small-cell self-organizing wireless networks_06732895

17
INVITED PAPER Small-Cell Self-Organizing Wireless Networks Moving to smaller and smaller cells may be the secret to attaining substantial improvements in spectrum efficiency. The authors take a closer look at the improvements achievable with small-cell self-organizing networks. By Albrecht J. Fehske , Ingo Viering , Jens Voigt , Cinzia Sartori , Simone Redana, and Gerhard P. Fettweis, Fellow IEEE ABSTRACT | Increasing the spatial reuse of frequency spec- trum by deploying more access points has historically been the most effective means to improve the capacity of any cellular communication network. Today’s mobile networks face a proliferation of data services and overall demand for data traffic that has been strongly increasing over several years. As a result, increasing network capacity through the deployment of small lower power nodes is of key importance for mobile network operators. Although such small access points are conceptually equivalent to conventional cellular base stations in many ways, the expected large number of small cells as well as their much more dynamic unplanned deployment raise a variety of challenges in the area of network management. This paper discusses such challenges and reviews state-of-the-art modeling as well as selected network management techniques. KEYWORDS | Capacity and coverage optimization; closed-loop optimization; energy saving management; interference mod- eling; long-term evolution (LTE); mobility load balancing; self- optimization; self-organizing network (SON); small cells; system modeling; wireless network planning and optimization OVERVIEW We outline a mathematical framework that allows to numerically assess the performance of networks featuring a large number of small access points as well as to develop algorithms for their self-organization. We introduce elementary network management techniques for mobility robustness optimization and mobility load balancing, which operate on the shorter scales of seconds. We further discuss techniques for coverage and capacity optimization and energy saving management which operate on the larger scales of hours. We highlight the specific challenges due to small-cell deployments in all discussions. I. INTRODUCTION In principle, the increasing demand for cellular network capacity may be addressed in three different ways: utilizing more frequency spectrum, achieving greater transmission spectral efficiency over the air interface, or increasing the network density, thereby increasing the spatial reuse of spectrum and lowering the distance between the transmitter and the receiver. Historically, network densification has improved network capacity much more (about 2.5 times) than the increase of spectrum and spectral efficiency combined [1]. Small-cell deploy- ments are, thus, widely believed to be also fundamental for drastically improving capacity and quality of service (QoS) in future cellular networks [2]. Moreover, they are also heralded as a means to reduce cellular networks’ power consumption and overall footprint [3]–[5]. Both advan- tages are mostly accredited to the fact that small-cell links enjoy, on average, better propagation conditions, which then allows for lower transmit powers, and thus, lower total power consumption. A. Types of Small Cells Various small-cell types and designs exist in the market as well as in the literature today. By now, a certain termi- nology has been, to an extent, established in both industry, Manuscript received August 7, 2013; revised January 10, 2014; accepted January 15, 2014. Date of publication February 5, 2014; date of current version February 14, 2014. The work was supported in part by the Government of the Free State of Saxony, Germany, within the Cool Silicon Cluster of Excellence under Contracts 29908/2794 and 14056/2367. A. J. Fehske and G. P. Fettweis are with the Electrical Engineering Department, Technical University of Dresden, 01069 Dresden, Germany (e-mail: [email protected]). I. Viering is with Nomor Research GmbH, 81541 Munich, Germany and also with Munich University of Technology, 85748 Munich, Germany. J. Voigt is with Actix GmbH, 01067 Dresden, Germany. C. Sartori and S. Redana are with Nokia Siemens Networks, 81541 Munich, Germany. Digital Object Identifier: 10.1109/JPROC.2014.2301595 0018-9219 Ó 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. 334 Proceedings of the IEEE | Vol. 102, No. 3, March 2014

Upload: alephcase

Post on 04-Dec-2015

14 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: Small-Cell Self-Organizing Wireless Networks_06732895

INV ITEDP A P E R

Small-Cell Self-OrganizingWireless NetworksMoving to smaller and smaller cells may be the secret to attaining substantial

improvements in spectrum efficiency. The authors take a closer look at the

improvements achievable with small-cell self-organizing networks.

By Albrecht J. Fehske, Ingo Viering, Jens Voigt, Cinzia Sartori, Simone Redana, and

Gerhard P. Fettweis, Fellow IEEE

ABSTRACT | Increasing the spatial reuse of frequency spec-

trum by deploying more access points has historically been the

most effective means to improve the capacity of any cellular

communication network. Today’s mobile networks face a

proliferation of data services and overall demand for data

traffic that has been strongly increasing over several years. As a

result, increasing network capacity through the deployment of

small lower power nodes is of key importance for mobile

network operators. Although such small access points are

conceptually equivalent to conventional cellular base stations

in many ways, the expected large number of small cells as well

as their much more dynamic unplanned deployment raise a

variety of challenges in the area of network management. This

paper discusses such challenges and reviews state-of-the-art

modeling as well as selected network management techniques.

KEYWORDS | Capacity and coverage optimization; closed-loop

optimization; energy saving management; interference mod-

eling; long-term evolution (LTE); mobility load balancing; self-

optimization; self-organizing network (SON); small cells; system

modeling; wireless network planning and optimization

OVERVIEW

We outline a mathematical framework that allows to

numerically assess the performance of networks featuring

a large number of small access points as well as to develop

algorithms for their self-organization. We introduceelementary network management techniques for mobility

robustness optimization and mobility load balancing,

which operate on the shorter scales of seconds. We further

discuss techniques for coverage and capacity optimization

and energy saving management which operate on the

larger scales of hours. We highlight the specific challenges

due to small-cell deployments in all discussions.

I . INTRODUCTION

In principle, the increasing demand for cellular network

capacity may be addressed in three different ways:

utilizing more frequency spectrum, achieving greater

transmission spectral efficiency over the air interface, or

increasing the network density, thereby increasing the

spatial reuse of spectrum and lowering the distancebetween the transmitter and the receiver. Historically,

network densification has improved network capacity

much more (about 2.5 times) than the increase of spectrum

and spectral efficiency combined [1]. Small-cell deploy-

ments are, thus, widely believed to be also fundamental for

drastically improving capacity and quality of service (QoS)

in future cellular networks [2]. Moreover, they are also

heralded as a means to reduce cellular networks’ powerconsumption and overall footprint [3]–[5]. Both advan-

tages are mostly accredited to the fact that small-cell links

enjoy, on average, better propagation conditions, which

then allows for lower transmit powers, and thus, lower total

power consumption.

A. Types of Small CellsVarious small-cell types and designs exist in the market

as well as in the literature today. By now, a certain termi-

nology has been, to an extent, established in both industry,

Manuscript received August 7, 2013; revised January 10, 2014; accepted

January 15, 2014. Date of publication February 5, 2014; date of current version

February 14, 2014. The work was supported in part by the Government of the

Free State of Saxony, Germany, within the Cool Silicon Cluster of Excellence under

Contracts 29908/2794 and 14056/2367.

A. J. Fehske and G. P. Fettweis are with the Electrical Engineering

Department, Technical University of Dresden, 01069 Dresden, Germany

(e-mail: [email protected]).

I. Viering is with Nomor Research GmbH, 81541 Munich, Germany and also with

Munich University of Technology, 85748 Munich, Germany.

J. Voigt is with Actix GmbH, 01067 Dresden, Germany.

C. Sartori and S. Redana are with Nokia Siemens Networks, 81541 Munich, Germany.

Digital Object Identifier: 10.1109/JPROC.2014.2301595

0018-9219 � 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

334 Proceedings of the IEEE | Vol. 102, No. 3, March 2014

Page 2: Small-Cell Self-Organizing Wireless Networks_06732895

e.g., [2] and [5], as well as research consortia, e.g., [6] and[7]. In this regard, it is usually distinguished between

femtocells, picocells, and microcells. In essence, these three

types differ in average transmit power and thus cell size.

While femtocells operate with only a few, up to hundred

milliwatts, the transmit power of a picocell may be up to 1 W,

whereas the transmit power of a microcell is commonly

between 2 and 5 W. In comparison, a conventional macrocell

typically radiates between 20 and 40 W per antenna.Accordingly, small-cell coverage ranges from a few tens of

meters for femtocells and up to 2 km for microcells, also

depending on the propagation environment. Typically, femto-

base stations (BSs) are employed indoors for private homes

and small offices and often serve a closed subscriber group. As

femto-BSs are home eNodeBs (HeNB), they are not usually

deployed by the operator. The framework and algorithms

outlined in this paper apply to all types of small cells. Thenumerical examples focus on microcells and picocells, which

are suitable for a larger range of deployments, e.g., in malls,

city centers, at road sites, or even in tunnels.

B. Small-Cell Deployment ProjectionsAs a result of the many advantages promised by small-

cell enhanced networks, market projections suggest a rapid

increase in shipping numbers as well as in the number ofsmall cells being deployed over the coming years. While

concrete figures still vary widely, recent surveys report

that many operators plan to set up between 3 and 5 micro- or

pico-BSs per macrocell. Peak numbers even state 13–20 picos

per macrocell, at least in urban environments [5].

In contrast to macrocells, the amount of traffic captured

by small cells is not strongly related to the size of their

coverage area because small BSs are typically placed atlocations with high traffic demand. Therefore, in addition to

merely considering the number of small cells deployed per

macrocell, it is crucial to consider the fraction of traffic

actually served by low-power access points. According to

Cisco’s Global Mobile Traffic Forecast [8], the fraction of

overall mobile traffic offloaded to both femtocells and WiFi

access points was about 30% in 2012 and will grow up to 46%

globally until 2017. Other analysts’ data suggest that incertain European and North American networks already

50%–60% of today’s data traffic coming from mobile devices

is being offloaded to WiFi access points [9]. Predictions by

Juniper research indicate that up to 60% mobile data traffic

may be offloaded to microcells and picocells as well as to

femtocells and WiFi as early as 2016 [10].

C. The Need for Self-OrganizationSmall cells feature common air interfaces such as WiFi

or 3GPP long-term evolution (LTE) and, apart from the

transmit power, are equivalent to macro-BSs. From a net-

work management perspective, however, there are two key

differences between small cells and conventional macro

cells. The first difference is that small cells come in much

larger numbers into wireless networks than macro cells. As a

consequence, manual processes for configuration and opti-mization are no longer feasible. Configuration and optimi-

zation begins with acquisition of an IP address, downloading

the correct software version, and then downloading the

correct parameter configuration from an operator’s data

base. The process continues with adjusting parameter config-

uration during operation based on measurements, key

performance indicators (KPIs), customer complaints, etc.

All those aspects must be self-organized in order to supportthe business case of small cells.

The second difference is that small cells are deployed

much more dynamically. Whereas macrocell deployments

are thoroughly planned, small cells are often commis-

sioned quickly whenever capacity need is detected and

without an elaborate planning phase. Since, in most cases,

small cells are capacity overlays, i.e., they do not provide

basic network coverage, operators may tend to switchthem off when capacity is not needed. Such changes of the

deployment require adaptation not only in the neighbor-

hood of the deactivated cell, but also, in some cases, in the

neighborhood of the neighborhood, which may consist of

both, small cells and macrocells. In general, the impact of

small-cell deployments on users connected to macrocells

must be kept minimal at all times.

As a result, small-cell networks not only required moreadaptations and reconfigurations due to more cells, but

also they feature more adaptations and reconfigurations

per cell. This requirement highlights the importance of

self-organizing network (SON) technologies specifically

for small-cell networks. Historically, SON features have

been proven to be very valuable already for traditional

macronetworks, to adapt, e.g., to changes in the environ-

ment, the traffic conditions, or the deployment [11].

D. SON ArchitecturesFor the purposes of this paper, we structure the

different SON technologies into two broad categories:

small-scale short-term techniques on the one hand and

large-scale long-term techniques on the other hand. Small-

scale short-term SON techniques need to react to problems

such as traffic imbalances or handover failures on timescales of seconds or even fractions of seconds. Further,

parameter adaptation is performed locally at the BSs and

based on information exchanged among neighbors. For this

reason, we also refer to small-scale short-term techniques as

a distributed self-organizing network (D-SON). Large-scale

and long-term network optimization typically aims at

improving capacity-centric objectives based on long-term

average input values. Corresponding algorithms jointly adjustparameters of an entire cluster of cells to daily traffic

variations rather than reacting on instantaneous traffic peaks

at single cells. Since parameters of several cells are optimized

jointly, which requires some central coordination, we also

refer to large-scale long-term term approaches as a central-

ized self-organizing network (C-SON). In practice, a C-SON

entity is a part of the operational support system (OSS) on

Fehske et al : Small-Cell Self-Organizing Wireless Networks

Vol. 102, No. 3, March 2014 | Proceedings of the IEEE 335

Page 3: Small-Cell Self-Organizing Wireless Networks_06732895

network management level (NML) or element managementlevel (EML) and handles multilayer, multivendor, and

multitechnology small-cell deployments.

Note that the categories D-SON and C-SON are to

some extent idealizing, and there are certainly algorithms

and processes that exist in the gray area between these two

extreme cases. The X2 interface may not be necessary for

all algorithms, especially in a C-SON context.

E. OutlineThe paper is structured as follows. Section II lays out a

framework for modeling and performance assessment of

small-cell SONs. This framework captures network dynamics

on time scales relevant for short-term as well as for long-term

network management approaches. Section III outlines basic

cell configuration procedures that are particularly affected by

deployment of large numbers of small cells. Sections IV andV outline D-SON and, respectively, C-SON techniques for

small-cell networks.

II . MATHEMATICAL FRAMEWORK

When studying a complex system, one ideally chooses a

level of abstraction that conveys all effects of interest with

as little complexity as possible. In case of self-organizingwireless networks, such abstraction involves choosing

adequate time scales of observation. Network management

events and procedures suitable for small-cell networks

occur on the order of several hundred milliseconds up to

several tens of seconds. Study of their effects, thus, re-

quires simulation of comparably long periods of real time,

i.e., several tens of minutes or above.

Common radio link models and tools accurately repre-sent effects and mechanisms such as fast fading and radio

resource management (i.e., the packet scheduler) on a

millisecond basis. Due to the high level of detail, the

numerical effort of studying longer periods with common

link level tools is prohibitively high. Moreover, small-cell

networks feature a large number of interacting access points,

which make the complexity problem even more pronounced.

In contrast, popular so-called snapshot-based tools capturevery long-term average effects, corresponding to periods on

the order of days or weeks. Key terms and effects such as

radio resource management, QoS perceived by the users, as

well as user mobility are not captured with enough accuracy

or are not represented at all as in case of mobility.

To allow performance evaluation of large-scale small-cell

networks, we present a modeling framework that strikes a

balance between the accuracy with which small-scale effects,specifically radio resource management, are represented and

the periods of time as well as the number of cells that can be

jointly investigated with feasible complexity. Emphasis is put

on capturing the signal-to-interference-and-noise ratio

(SINR) and throughputs over short periods of a few hundred

milliseconds as well as the resulting resource utilization of

the cells.

A. General Scenario DescriptionWe outline a mathematical model that allows to

numerically assess the performance of networks featuring a

large number of small access points as well as to be applied in

algorithms for self-organization of such networks. The model

considers the downlink of a cellular network consisting of

C base stations, corresponding to C cells, which cover a

regionR � R2. BSs may be either regular or small cells and

their locations can be arbitrarily chosen. Consider further acollection of users to be present in the network with user ulocated at position pu 2 R. We assume all signal propagation

effects between any location p 2 R and any cell c to be

collected in the propagation maps LcðpÞ. The maps Lc capture

both, effects due to the environment such as distance-

dependent pathloss and shadowing as well as effects due to

system design such as antenna height, antenna patterns, or

tilt angles. If some of the design parameters are adjustable(most prominently the antenna tilt angles), then all propa-

gation maps are, in general, depending on the corresponding

parameter values. Let Ptx;c denote the transmit power of BS c.

The power received at location p from BS c is then given as

the product of transmit power and path loss, i.e.,

Prx;cðpÞ ¼ LcðpÞ Ptx;c: (1)

Since users may be mobile, their locations, receive powers,

and related terms may depend on time. We note this

dependency by writing pðtÞ, Prx;cðtÞ, etc., only where neces-

sary. Note that in any case the propagation maps are assumed

to be static. We further assume in this framework that fast

fading is only considered by its average in the propagation

maps Lc.Note that, in practical systems, propagation maps need to

be created from measurements collected at the user

equipment (UE) and reported back to the base stations. In

this case, location of the terminal is either estimated by the

system or reported with measurements of, e.g., power

received from different BS. Alternatively, propagation maps

for simulations may be obtained from common path loss

models or, more accurately, ray-tracing tools.

B. Modeling SINR and ThroughputIn order to judge and compare the performance of

management or deployment solutions we lay out a series ofperformance metrics. As a baseline, we start with the SINR

as perceived by individual users and corresponding through-

put achieved. We define the SINR of user u connected to cell

c as the ratio of signal power and weighted interference-plus-

noise power, i.e.,

�u ¼Prx;cðpuÞP

d 6¼c �d Prx;dðpuÞ þ �(2)

Fehske et al : Small-Cell Self-Organizing Wireless Networks

336 Proceedings of the IEEE | Vol. 102, No. 3, March 2014

Page 4: Small-Cell Self-Organizing Wireless Networks_06732895

where term � denotes noise power. Factors �d denote theutilization of transmission resources, i.e., physical re-

source blocks (PRBs) in an LTE system. Several definitions

of the utilization �c are discussed in Section II-C.

Weighting the interference power received from each

cell with its current resource utilization yields a certain

dependency of the total interference on the different cell

utilizations: very active cells contribute more interference

power than less active ones. This formulation reflects acertain approximate average interference power that users

see over some period as follows. The time scale of interest

here corresponds to a few hundred milliseconds up to a

few seconds. On much smaller scales of only a few

milliseconds, scheduling algorithms assign resource ele-

ments to individual users. Thus, in reality, each resource

element (corresponding to a short time period) individu-

ally sees either full interference from some other cell or nointerference from that cell at all. Accordingly, an exact

formulation yields many different SINR values, depending

on the instants and durations at which a user is scheduled

and depending on which other cells are active. Tracking

these random events individually requires a rather compli-

cated queuing theoretic approach (e.g., [12]) and is not

advocated here. In such a model, however, and under certain

simplifying assumptions, the utilization �c corresponds to theprobability that BS c is transmitting and the sumP

d6¼c �d Prx;dðpuÞ reflects the interference power received

on average on time scales up to a few seconds. In contrast to

an exact approach, the model in (2) based on this notion of

average interference provides a simple and tractable

formulation of the SINR. Note that �u is also only an ap-

proximation of the average SINR, since averaging is

performed in the denominator. The same model has beenproposed in similar contexts independently in [13]–[15].

The data rate experienced by users is usually deter-

mined by a rate function, which maps the SINR at hand to

a feasible data rate. A popular example is Shannon’s

capacity formula. More realistic rate mappings are used in

practice, some of which are based on Shannon’s formula

and additional factors to account for signaling overheads

and practical modulation, coding as well as schedulingschemes [16]. Rate functions resulting from link level

simulations are also conceivable and commonly used. We

do not consider any specific rate function in this section.

Since SINRs vary on the millisecond scale, the corre-

sponding data rates vary accordingly. An approximate

average throughput seen by user u over periods of up to a

few seconds may conveniently be defined as the rate

corresponding to the approximate SINR in (2), i.e.,

Ru :¼ Rð�uÞ (3)

where Rð�Þ denotes the rate achievable per resource

element at a certain SINR. Note again that �u as well as Ru

are not exact but only approximate averages of the SINRand throughput per resource element, respectively.

C. Modeling Cell LoadThe notion of cell load is as fundamental for the QoS

perceived by individual users as its adequate definition is

controversial. Coincidentally, most of the controversy

relates to the interplay of load and the QoS associated with

different types of service. The load definitions discussed

subsequently are considered by the authors to be

applicable (and are indeed applied) to practical algorithm

implementations and underlie the results presented inSections IV and V.

1) Cell Load Based on Resource Utilization: The most

intuitive notion of cell load is the degree of utilization of

transmission resources, e.g., in LTE, the utilization of

PRBs. Let B denote the transmission resources available in

the cell (e.g., the number of PRBs) and Bu the average

number of resources assigned to user u over the period ofinterest, e.g., a few hundred milliseconds. The load of cell

c over that period is then given by the sum of all resources

used

�c ¼1

B

XUc

u¼1

Bu (4)

where Uc denotes the number of active users connected to

cell c. In (4), we consider a practical system where Bu is

measured over the period of interest and, since only a totalof B resources is available, the load is always at most equal

to 1. How to estimate �c in a simulation environment is

discussed in Section II-C2d. Many common applications

such as web browsing, e-mailing, or file downloads

generate so-called elastic or best effort traffic, which can

be delivered with arbitrarily high or low data rates. If

active best effort users are present during the period of

interest, the scheduler assigns all spare transmissionresources to them such that we have

PUc

u¼1 Bu ¼ B. In

this case, the cell load in (4) simply becomes

�c :¼ 1; Uc > 00; else.

�(5)

Note that this load formulation holds for any schedulingscheme, since it only involves counting the resources

occupied over a certain reporting period.

2) Virtual Load Based on User Satisfaction: Using a load

definition based merely on resource utilization as outlined

above bares certain disadvantages when designing algorithms

for traffic steering. Consider, for instance, a scenario in

Fehske et al : Small-Cell Self-Organizing Wireless Networks

Vol. 102, No. 3, March 2014 | Proceedings of the IEEE 337

Page 5: Small-Cell Self-Organizing Wireless Networks_06732895

which a single large background download is performed in asmall cell and at the same time several streaming applications

are serviced in an umbrella macrocell, which reaches its load

limit. Assume that transferring some streaming users to the

small cell (some users may be at the cell border) would

greatly enhance their QoS as well as the QoS of those re-

maining in the macrocell. Furthermore, reducing resources

assigned to the download does not significantly affect its QoS

due to the nature of elastic traffic. Unfortunately, the loaddefinitions in (5) suggest that also the small cell is fully

loaded since all resources are utilized for the download.

a) Target bit rates: This situation is avoided when

introducing a notion of user satisfaction into the definition of

load. On a basic level, satisfaction is represented by the

degree of which a certain minimum bit rate, say Ru for user u,

is provided. In case of elastic traffic, terms Ru represent an

equivalent minimum bit rate, whose value is commonlychosen to be small compared to, e.g., streaming applications

to reflect the less stringent QoS requirements of the former

compared to the latter. The values of Ru may be further

adapted according to the user’s subscription level (i.e., gold,

silver, bronze, etc.). Based on this idea, we estimate the

number of resources required for user u as the ratio of target

rate and achievable rate per resource, limited to some

maximum value Bmax, i.e.,

Bu :¼ minRu

Ru; Bmax

� �: (6)

The maximum amount of resources considered per user is

typically around 20% of the total, i.e., Bmax ¼ 0:2B. This

limit prevents individual users with bad channels, i.e.,

small Ru, from distorting the resource estimate Bmax byclaiming unnecessarily many resources.

b) Virtual load: A virtual cell load may then be

defined as the smallest fraction of resources required to

achieve all targets [13], [17], i.e.,

�c ¼1

B

XUc

u¼1

Bu: (7)

Note that �c differs from the actual load �c in two aspects:

• the value of �c may well exceed 1 if rate targets

cannot be met;

• in case �c G 1, the actual cell load �c may be largerthan � if spare resources are assigned to best effort

users, as explained above.

In our example above, assuming that the targets Ru are

properly chosen, the virtual load of the small cell servicing

the background download would be small, whereas the vir-

tual load of the umbrella cell would expected be high. Such

imbalance would potentially trigger handovers to the small

cell to improve the QoS seen by the streaming users. Corres-ponding schemes are outlined in Section IV-A2 and B2.

c) Exchanging load information between cells: Traffic

steering mechanisms dynamically balance load among

different cells. For this purpose, BSs must be aware of the

load of relevant neighbors, i.e., load information must be

exchanged between them. To aide load balancing, 3GPP

standards define the notion of compound available capacity

(CAC), which, if communicated to neighbors, indicates howmuch load a cell is willing to accept. Concrete definitions of

load are not standardized and left to equipment vendors. A

simple and intuitive definition of the CAC, which we use

subsequently, is the difference between virtual load and

some maximum value permitted, i.e.,

CACc ¼ �target � �c (8)

where �target is a design parameter indicating a load limit,

typically chosen below a value of 1 to leave some capacity

margin. Concrete traffic steering mechanisms are dis-

cussed in Section III.

d) Numerical Calculation: In practice, the throughputs

Ru are measured and virtual cell loads �c can be directly

computed. When system performance is assessed in

simulations, however, their values of Ru must be estimatedfrom SINR values according to the modeling framework

outlined in this section. In this case, the throughput Ru is a

function of the loads of all interfering cells [cf., (2) and (3)].

Assuming that transmission resources allocated to user u are

given by (6), i.e., Bu ¼ Bu, the cell loads �c themselves

depend on the throughputs Ru [cf., (4)]. Because of this

dependency, all cell loads �c must be jointly obtained as

solutions of a system of C nonlinear equations given by (4) forc ¼ 1 . . . C. Under mild conditions on the rate function Rð�Þ,a solution is found by applying a fixed-point iteration, where

the load of cell c in iteration k is given by the rule

�ðkþ1Þc ¼ min

1

B

XUc

u¼1

BðkÞu 1

" #

where BðkÞu is the resources allocation in iteration k, i.e.,corresponding to �ðkÞc . Further, (7) is guaranteed to have a

unique solution, such that the cell loads are always well

defined [14]. In contrast to (4), the minð�Þ operation is

needed here since the sum over the resources requiredPUc

u¼1 BðkÞu may become larger than the total B.

D. Modeling Long-Term Average Cell LoadThe models so far assumed scenarios with a fixed number

of users present in each cell. Key terms such as SINR, user

throughput, and load are defined as averages over a few

hundred milliseconds or a few seconds. These situations are

Fehske et al : Small-Cell Self-Organizing Wireless Networks

338 Proceedings of the IEEE | Vol. 102, No. 3, March 2014

Page 6: Small-Cell Self-Organizing Wireless Networks_06732895

instantaneous in the sense that during these short periodsusers can be assumed to be either active, i.e., a service re-

quest is being processed, or not, i.e., they do have data

pending.

If the behavior network is to be characterized over much

longer periods on the order of hours, clearly, a large number

of requests occur and are processed by the cells.

Corresponding network models then need to consider the

random nature of both, the instants of occurrences as well asrandom sizes of the requests. Such an approach leads to

queuing-theoretic formulations, which reflect the average

behavior of the network over longer periods and are suited to

design and evaluate mechanisms for slow adaptation of, e.g.,

antenna parameters at macro-BSs or on–off switching

strategies for small cells, which occur on the scale of half an

hour to a few hours. Those models also allow to evaluate the

stability of a network and are well suited to formulate large-scale long-term SON techniques. Mathematical models are

overviewed in the subsequent paragraphs, Section V then

discusses possible applications.

1) Long-Term Cell Capacity and Stability: In queuing

models, service requests are also referred to as data flows,

some of which may originate from the same user. For the

sake of analytical tractability, we consider only best efforttraffic where flows correspond to, e.g., individual web pages

requested during web browsing sessions. Thus, only elastic

data traffic is considered. Assume that flows are of random

size with mean S and occur randomly in time. Let �c denote

the mean number of requests per second that occur in cell c.

The product �cS then describes the average number of

megabits per second arriving at the cell and is called the

traffic intensity. Let the data per resource at location p in thecell be given by �RðpÞ. Instead of individual users, we consider

a spatial user distribution �c over the entire cell areaAc and

define the harmonic mean of the throughputs achievable in

the cell as

C :¼ZAc

�cðpÞBc �RðpÞ dp

264

375�1

: (9)

Term Bc refers to the total transmission resources available

to cell c (which may be equal to the total B) rather than the

resources Bu allocated to a single user u. For a so-calledM/G/1 processor sharing model, the cell is stable if and

only if this relation holds

�S G C: (10)

If (10) does not hold, on average, more service requests

arrive at the cell than depart from it, and their number

grows without bound. Since C is the smallest upper boundon the traffic intensity for which the cell is still stable, it is

referred to as the cell capacity. Moreover, the average

utilization of transmission resources over long periods, i.e.,

the long-term average cell load, is computed as the ratio of

traffic intensity and cell capacity, i.e., as

�� :¼ �S

C: (11)

2) Long-Term SINR and Rate: Computation of �� requires

information about the data rates �RðpÞ at all locations p. A

precise characterization, however, is not trivial. In case of

elastic traffic, the instantaneous resource utilization is

given by (5), i.e., the cell is either transmitting with full

power, or not at all. As a consequence, in case of C cells,

each flow could possibly see several out of 2C�1 different

SINRs values and corresponding data rates, according to allcombinations that some of the C � 1 interfering cells are

transmitting or not. A rigorous formulation of these effects

requires definition of a so-called coupled processor model

jointly for all C cells. Unfortunately, such a model renders

analytically intractable and different proposals for approx-

imating its behavior, specifically in the context when

wireless networks have been proposed [14], [18].

We discussed a similar situation already with respect tothe definition of the short-term SINR � in Section II-C2d,

and we will also use the same pragmatic approach from

[14] here. In the queuing theoretic formulation of a single

BS, say c, the utilization ��c also corresponds to the

probability that BS c is transmitting andP

d 6¼c ��dPrx;dðpÞdenotes the average interference power observed at

location p in cell c. A corresponding SINR at location pbased on average interference may then be defined as

��ðpÞ :¼ Prx;cðpÞPd6¼c ��dPrx;dðpÞ þ �

: (12)

Note that �� only approximates the average SINR at different

locations. In analogy to Section II-B, a corresponding

approximation of the average data rates is then obtained as

�RðpÞ :¼ R ��ðpÞð Þ (13)

where Rð�Þ denotes the rate function.

In analogy to the short-term average SINR �c, also the

long-term average ��c is given only implicitly when

assuming average interference [cf., (9) and (11)–(13)],

and cell loads must be obtained via a fixed point iteration,

as discussed in Section II-C2 above. The convergence of

this fixed point iteration approach follows from [19].

Fehske et al : Small-Cell Self-Organizing Wireless Networks

Vol. 102, No. 3, March 2014 | Proceedings of the IEEE 339

Page 7: Small-Cell Self-Organizing Wireless Networks_06732895

III . INTEGRATION OF SMALL CELLSINTO MACRONETWORKS

At a basic conceptual level small cells do not look differentfrom macrocells, especially from a terminal’s perspective.

Due to the expectedly much larger number of small cells,

however, their automated configuration is much more

relevant than it is for macrocells. A manual configuration of

each individual small cell would be costly and strongly ques-

tion its business case. Here, we outline only two configuration

mechanisms, namely, cell identification and automatic neigh-

bor relations. These are fundamental for any SON optimiza-tion technique as well as particularly challenging when large

numbers of small cells are deployed. For details on general

configuration mechanisms, we refer to [11] and [20].

A. Cell IdentificationFor easy and low-cost identification of BSs, in general,

each cell is assigned a nonunique local identifier, called

physical cell ID (PCI), in addition to a unique global cell

ID. If the PCI is not unique in a certain neighborhood, two

types of problems may arise:

• a PCI collision occurs if two neighboring cells have

the same PCI;

• a PCI confusion occurs if a cell has two neighborswith the same PCI.

Macrocells typically have about 10–20 macroneighbors

which must have mutually different PCIs. With LTE defining

a total number of 504 PCIs, the assignment without PCI

confusion can still be done manually. Unfortunately,

macrocells can have a much more complicated problem of

small-cell neighbors which renders proper assignment

complicated problem. Furthermore, dynamic addition andremoval of small cells due to more dynamic deployment or

on–off switching increase the need for automated PCI

assignment. For instance, the removal of a small cell may

connect previously isolated cells and risk both, PCI collision

as well as confusions. The resolution of the problem may

create further PCI problems requiring further PCI changes.

Adding a small cell seems to be simpler, as long as the

neighbor relations are known. These are discussed inSection III-B.

PCI assignment is described mathematically as a graph

coloring problem, which can be solved by graph-theoretical

methods [21].

Practical solutions reserve one set of PCIs for the small

cells leaving the rest for the macrocells. In many cases,

small cells are deployed in coordinated clusters with special

coordination inside the clusters. In this case, the reservoirof small-cell PCIs can be further split into smaller subsets.

Each cluster can be assigned a subset and allocates the PCIs

autonomously, whereas coordination of the subset alloca-

tion of the clusters is still needed. Those split options are

simple, however, they are obviously not optimal and may

lead to PCIs shortage in the presence of a large number of

small cells.

B. Automatic Neighbor RelationsA prerequisite for a handover or for any other interaction

between two neighboring cells is certainly that this neighbor

relation is known. While defining a neighbor relation is

trivial when schematically drawing a network of multiple

contiguous cells, it is much less obvious in a real network

where the coverage area of cells is influenced by reflection,

diffraction, and shadowing effects and is, in general,

noncontiguous. In addition, neighbor relations may occurand disappear with deployment changes, parameter changes,

or environmental changes. We can illustrate the problem of

unexpected (and unpredictable) neighbor relations even

with the widely used hexagonal network models. As a simple

mathematical description, we define two cells c and d as

neighbors if there is a significant areaA of locations p where

the received signals are similar, i.e.,

Prx;cðpÞ � Prx;dðpÞ�� �� G Dmax; with p 2 A: (14)

In Fig. 1, we have emulated the propagation effects by

increasing the standard deviation of the log-normal shadow-

ing in the macrocells (shadowing for the small cell is

neglected). Otherwise, we are using standard propagationassumptions as specified in [22], specifically an intersite

distance of 500 m, shadowing decorrelation distance of 50 m,

and tilts of 150. A single small cell is added to the area of

macrocell 1 in the center (brown color). We use Dmax ¼3 dB and declare the area A significant if it covers more

than 125 m2.

The left layout represents the 0-dB case (i.e., no

shadowing), and the right layout exemplarily shows 12-dBstandard deviation. With low standard deviation, the

macrocell 1 has exactly seven neighbors as expected, and

the small cell has one single neighbor, which is macrocell 1.

With increasing standard deviation, the number of neighbors

increases significantly. For instance, for 12-dB standard

deviation as in the right layout, four neighbors are detected in

the small cell, namely, cells 1, 26, 40, 21, and 41.

Those examples already show that planning the list ofneighbors can never be optimal since the propagation maps

used in the mathematical model in (14) are never known

with sufficient accuracy, e.g., coverage overshots are almost

impossible to predict. There will always be the risk of ending

up in too many neighbors with a generous planning, or with

missing neighbors with a more economic planning. Practi-

cally, the list of neighbors is permanently updated during the

operation. Measurements from the terminals are used forthis purpose. As soon as a UE reports a cell d which is not

known to the serving cell c as a neighbor, the serving cell ccan initiate a process to uniquely identify the cell d and

finally add it to the list of neighbors. Furthermore, an X2

interface would be set up between them. In parallel, it should

also be permanently monitored whether all entries in the list

of neighbors are still relevant, otherwise the list would grow

Fehske et al : Small-Cell Self-Organizing Wireless Networks

340 Proceedings of the IEEE | Vol. 102, No. 3, March 2014

Page 8: Small-Cell Self-Organizing Wireless Networks_06732895

and grow. Note that each neighbor would add additional

constraints when trying to coordinate with the neighbor-

hood. So, the tradeoff between adding and removing neigh-

bors is a significant challenge, in particular, for small cells.

IV. SMALL-SCALE SHORT-TERMSELF-ORGANIZATION

Reliable and efficient operation of small cells requires proper

handovers and cell reselections between a macrocell and a

small cell as well as between two small cells. Specifically, the

requirements are:

• proper mobility: avoidance of radio link failures(RLFs) and unnecessary handovers such as ping-

pongs (PPs);

• proper traffic steering: appropriate fractions of

traffic will be moved toward the small cells, and

potentially back to the macrocell.

The subsequent discussion on mobility and traffic steering

distinguishes two different small-cell operations: On the

same frequency band as the macrolayer (so-called intra-frequency small cells), or on a separate frequency layer

(so-called interfrequency small cells). In the latter case,

there is no interference between macrocells and small

cells. There are certainly multilayer deployments with

combinations of intrafrequency and interfrequency small

cells, however, for sake of simplicity, we only discuss these

two distinct cases.

A. Interfrequency Small CellsSubsequently, we discuss mobility robustness and traffic

steering procedures in case small cells and macrocells

operate on different frequency layers.

1) Mobility Robustness: Interfrequency mobility hand-

overs are typically triggered when the terminal reports that

the serving cell c is weak and another cell d on a different

frequency layer has sufficient signal strength. In order to

avoid reacting on measurement outliers we require that

the conditions be fulfilled for a certain period � called time

to trigger. The base station would trigger an interfrequencyhandover at time t0 if condition

McðtÞ G Tlow and MdðtÞ > Thigh (15)

is fulfilled for t0 � � G t G t0. Terms McðtÞ and MdðtÞ are

measurements of the signal strength Prx;cðpðtÞÞ, Prx;dðpðtÞÞfrom cells c and d available at time t, which are subject to

some delay and some error. Due to the missing interferencebetween cell c and cell d, thresholds Thigh and Tlow, as well as

� , are not too difficult to configure such that no RLFs or PPs

occur. In order to make an interfrequency measurement, the

terminal has to change its frequency oscillators, which costs

energy and time in which the terminal cannot receive data.

Hence, an important target is minimizing interfrequency

measurements, which is typically accomplished by

Fig. 1. Neighbor relations and shadowing.

Fehske et al : Small-Cell Self-Organizing Wireless Networks

Vol. 102, No. 3, March 2014 | Proceedings of the IEEE 341

Page 9: Small-Cell Self-Organizing Wireless Networks_06732895

performing them only when the signal of the serving cell isalready weak, i.e., close to Tlow.

2) Traffic Steering: Note that the rule in (15) does not

move any user to a small cell as long as the serving

macrocell has sufficient coverage, i.e., the UEs never

connect to a small cell if macrocoverage is sufficient. In

order to push traffic to small cells, we could try to adjust

parameters Tlow and Thigh to achieve some load balance.However, adjusting those thresholds to balance cell loads

and, at the same time, preserving mobility robustness with

only few interfrequency measurements is a rather

complicated task.

An alternative method is to observe and exchange load

information among neighboring cells and trigger dedicated

handovers for selected users individually. How to obtain

and communicate load information in terms of virtual loadand CAC has been introduced in Section II-C2. As long as

no load imbalance is detected, no further handover action

is required beyond the regular mobility handovers.

However, if load imbalance is detected, i.e., a well-loaded

cell has a less-loaded neighbor, the base station can initiate

additional handovers. In order to select appropriate users

for such a handover, the BS may follow internal rules such

as ‘‘keep fast users and voice users in the macrolayer’’ or‘‘prefer heavy users.’’ In order to guarantee a successful

handover, the cell needs to make sure that the candidate

UEs are well covered by the less-loaded neighbor. To this

end, the BS may request a single interfrequency measure-

ment from the candidate UEs, and afterwards execute only

noncritical handovers.

Fig. 2 illustrates the principles of a possible imple-

mentation. If the virtual load �c [as defined in (7)] of cell cexceeds a certain threshold �max it checks for interfre-

quency neighbor d with nonzero composite available

capacity defined in (8). A value CACd > 0 indicates that

this cell is willing to accommodate traffic. In case of load

imbalance, the overloaded cell c would set up a list L of

served candidate UEs which may potentially be handed

over to this target. This list may take into account UE

properties such as type of service, velocity, priority, etc.Listed UEs must perform an interfrequency measurement.

For the sake of simplicity, we assume that UE u reports the

SINR �du, expected after transfer to cell d. If �d

u is below a

threshold �min, the handover to cell d is considered too

risky for UE u. From the remaining UEs in the list, up to NUEs are handed over form cell c to d. Before checking the

load again, cell waits for a period �TS in order to give the

load measurements a fair chance to settle after executingthe handovers. Simulation results at the end of Section III

will show that already this simple mechanism achieves

very good performance.

For the sake of completeness, we would like to mention

that we have only discussed procedures for active UEs

which are connected to a cell. Fotiadis et al. [23] present

methods which allow to steer idle UEs as well.

B. Intrafrequency Small CellsThis section discusses mobility robustness and traffic

steering procedures when small cells and macrocells

operate on the same frequency layer.

1) Mobility Robustness: The major difference to the

previous case is clearly that cells of a single frequency layer

interfere with each other. Intrafrequency handovers are

triggered at time t0 if the target cell becomes better thanthe serving cell for a period of � , i.e., when condition

McðtÞ > MdðtÞ þ !c;d; with t0 � � G t G t0 (16)

is fulfilled. Note that (16) is in decibels. Interference

necessitates a much more exact handover timing becauseprior to the optimal handover time t0 the target cell is

hidden (i.e., well below) in the interference of the serving

cell and after t0 the serving cell will disappear in the

interference of the target cell. The choice of handover time

is controlled via parameter !c;d, called a cell individual

offset (CIO), which can be set individually for every cell

pair. Small values risk so-called too early handovers, since

the target may not be stable enough. Large values risk so-called too late handovers, since the source cell may

disappear before the handover is properly executed. Note

that there is another nonnegligible period of time required

for preparing the handover after it is triggered at t0, which

we do not consider here further.

a) User mobility in practice: Intuitively, a robust choice

of the CIO depends on the UE velocity, i.e., fast UEs require a

smaller CIO than slow ones. However, since McðtÞmeasuresthe signal strength Prx;cðpðtÞÞ, there is also a dependency on

the gradient of Prx;cð�Þ, specifically the pathloss change in the

direction of movement. Whereas totally uncorrelated user

mobility, i.e., independent Markovian walks, are often

assumed, the behavior in reality is strongly dictated by the

infrastructure. Buildings and streets typically force the

majority of the UEs to have a similar path of movement.

Fig. 2. Traffic steering algorithm.

Fehske et al : Small-Cell Self-Organizing Wireless Networks

342 Proceedings of the IEEE | Vol. 102, No. 3, March 2014

Page 10: Small-Cell Self-Organizing Wireless Networks_06732895

Uncorrelated motion would suggest that only a handoveroffset per user yields appreciable handover performance. In

practice, however, already a cell-pair-specific choice of the

CIO leads to significant improvements [24].

b) Event-triggered CIO optimization: Current state-of-

the-art (SotA) schemes adapt CIO based on observation of

the performance. UEs detect an RLF whenever a measure-

ment of the SINR � [cf., (2)] remains below a certain

threshold for a certain amount of time. PPs are defined as twoconsecutive successful handovers with a time period shorter

than �PP. We define a PP weight cpp, which configures how

severe PPs are considered compared to RLFs. A value

cpp ¼ 0 ignores PPs, and a value cpp ¼ 0 considers PPs to be

as severe as RLFs. Note that cpp may differ among operators

and depend on the concrete implementation. Each RLF is

analyzed via exchange of information between BSs via X2 to

determine the causing cell and whether the handover wastriggered too early or too late [11], [20]. Over a period long

enough to guarantee stable and sufficient statistics, each cell

c counts the number of too late handovers NðTLÞc;d , the number

of too early handovers NðTEÞc;d , and the number of PPs N

ðPPÞc;d

toward every neighbor d. After the period, CIOs are updated

according to the rule

!c;d ¼!c;d � 1 dB; if NðTLÞc;d > N

ðTEÞc;d þ cpp � NðPPÞ

c;d

!c;d ¼!c;d þ 1 dB; if NðTLÞc;d G N

ðTEÞc;d þ cpp � NðPPÞ

c;d :

(17)

Note that we focus on the scientific essence of the algorithm,

omitting further details for the sake of clarity. Concretemechanisms to analyze RLFs and exchange corresponding

information among BSs are defined in 3GPP standardization

and, together with (17), referred to as mobility robustness

optimization (MRO). Information exchange between BSs is

practically achieved via the so-called X2 interface [25], and

MRO can thus readily be applied to small cells as long as an

X2 interface is available.

c) Knowledge-based CIO optimization: Looking at theequations above, theoretically, optimal values of the CIOs

would not only require the exact knowledge of the

propagation maps PRxðpÞ, but also of the user paths pðtÞ.We would like to emphasize again that knowledge of the

velocity is not sufficient at all. Even if all this information

were well known, finding an optimal solution would be

highly challenging, and no approach is known so far [24].

The previously described practical solution assumes that, dueto infrastructure such as streets, paths pðtÞ will be very

similar for many users, and there is some stationarity in the

users’ movement.

d) Handovers between different cell types: Mobility

between a macrocell and a small cell is more challenging

than between macrocells. Whereas the pathloss situation is

typically quite symmetric between two macrocells, small

cells do appear and, more importantly, disappear much

quicker than macrocells along motion paths. Fig. 3 illustrates

the received powers from PRx;kðpÞ along a straight line

between two macrocells and a small cell. Due to their larger

distance, the slopes corresponding to macrocells are rather

flat at the intersections, whereas the small-cell slope is much

steeper. For this reason, outbound handovers, i.e., from small

cells to macrocells, are typically riskier and sensitive to RLFs.This behavior is also apparent in the subsequent simulation

results.

e) Numerical results: As before, we consider a hex-

agonal macroenvironment (intersite distance 500 m) with

some small cells. Infrastructure is taken into account by

letting 75 users move along streets at a speed of 30 km/h.

In addition, there are 600 slow users moving at a speed of

3 km/h into random directions within hotspots with radiiof 75 m around the small cells. Fig. 4 illustrates the scenario,

which has been constructed to pronounce mobility robust-

ness effects. Initially, a default !c;d ¼ 3 dB is used for all cell

pairs. During the simulation, we use standardized MRO

mechanisms to analyze RLFs and PPs, and then modify !c;d

according to (17). Fig. 5 depicts the number of RLFs and PPs

Fig. 3. Pathloss slopes for macrocells and small cells.

Fig. 4. Layout for MRO simulations.

Fehske et al : Small-Cell Self-Organizing Wireless Networks

Vol. 102, No. 3, March 2014 | Proceedings of the IEEE 343

Page 11: Small-Cell Self-Organizing Wireless Networks_06732895

per UE per minute after the algorithm has converged. Counts

for outbound handovers (a small cell to a macrocell),

inbound handovers (a macrocell to a small cell), and the total

of all handovers are shown separately.Without MRO, there is a significant number of RLFs and,

as previously indicated, there are many more outbound

problems than inbound problems (a factor of about 3).

Switching MRO on significantly reduces, in particular, the

outbound RLFs as well as the number of PPs. Increasing the

importance of PPs by increasing cPP reduces their number

while only slightly sacrificing on the RLFs side. We can

conclude that the classical MRO can significantly reducehandover problems for small-cell deployments as well,

although being originally designed for macrocells. Interest-

ingly, inbound and outbound problems are balanced when

MRO is in use.

We further emphasize that SotA MRO mechanisms

operate on quite limited information, For instance, they do

not distinguish between individual users and only work on

the majority of users in a cell. Current standardizationactivities aim at improving measurement procedures,

especially for small-cell deployments [26].

2) Traffic Steering: It appears quite challenging to move

traffic between macrocells and small cells on the same

frequency layer since most of the degrees of freedom are

required for mobility robustness, especially in cases with

fast velocities and steep shadowing slopes. However, thereare also scenarios with limited mobility challenges where

the cell boundary might be shifted toward highly loaded

cells. Typically, it is desired to push traffic into small cells.

Shifting the cell boundaries is often referred to as range

expansion. Looking at (16), we could move a boundary by

modifying CIOs: Increasing !c;d moves traffic toward cell c,

and decreasing !c;d moves traffic toward cell d.

In analogy to MRO, procedures for modifying CIOs dueto traffic are called mobility load balancing (MLB), a

crucial part of which is the exchange of load information.

Since both MRO and MLB procedures modify the same

parameters, their interaction must be handled with care.

Consider, for instance, outbound mobility: it is problematicdue to the steep pathloss slope of the small cells. An MRO

algorithm typically compensates for the steep pathloss slope

of small cells by decreasing!c;d in small cell c. Unfortunately,

this action decreases the range of small cells and pushes the

traffic to the macrocell, which may counteract MLB

activities.

a) Numerical results: The following simulation setup

reflects a typical small-cell deployment. We assume threenetwork layers: one 800-MHz macrolayer with intersite

distance of 500 m and 10-MHz bandwidth, a cosited

macrolayer at 2600 MHz with 20-MHz bandwidth, and a

layer of small cells sharing the 20-MHz with the macrolayer

at 2600 MHz. Twenty one hundred UEs (i.e., 100 per

macroarea) are moving in the network. Sixty six percent of

them are forced into circular hotspot areas of radius 60 m.

Four small cells are placed inside every hotspot area. The restis uniformly distributed in the network. Since we will focus

on traffic steering and load balancing, the users are moving at

3 km/h into random direction, i.e., we do not have significant

mobility problems in this scenario. Each user downloads a

file of size 1 MB, spends some time in idle mode, before

downloading the next file. Active users are served in a

resource fair way. Fig. 6 illustrates the network layout, and

colored areas are served by small cells, which do not neces-sarily cover entire hotspots.

Fig. 7 shows simulation results. The reference case ‘‘no

TS’’ is the simplest traffic steering policy where UEs are

instructed to prioritize the high capacity frequency (here

2600 MHz) when being in idle mode, i.e., all users will set

up their calls at 2600 MHz unless the coverage at 2600 MHz

Fig. 5. Mobility robustness optimization for HetNet.

Fig. 6. Layout for traffic steering simulations.

Fehske et al : Small-Cell Self-Organizing Wireless Networks

344 Proceedings of the IEEE | Vol. 102, No. 3, March 2014

Page 12: Small-Cell Self-Organizing Wireless Networks_06732895

is too weak. In simulation scenario ‘‘TS,’’ we move active UEs

between 2600 and 800 MHz according to the intercell

steering procedures described in Section IV-A. Finally, sce-nario ‘‘TS’’+‘‘MLB’’ adds MLB as described in Section IV-B,

i.e., the boundaries within 2600 MHz, in particular between

the macro cell and the small cell, are shifted via modifying

the CIOs. The results are given in terms of the average user

throughput, UE distribution, and the virtual load. The virtual

load is defined in (7) using a target bitrate of Ru ¼ 1 Mb/s.

In the reference case all users use the 2600-MHz layer.

Due to the small intersite distance, even this higher frequen-cy provides full coverage and 800 MHz is almost empty. Near

60% of the users are connected to small cells, which corre-

sponds to the fraction of the UEs located in small-cell areas.

In the TS scenario, UEs are pushed to the 800-MHz layer.

The number of users in macros at 800 MHz is smaller than at

2600 MHz due to smaller bandwidth, however, the

balancing effect is illustrated by the loads. As a consequence,

the user throughput averaged over all layers significantlyincreases by about 47%. The small-cell load remains moder-

ate, although the number of UEs is still large. Note that the

mechanism cannot push more traffic to the small cells since

the area served by small cells is limited.

Adding MLB mechanisms artificially increases the small-

cell areas by raising the CIOs between macro cells and small

cells on the 2600-MHz layer. The number of UEs served by

the small cell, the overall load balance, and the overallaverage throughput are all slightly increased. Further range

expansion does not appear feasible in this scenario, due to

very high intercell interference (a macro cell to a small cell)

in the expanded areas, which raises occurrence of RLFs and

lowers throughput.

C. Enhanced Intercell Interference CoordinationWe have observed that in a scenario as mentioned

above the offloading effect toward small cells is limited by

the macrointerference. Without this interference, there ismore potential to push traffic to the small cells (certainly

unless the small-cell deployment would be so optimal that

their natural coverage already led to a significant load).

Interference reduction may be achieved by orthogonalizing

macrocell and small-cell transmission, e.g., in time by

reserving individual time slots for small cells. Initially, this

approach worsens the situation in the macrocells since their

transmission resources are cut. Under certain circumstances,however, this investment pays off due to two effects. First,

the more small cells are deployed, the more cells benefit (via

reduced interference) from the investment of a single

macrocell. Second, in case of less interference from the

macrocells, MLB mechanisms are able to increase CIOs

much further and expand small-cell service into hot zone

areas.

D. Small-Cell Coverage Optimization AspectsFor the sake of completeness, we would like to list a

number of further aspects which have to be taken into ac-

count when deploying small cells in a macronetwork.

• It is often assumed that small cells must use as much

power as possible in order to attract as much traffic as

possible. In some cases, however, small-cell power

must be adjusted more carefully. The most promi-nent example is given by closed subscriber groups

where UEs outside a given location (e.g., a house or

an apartment) are not admitted to a small cell. In this

case, coverage should not leak outside this location.

Small cells in heavily crowded areas should also not

use maximum power to avoid overload.

• Similarly, the target values for the uplink power

control should not threaten the uplink of a macro-cell nearby.

• Coverage motivated small cells (e.g., indoor cells)

may relax the coverage requirements for macrocells.

Macrocells may exploit this aspect, e.g., by using

steeper tilts.

V. LARGE-SCALE LONG-TERMSELF-ORGANIZATION

Large-scale long-term C-SON algorithms optimize certain

configuration management parameters of a collection of a

cluster of neighboring cells. For SotA hardware and data base

technology, the time cycle for data exchange between the

network management and the C-SON entity is not faster than

15 min. Parameter changes happen on an hourly or even daily

basis. Typical C-SON use cases for small-cell deploymentsare the so-called coverage and capacity optimization (CCO)

and the energy saving management (ESM), both of which are

addressed by examples given in this section.

Long- and short-term techniques also have overlapping

use cases. In particular, MRO and MLB, as discussed in

Section IV for short-term adaptations, may be addressed

by long-term techniques as well. In this case, C-SON and

Fig. 7. Traffic steering simulation results.

Fehske et al : Small-Cell Self-Organizing Wireless Networks

Vol. 102, No. 3, March 2014 | Proceedings of the IEEE 345

Page 13: Small-Cell Self-Organizing Wireless Networks_06732895

D-SON also optimize the same parameters, for example, theCIOs introduced in (16). For this reason, a clean coexistence

of algorithms is not necessarily guaranteed and indepen-

dently running implementations may suggest conflicting

parameter changes, which necessitates a coordination entity

to resolve conflicts. Clearly, coordination may already be

required between SON algorithms operating on the time

same scale. One example is the conflict between short-term

CIO adaptations due to MRO targets on the one hand andMLB targets on the other hand, as outlined in Section IV-B2.

Various coordination approaches are found in [27] and [28].

In case of long-term adaptations, a joint handling of

several different use cases may be achieved not only via a

separate coordination process, but also through formulation

of a joint optimization problem [29]. In fact, one advantage of

long-term techniques lies in the availability of tractable

mathematical models (outlined in Section II-D), which allowto obtain parameter changes as solutions of certain

optimization problems. Such solutions are feasible in

practice since long-term averages of, e.g., receive powers

can be obtained with greater reliability than instantaneous

values. Furthermore, since only the long-term averages of

KPIs are optimized, short-term dynamic effects such as

mobility have lesser impact than in case of short-term

adaptations.On the downside, a major disadvantage of centralized

approaches, specifically with regard to small-cell deploy-

ments, is that they may become impossible to solve when a

very large number of BSs is to be jointly optimized.

In Section V-B, we briefly outline how to formulate

multicell optimization problems based on the models

previously discussed in Section II-D. Section V-C discusses

the application of the framework to jointly handle the usecases CCO and ESM mentioned above.

A. Input DataIn principle, a C-SON algorithm may observe the per-

formance of one or multiple D-SON implementations over

multiple layers, vendors, and technologies. Typical inputs are

different varieties of counters pertaining to performance and

failure management. Simple examples are the numbers oftoo early or too late handovers, as well as the number of

handover PPs introduced in Section IV-B1. These inputs are

commonly available on a per cell basis.

In contrast, information on a per user basis, so-called call

trace data, is becoming increasingly important. Call traces

are collected directly from the user terminals and contain,

among others, information related to the pathlosses with

respect to individual BSs, the SINR, and, most importantly,the position of the UEs.

B. Problem FormulationAccording to the underlying queuing-theoretic frame-

work, many QoS-related performance measures are mono-

tonic functions of the average cell load ��c defined in (11). For

instance, the throughput per resource experienced on

average by users at location p in cell c is given byð1� ��cÞ�RðpÞ. Similar expressions exist for the mean delay

in the cell or the average number of service requests that are

present at any point in time. These relations can be exploited

for QoS optimization. Due to monotonicity, reducing the

average cell load simultaneously improves all of these

performance measures, and, for this reason, the cell load is

well suited as a target for optimization. Multiple cells are

readily handled via a multicell objective function thatdepends on the average loads of many cells. A popular

choice is a so-called �-optimal objective [29]–[31]

%�ð��1; . . . ; ��CÞ ¼PC

c¼1

ð1� ��cÞ1��

�� 1; for � 6¼ 1,PC

c¼1� logð1� ��cÞ; for � ¼ 1

8<:

(18)

which must be minimized for some fixed parameter � � 0.Since ��c denotes the average resource utilization, term

1� ��c represents the average amount of resources

available at cell c. Thus, minimizing %� for � ¼ 0

maximizes the sum of available resources over all cells.

In case � ¼ 1, the corresponding product is maximized,

which leads to a proportional fair treatment of the

individual cells. For �!1, the minimum of the available

resources 1� ��c is maximized, which leads to egalitariansolutions where all loads ��c are equal.

The definition of the long-term average cell load ��c in

(11) incorporates many key system aspects: the cell area

Ac, the available transmission resources Bc, as well as the

location-specific traffic demand intensity �S�cðpÞ and

average data rates �RðpÞ provided by the individual radio

links. While the bandwidth and cell area may be seen as

free parameters,1 the average data rates �R depend on otherterms, such as transmit powers, antenna downtilts, or

other antenna parameters. Clearly, changing these char-

acteristics will also change the load. If we denote a vector

of such parameters and some feasible set by x and X ,

respectively, we can write %�ðxÞ ¼ %�ð��1ðxÞ; . . . ; ��CðxÞÞ,and a general parameter optimization problem is given by

minimizez

%�ðxÞ subject to x 2 X : (19)

The relation between parameters x and average load ��c is

very complicated in general and analytically hardly tractable.

For this reason, solving (19) is often accomplished via

heuristic search strategies such as, e.g., the Nelder and Mead

algorithm or Powell’s algorithm [32], [33].

1In practical setups, the serving areas Ac are set via handoverparameters, specifically the CIOs. When optimizing a model, as we targethere, advantageous cell areas are found first, and then translated intocorresponding CIO values.

Fehske et al : Small-Cell Self-Organizing Wireless Networks

346 Proceedings of the IEEE | Vol. 102, No. 3, March 2014

Page 14: Small-Cell Self-Organizing Wireless Networks_06732895

In practical systems, the information required about thereceive powers Prx;cðpÞ and average traffic demands �S�cðpÞat different locations p may be obtained from collecting call

traces over longer time periods of hours or days. Once this

information is available, SINRs, throughputs, and average

loads can be estimated according to the framework in

Section II-D. Improved parameter settings may then be

obtained from the model via search methods mentioned

above and then applied in the network. The clear advantageof this approach is that parameter settings are tested in the

model and need not be iteratively tested and improved in a

live network until convergence is completed, as is required

for the short-term methods in Section IV.

C. Numerical StudiesParameter vector x may contain different types of

parameters. In case of small-cell deployments, interestingchoices for x are the serving areas Ac associated with each

BS. Their adaptation allows small cells to extend coverage

into hotspots and serve more traffic. In addition, adaptation

of the transmission resources Bc available to individual small

cells as well as to macrocells can mitigate the degrading

effects of intercell interference, especially in areas where,

due to adjustment of cell areas, the serving cell does not

provide the strongest receive power.We study a scenario with three colocated macrocells and

seven small cells, each located close to a traffic hotspot

within the coverage area of the macrocells. Additional

macrocells provide a desired level of interference. Fig. 8

illustrates the deployment. We divide the amount of trans-

mission resources, e.g., the overall frequency bandwidth Binto 63 sub-bands, in each of which only individual macro-

BSs, or certain groups of small BSs, or combinations of thesemay be used for transmission. In LTE systems, a sub-band

consists of one or more PRBs and can be translated into an

PRB-based interference coordination scheme. Equivalently,

the bandwidth fragmentation can also be interpreted as a

time-based coordination scheme on small time scales. The

individual widths of the sub-bands as well as the serving areas

of all BSs are to be optimized according to (19). For a total of

ten BSs (seven small cells and three macrocells), the param-

eter vector is thus given as x ¼ fB0; B1;. .

.; B63;A1; . . . ;

A10g. We choose the objective function %1, i.e., bandwidths

and areas are chosen to maximize the product of availableresources in all cells. The concrete choice of groups of small

cells and the algorithmic solution of (19) in this case are

discussed in [34] and not further detailed here.

1) Use Case Coverage and Capacity Optimization: Some

major results with regard to capacity optimization are

illustrated in Fig. 9. Depicted are the QoS, measured as

the 5, 50, and 95 percentiles of the average user throughputtaken over the entire analysis area. The percentiles are shown

for two cases: first, for a reference case, denoted by full

frequency reuse (FFR), where all cells use only the common

band B0 and cell areas are assigned according to maximum

receive power; and second, after optimized cell areas and

bandwidths according to solutions of (19) are assigned,

which is denoted by OPT.

Throughput statistics are taken over the analysis areaonly. With increasing total traffic demand, the ratio between

regular and hotspot traffic densities increases from 1 : 1 to

1 : 30. We observe that for low total demand all percentiles

are at acceptable values, which means all locations are

serviced with acceptable bit rates. For increasing total traffic,

user throughputs decrease. In particular, the 5-percentile

(representing critical users at the cell edges) drops quite

dramatically by more than two orders of magnitude fromaround 10 Mb/s to below 100 kb/s. The decrease in user

throughput may be mitigated for all percentiles by adapting

cell areas and bandwidths accordingly.

2) Use Case Energy Saving Management: Besides alloca-

tion of suitable bandwidths, the optimization results also

indicate which small cells may be switched off to conserveFig. 8. Layout for C-SON traffic steering simulations [34].

Fig. 9. Quantiles of user throughput under FFR and under optimized

resource allocation and cell area assignment [34].

Fehske et al : Small-Cell Self-Organizing Wireless Networks

Vol. 102, No. 3, March 2014 | Proceedings of the IEEE 347

Page 15: Small-Cell Self-Organizing Wireless Networks_06732895

energy. Each cell only transmits in assigned bands and the

total amount of transmission resources corresponds to the

sum of the size of these bands. Feasible solutions to (19)may suggest that the total bandwidth allocated in

individual groups of small cells is zero (which also forces

their area to be zero). In this case, these cells are

effectively not transmitting and are switched off. Fig. 10

illustrates the application of the algorithm to the same

deployment in an ESM use case. Since the traffic intensity

varies hourly over the course of one day (blue curve), not

necessarily all small cells are required for service, andenergy may be saved by switching some of them off. If the

optimization is performed for each traffic intensity, i.e., on

an hourly basis, the number of active small cells may be

chosen as the number of small cells with nonzero

bandwidth assigned (red curve).

VI. SUMMARY

Although such small access points are conceptually equiva-

lent to conventional cellular base stations in many ways, theexpected large number of small cells as well as their much

more dynamic unplanned deployment raises a variety of

challenges in the area of network management. This paper

overviews these challenges and argues that they can only be

met by SON techniques.

We outline a mathematical framework that allows to

numerically assess the performance of networks featuring a

large number of small access points as well as to developalgorithms for their self-organization. Particular attention is

paid to the time scales on which management processes

happen. These are, in particular, on the order of either frac-

tions of seconds up to several seconds or on the order of

fractions of hours up to several hours. Both time scales are

not appropriately represented by very detailed link level

models or by simpler snapshot-based tools.

Besides reviewing basic network configuration tech-niques, we introduce elementary network management

techniques for mobility robustness optimization and mobility

load balancing, which operate on the shorter scales of

seconds. We further discuss techniques for coverage and

capacity optimization and energy saving management which

operate on larger scales of hours. We highlight the specific

challenges due to small-cell deployments in all discussions.h

Acknowledgment

The authors would like to thank H. Martikainen,

P. Fotiadis, M. Polignano, and P. Zanier of Nokia Siemens

Networks for providing the simulation results in Section IV

as well as J. Bartelt of Dresden University of Technology for

providing the simulation results in Section V.

REF ERENCE S

[1] V. Chandrasekhar, J. Andrews, andA. Gatherer, ‘‘Femtocell networks: A survey,’’IEEE Commun. Mag., vol. 46, no. 9, pp. 59–67,Sep. 2008.

[2] Small Cell Forum, ‘‘Small cellsVWhat’s thebig idea?’’ Tech. Rep. 6295097, Apr. 2012.

[3] H. Claussen, L. T. W. Ho, and F. Pivit,‘‘Leveraging advances in mobile broadbandtechnology to improve environmentalsustainability,’’ Telecommun. J. Australia,vol. 59, no. 1, pp. 1–18, 2009.

[4] R. Razavi and H. Claussen, ‘‘Urban small celldeployments: Impact on the network energyconsumption,’’ in Proc. IEEE WirelessCommun. Netw. Conf. Workshops, Apr. 2012,pp. 47–52.

[5] R. Webb, T. Wehmeier, and K. Dyer,‘‘Small Cells 2012 Integration andOptimisation,’’ Mobile Europe, Tech.Rep., 2012.

[6] EARTH, 2009. [Online]. Available:www.ict-earth.eu

[7] G. Auer, V. Giannini, C. Desset, I. Godor,P. Skillermark, M. Olsson, M. Imran,D. Sabella, M. Gonzalez, O. Blume, andA. Fehske, ‘‘How much energy is needed to

run a wireless network?’’ IEEE WirelessCommun., vol. 18, no. 5, pp. 40–49, Oct. 2011.

[8] Cisco, ‘‘Cisco visual networking index:Global mobile data traffic forecast update2012–2017,’’ Cisco Inc., Tech. Rep., 2012.

[9] RadioOpt GmbH, ‘‘RadioOpt,’’ 2012.

[10] N. Bhas, ‘‘Mobile data offload & onload,’’Juniper Research, Tech. Rep., Apr. 2012.

[11] S. Hamalainen, H. Sanneck, andC. Sartori, Eds., LTE Self-OrganizingNetworks. New York, NY, USA:Wiley, 2012.

[12] S. Borst, ‘‘User-level performance ofchannel-aware scheduling algorithms inwireless data networks,’’ IEEE/ACM Trans.Netw., vol. 13, no. 3, pp. 636–647, Jun. 2005.

[13] I. Viering, M. Dottling, and A. Lobinger, ‘‘Amathematical perspective of self-optimizingwireless networks,’’ in Proc. Int. Conf.Commun., 2009, DOI: 10.1109/ICC.2009.5198628.

[14] A. J. Fehske and G. P. Fettweis, ‘‘Aggregationof variables in load models for cellular datanetworks,’’ in Proc. Int. Conf. Commun.,Ottawa, ON, Canada, 2012, pp. 5102–5107.

[15] I. Siomina and D. Yuan, ‘‘Analysis of cell loadcoupling for LTE network planning and

optimization,’’ IEEE Trans. Wireless Commun.,vol. 11, no. 6, pp. 2287–2297, Jun. 2012.

[16] P. Mogensen, W. Na, I. Z. Kovacs,F. Frederiksen, A. Pokhariyal, K. I. Pedersen,T. Kolding, K. Hugl, and M. Kuusela, ‘‘LTEcapacity compared to the Shannon bound,’’ inProc. IEEE Veh. Technol. Conf. Spring, 2007,no. 1, pp. 1234–1238.

[17] R. Kwan, R. Arnott, R. Paterson,R. Trivisonno, and M. Kubota, ‘‘On mobilityload balancing for LTE systems,’’ in Proc.IEEE 72nd Veh. Technol. Conf. Fall, Sep. 2010,DOI: 10.1109/VETECF.2010.5594565.

[18] T. Bonald, S. Borst, N. Hegde, andA. Proutiere, ‘‘Wireless data performance inmulti-cell scenarios,’’ in Proc. Joint Int. Conf.Meas. Model. Comput. Syst., Jun. 2004, vol. 32,no. 1, pp. 378–380.

[19] R. D. Yates, ‘‘A framework for uplink powercontrol in cellular radio systems,’’ IEEE J. Sel.Areas Commun., vol. 13, no. 7, pp. 1341–1347,Sep. 1996.

[20] Third Generation Partnership Project(3GPP), ‘‘TR36.300 overall description,’’Tech. Rep., 2013.

[21] F. Ahmed, O. Tirkkonen, M. Peltomaki,J.-M. Koljonen, C.-H. Yu, and M. Alava,‘‘Distributed graph coloring forself-organization in LTE networks,’’

Fig. 10. Number of active small cells according to traffic demand

during the course of a day [34].

Fehske et al : Small-Cell Self-Organizing Wireless Networks

348 Proceedings of the IEEE | Vol. 102, No. 3, March 2014

Page 16: Small-Cell Self-Organizing Wireless Networks_06732895

J. Electr. Comput. Eng., vol. 2010, Jan. 2010,DOI: 10.1155/2010/402831.

[22] Third Generation Partnership Project (3GPP),‘‘Further advancements for {E-UTRA}physical layer aspects,’’ TR36.814, 2010.

[23] P. Fotiadis, M. Polignano, L. C. Gimenez,I. Viering, C. Sartori, A. Lobinger, andS. Redana, ‘‘Multi-layer traffic steering:RRC idle absolute priorities and potentialenhancements,’’ in Proc. IEEE Veh. Technol.Conf. Spring, Dresden, Germany, Jun. 2013,DOI: 10.1109/VTCSpring.2013.6692643.

[24] I. Viering, B. Wegmann, A. Lobinger,A. Awada, and H. Martikainen, ‘‘Mobilityrobustness optimization beyond Dopplereffect and WSS assumption,’’ in Proc. 8th Int.Symp. Wireless Commun. Syst., Nov. 2011,pp. 186–191.

[25] Third Generation Partnership Project (3GPP),‘‘TR36.423 X2 application protocol (X2AP),’’Tech. Rep., 2013.

[26] Third Generation Partnership Project (3GPP),‘‘TR 36.839 mobility enhancements inheterogeneous networks,’’ Tech. Rep., 2013.

[27] L. Schmelz, M. Amirijoo, A. Eisenblaetter,R. Litjens, M. Neuland, and J. Turk, ‘‘Acoordination framework for self-organisationin LTE networks,’’ in Proc. IFIP/IEEE Int.Symp. Integr. Netw. Manage., Dublin, Ireland,May 2011, pp. 193–200.

[28] T. Bandh, R. Romeikat, H. Sanneck, andH. Tang, ‘‘Policy-based coordination andmanagement of self-organizing-network(SON) functions,’’ in Proc. IFIP/IEEE Int.Symp. Integr. Netw. Manage., Dublin, Ireland,May 2011, pp. 827–840.

[29] A. Fehske, H. Klessig, J. Voigt, andG. Fettweis, ‘‘Concurrent load-awareadjustment of user association and antennatilts in self-organizing radio networks,’’IEEE Trans. Veh. Technol., vol. 62, no. 5,pp. 1974–1988, Jun. 2013.

[30] J. Mo and J. Walrand, ‘‘Fair end-to-endwindow-based congestion control,’’

IEEE/ACM Trans. Netw., vol. 8, no. 5,pp. 556–567, Oct. 2000.

[31] H. Kim, G. de Veciana, X. Yang, andM. Venkatachalam, ‘‘Distributedalpha-optimal user association and cell loadbalancing in wireless networks,’’ IEEE ACMTrans. Netw., vol. 20, no. 1, pp. 177–190,Feb. 2012.

[32] J. A. Nelder and R. Mead, ‘‘A simplex methodfor function minimization,’’ Comput. J., vol. 7,no. 4, pp. 308–313, Jan. 1965.

[33] M. J. D. Powell, ‘‘An efficient method forfinding the minimum of a function of severalvariables without calculating derivatives,’’Comput. J., vol. 7, no. 2, pp. 155–162,Feb. 1964.

[34] J. Bartelt, J. Voigt, A. Fehske, H. Klessig, andG. Fettweis, ‘‘Joint bandwidth allocationand small cell switching in heterogeneousnetworks,’’ in Proc. IEEE Veh. Technol. Conf.Fall, Las Vegas, NV, USA, Sep. 2013,DOI: 10.1109/VTCFall.2013.6692255.

ABOUT T HE AUTHO RS

Albrecht J. Fehske received the Dipl.-Ing. degree

and the Ph.D. degree in electrical engineering from

the Technical University of Dresden, Dresden,

Germany, in 2007 and 2014, respectively.

He is a Research Group Leader at the Vodafone

Chair at the Technical University of Dresden.

During his studies, he worked with the Mobile

and Portable Radio Research Group, Virginia

Polytechnic Institute and State University,

Blacksburg, VA, USA. He also worked at Vodafone

Group R&D, Newbury, U.K. His research focuses on energy efficiency as

well as self-organization and optimization in cellular networks. In 2013, he

cofounded AIRRAYS, a technology startup developing adaptive antenna

technology for mobile communications.

Ingo Viering received the Dipl.-Ing. degree in

electrical engineering from the University of

Technology Darmstadt, Darmstadt, Germany, in

1999 and the Dr.-Ing. degree in electrical engi-

neering from the University of Ulm, Ulm, Germany,

in 2003.

He spent a year as a Researcher with the

Telecommunications Research Center Vienna

(FTW), Vienna, Austria, in 2002, where he conducted

early measurements of the multiple-input–multiple-

output (MIMO) channel. He is a cofounder and CEO of Nomor Research

GmbH, Munich, Germany. Furthermore, since 2007, he has been a Senior

Lecturer atMunich University of Technology,Munich, Germany. His research

interests are system aspects of current and future communication systems,

including the detailed interaction of the multitude of features. He has filed

around 70 patents, published more than 60 scientific papers, and he is

actively contributing to 3GPP.

Dr. Viering was awarded the VDE Award in 2009 for the achievements

of Nomor Research.

Jens Voigt received the Diploma and the Ph.D.

degree from the Department of Electrical and

Computer Engineering, Dresden University of

Technology, Dresden, Germany, in 1995 and

2001, respectively.

In 2000, he cofounded Radioplan, a Dresden-

based specialist in cellular network simulation and

automatic cell optimization solutions, which was

acquired by Actix in 2006. Within Actix he worked

as a Principal Research Engineer and Project

Manager in various research and development projects. In 2013, Actix

was acquired by Amdocs, where he is now working as a Senior Technical

Architect. He has coauthored more than 30 publications and holds

multiple international patent families. With over 15 years of experience in

radio-frequency (RF) engineering, cellular network technology as well as

optimization algorithm research and product specification, his research

interests include various aspects of self-organizing network (SON)

technology, geolocation techniques, cellular network simulations, and

ray-tracing-based wireless channel simulations.

Cinzia Sartori received the Diploma in electrical

engineering from the University of Pavia, Pavia,

Italy.

She is a 5G Researcher at Nokia Solutions

Networks, Munich, Germany. Until mid-2013, she

was a SON Research Area Manager at Nokia

Siemens Networks, where she was responsible

for the self-organizing network (SON) research

field in radio research. She is the coeditor of LTE

Self-Organizing Network (New York, NY, USA:

Wiley, January 2012). Previously, she headed the RAN System Architec-

ture Network Telecom team in Nokia Siemens Networks (2007–2009),

which included 2G, 3G, WiMaX, and LTE call processing. Previously, she

worked for GTE and Siemens where she led the R&D team for SS7 and

Radio Resource Management software development.

Fehske et al : Small-Cell Self-Organizing Wireless Networks

Vol. 102, No. 3, March 2014 | Proceedings of the IEEE 349

Page 17: Small-Cell Self-Organizing Wireless Networks_06732895

Simone Redana received the M.Sc. and Ph.D.

degrees in electrical engineering from the Poli-

tecnico di Milano, Milan, Italy, in 2002 and 2005,

respectively.

In 2006, he joined Siemens Communication,

Milan, Italy, which merged with Nokia Networks in

2007 to become Nokia Siemens Networks and, in

2013, became Nokia Solutions and Networks.

Since 2008, he has been with NSN Germany,

where he is currently Head of Radio Research

team in Munich. He contributed to the relay concept design in the

European Union project WINNER II and the Eureka Celtic project

WINNER+ as well as he led the work package on advanced relay concept

design in the European Union project ARTIST4G. He contributed to the

business case analysis of relay deployments and to the standardization of

Relays for Long-Term Evolution (LTE) Release 10. He led research and

standardization projects on self-organizing network (SON) for LTE

Release 11. His current research interests are on new methods to access

more spectrum for international mobile telecommunications.

Gerhard P. Fettweis (Fellow, IEEE) received the

Dipl.-Ing. and Ph.D. degrees in electrical engineer-

ing under the guidance of Prof. H. Meyr from

Aachen University of Technology (RWTH), Aachen,

Germany, in 1986 and 1990, respectively.

From 1990 to 1991, he was a Visiting Scientist

at the IBM Almaden Research Center, San Jose, CA,

USA. From 1991 to 1994, he was responsible for

signal processor development at TCSI Inc.,

Berkeley, CA, USA. Since September 1994, he has

held the Vodafone Chair at the Technical University of Dresden (TUD),

Dresden, Germany. In 2009, he initiated the Leading Edge Cluster Cool

Silicon and led it until 2010. Since 2011, he has led the Collaborative

Research Center (CRC) Highly Adaptive Energy-efficient Computing

(HAEC) at TUD as its speaker. As chairman of TUDs focus area Information

Technology and Microelectronics, he coordinates the proposal for a

Cluster of Excellence within the German Excellence Initiative Center for

Advancing Electronics Dresden (cfAED) addressing different paths to

advance electronics beyond current/upcoming limitations of CMOS

technology.

Prof. Fettweis has been an elected member of the IEEE Solid State

Circuits Society Board (Administrative Committee) since 1999. He was

elected as IEEE Distinguished Lecturer for Solid State Circuit Society

(2009–2010) and for the Vehicular Technology Society (2011–2013). He

also serves on several supervisory boards, and on advisory committees

of companies and research institutes.

Fehske et al : Small-Cell Self-Organizing Wireless Networks

350 Proceedings of the IEEE | Vol. 102, No. 3, March 2014