Load Balancing in Downlink LTE Self-Optimizing Networks

Abstract

In this paper we present system level simulation results of a self-optimizing load balancing algorithm in a long-term-evolution (LTE) mobile communication system. Based on previous work, we evaluate the network performance of this algorithm that requires the load of a cell as input and controls the handover parameters. We compare the results for different simulation setups: for a basic, regular network setup, a non-regular grid with different cell sizes and also for a realistic scenario based on measurements and realistic traffic setup.

Full text

Page 1

Load Balancing in Downlink LTE
Self-Optimizing Networks
Andreas Lobinger∗, Szymon Stefanski†, Thomas Jansen‡, Irina Balan§
∗Nokia Siemens Networks, M¨ unchen, Germany, andreas.lobinger@nsn.com
†Nokia Siemens Networks, Wrocław, Poland, szymon.stefanski@nsn.com
‡Technical University of Braunschweig, Braunschweig, Germany, Jansen@ifn.ing.tu-bs.de
§Interdisciplinary Institute for Broadband Technology, Ghent, Belgium, Irina.Balan@intec.ugent.be
Abstract—In this paper we present system level simulation
results of a self-optimizing load balancing algorithm in a long-
term-evolution (LTE) mobile communication system. Based on
previous work [2][1], we evaluate the network performance of
this algorithm that requires the load of a cell as input and
controls the handover parameters. We compare the results for
different simulation setups: for a basic, regular network setup, a
non-regular grid with different cell sizes and also for a realistic
scenario based on measurements and realistic traffic setup.
I. INTRODUCTION
In existing networks, parameters are manually adjusted to
obtain a high level of network operational performance. In LTE
the concept of self-optimizing networks (SON) is introduced,
where the parameter tuning is done automatically based on
measurements. A challenge is to deliver additional perfor-
mance gain further improving network efficiency. The use of
load-balancing (LB), which belongs to the group of suggested
SON functions for LTE network operations, is meant to deliver
this extra gain in terms of network performance. For LB this is
achieved by adjusting the network control parameters in such a
way that overloaded cells can offload the excess traffic to low-
loaded adjacent cells, whenever available. In a live network
high load fluctuations occurs and they are usually accounted
for by overdimensioning the network during planning phase.
A SON enabled network, where the proposed SON algorithm
monitors the network and reacts to these peaks in load, can
achieve better performance by distributing the load among
neighbouring cells [10]. The load balancing algorithm aims
at finding the optimum handover (HO) offset value between
the overloaded cell and a possible target cell. This optimised
offset value will assure that the users that are handed over to
the target cell will not be returned to the source cell and thus
the load in the current cell is diminished. Simulations were
conducted for synthetic regular hexagonal and non regular
network layout as well as a realistic network scenario.
The work has been carried out in the EU FP7 SOCRATES
project [1].
II. DEFINITIONS AND METRIC
In [2] a mathematical framework for SON investigations on
the downlink is defined.
The main parts for this mathematical framework are recalled
below:
• A network layout by network nodes, eNodeB(s) (eNB),
defining a cell c at the position ? pc.
• A user u located at a position ? qu. In dynamic simulation
the user positions can be evolve over simulation time.
• A connection function X(u), following that a user u is
served by cell c = X(u) with the constraint (according to
the definition of LTE) that every user is connected exactly
to a single cell.
• A pathloss mapping L, LX(u)(? qu) defined by the po-
sitions of user u relative to a cell c, c = X(u). The
pathloss mapping takes all position-dependent channel
model effects into account, which are: distance-dependent
pathloss, shadowing and angle-dependent antenna pat-
terns.
• A cell load ρc, which defines the ratio of used resources
in LTE physical resource blocks (PRBs) to all available
resources in a given cell.
With the above framework and two additional terms i.e. N
as thermal noise and Pcas transmit power for a cell, we can
now define and evaluate for every user, in every time-step a
user specific signal to noise and interference ratio, SINRu.
SINRu=
Pc· LX(u)(? qu)
c?=X(u)ρc· Pc· Lc(? qu).
N +?
(1)
The evaluation complexity is reduced by the selection of a
single service definition for all users, a constant bit rate (CBR)
service. The CBR assumption used further on is 512kBit/s.
A. Load and throughput mapping
We assume that the best modulation coding scheme (MCS)
is used for a given SINR and the highest data rate R(SINR)
is achievable, this can be represented by Shannon formula.
R(SINRu) = log2(1 + SINRu)
(2)
However for better approximation to realistic MCS, the map-
ping function is scaled by a factor 0.6 and is bounded by max-
imum available bitrate (4.4 bps/Hz) and minimum required
SINR (-6.5 dB), a detailed description of which can be found
in [4]. Based on the achievable throughput at a given SINR,
the demanded data rate D and the transmission bandwidth
BW of one PRB (PRB bandwidth in LTE system is 180kHz),

Page 2

0
1
2
3
4
5
-7-5-3-11357
SINR [dB]
9111315 1719 21 2325
data rate R [bps/Hz]
0
2
4
6
8
10
12
14
16
18
required PRBs N
data rate required PRBs for 512 kb/s
Fig. 1.Required PRBs for transmission of 512 kbps as a function of SINR.
the amount of required PRBs N can be obtained from the
following:
Nu=
R(SINRu) · BW
Figure 1 presents the relationship between SINR, throughput
and required resources in DL transmission.
Du
(3)
B. Virtual load and unsatisfied users metric
In this document we investigate scenarios with different
load distributions including user u concentrations to overload
the serving cell. The overload situation occurs when the total
required number of PRBs Numay exceed the amount of the
total available resources in one cell MPRB. Thus we introduce
the virtual cell load, ? ρcwhich can be expressed as the sum of
c for user u
? ρc=
This metric allows us to exceed 100% of cell load (? ρc >1)
overloaded cells part of the users can not be served with the
required service quality and we call them as unsatisfied. The
total number of unsatisfied users in the whole network (users
being a sum of unsatisfied users in all cells, where number of
users in cell c is represented by Mc) can be written as:
?
Detailed explanation of the virtual load concept and unsatisfied
users metrics can be found in [2].
the required resources of all users u connected to the cell c
by the connection function X(u) which gives the serving cell
1
MPRB
·
?
u|X(u)=c
Nu.
(4)
and such situations indicate a scale of overload in cell c. In
z =
∀c
max
?
0,Mc·
?
1 −1
? ρc
??
(5)
C. Load estimation
Load balancing is achieved by handing over users from
the overloaded cells to those adjacent cells which are able
to accommodate additional load. After HO, the user’s SINR
condition at cell served by target eNB (TeNB) as well as
generated load is different than in previous cell served by
serving eNB (SeNB). The load transferred during LB operation
should not exceed the capacity reported as available by the
TeNB. This issue should be controlled by admission and con-
gestion control mechanisms in the TeNB. When the admission
and congestion control mechanisms reject LB HO requests,
there will be a significant increase in signaling overhead if
the requests are repeated. We propose a prediction method to
adress this for the load required at TeNB side, based on SINR
estimation after LB HO, taking into account UE measurements
like the reference signal received power (RSRP) and received
signal strength indicator (RSSI). For simplification we do
not consider additional factors related to the current load at
SeNB and TeNB which may have an impact on SINR. Signal
received by user u from cell c we can write as:
Sc= Pc· Lc(? qu)
(6)
Before LB HO the user u1 is connected to the SeNB,
SeNB = X(u1) with the strongest received RSRP signal
(signal S1in Figure 2 a). The RSRP signal from TeNB (S2
in Figure 2a) is a component of total interference as well as
signals originated from other eNBs (represented by I in Figure
2a). After HO of user u1to TeNB, received signal S1from the
previously serving SeNB now contributes to the interference
signal at u1whereas signal S2from TeNB is the wanted signal
(see Figure 2b).
S1
S2
I? i nt er f er ence f r om ot hereNB
a)
b)
SeNB
TeNB
SeNB
S1
S2
I? i nt er f er ence f r om ot hereNB
TeNB
u1
u1
u2
u3
u3
u2
Fig. 2.Signals received by UE a) before HO, b) after HO
We assume that during the time of HO execution, the user’s
u1 position ? qu1does not change significantly and hence we
also assume no changes of received signal power by the user
u1. We can extract signals S1 and S2 from the interference
part of SINRSeNB and SINRTeNB equations and after
combining them, this relation can be written as follows:
SINRu1,TeNB=
S2
S1
SINRu1,SeNB
+ S1− S2
.
(7)

Page 3

III. LOAD BALANCING ALGORITHM
In this investigation we are using the parameter “cell specific
offset” to force users to handover from the SeNB to a TeNB. It
is important to set this parameter carefully, so as not to exceed
the acceptable load at TeNB after the LB HO procedure. The
main goal of the presented Algorithm 1 is to find the optimum
HO offset that allows the maximum number of users to change
cell without any rejections by admission control mechanism at
TeNB side. Before applying the LB procedure, the SeNB needs
to create a list of potential targets for HO, collect measurement
reports from the served UEs as well as to collect the available
resources reports from neighbouring cells. These preparation
actions are included in steps 1 - 5 of Algorithm 1. For each
adjusted values of HO offset T, SeNB sorts the list of the
potential TeNB with respect to the number of possible LB
HOs. Subsequently, for given HO offset T and cell C from
the list L a load after HO, ? ρc, is estimated. If the predicted
offset to this cell is adjusted to the T value and virtual load
at SeNB ρSeNB is reduced by the amount generated by the
users handed over with this offset. Algorithm works until load
ρSeNBat SeNB is higher than accepted level ρThld,SeNBand
HO offset is below the maximum alowed value and neighbour
cells are able to accommodate additional load.
load ? ρcdoes not exceed acceptable level ρThldat TeNB, HO
Algorithm 1 HO offset based LB algorithm
Require: List L of potential Target eNB (TeNB) for LB HO
1. collect measurements from user:; RSRP to potential
TeNB,
2. group users corresponding to the best TeNB for LB HO
(criterion is the difference between SeNB and TeNB
measured signal quality),
3. obtain information from TeNB on available resources,
4. estimate a number of required PRBs after LB HO for each
user in the LB HO group,
5. T ⇐ 0
6. while ρSeNB> ρThld,SeNB∧ T < Tmaxdo
7.
i ⇐ 1
8.
T ⇐ T + step
9.
L ⇐ sort (TeNB according to number of users allowed
to LB HO with given T, descending order )
10.
while ρSeNB> ρThld,SeNB∧ i ≤ size(L) do
11.
C ⇐ L(i) {take next cell from list}
12.
estimate ? ρcafter HO for given T
ρSeNB ⇐ ρSeNB − ρSeNB,T {update load in
overloaded cell by substract handed over load}
15.
TC,u⇐ T
16.
end if
17.
i ⇐ i + 1
18.
end while
19. end while
20. adjust HO offsets TC
13.
14.
if ? ρc< ρThld,Cthen
-1000-5000 5001000
-1000
-800
-600
-400
-200
0
200
400
600
800
1000
X [m]
Y [m]
Base Station
antenna orientation
hotspot
users in hotspot
equal distributed users
-2000-100001000
-2000
-1500
-1000
-500
0
500
1000
1500
2000
X [m]
Y [m]
Fig. 3. Regular and non-regular simulation scenarios with hotspot path
IV. SIMULATION SCENARIOS
As already mentioned a standard LTE DL system of 10MHz
bandwidth is used with the simulation assumptions in the LTE
3GPP definitions [7]. For both the synthetic and real scenario a
simulation time-step of 500ms have been used. So all internal
signals, evaluations and updates are carried out with averaging
over this 500ms step. Also the LB algorithm is called in every
time-step.
A. Synthetic scenario
Following the standard simulation assumptions a simulation
setup with 19 sites in a regular hexagonal grid, 3 sectors
per site and 57 cells are defined. Additionally, a non-regular
grid with 12 sites, 3 sectors per site and 36 cells is used
as comparison taking real network effects like different cell
sizes, number of neighbor cells and interference situations into
account.
For employing localized higher load in the system a simula-
tion scenario is used here with a setup of background load in
all cells with a low number of users -so they are satisfied in any
position of the network- and an additional hotspot in which
new, additional users are dropped over time. The hotspot area
is moved over time on a path through the network as depicted
in Figure 3.
The channel model is defined in closed form [7], so the
pathloss mapping L is continuos. The movement of the users
(whether background or dropped in the hotspot) is a constant
velocity and random waypoint model.
B. Real scenario
The realistic reference scenario in the SOCRATES project
is an LTE network based on the real layout of the existing
2G and 3G macro networks provided by a network operator
(see [8] for more details). The scenario data includes the
following information: Network configuration, Pathloss data,
Clutter data, Heigth data, Traffic data, Mobility data.
An area of 72 km x 37 km is selected. The resulting
network comprising of 103 sites and 309 cells is shown
partly in Figure 4. For a smaller area of 1.5 km x 1.5 km
within the realistic scenario the detailed microscopic mobility
model SUMO (Simulation of Urban MObility) is used [11]
to introduce relevant user mobility. The mobile users in this
mobility model are cars and thus travel along streets. The

Page 4

Fig. 4.Realistic Scenario, area with basestation positions
mobility model incorporates traffic lights, multiple lanes, user
acceleration and braking, queueing in front of barricades and
overtaking of slower cars. In order to generate an overload
in the scenario a moving hotspot is defined by a bus with a
set of users moving together along a road. The distribution of
additional static users is based on traffic maps provided by the
same operator.
Based on operator measurements and path predictions us-
ing a ray-tracer a pathloss mapping L is now available for
background users (static + mobile) and the users in the
bus (hotspot). For this pathloss mapping only the 30 most
significant cells -including both the connected cell and the
interfering cells- are taken into account.
In Figure 5 the path of the bus is shown in more detail.
Gray shapes make an image of the city streets and magenta
lines determine the theoretical cells marked by numbers. Base
stations and antennas orientations are shown in the same way
as on Figure 3. Bus route is tagged by the brown colour. Circle
markers with the letters correspond to the points of HO of
more than 50% of the users in bus. Starting from the cell
#105; users in bus at point a ? are handovered to cell #104
it is writen as: a ? → 104. Following sequence describe the
HO scheme of the bus users during the 10 min simulation of
reference case: a ? → 104; b ? → 105; c ? → 103; d ? → 50;
e ? → 103; f ? → 50; g ? → 99; h ? → 50; i ? → 15; j ? → 97;
k ? → 15; l ? → 144.
V. RESULTS
A. Synthetic Scenarios
In Figures 6 and 7 timeline of the z-metric for each simula-
tion scenario with a reference system (red curve) versus load-
balancing (green curve) are depicted. The operating point here
-in both scenarios- is 5 users per cell in background (equally
dropped) and 40 users dropped in hotspot. The performance of
Fig. 5.Realistic Scenario, hotspot (bus) path
the load-balancing is -as expected- dependent on the hotspot
position and on the users positions within the hotspot. The
averaged unsatisfied users metric z is depicted as horizontal
line. In table I some more operating points and the average
results are shown as overview.
0100200300
time [s]
400500600
0
3
6
9
12
15
18
21
_
z = 7.68
_
z = 2.03
unsatisfied users z
reference systemLoad Balancig
Fig. 6.Load Balancing performance over time, regular network
0100200300
time [s]
400500 600
0
3
6
9
12
15
18
21
_
z = 9.78
_
z = 4.47
unsatisfied users z
reference systemLoad Balancig
Fig. 7.Load Balancing performance over time, non regular network
B. Real Scenario
In Figure 8 a timeline of the z-metric for the real simulation
scenario with a reference system (red curve) versus load-
balancing (green curve) is depicted. As descriped in IV-B part
of the users move together on a bus along a street. Different
than the operating point chosen for the synthetic scenarios
here a significant background load with 80 static and 80 slow

Page 5

TABLE I
SYNTHETIC SCENARIO, AVERAGE NUMBER OF UNSATISFIED USERS
¯ z reference system
regular
¯ z load-balancing
regularscenario non-reg.non-reg.
users in hotspot
30
40
50
60
2.8
7.7
13.6
22.0
3.2
9.8
16.8
23.1
0.3
2.0
6.4
15.5
0.7
4.5
10.6
17.4
moving users is used. In Figure 9 we can observe load (virtual
load as defined in Equation 4) variations during the first 300s
of the simulation. The markers in Figures 8 and 9 correspond
to Figure 5 and the description in section IV-B. Comparing
reference and load-balancing timelines we can see that the
load-balancing algorithm can redistribute load to neighboring
cells which significantly reduces overload in whole observed
period. In table II more operating operating points and average
results are shown as overview.
0 100 200300
time [s]
400 500600
0
20
40
60
80
100
120
abc
d
e
f
g
h
i
j
k
l
_
z = 63.4
_
z = 47
unsatisfied users z
reference systemLoad Balancig
Fig. 8. Load Balancing performance over time, realistic scenario
TABLE II
REALISTIC SCENARIO, AVERAGE NUMBER OF UNSATISFIED USERS
users in bus
20
30
40
50
60
¯ z reference system
¯ z load-balancing
47.0
53.8
63.4
74.3
86.6
30.6
37.8
47.0
58.8
68.2
VI. CONCLUSION AND OUTLOOK
In this paper we describe a simulation campaign of a
load-balancing algorithm. The algorithm evaluates the load
condition in a cell and the neighbouring cells and estimates the
impact of changing the HO parameters in order to improve the
overall performance of the network. We propose a method for
load estimation after HO would occur, which is based on SINR
prediction and utilise UE measurements. The efficiency of LB
algorithm was checked in simulation of synthetic networks
layout as well as for a part of real network in which load
situation changes dynamically.
An overall load-balancing gain is visible in all simulated
scenarios in the provided timelines and tables; leading to
the general conclusion that the average number of satisfied
users can be improved with load-balancing. The possible gain
depends on the local load situation and the available capacity
0 255075 100 125 150 175 200 225 250 275 300
time [s]
Load Balancing
0
2
4
6
8
10
virtual load
Reference
a b cde
cell #103
cell #104
cell #105
02550 75 100 125 150 175 200 225 250 275 300
time [s]
0
2
4
6
8
10
virtual load
ab cde
cell #103
cell #104
cell #105
Fig. 9.Virtual load ? ρ over time in seperate cells, realistic scenario
in the neighbouring cells. The algorithm relies on capacity
available in the neighbouring cells around overloaded cells
and the proposed load-balancing algorithm can redistribute
the load by changing the HO parameters; If no capacity is
available around, the network parameters and so the network
performance are left unchanged.
REFERENCES
[1] SOCRATES, Self-optimisation and self-configuration in wireless net-
works, European Research Project, http://www.fp7-socrates.eu.
[2] I. Viering, M. D¨ ottling, A. Lobinger, A mathematical perspective of
self-optimizing wireless networks, IEEE International Conference on
Communications 2009 (ICC), Dresden, Germany, May 2009.
[3] 3GPP, Self-configuring and self-optimizing network use cases and solu-
tions, Technical Report TR 36.902, available at http://www.3gpp.org.
[4] 3GPP, 3rd Generation Partnership Project;Technical Specification Group
Radio Access Network;Evolved Universal Terrestrial Radio Access (E-
UTRA);Radio Frequency (RF) system scenarios (Release 8), Technical
Report TR 36.942, available at http://www.3gpp.org.
[5] M. Amirijoo, R. Litjens, K. Spaey, M. D¨ ottling, T. Jansen, N. Scully, and
U. T¨ urke, ”Use Cases, Requirements and Assessment Criteria for Future
Self-Organising Radio Access Networks,” Proc. 3rd Intl. Workshop on
Self-Organizing Systems, IWSOS 08, Vienna, Austria, December 10-12,
2008.
[6] Next Generation Mobile Networks, Use Cases related to Self Organising
Network, Overall Description, available at http://www.ngmn.org.
[7] 3GPP, Physical Layer Aspects for evolved Universal Terrestrial Ra-
dio Access (E-UTRAN), Technical Report TR 25.814, available at
http://www.3gpp.org.
[8] N. Scully, J. Turk, R. Litjens, U. T¨ urke, M. Amirijoo, T. Jansen, L.
Schmelz, ”Review of use cases and framework II”, Deliverable 2.6 EU-
Projekt SOCRATES (FP7-2007-INFSO-ICT-216284), March, 2009
[9] 3GPP, Evolved Universal Terrestrial Radio Access Network (E-UTRAN);
X2 Application Protocol (X2AP), Technical Specification TS 36.423,
available at http://www.3gpp.org.
[10] U. T¨ urke, and M. Koonert, ”Advanced site configuration techniques
for automatic UMTS radio network design,” Proc. Vehicular Technology
conference VTC 2005 Spring, vol. 3, pp. 1960-1964, Stockholm, Sweden,
May 2005.
[11] H. Schumacher, M. Schack and T. K¨ urner, ”Coupling of Simulators for
the Investigation of Car-to-X Communication Aspects” 7th COST2100
Management Committee Meeting, TD(09)773, Braunschweig, Germany,
February 2009.