You are on page 1of 20

Si

Sistemski
ki AAspekti
kiu
Tlk
Telekomunikacijama
ik ij

mugdim.bublin@siemens.com
A Cost Function Approach
pp to
Self - Organizing RRM
in Mobile Cellular Networks

mugdim.bublin@siemens.com
Overview

Motivation
M ti ti

Principles of Self-Organization

Our Approach: Cost function with adaptive weights

Examples: DCA & Scheduler

Summary andd Challenges


h ll

mugdim.bublin@siemens.com
Moti ation
Motivation

U examples
Use l from
f nature
t for
f ththe d
design
i off 4G mobile
bil networks
t k

mugdim.bublin@siemens.com
Principles of Self-Organization
Self Organization

Global behavior emerges from local behavior of autonomous distributed "Agents"


Agents

Simple Rules

Usage of Local Information

p
Near Optimal Global Behavior

Adaptive and Robust to Changes

mugdim.bublin@siemens.com
SON ffor LTE

Self-Organizing Networks (SON)


Objective: decrease OPEX/CAPEX related to network configuration, operation, optimization
Main functionalities:
Self-Configuration (Plug & Play of new eNodeB)
Self-Healing (e.g. Cell Outage Compensation)
Self-Optimization* (e.g. Mobility Load Balancing, Handover Optimization, Energy
Saving Management, etc.)
Rel. 8/9/10 focused on SON for LTE
Rel
Rel. 11 addresses 3G and inter-RAT SON (Radio Access Technology)

OA&M SON function


related
Decision indicators Statistical
Analysis
* Self-Optimization

Setting of Configuration Performance measurement


parameters reports,
Alarm information,, etc.
eNodeB

mugdim.bublin@siemens.com
P
Proportional
ti lFFair
i Scheduler
S h d l

Schedule the user


with the highest
ratio:

Rk = possible rate of Ch
hannel gain
user k
Tk = average
throughput of user k
ti
time

mugdim.bublin@siemens.com
General Schedulingg Minimize the Cost Function

Minimize: CF = WPL PL + WI I - C

"PropFair" Scheduling: WPL = 1, WI = 1, C = -k*R =>CF = PL + I + k*R


BestCQI Scheduling: WPL = WI = 1, C = 0 =>> CF = PL + I.
"BestCQI"
"Round Robin": WPL = WI = 0, C = T (scheduling time) => CF = -T

Fair scheduling:
MS1 MS1 MS2 MS1 MS2 MS1 MS1

Greedy scheduling:
MS2 MS2 MS2 MS2 MS1 MS2 MS2 MS2

Round-Robin scheduling:
MS2 MS1 MS2 MS1 MS2 MS1 MS2
MS1

mugdim.bublin@siemens.com
DCA: State of the Art

H to
How t allocate
ll t channels
h l to
t cells
ll and
d users??

Dynamic Channel Allocation (DCA)


Simple heuristic rules:
Take channel Randomly
Take
T k channel
h l with ith Minimum
Mi i IInterference
t f
Take channel with Maximum Interference below a threshold

But
No common theoretical basis
The same algorithms
g used for different network states

mugdim.bublin@siemens.com
Model

Encode heuristic rules into the Cost Function


Users in UL, BS in DL minimize Cost Function (CF):

A possible Cost Function : CF = |WPL PL + WII C|


PL Path-loss in dB
I - Interference in dBm
WPL, WI and
d C Parameters,
P t " i ht "
"weights"

BS set parameters WPL WI and C according to cell state


state = f(number of users, QoS, Interference)

mugdim.bublin@siemens.com
Example
p I: DCA

Users: Take channel(s) which Minimize Cost Function:


CF(channel) = |WPL PL + WI I C|

Cells set weights WPL, WI and C according to load for example:


Low load: All channels are equally "good".
Average Interference - Random DCA: WPL = WI = 0,
0 C = random

Low-Medium load: Neighbor cells can use different channels for most of the time.
A id IInterference
Avoid f - Minimum
Mi i Interference
I f DCA:
DCA WPL = 0, 0 WI = 1,
1 C isi a large
l negative
i
number

High load: Provide as much users as possible with "good enough" channels. Scale
Interference - Reuse Partitioning DCA: WPL = WI = 1, C ~ CIR -threshold

mugdim.bublin@siemens.com
Min I DCA Channel Segregation

Take channel with Minimum average interference Ia,


I a[n] = Ia[n-1] + (1-)Iactual[n]

Begin of simulation
10 seconds after simulation begin
20 70
BS_1 BS_1
18 BS_2 BS_2
BS_3 60 BS_3
16

14 50
of usage

Percent of usage
12
40
Percent o

10
30
8

6 20

4
10
2

0 0
(0, 0) (0, 1) (0, 2) (0, 3) (0, 4) (0, 5) (0, 6) (0, 7) (0, 0) (0, 1) (0, 2) (0, 3) (0, 4) (0, 5) (0, 6) (0, 7)
Channel (f, TS) Channel (f, TS)

mugdim.bublin@siemens.com
What to do for Higher Load?

10 s after sim ulation begin (high load)


30
BS_1
BS 2
BS_2
BS_3
25

20
ge
Percent of usag

15
P

10

0
(0, 0) (0, 1) (0, 2) (0, 3) (0, 4) (0, 5) (0, 6) (0, 7)
Channel (f, TS)

mugdim.bublin@siemens.com
ARP: Autonomous Reuse Partitioning

150

Cell 1
140

Ch 1

ath-loss [dB]
130
Cell 2 Ch2

Chl 3 120

Pa
110

100
-100 -95 -90 -85 -80
Interference [dBm ]

Minimize Cost Function: CF(channel) = | PL + I CIRthreshold|

mugdim.bublin@siemens.com
Performance by adaptive weight settings

Minimize Cost Function: CF = | PL + I C(load)|

60
0*dropped

50
atisfied + 10

40

30
GoS = unsa

Random
20 Min I
Cost Function
G

ARP
10
0 1 2 3 4 5 6
number of users [thousand]

mugdim.bublin@siemens.com
Examples of Scheduling & DCA

automatically adapt weights according


to for
f example:l LMS algorithms:
l h

CF
W ( n 1) W ( n ) Costs
W

"Optimal"
p DCA/Schedulingg
combinations might arise

mugdim.bublin@siemens.com
Summary

Agents Cells,
Distributed "Agents" Cells Users
Users, Services

Signalization reduction Local information in each cell

Simple Rules Random, Min I, ARP or combinations of them

p
Near Optimal Global Behavior Better than anyy of the investigated
g algorithms
g

Adaptive and Robust to Changes "Always best" by parameter settings

Significant complexity reduction: O(N) instead of O(N!)/O(NK)

mugdim.bublin@siemens.com
Challenges

Form(s) of the Cost Function

Automatic "weight" settings

Interplay with other RRM algorithms and Link Adaptation

Inter
Inter-system/inter-mode
system/inter mode RRM

How to enforce cooperation: Standardization or Game Theory?

mugdim.bublin@siemens.com
Pi j ?
Pitanja?

mugdim.bublin@siemens.com
H l na paznji!
Hvala ji!

mugdim.bublin@siemens.com

You might also like