000 6000 7000 80000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Act

ivat

ion

Dec

isio

n (0

=O

FF

,1=

ON

)

Time (s)

Always OFFAlways ONAAS SON

Activation decisions

0 1000 2000 3000 4000 5000 6000 7000 80000

5

10

15

20

25

30

Mea

n U

ser

Thr

ough

put (

Mbp

s)

Time (s)

Always OFFAlways ONAAS SON

User performance



VeSn with bandwidth sharing

Description

Total frequency bandwidth split between inner and outer sectors.

Transmit power per Hz does not change.

Advantages

No Inter-cell interference between inner and outer cells.

Increased transmit power for each user compared to reuse one.

Increased antenna gain for inner cell users.

Drawbacks

Reduced capacity because there is no reuse.

Problem: Which sharing proportions for the frequency bandwidth?



Optimal bandwidth sharing

Parameter: δ - proportion of bandwidth allocated to inner cell.

Criteria: Alpha-fair utility of all users throughputs

Uα(δ) =

∑u∈Ui

log(δRu

)+∑

u∈Uolog((1− δ)Ru

)α = 1∑

u∈Ui

(δRu)1−α

1−α +∑

u∈Uo

((1−δ)Ru)1−α

1−α α 6= 1

If α = 1, optimal parameter in closed form:

δ =Ni

Ni + No(13)

with Ni = |Ui | and No = |Uo |.For general α:

δ[k + 1] = δ[k] + ε∂Uα(δ[k])

∂δ



VeSn with bandwidth sharing performance

Network layout

0 2 4 6 8 100

10

20

30

40

50

60

70

80

90

100

Arrival rates (users/s)

Load

s (%

)

VS bandwidth sharingVS reuse oneBaselineSON controller

Maximum loads

0 2 4 6 8 100

5

10

15

20

25

30

35


Mea

n U

ser

Thr

ough

puts

(M

bps)



0 2 4 6 8 100

2

4

6

8

10

12


Cel

l Edg

e U

ser

Thr

ough

puts

(M

bps)




Overview

1 Introduction

2 Background



5 SON Coordination



Active Antenna Systems Virtual Sectorization

Virtual Sectorization (ViSn) description

Evolution of vertical sectorization.

Spatial separation of beam (both vertically and horizontally) using antennaarrays.

Conservation of total power budget leading to resource allocation problems.

Can be implemented with reuse one or frequency sharing (as in VeSn).

Antenna Array Network layout


Active Antenna Systems Virtual Sectorization

ViSn performance results

00:00 00:50 01:40 02:30 03:20 04:100

1

2

3

4

5

6

Time (HH:MM)

Arr

ival

rat

es (

user

s/s)

λh

λTotal

Traffic profile

00:00 00:50 01:40 02:30 03:20 04:1030

40

50

60

70

80

90

Time(HH:MM)

Load

s (%

)

ViSn bandwidth sharingViSn reuse oneBaseline (No ViS)

Maximum loads

00:00 00:50 01:40 02:30 03:20 04:108

9

10

11

12

13

14

15

16

17

18

Time(HH:MM)

Mea

n U

ser

Thr

ough

puts

(M

bps)



00:00 00:50 01:40 02:30 03:20 04:100.5

1

1.5

2

2.5

3

3.5

4

4.5

Time(HH:MM)

Cel

l Edg

e U

ser

Thr

ough

puts

(M

bps)




Overview

1 Introduction

2 Background



5 SON Coordination



Active Antenna Systems Multilevel Beamforming

Introduction to multilevel beamforming

Challenging topic in Massive MIMO community

State of the art: channel matrix estimation and inversion

Goal: low complexity processing for beamforming in TDD & FDD and lowfeedback in case of FDD.

Our approach: Extend codebook idea to integrate coverage aspect =⇒ Beamplanning.



Multilevel beamforming idea

Example of beam hierarchy

Design the codebook hierarchically.

Find the best beam available bynavigating iteratively through thecodebook.

Tree search (logarithmic complexity)



Beam planning for each type of environment

Dense Urban Rural

Problem: automatic generation of beam codebook given basic cellinformation (size, antenna height, etc.). (future work)



Multilevel beamforming performance

0 5 10 15 20 25 30 35 40 450

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

User throughputs (Mbps)

CD

F

Empirical CDF

No beamforming(baseline)Hierarchical beamformingOptimal beamforming

User throughput CDFs comparison

Data rate function: R = min(Rmax , η log2(1 + SINR))


Overview

1 Introduction

2 Background



5 SON CoordinationConceptProblem formulation and SolutionUse case



SON Coordination

SON model as control loops

Generic formulation of stochastic approximation algorithms

θn+1 = θn + ε(F (θn) + Mn) (14)

where

θ: parameters,F (·): search directions,Mn: noise.

Mean behavior described by the limiting Ordinary DifferentialEquation (ODE):

θ = F (θ) (15)


SON Coordination Concept

Stability and coordination

Jacobian of F (θ) defined as Gθ = JF (θ) where

Gθ(i , j) =∂Fi (θ)

∂θj(16)

Rosen’s sufficient condition for stability:

Theorem

If the matrix Gθ + GTθ is negative definite for every θ in

∏Nj=1 Sj , then θ = F (θ)

has a unique equilibrium point and it is asymptotically stable in∏N

j=1 Sj .

Local stability given by linearization: F (θ) = Aθ then AT + A negativedefinite is a sufficient condition.

Coordination idea: Apply a coordination matrix C (obtaining θ = CF (θ))such that (CA)T + CA is negative definite.


SON Coordination Problem formulation and Solution

Coordination matrix computation

minimize ‖C + A−1‖Fs.t. (CA)T + CA ≺ 0;C ∈ C

(17)

where

‖.‖F is the Frobenius norm.

C: the set of coordination matrices satisfying the system constraints.

θi = Fi (θ)www�θi =

∑j

Ci,jFj(θ)

System constraints


SON Coordination Use case

Use case description

SON function Parameters KPIs

Load balancing Transmit pilot powerPpilot

Load of the cell ρ

Outage Probabilityminimization

Transmit data powerPTCH

Coverage probability ofthe cell K

Blocking Rateminimization

Admission threshold AT Blocking rate of the cellb


SON Coordination Use case

Performance results

500 1000 1500 2000 2500 3000 35000

10

20

30

40

50

60

70

80

90

100

Time (s)

BS

Loa

ds (

%)

BS 1 without coordinationBS 2 without coordinationBS 3 without coordinationBS 1 with coordinationBS 2 with coordinationBS 3 with coordination

Loads

500 1000 1500 2000 2500 3000 35000

10

20

30

40

50

60

70

80

90

100

Time (s)O

utag

e pr

obab

ility

(%

)

BS 1 without coordinationBS 2 without coordinationBS 3 without coordinationBS 1 with coordinationBS 2 with coordinationBS 3 with coordination

Coverage probabilities


Overview

1 Introduction

2 Background



5 SON Coordination

6 Conclusion & PerspectivesConclusionPerspectives


Conclusion & Perspectives Conclusion

Conclusion

SON algorithms for small cells

Load balancing (with constrained backhaul)Alpha-fair Interference coordination

SON algorithms for active antenna systems

VeSn feature activationAlpha-fair bandwidth sharing for VeSn and ViSnBeam selection algorithm for multilevel beamforming

Which is the best option?

With low cost backhaul or for non-line-of-sight coverage areas: Small CellsOthers: Active Antenna Systems (multilevel beamforming)

Systematic SON coordination framework

Tested for load balancing with interference coordination in small cells scenario.


Conclusion & Perspectives Perspectives

Perspectives

I Extend algorithms to other use cases: D2D, energy saving, etc.

I Backhaul-aware SON functions

I Multi-armed bandits for AAS features activation

I Beam planning automation and application to more use cases

I Which α in α-fair utilities?

I Coordination of highly non-linear systems of SON functions


Thank you!Questions are welcome.

optimization and self-optimization for lte networks · pdf fileoptimization and...

Documents