ph. d. defense

34
22/2/7 | Institute of Telecommunications | Area of Communications Engineering Downlink Adaptive Resource Allocation for a Multi-user MIMO OFDM System with and without Fixed Relays Ph. D. Defense M. Sc. Ying Zhang Reviewers: Prof. Dr.-Ing. Anja Klein Prof. Dr.-Ing. Dr. rer. nat. Holger Boche Examiners: Prof. Dr.-Ing. Peter Meißer Prof. Dr.-Ing. Han Eveking Institute for Telecommunications / Area of Communications Engineering Department for Electrical Engineering and Information Technology Darmstadt University of Technology

Upload: tristram-davies

Post on 31-Dec-2015

25 views

Category:

Documents


1 download

DESCRIPTION

Downlink Adaptive Resource Allocation for a Multi-user MIMO OFDM System with and without Fixed Relays. Ph. D. Defense. M. Sc. Ying Zhang Reviewers: Prof. Dr.-Ing. Anja Klein Prof. Dr.-Ing. Dr. rer. nat. Holger Boche Examiners: Prof. Dr.-Ing. Peter Meißer Prof. Dr.-Ing. Han Eveking - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering

Downlink Adaptive Resource Allocation for a Multi-user MIMO OFDM System with and without Fixed RelaysPh. D. Defense

M. Sc. Ying Zhang

Reviewers:Prof. Dr.-Ing. Anja Klein

Prof. Dr.-Ing. Dr. rer. nat. Holger BocheExaminers:

Prof. Dr.-Ing. Peter MeißerProf. Dr.-Ing. Han Eveking

Institute for Telecommunications / Area of Communications Engineering Department for Electrical Engineering and Information Technology

Darmstadt University of Technology

Page 2: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 2

Challenges: 3-dimensional

resources

Two-hop communication

Interference among access points (APs), including BS and RNs

Introduction (1): Downlink Multi-user OFDM MIMO System with and without Fixed Relays

OFDM (Orthogonal frequency division multiplexing)

MIMO (Multiple-input multiple-output)

BS: MT tx antennas; K UTs: K·MR rx antennas

Fixed relays (RN)

fre

qu

en

cy

time1

N

Downlink multi-user: base station (BS) transmits, K user terminals (UTs) receive.

space

time

1MT space

fre

qu

en

cy

time

BSRN

signal

interference

fre

qu

en

cy

time

space

timeMT < K·MR

N sub-carriers

Page 3: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 3

Introduction (2): Adaptive Resource Allocation

Adaptive FDMA Channel fading is time-varying, frequency-

selective and independent among users. Always allocate best resources to users.

Adaptive SDMA Interference among co-located users is

proportional to their spatial correlation. Always allocate users with sufficiently low

spatial correlation together.

Dynamic resource reuse among multiple APs Inter-AP interference is time-varying,

frequency-selective. Reuse resource, i.e. multiple APs use the

same resource, when the interference is sufficiently low.

UT 1UT 2UT 3

freq.

channel gain

1 N

hk: channel vectorwk: antenna vector

Signal S1 = |h1T w1|2

Interference I1 = |h1T w2|2

max S1 w1=h1* I2 = h1

T h2

AP

UT1

UT2

w1

w2

AP

high interference

low inerference

AP1

AP2

AP3

Page 4: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 4

Outline

Adaptive Resource Allocation in a Single CellJoint adaptive FDMA/SDMA

Power minimization with user rate constraints Rate maximization with user fairness constraints

Signaling Overhead for Adaptive Resource Allocation Optimization of Chunk Dimension Optimization of Chunk Update Interval Resource Allocation with Reduced Channel Feedback Optimization of Bandwidth Request Transmission

Adaptive Resource Allocation in a Relay-enhance CellTwo-level adaptive resource allocation Construction of logic beams Grouping of logic beams Resource allocation among logic beams

Page 5: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 5

Adaptive Resource Allocation in a Single Cell- Joint Adaptive FDMA/SDMA

State of the art

Adaptive FDMA in OFDMA system:

no spatial dimension

Adaptive SDMA in narrow-band MIMO system:

no frequency dimension

Contribution 1: propose low-complexity algorithm performing joint optimization of adaptive FDMA/SDMA for power minimization problem and rate maximization problem.

Spatial correlation is frequency selective.

Joint optimization of adaptive- FDMA and SDMA is required.

Assumption: using chunk as basic resource unit

freq

uenc

y

time

space

Layer 1

Layer 2

Layer 3

Chunk

Page 6: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 6

Joint Adaptive FDMA/SDMA- Power minimization problem (1)

Power minimization problem:

Solutions: Optimal solution: exhaustive search by integer linear programming huge complexity

Sub-optimal solution: Successive Bit Insertion (SBI) low complexity

a) Initialization:

b) Each iteration: where

c) End condition:

k: user index

n: chunk index

Rk: min. data rate requirement

rk,n: allocated data rate

Pk,n: required transmit power

cost for granting a given rate increase Δr to user k on chunk n

Page 7: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 7

Joint Adaptive FDMA/SDMA- Power minimization problem (2)

Three variants for the cost function Original

With Priority

Power increase required to increase the data rate of user k on chunk n by Δr

relative allocated rate

Weighted priority (WP)

First priority (FP)

Prioritize the user whose allocated rate is far away from the minimum requirement Rk.

Original variant is the worst. FP variant is better than WP variant

and approaches the optimal solution.

Page 8: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 8

Joint Adaptive FDMA/SDMA- Rate maximization problem (1)

Rate maximization problem: to maximize sum data rate while satisfying user fairness properties under power constraint. Equal power distribution over chunks: User fairness strategies

Proportional Fair Strategy Max-Min Fair Strategy

Criteria

Solution

Weighted Prioirty (WP): First Prioirty (FP):

Notes: average data rate update

average data rate update after each iteration

Page 9: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 9

Joint Adaptive FDMA/SDMA- Rate maximization problem (2)

To further reduce the complexity assume equal power sharing among users served on the same chunk

Successive User Insertion (SUI)

Initialization:

Each iteration:

End condition: sum rate cannot be increased without violating the power constraints

set of users served on chunk n

where

Page 10: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 10

Joint Adaptive FDMA/SDMA- Rate maximization problem (3)

Bit-Ins outperforms User-Ins by adaptive power loading among different spatial layers.

WP improve user satisfaction at the expense of total cell throughput compared to FP.

Suc. Bit Ins. vs. Suc. User Ins. Proportional Fair (WP) vs. Max-min Fair (FP)

FP + SBI FP + SUI WP + SUI

Average user throughput (Mbps)

5.42 4.50 7.35

Page 11: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 11

Joint Adaptive FDMA/SDMA- Rate maximization problem (4)

SUI performs worse than SBI due to two factors:• No power adaptation among users• Discrete rate adaptation

Page 12: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 12

Joint Adaptive FDMA/SDMA- Rate maximization problem (5)

Disjoint (State of the art):

• Allocate chunks in arbitrary order, e.g. one after another in order

• For the given chunk, select the user that minimizes cost function

FP for max-min fairness: Joint approach achieves around 35% more throughput than disjoint approach.

WP for proportional fairness: The performance gain of joint approach over disjoint one is small.

Joint vs. Disjoint FDMA/SDMA

Page 13: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 13

Signaling Overhead - Optimization of Chunk Dimension (1)

AP needs to inform users the results of adaptive resource allocation, resulting in additional signaling overhead

2 2 ctl

sub symb

log logovh

K R r

n n

No. of trans. mode

No. of users

Rate of signaling (bits/symbol)

Chunk dimension: nsub subcarriers by nsymb symbols

Contribution 2: Analytically deriving the relationship between the performance of adaptive resource allocation and the chunk dimension so as to derive the optimal chunk dimension.

• Increasing chunk dimension reduces signaling overhead;

• Decreasing chunk dimension enhances performance of adaptive resource allocation.

Trade-off exists in choosing chunk dimension:

Page 14: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 14

1 5

2

3 m

4

Signaling Overhead - Optimization of Chunk Dimension (2)

Roadmap of the analytical derivation

2( )kmh

Mean value within m-th chunk

mean variance

Channel coefficient 0 1

0 Ω

Channel coefficient

within m-th chunk

1- Ω( ),k

n th

( )kmh

( ),k

n th

( )kmh

The user with highest performance is equivalent to highest

ymbsub 11( ) ( )

,0 0sub symb

1 snnk k

m n tn t

h hn n

t

n

2( )mkmh PDF of , , can then be calculated through order statistics.

Assumptions in the analytical derivation: An OFDMA system with one AP, K users and N subcarriers Channel coefficients modeled as a stationary two dimensional zero-mean Gaussian process, whose

variance is set to one without loss of generality Perfect channel knowledge known at the AP Performance evaluated in terms of Shannon capacity

Page 15: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 15

Signaling Overhead - Optimization of Chunk Dimension (3)

Ceff,max = 3.06 bits/s/Hz at (8,18)

Max delay spread: 3.2s; Velocity: 100km/h

Optimal chunk dimension is 8 sub-carriers by 18 OFDM symbols.

Page 16: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 16

Signaling Overhead - Optimization of Channel Update Interval (1)

Tup,opt =arg max ρ(Tup)

Contribution 3: Derive the optimal channel update interval which maximizes the effective throughput.

β: Signaling for channel knowledge update

Tup: Channel update interval

0up

1 upTT

Fixed Allocation

Tup0

too less channelknowledge

no adaptation gain

too muchoverhead

• Increasing Tup reduces signaling overhead

• Decreasing Tup enhances adaptive resource allocation.

Channel is time-variant due to mobility

Channel knowledge shall be periodically updated.

Trade-off exits in choosing channel update interval Tup:

Effective Throughput:

Page 17: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 17

Signaling Overhead - Optimization of Channel Update Interval (2)

frame duration

coherence time

0 5 10 1560

70

80

90

Update Interval Tup

[Frame]

Th

rou

gh

pu

t 0 [M

bp

s]

5 km/h10 km/h20 km/h

0 0.1 0.250

60

70

80

90

Th

rou

gh

pu

t 0 [M

bp

s]

Relavtive update interval T'up

5 km/h10 km/h20 km/h

Update interval relative to speed of the channel time variability:

0 'upTApproximate by a linear curve suitable for arbitrary velocity based on the numerical results:

Page 18: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 18

Signaling Overhead - Optimization of Channel Update Interval (3)

Optimal update interval: up,optT v

Page 19: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 19

Signaling Overhead - Allocation with Reduced Channel Feedback (1)

Contribution 4: perform joint adaptive FDMA/SDMA based on long-term CSI and short-term CQI.

Long-term Generalized Eigenbeamforming Opportunistic beamforming/SDMA

Assumption: long-term CSI

No adaptation in time- and freq.- domains.

Pre-determine Q beams wq

Each user selects the best beam.Allocate each beam to the best user.

No Adaptive switching between with and w/o SDMA.

Less signaling

Assumption: short-term CQIown signal

interference to others

Different levels of channel knowledge:short-term CSI (channel state information): full channel matrix short-term CQI (channel quality matrix): channel gain long-term CSI: channel covariance matrixlong-term CQI: average channel gain – average SNR

State of the art

Page 20: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 20

Signaling Overhead - Allocation with Reduced Channel Feedback (2)

lt / trk k kR R R

2

,1

lt2,

lt2,

,SINR

K

kjjjk

Hjjnk

kkHkknk

nk

pa

pa

wRw

wRw H

rkrknknk

Hkkkk

a ,,,,

lt

ˆ

eig

VH

VVR

(a) Assuming single receive antenna (b) Approximate the channel matrix by the first r dominant eigenvector

, , , , tr Hk n k n k n k na H H H

SINR Underestimated

SINR overestimated

Conservative resource allocation Aggressive resource allocation

Calculation of Actual SINR: require full channel feedback — short-term CSI

Prediction of SINR for long-term Generalized Eigenbeamforming : Available channel knowledge:

long-term CSI

short-term CQI

Proposed methods:

Page 21: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 21

Signaling Overhead - Allocation with Reduced Channel Feedback (3)

Conservative resource allocation is beneficial than aggressive resource allocation.

0 5 10 15 20 25

4

8

12

16

Average SNR [dB]

Ave

rag

e u

ser

thro

ug

hp

ut

[Mb

ps]

GoBGenEigBF-IDEALGenEigBF-CONSZFBFGenEigBF-AGGR

GoB: less adaptive, more accurate SINR estimation better in high SNR region.

GenEigBF: more adaptive, less accurate SINR estimation better in low-middle SNR region.

CONS

AGGR

error = SINR-SINRest [dB]

SNR =15 dB

CC

DF

(err

or)

Page 22: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 22

Signaling Overhead - Optimization of Bandwidth Request Transmission (1)

Performance: average delay - the time difference between the arrival and the successful transmission of the BW-REQ.

Contribution 5: (a) derivative the performance of random access in a frame-based system under the assumption that the arrival of BW-REQ is modeled as Bernoulli process; (b) propose a novel user grouping approach which improves the performance of random access.

Users ask for allocation of resources by sending a bandwidth request (BW-REQ).

State of the art: random access in WLAN (wireless local access network) under the assumption that users always have data to transmit has been well-studied by Bianch.

Two typical approaches:

User 1 User 2 User 31. Polling

2. Random Access

confliction

Average delay = frame

No. of users

N TOs UL Data

uplink frame

Each BW-REQ is transmitted in one transmission opportunity (TO).

Widely used scheme: slotted-Aloha with truncated binary exponential back-off (TBEB) algorithm

Page 23: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 23

Signaling Overhead - Optimization of Bandwidth Request Transmission (2)

Bianch models slotted-Aloha with TBEB algorithm with Markov Chain.

m: maximum back-off stageW0: initial window size.

Slotted-Aloha with TBEB algorithm has two parameters:

In stage i, a back-off counter ci between 0 to Wi-1 is chosen, Wi=2iW0. The back-off counter ci indicates the

number of TOs the user has to wait before a transmission attempt.

When a collision happens, the user goes to stage i+1 unless it reaches the maximum stage.

In case of successful transmission, the user goes to stage 0.

Page 24: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 24

Signaling Overhead - Optimization of Bandwidth Request Transmission (3)

2 11N - 11 110

1

Successful transmission

λ 1-λ...

Introduce two kinds of idle states in the Markov chain.

Users know whether there is a collision:

• immediately after the transmission in WLAN

• at the beginning of the next frame in frame-based system

Bianch’s analysis assumes that users always have some packets to transmit.

The arrival of the BW-REQ is modeled as Bernoulli process with parameter λ, i.e. a new BW-REQ occurs with probability λ in every frame at the beginning of each frame.

If a user transmits in the n-th TO, it won’t immediately know until the next frame and thus the back-off process will stop for (N-n)-th TOs.

After successful transmission, when there is no new BW-REW coming, the back-off process will stop for N TOs.

21 1N -1

1 N

11

Transmission attempt

1...

...

Page 25: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 25

Signaling Overhead - Optimization of Bandwidth Request Transmission (4)

Analytical results meet the simulation results pretty well.

Page 26: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 26

Signaling Overhead - Optimization of Bandwidth Request Transmission (5)

Observations: the performance is almost the same when the ratio between the number of users and the number of TOs is constant.

Proposal:

• Divide users into G groups, such that users in the same group have similar channel quality.

• Divide resources into G groups such that

Problem: resources are not efficiently used

• The resource required for BW-REQ transmission depends on the used data rate.

• Users support different data rate according to their channel quality.

• Each TO should be large enough for the transmission using the lowest data rate.

Kg (No. of users in group g)= constant

Ng (No. of TOs in group g)

Page 27: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 27

Signaling Overhead - Optimization of Bandwidth Request Transmission (6)

More TOs and so less delay, given the same amount of resources;

Alternatively, less resources are required to provide the same number of TOs.

Comparing my proposal to the conventional method:

Example: 144 symbols for BW-REQ Tx, 48-bits BW-REQ;12 users support 2 bits/symbol, 12 users only support 1 bits/symobl.

Conventionally, 1 bits/symbol 48 symbols / BW-REQ 3 TOs

In my proposal, 96 symbols using 2 bits/symbol 2 TOs48 symbols using 2 bits/symbol 2 TOs

K/N = 8

N1/K1 = N2/K2 = 6

Page 28: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 28

Adaptive Resource Allocation in a REC - Two-level adaptive resource allocation

State of the Art: Centralized approach adaptive resource allocation for the whole

REC is carried out by the BS high complexity huge signaling overhead additional delay due to two-hop communication

Distributed approach adaptive resource allocation is carried out by individual AP for its serving users independently

inter-AP interference is unpredictable

Resource Partitioning performed by BS on longer time scale, of few millisecs, to dynamically partition the resources among APs (i.e. BS and RNs) within a REC according to average traffic load and interference scenario

Resource Scheduling performed by each AP in each sub-cell on a shorter time-scale, of less than 1 ms, in order to obtain multi-user diversity through frequency-adaptive resource allocation

Contribution 6: Propose a two-level adaptive resource allocation

BS

RN

Page 29: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 29

Two-level adaptive resource allocation- dynamic logic beam

Proposed: dynamic logic beam Beam construction:

identification of groups of spatially correlated AP-UT links at each AP (BS or RN). Each group of links is referred to as beam;

Beam grouping:

identification of spatially uncorrelated beams which are allowed to share the same time-frequency resources, i.e. chunks.

Resource partitioning:

assign resources to BS-RN links and logic beams (BS/RN-UT links):

Reference: sectorization — fixed logic beam the sub-cell of each RN is divided into fixed sectors; fixed sets of spatially uncorrelated sectors are allowed to

share the same chunks in the spatial domain.

logic beam

sector

Page 30: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 30

Two-level adaptive resource allocation- Beam construction

11 22 33 44 55

1 21 2 44 55

55

11 22 33 44 55

1 21 2 3333 44 55

55

1 21 21 21 24 5555

3

3333

1 21 21 21 24

1

2

3

4

5

AP

Beam construction: Initialization: each user constitues a beam Each iteration: combining two beams with highest spatial correlation together End: spatial correlation of any two beams is sufficiently low

iterations

Definition of spatial correlation between beams A and B:

Spatial correlation between users i and j

beam A

beam B

Page 31: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 31

Two-level adaptive resource allocation- Beam grouping

Capacity increases

A B C D

A D

A D B

E

A D B C

A D C C E

Capacity increases

Capacity increases

Capacity decreases

jiBjAi

BA II /,

/ max

X

Beam grouping: successively add best beam until group capacity decreases

Group capacity / group data rate: Rate requirement of user i

achievable rate of user i

iterations

Page 32: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 32

Two-level adaptive resource allocation- Resource partitioning

Resources allocated to the first hops and the second hops should be balanced.

Chunk-by-chunk balancing (CBC)

1. Generate one beam group

2. Allocate one chunk for the group

3. Reserve resources for the first-hop links

4. Repeat 2-3 till at least one beam in the group is completely allocated

5. Repeat Steps 1-4 till

a) All beams are completely allocated, or

b) No resource is left

Iterative independent balancing (IIB)

1. Allocation for beams: calculate the amount of resources required to completely allocate all the beams

2. Allocation for first-hop links: calculate the amount of resources required for the first-hop links

3. If the sum of the required resources > the available resources,

a) proportionally scaling down the required data rate of each user

b) Go back to step 1Completely allocated: allocated rate ≥ required rate

End-to-end throughput is equal to the minimum between that of the first hop (BS-RN links) and that of the second-hop (BS/RN-UT links).UT BS RN UT

single hopend-to-endconnection

second hop

access link relay link access link

first hop

end-to-endconnection

Page 33: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 33

Two-level adaptive resource allocation- Simulation Results

Dynamic approach achieves higher cell throughput and better user fairness.

CCB achieves high cell throughput, but IIB guarantees user fairness.

Jain’s fairness index:Jain’s fairness index:

2

1

2

1

( )

n

ii

n

ii

xf x

n x

1( ) 1f x

n

unfair totally fair

Page 34: Ph. D. Defense

23/4/19 | Institute of Telecommunications | Area of Communications Engineering 34

Summary

This work proposes low-complexity sub-optimal algorithms for joint optimization of

adaptive FDMA and SDMA; analytically derives the performance of adaptive FDMA as a function of

the chunk dimension which facilitates the optimization of the chunk dimension;

investigates the optimal update interval of channel knowledge; applies joint adaptive FDMA and SDMA to a system when only long-term

CSI and short-term CQI are available at the transmitter; analytically derives the performance of random access and proposes a

grouping mechanism which enables more efficient usage of resources; presents a hierarchical approach for the adaptive resource allocation in a

relay-enhanced cell which achieves high adaptation gain at low signaling overhead.