Mobility Management and Resource Allocation towards 5G ...

Post on 10-May-2022

4 views 0 download

Transcript of Mobility Management and Resource Allocation towards 5G ...

Mobility Management and Resource Allocationtowards 5G Radio Access Networks (RANs)

Konstantinos AlexandrisEmail: konstantinos.alexandris@eurecom.fr

PhD thesis defenceCommunication Systems Dep., EURECOM

March 9, 2018

Konstantinos Alexandris 1 / 52

Outline

1 Introduction

2 Mobility (HO) management in LTE/LTE-A and beyond

3 Multi-connectivity resource allocation towards 5G RANs

4 Conclusion & Future directions

Konstantinos Alexandris 2 / 52

Introduction

Konstantinos Alexandris 3 / 52

5G: Ushering a new era

2020: 50B devices to be prevalent

Konstantinos Alexandris 4 / 52

5G: Ushering a new era

2020: 50B devices to be prevalent

5G architecture

X SCs → UDNs

X New paradigms

H2M, M2M

X New technologies

Massive MIMOmm-wave

X Network management

SDN, MEC

Konstantinos Alexandris 4 / 52

5G: Ushering a new era

2020: 50B devices to be prevalent

What are the 5G requirements?

5G Services: xMBB, uRLLC, mMTC

Konstantinos Alexandris 4 / 52

5G: Ushering a new era

2020: 50B devices to be prevalent

What are the 5G requirements?

5G Services: xMBB, uRLLC, mMTC

5G Use cases

X Media and Broadband

X E-health

X IoT

X V2X

Konstantinos Alexandris 4 / 52

5G: Ushering a new era

2020: 50B devices to be prevalent

What are the 5G requirements?

5G Services: xMBB, uRLLC, mMTC

5G Use cases

X Media and Broadband

X E-health

X IoT

X V2X

Konstantinos Alexandris 4 / 52

5G: Ushering a new era

2020: 50B devices to be prevalent

What are the 5G requirements?

5G Services: xMBB, uRLLC, mMTC

5G Use cases

X Media and Broadband

X E-health

X IoT

X V2X

Konstantinos Alexandris 4 / 52

5G: Ushering a new era

2020: 50B devices to be prevalent

What are the 5G requirements?

5G Services: xMBB, uRLLC, mMTC

5G Use cases

X Media and Broadband

X E-health

X IoT

X V2X

Konstantinos Alexandris 4 / 52

5G Networks

Macrocell

Wi-Fi

LTE

UDN

CN

Internet

RNsD2D

D2D D2D

V2X

D2D

SDN

LTE

D2D

Massive MIMO

D2D

D2D

MC

SC

SC

Multi-connectivity

Multi-connectivity

C-RAN

Konstantinos Alexandris 5 / 52

From 4G to 5G: Contributions

New algorithms

Centralized architectureLTE/LTE-A RAN control

Mobility mgt (RRC)

X2 Handover (HO)Load + HO in HetNets

Multi-connectivity resourceallocation (MAC)

Air + BH limitationsOpportunistic scheduling

Konstantinos Alexandris 6 / 52

From 4G to 5G: Contributions

New algorithms

Centralized architectureLTE/LTE-A RAN control

Mobility mgt (RRC)

X2 Handover (HO)

Load + HO in HetNets

Multi-connectivity resourceallocation (MAC)

Air + BH limitationsOpportunistic scheduling

Konstantinos Alexandris 6 / 52

From 4G to 5G: Contributions

New algorithms

Centralized architectureLTE/LTE-A RAN control

Mobility mgt (RRC)

X2 Handover (HO)Load + HO in HetNets

Multi-connectivity resourceallocation (MAC)

Air + BH limitationsOpportunistic scheduling

Konstantinos Alexandris 6 / 52

From 4G to 5G: Contributions

New algorithms

Centralized architectureLTE/LTE-A RAN control

Mobility mgt (RRC)

X2 Handover (HO)Load + HO in HetNets

Multi-connectivity resourceallocation (MAC)

Air + BH limitationsOpportunistic scheduling

Konstantinos Alexandris 6 / 52

X2 Handover in LTE/LTE-A

Konstantinos Alexandris 7 / 52

X2 Handover in LTE/LTE-A

MME

X2eNB0 eNB1

S1-U S1-U

S-GW

S1-MME S1-MME

S11

UEUE

UEUE

UE

S1 vs X2

HO: Procedure to transfer a UE andits context from source to targeteNB

X2 HO: Reduce EPC signaling loadby 6x!

Goals & Challenges:

Study of X2 handover parameters

Handover latency evaluation

Implementation in OpenAirInterface

Contrary to prior-art:

Open to experimenters community

HO parametirization

Flexibility + Accessibility

Konstantinos Alexandris 8 / 52

X2 Handover in LTE/LTE-A

MME

X2eNB0 eNB1

S1-U S1-U

S-GW

S1-MME S1-MME

S11

UEUE

UEUE

UE

S1 vs X2

HO: Procedure to transfer a UE andits context from source to targeteNB

X2 HO: Reduce EPC signaling loadby 6x!

Goals & Challenges:

Study of X2 handover parameters

Handover latency evaluation

Implementation in OpenAirInterface

Contrary to prior-art:

Open to experimenters community

HO parametirization

Flexibility + Accessibility

Konstantinos Alexandris 8 / 52

X2 Handover experimentation in oaisim

Run in OAI emulator

3GPP TS 36.331, 36.423PHY/MAC/RLC/PDCP/RRC

UE

MovingeNB 0

eNB 1

Fixed location

(2680,4800)

Fixed location

(4000,4800)

Mobile location

(1800,4800)↔(4700,4840)

Network components

2 eNBs: source/target1 UESISO link (Tx/Rx)

Grid topology

eNBs: Fixed positionUE: Mobility tracesStraight line movement

Traffic:

UL/DL UDP traffic

Configuration:

.config/.xml filesSpecifically: RF, protocol,mobility, traffic params

Output:

pcap/log files, messagesignalling

Konstantinos Alexandris 9 / 52

X2 Handover experimentation in oaisim

Run in OAI emulator

3GPP TS 36.331, 36.423PHY/MAC/RLC/PDCP/RRC

UE

MovingeNB 0

eNB 1

Fixed location

(2680,4800)

Fixed location

(4000,4800)

Mobile location

(1800,4800)↔(4700,4840)

Network components

2 eNBs: source/target1 UESISO link (Tx/Rx)

Grid topology

eNBs: Fixed positionUE: Mobility tracesStraight line movement

Traffic:

UL/DL UDP traffic

Configuration:

.config/.xml filesSpecifically: RF, protocol,mobility, traffic params

Output:

pcap/log files, messagesignalling

Konstantinos Alexandris 9 / 52

X2 Handover experimentation in oaisim

Run in OAI emulator

3GPP TS 36.331, 36.423PHY/MAC/RLC/PDCP/RRC

UE

MovingeNB 0

eNB 1

Fixed location

(2680,4800)

Fixed location

(4000,4800)

Mobile location

(1800,4800)↔(4700,4840)

Network components

2 eNBs: source/target1 UESISO link (Tx/Rx)

Grid topology

eNBs: Fixed positionUE: Mobility tracesStraight line movement

Traffic:

UL/DL UDP traffic

Configuration:

.config/.xml filesSpecifically: RF, protocol,mobility, traffic params

Output:

pcap/log files, messagesignalling

Konstantinos Alexandris 9 / 52

X2 Handover experimentation in oaisim

Run in OAI emulator

3GPP TS 36.331, 36.423PHY/MAC/RLC/PDCP/RRC

UE

MovingeNB 0

eNB 1

Fixed location

(2680,4800)

Fixed location

(4000,4800)

Mobile location

(1800,4800)↔(4700,4840)

Network components

2 eNBs: source/target1 UESISO link (Tx/Rx)

Grid topology

eNBs: Fixed positionUE: Mobility tracesStraight line movement

Traffic:

UL/DL UDP traffic

Configuration:

.config/.xml filesSpecifically: RF, protocol,mobility, traffic params

Output:

pcap/log files, messagesignalling

Konstantinos Alexandris 9 / 52

X2 Handover experimentation in oaisim

Run in OAI emulator

3GPP TS 36.331, 36.423PHY/MAC/RLC/PDCP/RRC

UE

MovingeNB 0

eNB 1

Fixed location

(2680,4800)

Fixed location

(4000,4800)

Mobile location

(1800,4800)↔(4700,4840)

Network components

2 eNBs: source/target1 UESISO link (Tx/Rx)

Grid topology

eNBs: Fixed positionUE: Mobility tracesStraight line movement

Traffic:

UL/DL UDP traffic

Configuration:

.config/.xml filesSpecifically: RF, protocol,mobility, traffic params

Output:

pcap/log files, messagesignalling

Konstantinos Alexandris 9 / 52

X2 Handover experimentation in oaisim

Run in OAI emulator

3GPP TS 36.331, 36.423PHY/MAC/RLC/PDCP/RRC

UE

MovingeNB 0

eNB 1

Fixed location

(2680,4800)

Fixed location

(4000,4800)

Mobile location

(1800,4800)↔(4700,4840)

Network components

2 eNBs: source/target1 UESISO link (Tx/Rx)

Grid topology

eNBs: Fixed positionUE: Mobility tracesStraight line movement

Traffic:

UL/DL UDP traffic

Configuration:

.config/.xml filesSpecifically: RF, protocol,mobility, traffic params

Output:

pcap/log files, messagesignalling

Konstantinos Alexandris 9 / 52

OAI X2 HO UE measurements

0 50 100 150 200 250 300−138

−136

−134

−132

−130

−128

−126

time (ms)

Filt

ered

RS

RP

(dB

m)

Filtered RSRP measurements with S=2

eNB0−S=2−realeNB1−S=2−realeNB0−S=2−smoothedeNB1−S=2−smoothed

Serving cell

Target cell

Handover is triggered

0 50 100 150 200 250 300−138

−136

−134

−132

−130

−128

−126

time (ms)

Filt

ered

RS

RP

(dB

m)

Filtered RSRP measurements with S=5

eNB0−S=5−realeNB1−S=5−realeNB0−S=5−smoothedeNB1−S=5−smoothed

Target cellServing cell

Handover is triggered

Handover criterion:

r dBmn [k] + S > r dBm

s [k],

where S = ofn + ocn − ofs − ocs − hys − off . (HO parameters)

Handover triggering: S > 0: high signaling overheadS < 0: high probability of failure

Konstantinos Alexandris 10 / 52

OAI X2 HO network measurements

10

15

20

25

30

35

40

45

50

55

60

UE (connected to idle) UE (idle to connected)

X2 handover measurements

Del

ay (

ms)

Handover delay:

DelayHO = TBefore HO + THO Preparation + THO Execution + THO Completion + TMargin︸ ︷︷ ︸detach time ≤ 65ms

Remarks:

Measured detach time adheres to the 3GPP standards

Konstantinos Alexandris 11 / 52

RF X2 Handover experimentation in OAI

X Real-world OAI X2 handover RF testbed measurements!

Konstantinos Alexandris 12 / 52

OAI RF X2 HO UE measurements (1/2)

Network information and RSRP measurements before/after the HO process

Konstantinos Alexandris 13 / 52

OAI RF X2 HO network measurements (2/2)

UE

MovingeNB 0

eNB 1

Different type of experiments:

Exp. Param. Value Param. Value#1 prach config index 0 Ch. atten. Ch0: 12, Ch1:16#2 prach config index 14 Ch. atten. Ch0: 12, Ch1:16#3 prach config index 0 Ch. atten. Ch0: 12, Ch1:10#4 prach config index 14 Ch. atten. Ch0: 12, Ch1:10

X2 handover measurements

1 2 3 4

Experiments

0

20

40

60

80

100

120

Del

ay (

ms)

UE (connected to idle)

UE (idle to connected)

Remarks:

prach config index=14 reduces delay: more chances to PRACH detection byeNB (Exp. 1/2 & Exp. 3/4)

Interference in DL is varying: additional delay to detect RRC/RAR messages

Channel attenuation, i.e., Ch0>Ch1 vs Ch1>Ch0 (Exp. 1/3 & Exp. 2/4)

Measured detach time > 65ms: Optimization is needed for 3GPP compliance

Konstantinos Alexandris 14 / 52

Take away messages

� Handover is an “expensive” process → Delay cost

� UE synchronization overwhelms the network in terms of delay

� Contention-free preamble process can be used to help inreducing such latency in the network

� S-like offsets can impact the HO triggering

� Such offsets can be expressed as functions of different metricsprovided by SDN-like mobility mgt schemes

Konstantinos Alexandris 15 / 52

Load-aware Handover Decisionalgorithm in Next-generation HetNets

Konstantinos Alexandris 16 / 52

Handover (HO) in HetNets

3GPP X2 HO applies a conventionalRSS-based algorithm

Prior art:

Beyond 3GPP legacyconventionalities

Other methods based on:

1 speed2 distance3 interference mgt

What about asymmetrical transmissionpower environments in HetNets?

3GPP Rel.9 introduces SCs asHeNBs

Goals and Challenges:

Macrocell (MC) BSs deteriorate thearea in terms of power

RSS-based HO: UE stays connectedto overloaded MCs whileunderloaded picocells are around

Scale-down power factors have been

proposed based on distance criteria:

No consideration of QoSmetrics (i.e., delay,throughput)

Konstantinos Alexandris 17 / 52

Handover (HO) in HetNets

3GPP X2 HO applies a conventionalRSS-based algorithm

Prior art:

Beyond 3GPP legacyconventionalities

Other methods based on:

1 speed2 distance3 interference mgt

What about asymmetrical transmissionpower environments in HetNets?

3GPP Rel.9 introduces SCs asHeNBs

Goals and Challenges:

Macrocell (MC) BSs deteriorate thearea in terms of power

RSS-based HO: UE stays connectedto overloaded MCs whileunderloaded picocells are around

Scale-down power factors have been

proposed based on distance criteria:

No consideration of QoSmetrics (i.e., delay,throughput)

Konstantinos Alexandris 17 / 52

Handover (HO) in HetNets

3GPP X2 HO applies a conventionalRSS-based algorithm

Prior art:

Beyond 3GPP legacyconventionalities

Other methods based on:

1 speed2 distance3 interference mgt

What about asymmetrical transmissionpower environments in HetNets?

3GPP Rel.9 introduces SCs asHeNBs

Goals and Challenges:

Macrocell (MC) BSs deteriorate thearea in terms of power

RSS-based HO: UE stays connectedto overloaded MCs whileunderloaded picocells are around

Scale-down power factors have been

proposed based on distance criteria:

No consideration of QoSmetrics (i.e., delay,throughput)

Konstantinos Alexandris 17 / 52

Handover (HO) in HetNets

3GPP X2 HO applies a conventionalRSS-based algorithm

Prior art:

Beyond 3GPP legacyconventionalities

Other methods based on:

1 speed2 distance3 interference mgt

What about asymmetrical transmissionpower environments in HetNets?

3GPP Rel.9 introduces SCs asHeNBs

Goals and Challenges:

Macrocell (MC) BSs deteriorate thearea in terms of power

RSS-based HO: UE stays connectedto overloaded MCs whileunderloaded picocells are around

Scale-down power factors have been

proposed based on distance criteria:

No consideration of QoSmetrics (i.e., delay,throughput)

Konstantinos Alexandris 17 / 52

System assumptions

Air-interface model

RSRP signal:

The signal at time k is:

rdBmi [k] , PdBm

Tx ,i︸ ︷︷ ︸Tx power

+ PdBL,i (d

ki )︸ ︷︷ ︸

Pathloss

+ ψdBi [k]︸ ︷︷ ︸

Shadowing

where i ∈ {m, pj} and ψdBi [k] ∼ N

(0, σ2

dB,i/ξ2)

.

EMA (L3) filtering:

The output signal is:

rdBmi [k] , (1− α)rdBm

i [k − 1] + αrdBmi [k]

where α , 2−q/2 and q ∈ N.

Traffic model

Network users:

a) Static users (SU):

Active users (AU): on-going traffic

Disconnected (DU): switched-off

b) Mobile users (MU):

Always active

M/G/1/PS system is adopted

non-GBR traffic/flow size with mean Y

New flows total arrival rate is:

λki = λ

(Nk

MU,i + NkAU,i

)

Konstantinos Alexandris 18 / 52

Load-aware (LA) decision algorithm

Input: Dkm, Dk

pj, rp,th

Output: user cell association

Proposed algorithm:

j∗ = arg maxj

rpj

if(rpj∗ [k] > rp,th

)then

if rpj∗ [k] > f(Dk

m, Dkpj∗

)rm[k]

thenconnect to picocell

elseconnect to macrocell

end ifend if

Assuming M/G/1/PS ⇒

Predicted avg delay: Dki =

1

µki − λk

i

,

for a service rate µki .

Konstantinos Alexandris 19 / 52

Load-aware (LA) decision algorithm

Input: Dkm, Dk

pj, rp,th

Output: user cell association

Proposed algorithm:

j∗ = arg maxj

rpj

if(rpj∗ [k] > rp,th

)then

if rpj∗ [k] > f(Dk

m, Dkpj∗

)rm[k]

thenconnect to picocell

elseconnect to macrocell

end ifend if

Assuming M/G/1/PS ⇒

Predicted avg delay: Dki =

1

µki − λk

i

,

for a service rate µki .

The handover rule is:

rpj∗ [k] > f(Dk

m, Dkpj∗

)rm[k]

Konstantinos Alexandris 19 / 52

Load-aware (LA) decision algorithm

Input: Dkm, Dk

pj, rp,th

Output: user cell association

Proposed algorithm:

j∗ = arg maxj

rpj

if(rpj∗ [k] > rp,th

)then

if rpj∗ [k] > f(Dk

m, Dkpj∗

)rm[k]

thenconnect to picocell

elseconnect to macrocell

end ifend if

Assuming M/G/1/PS ⇒

Predicted avg delay: Dki =

1

µki − λk

i

,

for a service rate µki .

The handover rule is:

rpj∗ [k] > f(Dk

m, Dkpj∗

)rm[k]

Exponential family is selected for f (·):

f(Dk

m, Dkpj

), exp

[−cDk

m/Dkpj

], c ≥ 1

Konstantinos Alexandris 19 / 52

Load-aware (LA) decision algorithm

Input: Dkm, Dk

pj, rp,th

Output: user cell association

Proposed algorithm:

j∗ = arg maxj

rpj

if(rpj∗ [k] > rp,th

)then

if rpj∗ [k] > f(Dk

m, Dkpj∗

)rm[k]

thenconnect to picocell

elseconnect to macrocell

end ifend if

Assuming M/G/1/PS ⇒

Predicted avg delay: Dki =

1

µki − λk

i

,

for a service rate µki .

The handover rule is:

rpj∗ [k] > f(Dk

m, Dkpj∗

)rm[k]

The f (·) ∈ [0, 1] so as to force the UEto connect in the picocell, iff:

Dkm � Dk

pj∗

Konstantinos Alexandris 19 / 52

Simulations (1/2)

0.01 0.02 0.04 0.06 0.08 0.1 0.120

0.1

0.2

0.3

0.4

0.5

λ

Averagedelay(sec)

Average delay for different NSU,m and NSU,pj = 10

Dp

Dm

NSU,m = 200

NSU,m = 400NSU,m = 600

0.01 0.02 0.04 0.06 0.08 0.1 0.120

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

λ

Cellassignmen

tprobability

Algorithms performance for different NSU,m and NSU,pj = 10

LA-Pr(u ∈ p)

LA-Pr(u ∈ m)

CONV-Pr(u ∈ p)

CONV-Pr(u ∈ m)

NSU,m = 400

NSU,m = 200

NSU,m = 600

Assumption: 1 mobile user (moves in 2D-RW) and 3 picocells

Remarks:

Comparison with the conventional (RSS-based) HO algorithm:

r dBmpj [k] > r dBm

m [k] + ∆ ∧ r dBmm [k] < r dBm

m,th

As the average delay gets sharper, Pr(u ∈ p)→ 1

Conventional HO keeps UE connected to the overloaded MC

Konstantinos Alexandris 20 / 52

Simulations (2/2)

50 60 80 100 120 140 160 1800

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

dm(m)

Picocellassignmen

tprobability

Algorithms performance for NSU,m = 200 and NSU,pj = 10

LA-λ = 0.05LA-λ = 0.08LA-λ = 0.1LA-λ = 0.11LA-λ = 0.12DISTCONV

λ 0.05 0.08 0.1 0.11 0.12

¯DDIST/¯DLA 0.8861 0.8141 0.8798 1.3066 4.5235

¯DCONV/¯DLA 1.0596 1.1701 1.4790 2.4073 8.8547

Load ↑: DIST HO: ∼ 4CONV HO: ∼ 8

Assumption: 1 mobile user (moves across a line) and 1 picocell

Remarks:

Distance-based HO (DIST HO):Up to a distance threshold, UE is connected tooverloaded MC (Load ↑-Delay ↑)Load ↑: LA HO associates faster the UE with the picocell (Delay ↓)Load ↓: LA HO keeps UE to MC compared to the DIST HO

Konstantinos Alexandris 21 / 52

Take away messages

� LA algorithm overcomes the HetNets power assymetries

� Such algorithm considers both the user (i.e., RSS) andnetwork (i.e., service delay) perspectives

� Conventional RSS and distance-based HO algorithms areinferior to the proposed scheme, esp. in high MC load

Konstantinos Alexandris 22 / 52

Multi-connectivity resource allocationin evolved LTE

Konstantinos Alexandris 23 / 52

Multi-connectivity in evolved LTE

Beyond HO: Seamless mobility

Single vs Multi-connectivity:

3GPP TS36.331: Legacy SC

Bandwidth and connectivityconstraintsOnly one RATNo seamless mobilitySet bounds on users QoS

Multi-connectivity framework

Wi-Fi

5G

4G LTE

Across any access node

Simultaneous connections in several technologies

Across Radio Access

Technologies

What about multi-connectivity in evolvedLTE?

X Component of 5G New Radio: DCextension (3GPP TS36.842)

X Multiple cell connections: single ormulti-RAT

X Several bands support

Goals & Challenges:

Optimized capacity coverage

Reliable high-speed data delivery

Effective resource utilization

Contrary to prior-art:

Users QoS requirements are

considered

Konstantinos Alexandris 24 / 52

Multi-connectivity in evolved LTE

Beyond HO: Seamless mobility

Single vs Multi-connectivity:

3GPP TS36.331: Legacy SC

Bandwidth and connectivityconstraintsOnly one RATNo seamless mobilitySet bounds on users QoS

Multi-connectivity framework

Wi-Fi

5G

4G LTE

Across any access node

Simultaneous connections in several technologies

Across Radio Access

Technologies

What about multi-connectivity in evolvedLTE?

X Component of 5G New Radio: DCextension (3GPP TS36.842)

X Multiple cell connections: single ormulti-RAT

X Several bands support

Goals & Challenges:

Optimized capacity coverage

Reliable high-speed data delivery

Effective resource utilization

Contrary to prior-art:

Users QoS requirements are

considered

Konstantinos Alexandris 24 / 52

Multi-connectivity in evolved LTE

Beyond HO: Seamless mobility

Single vs Multi-connectivity:

3GPP TS36.331: Legacy SC

Bandwidth and connectivityconstraintsOnly one RATNo seamless mobilitySet bounds on users QoS

Multi-connectivity framework

Wi-Fi

5G

4G LTE

Across any access node

Simultaneous connections in several technologies

Across Radio Access

Technologies

What about multi-connectivity in evolvedLTE?

X Component of 5G New Radio: DCextension (3GPP TS36.842)

X Multiple cell connections: single ormulti-RAT

X Several bands support

Goals & Challenges:

Optimized capacity coverage

Reliable high-speed data delivery

Effective resource utilization

Contrary to prior-art:

Users QoS requirements are

considered

Konstantinos Alexandris 24 / 52

Multi-connectivity in evolved LTE

Beyond HO: Seamless mobility

Single vs Multi-connectivity:

3GPP TS36.331: Legacy SC

Bandwidth and connectivityconstraintsOnly one RATNo seamless mobilitySet bounds on users QoS

Multi-connectivity framework

Wi-Fi

5G

4G LTE

Across any access node

Simultaneous connections in several technologies

Across Radio Access

Technologies

What about multi-connectivity in evolvedLTE?

X Component of 5G New Radio: DCextension (3GPP TS36.842)

X Multiple cell connections: single ormulti-RAT

X Several bands support

Goals & Challenges:

Optimized capacity coverage

Reliable high-speed data delivery

Effective resource utilization

Contrary to prior-art:

Users QoS requirements are

considered

Konstantinos Alexandris 24 / 52

System assumptions

Air-interface

Downlink (DL):

Physical data rate:

RDbj ,ui

= BDbj︸︷︷︸

max #PRBs

· WDbj︸︷︷︸

BW per PRB

· log2(1 + SINRDbj ,ui︸ ︷︷ ︸

DL SINR

)

DL SINR based on:

RSRPDbj ,ui

: Pathloss + Shadowing + Antenna gain

Uplink (UL):

Same definitions hold for uplink + Power control

PU,dBmui ,bj

= min( Pmax,dBmui︸ ︷︷ ︸

max UE power

, PdBm0 + α · LU,dB

ui ,bj︸ ︷︷ ︸pathloss

)

a, P0: power control parameters

BSs

UEsu1

u3

u2

u4

b1 b3b2

Connection and Traffic

Multi-connectivity: UL/DL LTE FDD SISO

UE association: min(SINRUui ,bj

, SINRDbj ,ui

) >

threshold︷ ︸︸ ︷SINRth

Active UEs: Mobile and connected to multiple cells

Konstantinos Alexandris 25 / 52

System assumptions

Air-interface

Downlink (DL):

Physical data rate:

RDbj ,ui

= BDbj︸︷︷︸

max #PRBs

· WDbj︸︷︷︸

BW per PRB

· log2(1 + SINRDbj ,ui︸ ︷︷ ︸

DL SINR

)

DL SINR based on:

RSRPDbj ,ui

: Pathloss + Shadowing + Antenna gain

Uplink (UL):

Same definitions hold for uplink + Power control

PU,dBmui ,bj

= min( Pmax,dBmui︸ ︷︷ ︸

max UE power

, PdBm0 + α · LU,dB

ui ,bj︸ ︷︷ ︸pathloss

)

a, P0: power control parameters

Connection and Traffic

Multi-connectivity: UL/DL LTE FDD SISO

UE association: min(SINRUui ,bj

, SINRDbj ,ui

) >

threshold︷ ︸︸ ︷SINRth

Active UEs: Mobile and connected to multiple cells

BSs

UEsu1

u3

u2

u4

b1 b3b2

U2U

Traffic type: User-to-user (U2U) traffic

Traffic routing: Locally routed via a common BS (2-hop)

2-hop routing offloads backhaul (EPC)

Public safety networks isolated BSs

Close community (U2U) apps: video sharing

Traffic requested rate: Rui ,uqE2E app target rate

Konstantinos Alexandris 25 / 52

Resource allocation

xU or Dbj ,(ui ,uq)

: Pct. of allocated PRBs

Utility functions:

PF: Proportional fairness

Φ (x) = log (x)

UPF: PF + QoS (R target rate)

Φ (x) = log

S(x,γ,R)︷ ︸︸ ︷(1

1 + e−γ(x−R)

)

γ impacts the shape of sigmoid

Non-linear: users perspective (QoS)

Linear: network perspective

0 1 2 3 4 5 6 7 8 9 10

106

0

0.2

0.4

0.6

0.8

1

Objective function:

U(

xUui ,uq

, xDui ,uq

),

Φ

∑bj∈B

Q

(xUbj ,(ui ,uq)

, xDbj ,(ui ,uq)

)where

Q

(xUbj ,(ui ,uq)

, xDbj ,(ui ,uq)

),

min

(xUbj ,(ui ,uq)

RUui ,bj

, xDbj ,(ui ,uq)

RDbj ,uq

)

Konstantinos Alexandris 26 / 52

Problem formulation

NUM problem is defined as:

maxX

∑(ui ,uq )∈C

U(

xUui ,uq

, xDui ,uq

)

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1-C2: maximum PRBs at BS (UL/DL)

C3-C4: maximum PRBs at UE (UL/DL)

C5: Power control in UL

Objective is non-differentiable!

Contains the min (·) function

Concave but non-differentiable

Transformation needs to be applied

The problem is shown to be convex

Konstantinos Alexandris 27 / 52

Problem formulation

NUM problem is defined as:

maxX

∑(ui ,uq )∈C

U(

xUui ,uq

, xDui ,uq

)s.t.

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1-C2: maximum PRBs at BS (UL/DL)

C3-C4: maximum PRBs at UE (UL/DL)

C5: Power control in UL

Objective is non-differentiable!

Contains the min (·) function

Concave but non-differentiable

Transformation needs to be applied

The problem is shown to be convex

Konstantinos Alexandris 27 / 52

Problem formulation

NUM problem is defined as:

maxX

∑(ui ,uq )∈C

U(

xUui ,uq

, xDui ,uq

)

s.t. C1:∑

(ui ,uq )∈Cbj

xUbj ,(ui ,uq) ≤ 1, ∀ bj ,

C2:∑

(ui ,uq )∈Cbj

xDbj ,(ui ,uq) ≤ 1, ∀ bj ,

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1-C2: maximum PRBs at BS (UL/DL)

C3-C4: maximum PRBs at UE (UL/DL)

C5: Power control in UL

Objective is non-differentiable!

Contains the min (·) function

Concave but non-differentiable

Transformation needs to be applied

The problem is shown to be convex

Konstantinos Alexandris 27 / 52

Problem formulation

NUM problem is defined as:

maxX

∑(ui ,uq )∈C

U(

xUui ,uq

, xDui ,uq

)

s.t. C1:∑

(ui ,uq )∈Cbj

xUbj ,(ui ,uq) ≤ 1, ∀ bj ,

C2:∑

(ui ,uq )∈Cbj

xDbj ,(ui ,uq) ≤ 1, ∀ bj ,

C3:∑bj∈B

∑uq∈Dui

xUbj ,(ui ,uq)B

Ubj≤ BU

ui, ∀ ui ,

C4:∑bj∈B

∑uq∈Sui

xDbj ,(uq ,ui )

BDbj≤ BD

ui, ∀ ui ,

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1-C2: maximum PRBs at BS (UL/DL)

C3-C4: maximum PRBs at UE (UL/DL)

C5: Power control in UL

Objective is non-differentiable!

Contains the min (·) function

Concave but non-differentiable

Transformation needs to be applied

The problem is shown to be convex

Konstantinos Alexandris 27 / 52

Problem formulation

NUM problem is defined as:

maxX

∑(ui ,uq )∈C

U(

xUui ,uq

, xDui ,uq

)

s.t. C1:∑

(ui ,uq )∈Cbj

xUbj ,(ui ,uq) ≤ 1, ∀ bj ,

C2:∑

(ui ,uq )∈Cbj

xDbj ,(ui ,uq) ≤ 1, ∀ bj ,

C3:∑bj∈B

∑uq∈Dui

xUbj ,(ui ,uq)B

Ubj≤ BU

ui, ∀ ui ,

C4:∑bj∈B

∑uq∈Sui

xDbj ,(uq ,ui )

BDbj≤ BD

ui, ∀ ui ,

C5:∑bj∈B

∑uq∈Dui

xUbj ,(ui ,uq)P

Uui ,bj

BUbj≤ Pmax

ui, ∀ ui .

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1-C2: maximum PRBs at BS (UL/DL)

C3-C4: maximum PRBs at UE (UL/DL)

C5: Power control in UL

Objective is non-differentiable!

Contains the min (·) function

Concave but non-differentiable

Transformation needs to be applied

The problem is shown to be convex

Konstantinos Alexandris 27 / 52

Problem formulation

NUM problem is defined as:

maxX

∑(ui ,uq )∈C

U(

xUui ,uq

, xDui ,uq

)

s.t. C1:∑

(ui ,uq )∈Cbj

xUbj ,(ui ,uq) ≤ 1, ∀ bj ,

C2:∑

(ui ,uq )∈Cbj

xDbj ,(ui ,uq) ≤ 1, ∀ bj ,

C3:∑bj∈B

∑uq∈Dui

xUbj ,(ui ,uq)B

Ubj≤ BU

ui, ∀ ui ,

C4:∑bj∈B

∑uq∈Sui

xDbj ,(uq ,ui )

BDbj≤ BD

ui, ∀ ui ,

C5:∑bj∈B

∑uq∈Dui

xUbj ,(ui ,uq)P

Uui ,bj

BUbj≤ Pmax

ui, ∀ ui .

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1-C2: maximum PRBs at BS (UL/DL)

C3-C4: maximum PRBs at UE (UL/DL)

C5: Power control in UL

Objective is non-differentiable!

Contains the min (·) function

Concave but non-differentiable

Transformation needs to be applied

The problem is shown to be convex

Konstantinos Alexandris 27 / 52

Simulations (1/3): Single vs Multi-connectivity

BSs

UEsu1

u3

u2

u4

b1 b3b2

Performance metric

User pairs aggregated rate:

ζui ,uq ,∑bj∈B

Q

(x?Ubj ,(ui ,uq), x

?Dbj ,(ui ,uq)

)

where

x?Ubj ,(ui ,uq)

, x?Dbj ,(ui ,uq)

: optimal for UL/DL

PF for different load cases (#UEs)

Comparison of Single/Multi-connectivityUE number Performance Single- Multi-in BS z/z/z metric connected connected

Connected BS 1 2.07Under-loaded Connected UE pairs 6 17.52case: 2/2/2 Aggregated user rate 0.99 Mbps 20.04 Mbps

Connected BS 1 1.34Uneven-loaded Connected UE pairs 34 49.95

case: 6/2/2 Aggregated user rate 11.68 Mbps 46.73 MbpsConnected BS 1 1.45

Over-loaded Connected UE pairs 90 162.22case: 6/6/6 Aggregated user rate 55.04 Mbps 57.01 Mbps

Remarks:

X Connectivity increases with the loaddecrement

X Higher number of UE pairs: trafficdiversity

X Gain in aggregated rate

X Advantages to both user andnetwork perspective

Konstantinos Alexandris 28 / 52

Simulations (2/3): Performance analysis of PF & UPF

Performance metric

Satisfaction ratio:

Mui ,uq= Prob

{ζui ,uq ≥ Rui ,uq

}.

Unsatisfied normalized error:

Eui ,uq =

∥∥∥∥∥ ζui ,uq − Rui ,uq

Rui ,uq

∥∥∥∥∥ , if ζui ,uq < Rui ,uq,

0 , o/w.

PF vs UPF: A trade-off between networkaggregated rate and users satisfaction

PF provides fairness and maximizesthe network aggregated rate

UPF extends PF and takes intoaccount the users QoS

Multi-connectivity case with 4 UEs/BS

QoS metrics comparison of PF and UPFMetric Requested rate PF problem UPF problem

0.1Mbps 68.72% 91.39%0.5Mbps 42.07% 58.03%

Satisfaction 1Mbps 25.15% 35.33%ratio 5Mbps 6.42% 23.10%

10Mbps < 1% <1%0.1Mbps 0.2122 0.0625

Unsatisfied 0.5Mbps 0.4131 0.2317normalized 1Mbps 0.5483 0.3906

error 5Mbps 0.7949 0.660710Mbps 0.8833 0.8441

Remarks:

UPF redistributes the resources toboost the user pairs satisfaction

Both of QoS metrics are decreasingwith the requested rate increment

Even in high R, UPF satisfies moreuser pairs (constrained by resources)

Konstantinos Alexandris 29 / 52

Simulations (3/3): Impact on γ on UPF

PDF plot of ζ for the UPF utility function with several γ

Remarks:

Sigmoid function approximates the step function (in UPF) when γ ↑Linear (PF case) vs Step (ideal UPF case) → (more/less) Network aggregatedrate vs (worse/better) user pairs QoS

Consequently:

X QoS requirements are more fulfilled with the increment of γ

X PDF tail decrement: less network aggregated rate-better users QoS

Konstantinos Alexandris 30 / 52

Take away messages

� Multi-connectivity outperforms the legacy single-connectivity

� Multi-connectivity can boost the aggregated rate, especially inunder-loaded scenarios

� UPF can increase the users satisfaction ratio when they areavailable network resources compared to PF

� Operators can impact the users network performance via UPFfunction shape (network vs user perspective)

Konstantinos Alexandris 31 / 52

Multi-connectivity resource allocationwith limited backhaul capacity in

evolved LTE

Konstantinos Alexandris 32 / 52

System model-Resource allocation

Air-interface

Uplink/Downlink (UL/DL):

Carrier frequency: Inter-frequency deployment

SINR: UL/DL based on RSRP

UL: Power control

Connection and Traffic

Multi-connectivity: UL/DL LTE FDD SISO

UE association: min(SINRUi,j , SINRD

j,i ) >

threshold︷ ︸︸ ︷SINRth

Active UEs: Connected to multiple cells

BSs

UEsu1

u3

u2

u4

b1 b3b2

Core network

Internet

Backhaul

links

xU or Dj,q : Pct. of allocated PRBs

Utility functions: PF + UPF

Objective functions:

UL:

U1

(xUi

), Φ

∑bj∈B

xUi,jR

Ui,j

DL:U2

(xDq

), Φ

∑bj∈B

xDj,qR

Dj,q

Konstantinos Alexandris 33 / 52

System model-Resource allocation

Air-interface

Uplink/Downlink (UL/DL):

Carrier frequency: Inter-frequency deployment

SINR: UL/DL based on RSRP

UL: Power control

Connection and Traffic

Multi-connectivity: UL/DL LTE FDD SISO

UE association: min(SINRUi,j , SINRD

j,i ) >

threshold︷ ︸︸ ︷SINRth

Active UEs: Connected to multiple cells

Traffic type: From/to (UL/DL) remote server traffic

Backhaul network: Star topology

Traffic requested rate: Ri app target rate

BSs

UEsu1

u3

u2

u4

b1 b3b2

Core network

Internet

Backhaul links

UL/DL

xU or Dj,q : Pct. of allocated PRBs

Utility functions: PF + UPF

Objective functions:

UL:

U1

(xUi

), Φ

∑bj∈B

xUi,jR

Ui,j

DL:U2

(xDq

), Φ

∑bj∈B

xDj,qR

Dj,q

Konstantinos Alexandris 33 / 52

Problem formulation

NUM problem is defined as:

maxX

∑ui∈S

U1

(xUi

)+∑

uq∈DU2

(xDq

)

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1-C2: maximum PRBs at BS (UL/DL)

C3-C4: maximum PRBs at UE (UL/DL)

C5: Power control in UL

C6-C7: maximum BH capacity (UL/DL)

Objective is differentiable!

Concave function + Linear constraints

Convex problem→Interior point method

Konstantinos Alexandris 34 / 52

Problem formulation

NUM problem is defined as:

maxX

∑ui∈S

U1

(xUi

)+∑

uq∈DU2

(xDq

)s.t.

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1-C2: maximum PRBs at BS (UL/DL)

C3-C4: maximum PRBs at UE (UL/DL)

C5: Power control in UL

C6-C7: maximum BH capacity (UL/DL)

Objective is differentiable!

Concave function + Linear constraints

Convex problem→Interior point method

Konstantinos Alexandris 34 / 52

Problem formulation

NUM problem is defined as:

maxX

∑ui∈S

U1

(xUi

)+∑

uq∈DU2

(xDq

)s.t. C1:

∑ui∈S,

(ui ,bj

)∈E

xUi,j ≤ 1, ∀ bj ∈ B,

C2:∑

uq∈D,(bj ,uq

)∈E

xDj,q ≤ 1, ∀ bj ∈ B,

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1-C2: maximum PRBs at BS (UL/DL)

C3-C4: maximum PRBs at UE (UL/DL)

C5: Power control in UL

C6-C7: maximum BH capacity (UL/DL)

Objective is differentiable!

Concave function + Linear constraints

Convex problem→Interior point method

Konstantinos Alexandris 34 / 52

Problem formulation

NUM problem is defined as:

maxX

∑ui∈S

U1

(xUi

)+∑

uq∈DU2

(xDq

)s.t. C1:

∑ui∈S,

(ui ,bj

)∈E

xUi,j ≤ 1, ∀ bj ∈ B,

C2:∑

uq∈D,(bj ,uq

)∈E

xDj,q ≤ 1, ∀ bj ∈ B,

C3:∑

bj∈B,(ui ,bj

)∈E

xUi,jB

Uj ≤ MU

i , ∀ ui ∈ U,

C4:∑

bj∈B,(bj ,uq

)∈E

xDj,qB

Dj ≤ MD

q , ∀ uq ∈ U,

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1-C2: maximum PRBs at BS (UL/DL)

C3-C4: maximum PRBs at UE (UL/DL)

C5: Power control in UL

C6-C7: maximum BH capacity (UL/DL)

Objective is differentiable!

Concave function + Linear constraints

Convex problem→Interior point method

Konstantinos Alexandris 34 / 52

Problem formulation

NUM problem is defined as:

maxX

∑ui∈S

U1

(xUi

)+∑

uq∈DU2

(xDq

)s.t. C1:

∑ui∈S,

(ui ,bj

)∈E

xUi,j ≤ 1, ∀ bj ∈ B,

C2:∑

uq∈D,(bj ,uq

)∈E

xDj,q ≤ 1, ∀ bj ∈ B,

C3:∑

bj∈B,(ui ,bj

)∈E

xUi,jB

Uj ≤ MU

i , ∀ ui ∈ U,

C4:∑

bj∈B,(bj ,uq

)∈E

xDj,qB

Dj ≤ MD

q , ∀ uq ∈ U,

C5:∑

bj∈B,(ui ,bj

)∈E

xUi,jB

Uj PU

i,j ≤ Pmaxi , ∀ ui ∈ U,

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1-C2: maximum PRBs at BS (UL/DL)

C3-C4: maximum PRBs at UE (UL/DL)

C5: Power control in UL

C6-C7: maximum BH capacity (UL/DL)

Objective is differentiable!

Concave function + Linear constraints

Convex problem→Interior point method

Konstantinos Alexandris 34 / 52

Problem formulation

NUM problem is defined as:

maxX

∑ui∈S

U1

(xUi

)+∑

uq∈DU2

(xDq

)s.t. C1:

∑ui∈S,

(ui ,bj

)∈E

xUi,j ≤ 1, ∀ bj ∈ B,

C2:∑

uq∈D,(bj ,uq

)∈E

xDj,q ≤ 1, ∀ bj ∈ B,

C3:∑

bj∈B,(ui ,bj

)∈E

xUi,jB

Uj ≤ MU

i , ∀ ui ∈ U,

C4:∑

bj∈B,(bj ,uq

)∈E

xDj,qB

Dj ≤ MD

q , ∀ uq ∈ U,

C5:∑

bj∈B,(ui ,bj

)∈E

xUi,jB

Uj PU

i,j ≤ Pmaxi , ∀ ui ∈ U,

C6:∑

ui∈S,(ui ,bj

)∈E

xUi,jR

Ui,j ≤ CU

h,j , ∀ bj ∈ B,

C7:∑

uq∈D,(bj ,uq

)∈E

xDj,qR

Dj,q ≤ CD

h,j , ∀ bj ∈ B.

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1-C2: maximum PRBs at BS (UL/DL)

C3-C4: maximum PRBs at UE (UL/DL)

C5: Power control in UL

C6-C7: maximum BH capacity (UL/DL)

Objective is differentiable!

Concave function + Linear constraints

Convex problem→Interior point method

Konstantinos Alexandris 34 / 52

Problem formulation

NUM problem is defined as:

maxX

∑ui∈S

U1

(xUi

)+∑

uq∈DU2

(xDq

)s.t. C1:

∑ui∈S,

(ui ,bj

)∈E

xUi,j ≤ 1, ∀ bj ∈ B,

C2:∑

uq∈D,(bj ,uq

)∈E

xDj,q ≤ 1, ∀ bj ∈ B,

C3:∑

bj∈B,(ui ,bj

)∈E

xUi,jB

Uj ≤ MU

i , ∀ ui ∈ U,

C4:∑

bj∈B,(bj ,uq

)∈E

xDj,qB

Dj ≤ MD

q , ∀ uq ∈ U,

C5:∑

bj∈B,(ui ,bj

)∈E

xUi,jB

Uj PU

i,j ≤ Pmaxi , ∀ ui ∈ U,

C6:∑

ui∈S,(ui ,bj

)∈E

xUi,jR

Ui,j ≤ CU

h,j , ∀ bj ∈ B,

C7:∑

uq∈D,(bj ,uq

)∈E

xDj,qR

Dj,q ≤ CD

h,j , ∀ bj ∈ B.

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1-C2: maximum PRBs at BS (UL/DL)

C3-C4: maximum PRBs at UE (UL/DL)

C5: Power control in UL

C6-C7: maximum BH capacity (UL/DL)

Objective is differentiable!

Concave function + Linear constraints

Convex problem→Interior point method

Konstantinos Alexandris 34 / 52

Simulations (1/5): Performance metrics

Simulation scenarios

3GPP TR 25.927:

Scenario A-Empty cell: 0/z/z

BSs

UEsz

b1 b3b2

Core network

Internet

Backhaul

links

more

resourcesz

Konstantinos Alexandris 35 / 52

Simulations (1/5): Performance metrics

Simulation scenarios

3GPP TR 25.927:

Scenario A-Empty cell: 0/z/zScenario B-Loaded cell: z/z/z

BSs

UEsz

b1 b3b2

Core network

Internet

Backhaul

links

zz

Konstantinos Alexandris 35 / 52

Simulations (1/5): Performance metrics

Simulation scenarios

3GPP TR 25.927:

Scenario A-Empty cell: 0/z/zScenario B-Loaded cell: z/z/z

BSs

UEsz

b1 b3b2

Core network

Internet

Backhaul

links

zz

Performance metrics

Network aggregated rate:

RD ,∑

uq∈D

∑bj∈B

(xDj,q

)?RDj,q

User satisfaction ratio:

SD , Prob

∑bj∈B

(xDj,q

)?RDj,q ≥ RD

q

where x?D

j,q : optimal for DL (id. in UL)

Konstantinos Alexandris 35 / 52

Simulations (1/5): Performance metrics

Simulation scenarios

3GPP TR 25.927:

Scenario A-Empty cell: 0/z/zScenario B-Loaded cell: z/z/z

BSs

UEsz

b1 b3b2

Core network

Internet

Backhaul

links

zz

Performance metrics

Network aggregated rate:

RD ,∑

uq∈D

∑bj∈B

(xDj,q

)?RDj,q

User satisfaction ratio:

SD , Prob

∑bj∈B

(xDj,q

)?RDj,q ≥ RD

q

where x?D

j,q : optimal for DL (id. in UL)

Use cases:

, Infinite BH

/ Finite BH

Remarks:

X Single vs Multi cells: Cell diversity

X BH limitations

Konstantinos Alexandris 35 / 52

Simulations (2/5): Infinite BH-0/z/z Scenario

Single-PF-0/2/2 Multi-PF-0/2/2 Single-UPF-0/2/2 Multi-UPF-0/2/2400500600700800900

Dat

a ra

te (

Mbp

s) DL aggregated rate comparison of Scenario A

Single-PF-0/4/4 Multi-PF-0/4/4 Single-UPF-0/4/4 Multi-UPF-0/4/4400500600700800900

Dat

a ra

te (

Mbp

s) DL aggregated rate comparison of Scenario A

Single-PF-0/2/2 Multi-PF-0/2/2 Single-UPF-0/2/2 Multi-UPF-0/2/20

20406080

100

Per

cent

age

(%) DL user satisfaction ratio comparison of Scenario A

Single-PF-0/4/4 Multi-PF-0/4/4 Single-UPF-0/4/4 Multi-UPF-0/4/40

20406080

100

Per

cent

age

(%) DL user satisfaction ratio comparison of Scenario A

Remarks:

Downlink-Params: z =2 or 4 users, CDh,j =∞

Multi vs Single connectivity

Empty cell scenario: Utilizes more resourcesSuperior in network aggregated rate and user satisfaction

Further gain: PF vs UPF

Network aggregated rate: PF outperforms UPF (Aggegated rate boost)User satisfaction: UPF outperforms PF (Targets on users QoS)

Konstantinos Alexandris 36 / 52

Simulations (3/5): Infinite BH-z/z/z Scenario

Single-PF-2/2/2 Multi-PF-2/2/2 Single-UPF-2/2/2 Multi-UPF-2/2/2600

700

800

900

Dat

a ra

te (

Mbp

s) DL aggregated rate comparison of Scenario B

Singel-PF-4/4/4 Multi-PF-4/4/4 Singel-UPF-4/4/4 Multi-UPF-4/4/4650

750

850

950

Dat

a ra

te (

Mbp

s) DL aggregated rate comparison of Scenario B

Single-PF-2/2/2 Multi-PF-2/2/2 Single-UPF-2/2/2 Multi-UPF-2/2/20

20406080

100

Per

cent

age

(%) DL user satisfaction ratio comparison of Scenario B

Single-PF-4/4/4 Multi-PF-4/4/4 Single-UPF-4/4/4 Multi-UPF-4/4/40

20406080

100

Per

cent

age

(%) DL user satisfaction ratio comparison of Scenario B

Remarks:

Downlink-Params: z =2 or 4 users, CDh,j =∞

Multi vs Single connectivity

Loaded cell scenario: No extra resources to exploitNetwork aggregated rate: no such difference

Further gain: PF vs UPF

UPF can still increase the user satisfaction rateQoS aware-Better reshuffling the resources!

Konstantinos Alexandris 37 / 52

Simulations (4/5): Finite BH-0/z/z vs z/z/z Scenario

150 200 250 300 350 400Backhaul Capacity (Mbps)

200

300

400

500

600

700

800

Avg

dat

a ra

te (

Mbp

s)

DL average aggregated rate of Scenario A

Single-PF-0/2/2Multi-PF-0/2/2Single-UPF-0/2/2Multi-UPF-0/2/2

150 200 250 300 350 400Backhaul Capacity (Mbps)

400

500

600

700

800

Avg

dat

a ra

te (

Mbp

s)

DL average aggregated rate of Scenario B

Single-PF-2/2/2Multi-PF-2/2/2Single-UPF-2/2/2Multi-UPF-2/2/2

Remarks:

Downlink-Params: z =2 users, CDh,j <∞

Aggregated rate converges when BH capacity>300 Mbps

0/z/z Scenario: Multi-connectivity outperforms the single one

Even with limited BH (<300 Mbps) ⇒ i.e., more resources are utilized

z/z/z Scenario: Slightly better (Multi vs Single)-No extra resources

PF vs UPF: UPF in Multi performs the worst to satisfy users QoSTrade-off: Network and User perspective

Konstantinos Alexandris 38 / 52

Simulations (5/5): Finite BH-0/z/z vs z/z/z Scenario

DL user satisfaction ratio

Backhaul Single Multi Single MultiCapacity PF PF UPF UPF(Mbps) 0/2/2 2/2/2 0/2/2 2/2/2 0/2/2 2/2/2 0/2/2 2/2/2

150 ∼20 ∼20 63 ∼20 ∼20 ∼20 81 ∼20

200 49 ∼50 95 ∼50 51 54 100 55250 72 65 95 68 83 80 100 91300 81 72 95 73 90 84 100 93>350 81 72 95 73 90 84 100 93

Remarks:

Downlink-Params: z =2 users, CDh,j <∞

Single to Multi: Boosts the user satisfaction: esp. in 0/z/z

Gain in PF: 3-31% (esp. in limited BH)

PF to UPF: Accelerates more the performance: esp. in z/z/z

Additional gain: 8-16% (esp. in limited BH)

Operators can adjust their policy to such trade-offs!

Konstantinos Alexandris 39 / 52

Take away messages

� Multi-cell approach achieves better network aggregated rate inempty cell scenario

� Multi-connectivity boost user satisfaction even in loadedscenarios with limited BH capacity

� In such scenarios, UPF can offer additional gain to users QoSre-utilizing better the network resources

� Network vs User perspective (Network aggregated rate vs UserSatisfaction) trade-off determines the intended policy

Konstantinos Alexandris 40 / 52

Opportunistic scheduling undermulti-connectivity with limited

backhaul capacity

Konstantinos Alexandris 41 / 52

System model-Resource allocation

Air-interface

Downlink (DL):

Carrier frequency: Inter-frequency deployment

SINR: DL based on RSRP

Channel model: Large/Small scale fading

Connection and Traffic

Multi-connectivity: DL LTE FDD SISO

UE association:1

|K|

∑k∈K

SINRDj,k,i >

threshold︷ ︸︸ ︷SINRth

Active UEs: Connected to multiple cells

BSs

UEsu1

u3

u2

u4

b1 b3b2

Core network

Internet

Backhaul

links

Resource allocation

Opportunistic scheduling

Exploit channel variability (fast fading)

Realistic situation: Discrete resource blocks

Formulate an optimization problem with binary

variables (0/1)!

No relaxation applies: Problems in rounding

Konstantinos Alexandris 42 / 52

System model-Resource allocation

Air-interface

Downlink (DL):

Carrier frequency: Inter-frequency deployment

SINR: DL based on RSRP

Channel model: Large/Small scale fading

Connection and Traffic

Multi-connectivity: DL LTE FDD SISO

UE association:1

|K|

∑k∈K

SINRDj,k,i >

threshold︷ ︸︸ ︷SINRth

Active UEs: Connected to multiple cells

Traffic type: From (DL) remote server traffic

Backhaul network: Star topology

Traffic requested rate: Ri app target rate

BSs

UEsu1

u3

u2

u4

b1 b3b2

Core network

Internet

Backhaul links

DL

Resource allocation

Opportunistic scheduling

Exploit channel variability (fast fading)

Realistic situation: Discrete resource blocks

Formulate an optimization problem with binary

variables (0/1)!

No relaxation applies: Problems in rounding

Konstantinos Alexandris 42 / 52

Problem formulation

xDj,k,i ∈ {0, 1}: Binary variable (PRB)

NUM problem is defined as:

maxxDi∈{0,1}|B|×|K|

∑ui∈U

U(

xDi

)

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1: maximum PRBs at BS

C2: maximum PRBs at UE

C3: sub-channel exclusivity at UE

Combinatorial problem difficult to be solved!

Non-linear Integer Programming

Sub-modularity argument

Find efficient algorithms to solve it!

Any performance guarantee?

Konstantinos Alexandris 43 / 52

Problem formulation

xDj,k,i ∈ {0, 1}: Binary variable (PRB)

NUM problem is defined as:

maxxDi∈{0,1}|B|×|K|

∑ui∈U

U(

xDi

)s.t.

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1: maximum PRBs at BS

C2: maximum PRBs at UE

C3: sub-channel exclusivity at UE

Combinatorial problem difficult to be solved!

Non-linear Integer Programming

Sub-modularity argument

Find efficient algorithms to solve it!

Any performance guarantee?

Konstantinos Alexandris 43 / 52

Problem formulation

xDj,k,i ∈ {0, 1}: Binary variable (PRB)

NUM problem is defined as:

maxxDi∈{0,1}|B|×|K|

∑ui∈U

U(

xDi

)s.t.

C1:∑ui∈U

∑k∈K

xDj,k,i ≤ BD

j , ∀ bj ∈ B,

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1: maximum PRBs at BS

C2: maximum PRBs at UE

C3: sub-channel exclusivity at UE

Combinatorial problem difficult to be solved!

Non-linear Integer Programming

Sub-modularity argument

Find efficient algorithms to solve it!

Any performance guarantee?

Konstantinos Alexandris 43 / 52

Problem formulation

xDj,k,i ∈ {0, 1}: Binary variable (PRB)

NUM problem is defined as:

maxxDi∈{0,1}|B|×|K|

∑ui∈U

U(

xDi

)s.t.

C1:∑ui∈U

∑k∈K

xDj,k,i ≤ BD

j , ∀ bj ∈ B,

C2:∑bj∈B

∑k∈K

xDj,k,i ≤ MD

i , ∀ ui ∈ U,

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1: maximum PRBs at BS

C2: maximum PRBs at UE

C3: sub-channel exclusivity at UE

Combinatorial problem difficult to be solved!

Non-linear Integer Programming

Sub-modularity argument

Find efficient algorithms to solve it!

Any performance guarantee?

Konstantinos Alexandris 43 / 52

Problem formulation

xDj,k,i ∈ {0, 1}: Binary variable (PRB)

NUM problem is defined as:

maxxDi∈{0,1}|B|×|K|

∑ui∈U

U(

xDi

)s.t.

C1:∑ui∈U

∑k∈K

xDj,k,i ≤ BD

j , ∀ bj ∈ B,

C2:∑bj∈B

∑k∈K

xDj,k,i ≤ MD

i , ∀ ui ∈ U,

C3:∑ui∈U

xDj,k,i ≤ 1, ∀ bj ∈ B, k ∈ K.

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1: maximum PRBs at BS

C2: maximum PRBs at UE

C3: sub-channel exclusivity at UE

Combinatorial problem difficult to be solved!

Non-linear Integer Programming

Sub-modularity argument

Find efficient algorithms to solve it!

Any performance guarantee?

Konstantinos Alexandris 43 / 52

Problem formulation

xDj,k,i ∈ {0, 1}: Binary variable (PRB)

NUM problem is defined as:

maxxDi∈{0,1}|B|×|K|

∑ui∈U

U(

xDi

)s.t.

C1:∑ui∈U

∑k∈K

xDj,k,i ≤ BD

j , ∀ bj ∈ B,

C2:∑bj∈B

∑k∈K

xDj,k,i ≤ MD

i , ∀ ui ∈ U,

C3:∑ui∈U

xDj,k,i ≤ 1, ∀ bj ∈ B, k ∈ K.

User rate constrained by BH capacity:

RDi ,

∑bj∈B

∑k∈K

(xDj,k,i

)?RDj,k,i×

min

CDh,j∑

ui∈U∑

k∈K

(xDj,k,i

)?RDj,k,i

, 1

Objective function: Maximize network utility

The proposed utilities Φ(·) are applied

Allocation can be based on PF or UPF

Constraints:

C1: maximum PRBs at BS

C2: maximum PRBs at UE

C3: sub-channel exclusivity at UE

Combinatorial problem difficult to be solved!

Non-linear Integer Programming

Sub-modularity argument

Find efficient algorithms to solve it!

Any performance guarantee?

Konstantinos Alexandris 43 / 52

Proposed algorithm

The proposed optimization problem is NP-hard!

The optimization problem is proved to be with:

submodular and monotone function3 matroid constraints

Theorem (Alexandris et al. 2018)

Let OPT be the optimal solution of the formulated problem andS? be the output of the greedy algorithm. Then, it holds that

f (S?) ≥ 1

4OPT.

A greedy algorithm exists with an explicit bound on theoptimal solution

S? ∈ {0, 1}B×K×U the set of the exported solution

Konstantinos Alexandris 44 / 52

Simulations

100 120 140 160 180 200Backhaul Capacity (Mbps)

0

0.05

0.1

0.15

Avg

uns

atis

fied

n

orm

err

or

Non-opportunistic UPF-Multi (Min SINR) (Avg 1.76 connections per UE)Non-opportunistic UPF-Multi (Avg SINR) (Avg 1.89 connections per UE)Opportunistic UPF-Multi (Avg 2.19 connections per UE)

Permormance metric

Unsatisfied norm error: EDi =

∥∥∥∥∥R

Di − RD

i

RDi

∥∥∥∥∥ , if RDi < RD

i ,

0 , o/w.

Remarks:

Downlink-Params: z =2 users, z/z/z Scenario

Opportunistic scheduling increases connectivity

Nominal case: MIN SINR- Non-opportunistic, i.e., mink∈K

(SINRD

j,k,i

)QoS: Opportunistic outperforms the non-opportunitsic one even BH is limited

Konstantinos Alexandris 45 / 52

Take away messages

� Opportunistic scheduling exploits channel variability

� Discrete PRBs: NP-hard problem

� Problem is proved to be with sub-modular monotone functionand matroid constraints

� Greedy algorithm performs well with a guarantee solutionbound

� Non-opportunistic schemes perform worse in terms of usersQoS with limited BH

Konstantinos Alexandris 46 / 52

Conclusion

Konstantinos Alexandris 47 / 52

Summary

� X2 HO implementation in OAI emulator + real RF testbed

� Load-aware HO algorithm in next-generation HetNets

� Multi-connectivity resource allocation towards 5G

� Multi-connectivity resource allocation with limited BH

� Opportunistic scheduling in multi-connectivity with limited BH

Konstantinos Alexandris 48 / 52

Conclusion

Lessons learnt:

– Centralized mobility management can handle better HOprocess towards next-generation networks

– Transfer and service delay cost can be reduced with QoS-awarecross-layer mechanisms

– Centralized multi-connectivity resource allocation boostsnetwork aggregated rate as well as users QoS

– Multi-connectivity offers seamless mobility

– Connectivity failure: No need to re-establish connection– Support of user association towards 5G

– Such mobility and resource management centralized schemescan be directly applicable to SDN technology

Konstantinos Alexandris 49 / 52

Future directions

– Joint resource allocation and user association inmulti-connectivity under users QoS and BH limitations

– Resource allocation and beamforming in multi-connectednetworks

– Autonomous self-backhauling mesh network support undermulti-connectivity moving cells scenarios

– U2U-U2N-N2U: Local/Core routing

– Exploring multi-connectivity in disaggegated RAN

– C-RAN: CU-DU-RU

Konstantinos Alexandris 50 / 52

Publications

1 K. Alexandris, C.-Y. Chang, N. Nikaein and T. Spyropoulos “Utility-basedOpportunistic Scheduling under Multi-Connectivity with Limited BackhaulCapacity”, IEEE Wireless Communication Letter, 2018, under review.

2 K. Alexandris, C.-Y. Chang, N. Nikaein and T. Spyropoulos “Multi-ConnectivityResource Allocation with Limited Backhaul Capacity in Evolved LTE”, in 2018IEEE Wireless Communications and Networking Conference (WCNC), April2018, Barcelona, Spain, to appear.

3 K. Alexandris, C.-Y. Chang, K. Katsalis, N. Nikaein and T. Spyropoulos“Utility-Based Resource Allocation under Multi-Connectivity in Evolved LTE”,in 2017 IEEE 86th Vehicular Technology Conference (VTC-Fall), September2017, Toronto, Canada.

4 K. Alexandris, N. Nikaein, R. Knopp and C. Bonnet, “Analyzing X2 Handoverin LTE/LTE-A”, WINMEE 2016, Wireless Networks: Measurements andExperimentation, May 2016, Arizona State University, Tempe, Arizona, USA,invited.

5 K. Alexandris, N. Sapountzis, N. Nikaein and T. Spyropoulos, “Load-awareHandover Decision Algorithm in Next-generation HetNets”, in 2016 IEEEWireless Communications and Networking Conference (WCNC), April 2016,Doha, Qatar.

Konstantinos Alexandris 51 / 52

Questions?

Thank You!!!

Konstantinos Alexandris 52 / 52

Appendix

Konstantinos Alexandris 1 / 14

X2 Handover process

X2 HO request ACK

X2 HO request

Measurement report

Ho admission & resource

setup

SN Status transfer

Path switch request

HO decision

UE

Path switch request ACK

Release resource

HO complete (RRC connection reconfiguration complete)

RRC connected

RRC connected

RRC idle

RRC connected

Before handover

Handover preparation

Handover completion

Handover execution

RRC connected

After handover

RRC idle

RRC idle

RACH

UL grant

HO command (RRC connection reconfiguation)

Source eNB Target eNB EPC

Konstantinos Alexandris 2 / 14

OAI RF X2 Handover testbed schematic

Konstantinos Alexandris 3 / 14

Network topology

We assume a macrocell and a bunch of picocells at givendistances Dj from it, where j ∈ {1, . . . ,P}

Konstantinos Alexandris 4 / 14

Delay prediction

Assuming stationary M/G/1/PS:

Dki =

1

µik − λk

i

,

Service rate: µki = Rk

i /Y

The expected average rate is:

Rki =

NkMU,i Ci +

NkAU,i∑l=1

C (dl,i )

NkMU,i + Nk

AU,i

0 200 400 600 800 10000

200

400

600

dm(m)

Cb(dm)(M

bps)

Average Capacity bounds in macrocell

Lower boundUpper bound

0 50 100 150 200100

200

300

400

dpj(m)

Cb(dpj)(M

bps)

Average Capacity bounds in picocell

Lower boundUpper bound

Active users: Fixed capacity

C(dl,i ) ,[CL(dl,i ) + CU(dl,i )

]/2

Moving users: Average out distance

Ci =(C iL + C i

U

)/2

Konstantinos Alexandris 5 / 14

SDN framework

1 Controller tier: a) receives the respective λki and µki from BSs, b) computes

and sends the Dki to all BSs

2 Network tier: a) send the respective λki and µki , b) receive the Dki , c) send the

Dki to the UE

3 User tier: Each time k, the UE: a) receives the delays Dki , b) triggers the HO

procedure

Konstantinos Alexandris 6 / 14

Network topology

We consider an area L ⊂ R2 served by a set of BSsB = {b1, · · · , b|B|} and a set of UEs U = {u1, · · · , u|U|}

BSs

UEsu1

u3

u2

u4

b1 b3b2

Konstantinos Alexandris 7 / 14

Problem transformation

maxX ,Z

∑(ui ,uq )∈C

U

(zui ,uq

)

s.t. C1: zbj ,(ui ,uq) ≤ xUbj ,(ui ,uq)R

Uui ,bj

, ∀ bj ,

C2: zbj ,(ui ,uq) ≤ xDbj ,(ui ,uq)R

Dbj ,uq

, ∀ bj ,

C3:∑

(ui ,uq )∈Cbj

xUbj ,(ui ,uq) ≤ 1, ∀ bj ,

C4:∑

(ui ,uq )∈Cbj

xDbj ,(ui ,uq) ≤ 1, ∀ bj ,

C5:∑bj∈B

∑uq∈Dui

xUbj ,(ui ,uq)B

Ubj≤ BU

ui, ∀ ui ,

C6:∑bj∈B

∑uq∈Sui

xDbj ,(uq ,ui )

BDbj≤ BD

ui, ∀ ui ,

C7:∑bj∈B

∑uq∈Dui

xUbj ,(ui ,uq)P

Uui ,bj

BUbj≤ Pmax

ui, ∀ ui .

Objective function:

U(xUui ,uq

, xDui ,uq

) ,

Φ(∑bj∈B

Q(xUbj ,(ui ,uq), x

Dbj ,(ui ,uq))︸ ︷︷ ︸

zbj ,(ui ,uq)∈Z

)

Constraints:

min (·) is replaced by C1, C2 using the auxiliaryvariable z in the objective

Convexity:

U(

zui ,uq

)= Φ

(1T|B|zui ,uq

)Φ (·) is concave→ the objective is concave

Constraints are linear

Problem is convex!

Konstantinos Alexandris 8 / 14

Simulations-Performance analysis of PF & UPF

CDF plot of PF/UPF utility functions with different R

(a) R = 1 Mbps (b) R = 5 Mbps

(c) R = 10 Mbps

Remarks:

UPF outperforms PF in all cases: provides less network aggregated rateimproving user pairs QoS

Even in high requested rate in (c) still performs better (UPF CDF tail)

PF has the same CDF among different requested rates: QoS unaware

Konstantinos Alexandris 9 / 14

Network topology

We consider an area L ⊂ R2 served by a set of BSsB = {b1, · · · , b|B|} and a set of UEs U = {u1, · · · , u|U|}

BSs

UEsu1

u3

u2

u4

b1 b3b2

Core network

Internet

Backhaul

links

Konstantinos Alexandris 10 / 14

Simulations-Infinite BH-z/z/z Scenario

Single-PF-2/2/2 Multi-PF-2/2/2 Single-UPF-2/2/2 Multi-UPF-2/2/2500

600

700

800

Dat

a ra

te (

Mbp

s) UL aggregated rate comparison of Scenario B

UL user satisfaction ratio comparison of Scenario B

Single-PF-2/2/2 Multi-PF-2/2/2 Single-UPF-2/2/2 Multi-UPF-2/2/20

20406080

100

Per

cent

age

(%)

Remarks:

Uplink-Params: z =2 users, CUh,j = CD

h,j =∞Unallocated resources in each BS due to UL power control (C5constraint)

Multi-cell approach can make use of those resources

Higher gains in network aggregated rate + user satisfaction

Konstantinos Alexandris 11 / 14

Simulations-Dynamic BH

Single-PF-0/2/2 Multi-PF-0/2/2 Single-UPF-0/2/2 Multi-UPF-0/2/2200

400

600

800

1000

Dat

a ra

te (

Mbp

s) DL aggregated rate comparison of Scenario A

Single-PF-2/2/2 Multi-PF-2/2/2 Single-UPF-2/2/2 Multi-UPF-2/2/2500

600

700

800

900

Dat

a ra

te (

Mbp

s) DL aggregated rate comparison of Scenario B

DL user satisfaction ratio comparison of Scenario A

Single-PF-0/2/2 Multi-PF-0/2/2 Single-UPF-0/2/2 Multi-UPF-0/2/20

20406080

100

Per

cent

age

(%)

DL user satisfaction ratio comparison of Scenario B

Single-PF-2/2/2 Multi-PF-2/2/2 Single-UPF-2/2/2 Multi-UPF-2/2/20

20406080

100

Per

cent

age

(%)

Remarks:

Downlink-Params: z =2 users, CDh,j ∼ U [150, 350] Mbps

Multi vs Single connectivity

Loaded cell scenario: No extra resources to exploit compared to emptycell oneNetwork aggregated rate: no such difference

Further gain: PF vs UPF

UPF can still increase the user satisfaction ratePresented results closely lie in the range of finite BH gains

Konstantinos Alexandris 12 / 14

Simulations-Utilization ratio-Dynamic BH

Utilization Single Multi Single MultiRatio PF PF UPF UPF(%) 0/2/2 2/2/2 0/2/2 2/2/2 0/2/2 2/2/2 0/2/2 2/2/2

Air-interface 57 90 94 98 57 90 93 98

Backhaul 63 92 95 93 63 92 93 91

Remarks:

Downlink-Params: z =2 users, CDh,j ∼ U [150, 350] Mbps

Multi-connectivity utilizes better the unallocated resources both in air-interfaceand BH

Unlimited BH: Air-interface resources are fully used (100%) and BH utilizationratio is <100%

Limited BH: Air-interface resources are not fully used (< 100 %) and BHutilization ratio is 100 %

Konstantinos Alexandris 13 / 14

Simulations-Opportunistic Scheduling

120 140 160 180 200 220 240 260 280 300Backhaul Capacity (Mbps)

50

60

70

80

90

100

Avg

use

r sa

tisfa

ctio

n (%

)

PF-Multi with avg 2.21 connections per UEUPF-Multi with avg 2.49 connections per UE PF-Single with avg 1.00 connection per UEUPF-Single with avg 1.00 connection per UE

120 140 160 180 200 220 240 260 280 300Backhaul Capacity (Mbps)

50

60

70

80

90

100

Avg

use

r sa

tisfa

ctio

n (%

)

PF-Multi with avg 2.01 connections per UEUPF-Multi with avg 2.19 connections per UE PF-Single with avg 1.00 connection per UEUPF-Single with avg 1.00 connection per UE

Remarks:

Downlink-Params: z =2 users, CUh,j = CD

h,j <∞

0/z/z Scenario: Multi-connectivity outperforms the single one

Even with limited BH ⇒ i.e., more resources are utilizedConnectivity gain: multiple connection

z/z/z Scenario: Better (Multi vs Single)-No extra resources

PF vs UPF: UPF in Multi performs the best to satisfy users QoSAdditional gain: single to multi-cell comparisonTrade-off: Network and User perspective

Konstantinos Alexandris 14 / 14