06139780
-
Upload
gio-zakradze -
Category
Documents
-
view
217 -
download
0
description
Transcript of 06139780
-
Performance of LTE SON Uplink Load Balancing in
Non-Regular Networks
Jussi Turkka
Department of Communications Engineering
Tampere University of Technology
Tampere, Finland
Timo Nihtil
Magister Solutions Ltd
Helsinki, Finland
Ingo Viering
Nomor Research GmbH
Munich, Germany
AbstractThis paper presents a performance evaluation of an
uplink load balancing algorithm for Long-Term Evolution (LTE)
Self-Organizing Networks (SON) in a non-regular network
layout and shows how cell sizes affect the performance of the
algorithm. The proposed algorithm solves local overload
situations by handing over users to neighboring cells and
adjusting power control settings. However, the non-regular
network layout and the varying cell size can limit the expected
gains of the algorithm in some situations due to the chosen uplink
load balancing strategies and the limitations of the LTE uplink
radio access technique. The performance evaluation is done by
using a fully dynamic LTE system simulator which is capable to
model user and network characteristics, mobility and radio
resource management (RRM) algorithms accurately.
Keywords - Radio Network Optimization; Non-regular Network
Layout; Self-Organizing Networks; Uplink Load Balancing;
I. INTRODUCTION
A rapid evolution of cellular networks and an increased
capacity demand has led to a situation where network
operators need to maintain large multi-vendor radio access
networks. The burden of operating and maintaining a complex
network infrastructure has caused a need to develop automated
solutions for network deployment, operation and optimization
which would reduce the operational expenditures and at the
same time improve the perceived end-user quality-of-service
(QoS). Self-organizing networks and Minimization of Drive
Tests solutions are currently researched by the network
vendors in the 3rd Generation Partnership Project (3GPP)
making possible the solutions for the network deployment
automation.
A target of the SON work item in the 3GPP is to define the
necessary measurements, procedures and interfaces to support
the self-configuration, self-optimization and self-healing use
cases which can dynamically affect the network operation, and
therefore, improve the network performance and reduce the
manual operation efforts [1]. The SON use cases in [1] target
to a coverage and capacity optimization, energy savings
optimization, interference reduction, automatic configuration
of physical cell identity, mobility robustness optimization,
mobility load balancing optimization, random access channel
optimization, automatic neighbor relations configuration, and
inter-cell interference coordination. However, the actual
solutions are not discussed in the 3GPP and the algorithms are
usually left to be vendor specific solutions.
A mathematical framework of one possible SON load
balancing (LB) algorithm for a downlink direction was
presented in [2] and the performance was evaluated later in
[3]. The downlink load balancing algorithm resulted in a better
network performance and an improved user happiness which
means that more users were able to achieve the given service
requirement e.g., the average bitrate requirement. Moreover,
the uplink load (UL) balancing algorithm was presented in [4].
The uplink enhancements were related to the limitations of the
maximum number of simultaneously scheduled users and
power limitation of the uplink direction. The solutions in [4]
were related to a better understanding of the overload and the
better control of initiating LB handovers (HO) together with
UL specific load adaptive power control (LAPC) which was
introduced in [5].
In this article we study the uplink load balancing
performance together with the proposed enhancements in a
non-regular cellular layout which was not analyzed in [4].
Moreover, the gain in non-regular layout is compared with the
regular hexagonal network layout with inter-site distance of
500 and 1732 meters. This paper is organized as follow.
Section II describes the uplink load balancing algorithm. In
Section III, the simulation parameters and assumptions are
described. Finally, section IV concludes the article.
II. ALGORITHM
A. Algorithm Description
The basic principle of the load balancing algorithm is to
handover a selected user (UE) from a stronger but overloaded
cell to a weaker but underloaded cell as described in [2-4]. By
doing so, the UE QoS is improved in both cells. This can
happen if the underloaded cell with a weaker signal strength
can allocate more physical resource blocks (PRB) to the
handovered UE than what was allocated in the overloaded
cell, and if the scheduler of the overloaded cell can allocate
the released resources to make some of the existing unhappy
UEs happier.
2011 IEEE 22nd International Symposium on Personal, Indoor and Mobile Radio Communications978-1-4577-1348-4/11/$26.00 2011 IEEE 162
-
The algorithm for the downlink load balancing is described
in [4]:
1. Collect measurements during a measurement period. 2. Detect an overload situation. 3. Find the best LB HO candidate. 4. Execute a LB HO. 5. Optimize a HO offset towards the LB HO candidate
cell.
In uplink, the functionality is similar but there are some
differences for detecting the overload and finding the best
neighboring cell for the load balancing handover as shown in
[4]. In addition to the handover based load balancing, the
uplink loading can be adjusted by tuning the power control
parameters [5], and therefore, the algorithm performance
depends on how much emphasis is put on these two different
load balancing strategies.
A virtual load c in a cell c is the sum of served UE component loads during the measurement period as described
in [4]:
,
1
,cN
u cuc
u u c
ARBGBR
BR SRB
(1)
where the component load of UE u inside the sum is a product
of two terms. The first term is a ratio of the guaranteed bitrate
GBRu and the realized bitrate BRu. The latter term is a ratio of
the allocated resources ARBu,c and total amount of schedulable
resource SRBc. The variable Nc is the total number of served
UEs during the measurement period. The cell c is overloaded
in case pc > 1 as explained in [2]. The definition of the
overload works well in downlink since all the PRBs can be
allocated always and therefore the users are unhappy mainly
due to the lack of resources. However, in uplink direction,
there are other reasons for UE unhappiness which cannot be
detected by using (1) as a sole overload indicator as explained
in [4]. The other reasons for the UE unhappiness are a UE
power limitation and a network control channel limitation.
B. Uplink limitations
If the UE is in a power limitation, then the available
transmission power, which is limited by the UE capabilities,
and the power control mechanism defines the maximum
available transmission bandwidth e.g., the number of ARBu,c
per transmission time interval (TTI). The assumption in [4] is
that the LB handover cannot improve the QoS of the power
limited UE if it is connected to the strongest cell unless the
weaker cell can schedule the UE more often. This can limit the
number of UEs which can be LB handovered if the UE QoS is
to be guaranteed to remain at least the same.
In case of the maximum scheduled users (MSU) limitation
as described in [4], the available Physical Downlink Control
Channel (PDCCH) resources limit the maximum number of
simultaneously scheduled user. This limitation together with
the power limitation can result in a situation where the cell is
underloaded in terms of PRB usage and yet there are many
UEs which cannot achieve the service requirement target such
as GBRu. In this study these UEs are called unhappy UEs.
This means that the sum in the latter part of the product in
(1) is never 1 even though the resources are always shared
with the maximum number of UEs in every TTI. If some of
the UEs are power limited and unable to allocate all the
available bandwidth to be happy, then there are unhappy users
at the same time with unused PRB resources. Therefore, the
first part of the product in (1) is larger than one but at the same
time the latter part of the product is smaller than 1. This can
results in an unhappiness even if the virtual load is smaller
than 1. Hence, the scheduler should never schedule the
resources to several power limited UEs at the same time since
this results easily in a large amount of unused PRBs.
C. Load Adaptive Uplink Power Control
The idea of the load adaptive uplink power controlling is to
adjust the available transmission power per PRB according to
the interference which tends to increase along the traffic
loading [5]. The algorithm adjusts dynamically the static
power offset P0 depending on the network loading by allowing
users to transmit data with the lower power per PRB in the
underloaded cells. This allows wider bandwidth allocations
which results in fewer problems due to MSU limitation e.g.
fewer resources are wasted if many power limited users are
scheduled simultaneously. However, there is a danger that in
large cells, the P0 is reduced too much which can drown cell
edge UEs to noise e.g. the power per PRB in transmission is
so small that the received PRBs power at eNB falls below the
thermal noise thresholds. This can limit the usage of LAPC in
load balancing and was the main reason why the uplink load
balancing and LAPC performance was analyzed in non-
regular layout consisting of small and large cells.
The LTE power control algorithm is specified in [6]. In this
study, a simplified version of the algorithm was used
excluding the transport format and power correction offsets as
shown in (2)
0 10min , 10log ,MAXP P P M PL (2)
Where the P is the power per TTI defining the maximum
bandwidth allocation the UE is capable to transmit. The PMAX
is the maximum available transmission power. The M is the
bandwidth allocation expressed in number of PRBs. The P0 is
the power offset and is a variable defining the degree of the fractional path loss compensation. The variable PL is the
estimation of the downlink path loss.
In this study, the dynamic LAPC adjustment of the P0 was
made based on the following equation (3)
0, 0, 1010log ,initial
c c cP P (3)
where the variable 0,
initial
cP is the initial P0 of the cell c and it is
adjusted based on the virtual loading of the cell. In case the 163
-
virtual loading is small, the P0,c is reduced resulting smaller
power allocation per PRB. How should the initial P0 be
configured in the first place? In regular hexagonal layouts with
inter-site distance (ISD) of 500 and 1732 meters the
assumption for a reasonable initial P0,c is -52 dBm. In the non-
regular layout, the initial P0 is based on the cell size and the
estimation of the path loss and a 5 percentile power limitation
rule which assumes that only the 5% of the users allocate all
the available power to a single PRB as shown in (4)
0, ,95% ,initial
c MAX cP P PL (4)
where the variable PLc,95% is the estimation of the cell pathloss
distribution 95 percentile point e.g., the path loss at the cell
edge.
III. SIMULATION AND MODELLING ASPECTS
A. Simulation Tool
The results in this paper are derived by using a fully
dynamic system simulation tool modeling E-UTRAN LTE
release 8 in downlink and uplink and it has been used in
several other publications as in [4,7]. The simulator maps link
level SINR to system level following the methodology in [8].
Both the downlink and the uplink can be modeled in an
OFDM symbol resolution with several radio resource
management, scheduling, mobility, handover and traffic
modeling functionalities. Simulation parameters are based on
the 3GPP specifications defining the used bandwidth, center
frequency, network topology, and radio environment [9].
B. Simulation Scenario
Three different network layouts are used in the simulations.
The performance of two regular hexagonal layouts with 500
and 1732 meters constant ISD were compared with the non-
regular Springwald layout which is described [10]. The
simulations consisted of a calibration simulation and a load
balancing simulation with a moving traffic hotspot. The
calibration simulations were done with and without the load
adaptive power controlling but the HO based load balancing
and the traffic hotspots were excluded. In the load balancing
simulation, the route of the moving traffic hotspot was
planned to go near the estimated cell edges without
overlapping the cell borders directly as depicted in Fig.1. The
overlapping route would balance the loading automatically
between the neighboring cells. The largest cells were located
in the beginning and the end of the route. The smallest cells
were in the middle of the route. In the regular and the non-
regular case, there was an additional tier of interfering cells
causing background interference to the outer tier cells. This
results in a similar kind of interference conditions to all
simulated cells.
In this paper only the uplink direction is analyzed. The
maximum number of the users in the simulation area was set
according to the offered background load target which
depends on the GBRu,c. The UE traffic profile was set to a
constant bit rate (CBR) and was 64, 128 or 256 kbps. The call
length was random and varied being approximately 23
seconds in average. Other simulation parameters are shown in
the Table I and the Table II.
IV. RESULTS
A. Calibration Simulation Results
Table III summarizes the calibration simulation results. The
offered background load was adjusted based on the sum of
users per cell with the selected CBR traffic model. As seen in
the Table III, all the users are happy in the dense ISD 500
network with the chosen bitrates and the LAPC can
Fig. 1. Springwald scenario with traffic hotspot route.
TABLE I
SCENARIO SPECIFIC SIMULATION PARAMETERS
Parameter Value
Regular layout 19 sites, each with 3 sectors
Non-regular layout Springwald
Antenna Type 65 deg, 14 dB gain
ISD 500 m / 1732m / varying
Pathloss model PL = 128.1 + 37.6*LOG10(Rkm)
Penetration loss 20 dB
Center frequency 2 GHz
Shadowing std 8 dB
Channel model 3GPP Typical Urban
Mobility model 3 km/h pedestrian
eNodeB max TX power 46 dBm
UE max TX power 23 dBm
Bandwidth 10 Mhz
Frequency reuse factor 1
TABLE II SIMULATION SPECIFIC PARAMETERS
Parameter Value
System LTE-FDD Rel.8
Simulation Time 71.5 seconds (1,000,000 symbols)
Hybrid ARQ yes
ARQ no
Link Adaptation Both inner and outer loop
BLER target 0.2
Channel sounding yes
Packet Scheduler TD-PF/FD-ATB
Handovers Sliding window size: 200ms
Handover margin: 3 dB
Traffic Model CBR 64/128/256 kbps
Average number of users
per sector
Depends on CBR
164
-
compensate the unhappiness in case of the higher bitrates by
adjusting the P0 according to the load. The initial P0 settings
were introduced in section II. In sparse ISD 1732 network,
there are unhappy UEs in 64 and 256 kbps CBR services.
LAPC compensation reduces the median P0 to -58.5 dBm for
the low bitrates reducing the unhappiness from 9.6% to 3.1%.
However, in case of 256 kbps bit rate the median P0 is -59.2
dBm but the LAPC compensation cannot reduce the
unhappiness from the 25.3%. In the Springwald, the
performance is a mixture of the dense and the sparse network
layouts due to the varying cell sizes as described in [10].
Table IV shows an example of the estimation of the
pathloss threshold with different P0 settings for three CBR
traffic classes. The PRB requirement M per TTI is calculated
based on robust QPSK modulation with 1/3 coding rate
assuming 0.3 BLER as shown in (5).
1000,
12 14 1
kbps
TTI
GBRM
B CR BLER N
(5)
where the variable B is a bits per symbol and CR is a code rate
of the chosen modulation and coding scheme (MCS). The
term (1-BLER)*NTTI indicates how many TTIs per second can
carry the actual user data at maximum. If the UE is scheduled
always then the NTTI equals to 1000 TTIs per second. It is
worth of noting that the last term is affected by a factor which
depends on the number of the UEs exceeding the MSU
threshold e.g. the probability that the UE can be scheduled
always decreases if there are more active UEs in the cell than
what is the MSU limitation. However, the behavior of this
depends on the scheduler as well.
The pathloss requirement is calculated based on (2)
assuming that the variable is 0.6 and the UEs are using the maximum power to allocate M PRBs always, and therefore, it
estimates the pathloss threshold for the UE power limitation.
The cell edge pathloss based on the 95% criteria in ISD 500,
ISD 1732 and Springwald is 120.5 dB, 133.5 dB and 137.5 dB
as described in [10]. As seen in (5) and Table IV, there are at
least two ways to improve the path loss region threshold. One
way is to use a better MCS which results in a smaller
bandwidth allocation M. Another way is to reduce the P0 if
possible. However, if the P0 is reduced too much, then power
allocation per PRB and signal to interference ratio (SINR)
decreases. This results in a more robust MCS selection and a
wider bandwidth allocation requirement M for the guaranteed
bit rate.
Table IV indicates that the ISD 500 cell edge threshold of
120.5 dB can be compensated with the LAPC if the GBR is
256 kbps just by adjusting the P0. By reducing the P0, the
pathloss requirement for the service is set to a smaller value
that the cell edge threshold of 120.5 dB. In the ISD 1732 case,
the 64 kbps service is already power limited since the pathloss
requirement with initial P0 is larger than the cell edge
threshold of 133.5 dB. To guarantee the 256 kbps service with
4 PRBs bandwidth allocation at the cell edge of a large cell
requires the P0 be smaller than -65.5 dBm according to the cell
edge pathloss and (2). However, the P0 cannot be reduced too
much or otherwise the UEs would drown to the noise because
the SINR gets too small for the chosen MCS [11]. Therefore,
the usage of the LAPC is limited in the large cells resulting in
smaller gains of UE happiness.
B. Traffic Hotspot Simulation Results
Table V shows the overall simulation results for the uplink
load balancing performance in the Springwald. The
simulations were run with three optimization strategies:
a reference case (no LB nor LAPC)
LAPC without HO LB optimization
LAPC and HO LB joint optimization. The route of the traffic hotspot (HS) is depicted in Fig.1. The
UE traffic profiles were 64, 128 and 256 kbps. The overall
results indicate that LAPC and load balancing can provide
gain for the UE happiness. For 256 kbps case, the gain is 3.6%
decreasing the unhappiness from 12% to 8.4%. In other cases,
the gain was approximately 1% but the unhappiness was rather
low as well. However, even if the overall gain in Springwald
is small, there are regions where the LAPC and UL HO load
balancing provided significant improvements as depicted in
Fig.2.
In Fig.2, the blue regions are areas where the UL LB and
LAPC cannot improve the UE happiness and the red regions
indicates areas where UEs are unhappy only in reference case
but not anymore after taking UL LB and LAPC into use. The
leftmost illustration presents the 128 kbps results. In that case,
nearly all unhappy users benefit the usage of the load
balancing in the region of small cells. However, in the region
of the large cells, the load balancing is not able to help the
unhappy users. This is obvious, since the path loss conditions
in the large cells are quite demanding as described in [10].
The rightmost illustration in Fig.2 presents the 64 kbps
results. In that case, there are less unhappy users and the
TABLE III
CALIBRATION SIMULATION RESULTS
CBR 64 kbps 2MB offered load, 32 UEs per Cell
Scenario ISD 500 ISD 1732 Springwald
P0 Scheme Ref LAPC Ref LAPC Ref LAPC
Total no of HOs 966 964 909 890 2609 2573
No of started calls 3844 3844 3752 3823 7797 7805
Unhappy users, UL % 0.0% 0.0% 9.6% 3.1% 1.7% 1.2%
Power limited UEs % 0.0% 0.0% 1.9% 0.9% 0.5% 0.3%
Times MSU limited 1 1 851 413 63 21
CBR 256 kbps 2MB offered load, 8 UEs per Cell
Scenario ISD 500 ISD 1732 Springwald
P0 Scheme Ref LAPC Ref LAPC Ref LAPC
Total no of HOs 440 434 477 516 1444 1393
No of started calls 1843 1846 1676 1682 3653 3676
Unhappy users, UL % 1.2% 0.0% 25.3% 24.4% 7.1% 5.9%
Power limited UEs % 0.6% 0.0% 17.5% 17.0% 3.6% 3.0%
Times MSU limited 0 0 15 9 0 0
TABLE IV
PATHLOSS THRESHOLD FOR POWER LIMITATION
P0=-52 P0=-57 P0=-62
GBRkbps PRB Req. Pathloss Req. Pathloss Req. Pathloss Req.
64 kbps 1 125 dB 133 dB 142 dB
128 kbps 2 120 dB 128 dB 137 dB
256 kbps 4 115 dB 123 dB 132 dB
165
-
remaining ones were concentrated only in the beginning of the
route which was surrounded by the large cells 5, 23, 25 and 27
and the cell edge path loss was approximately 140 dB. If that
situation is compared with the large ISD regular scenario with
cell edge path loss of 137.5 dB in Table IV, then it can be
concluded that the handover based UL load balancing and the
LAPC cannot improve the network performance.
C. Deployment considerations in Radio Networks
Based on these results and [4], one can conclude that UL
load balancing provides gain for UE happiness in dense
networks for a wide range of traffic profiles. Moreover, it was
shown that the cell edge UEs at large cells cannot benefit of
using the handover based load balancing especially in case
there is a traffic hotspot near the cell borders. In practical
network deployments, the UE mix is likely to be consisting of
indoor and outdoor users. Therefore, in large cells the indoor
users with stationary or smaller velocities cannot be
handovered because they suffer the building penetration loss
and the power and bandwidth limitation due to the larger
signal attenuation. On the other hand, the outdoor users with
smaller path loss and variety of velocities can be handovered
to tackle the overloading problem which makes the load
balancing useful in dense as well in sparse networks.
However, the effects of the higher UE mobility were not taken
into account in this study.
V. CONCLUSION
This paper shows a performance evaluation of the LTE
uplink load balancing algorithm in the non-regular radio
network consisting of small and large cells. Two load
balancing strategies were compared. The LAPC strategy
without the handover based load balancing. Secondly, the
effect of LAPC strategy together with the handover based load
balancing. Based on the results, it can be concluded that in
large cells, the users are power limited and only a little gain
can be achieved either by using the LAPC or the handover
based LB. Therefore, the overall gain in non-regular layout is
smaller compared with the gains in certain spatial regions
where the load balancing improves the performance quite
much. However, this spatial gain is difficult to see, if one is
only observing the global statistics of the network.
VI. ACKNOWLEDGEMENTS
The author would like to thank colleagues from Nokia,
Nokia Siemens Networks, Magister Solutions Ltd, Jyvskyl
University and the Radio Network Group at Tampere
University of Technology for their constructive criticism,
comments and support with the work.
REFERENCES
[1] 3GPP TR 36.902, Self-configuring and Self-optimization (SON) network use cases and solutions, ver. 9.2.0, June 2010.
[2] I. Viering, M. Dttling and A. Lobinger, A mathematical perspective of self-optimizing wireless networks, in proceedings of the IEEE International Conference on Communications, Dresden, Germany, June
2009.
[3] A. Lobinger et al., Load Balancing in Downlink LTE self-optimizing networks, in proceedings of 71th IEEE Vehicular Technology Conference (VTC 2010-Spring), Taipei, Taiwan, 2010.
[4] T. Nihtil & J. Turkka, Performance of LTE Self-Optimizing Networks Uplink Load Balancing, accepted to 73th IEEE Vehicular Technology Conference (VTC 2011-Spring), Budapest, Hungary, 2011.
[5] R. Mller et al., Enhancing uplink performance in UTRAN LTE networks by load adaptive power control, European Transactions on Telecommunications, 2010, DOI:10.1002/ett.1426.
[6] 3GPP TS 36.213, Evolved Universal Terrestrial Radio Access (E-UTRA); Physical Layer Procedures, ver. 9.2.0, June 2010.
[7] P. Kela et al., Dynamic Packet Scheduling Performance in UTRA Long Term Evolution in Downlink, in conference proceeding of ISWPC2008, 2008.
[8] K. Brueninghaus et al., Link performance models for system level simulations of broadband radio access systems, in proceedings of the Personal, Indoor and Mobile Radio Communications (PIMRC05), vol. 4, September 2005.
[9] 3GPP TR 25.814, Physical Layer Aspects for Evolved UTRA, version 7.1.0, September 2006.
[10] J. Turkka & A. Lobinger, Non-regular Layout for Cellular Network System Simulations, in proceeding of PIMRC 2010, Istanbul, September 2010.
[11] J. Turkka and J. Puttonen, Using LTE Power Headroom Report for Coverage Optimization, in proceedings of 74th IEEE Vehicular Technology Conference (VTC 2011-Fall), San Francisco, USA,
September 2011.
TABLE V LOAD BALANCING RESULTS IN SPRINGWALD
CBR 256 kbps Ref. LAPC LAPC + LB
Offered load ( BG ) kbps 2048 MB (8 MT x 256 kbps)
Offered load ( HS ) kbps 12800 MB (50 MT x 256 kbps)
Number of Handovers 1777 1811 2461
Number of LB Handovers 0 0 559 (23%)
No of started calls 4197 4255 4244
Unhappy users, UL 12.0% 9.9% 8.4%
Power limited UEs 3.2% 2.0% 2.4%
CBR 128 kbps Ref. LAPC LAPC + LB
Offered load ( BG ) kbps 1536 MB ( 12 MT x 128 kbps)
Offered load ( HS ) kbps 7680 MB ( 60 MT x 128 kbps)
Number of Handovers 2563 2495 3544
Number of LB Handovers 0 0 857 (24%)
No of started calls 6545 6551 6556
Unhappy users, UL 4.8% 3.7% 3.2%
Power limited UEs 0.5% 0.4% 0.4%
CBR 64 kbps Ref. LAPC LAPC + LB
Offered load ( BG ) kbps 1536 MB ( 24 MT x 64 kbps)
Offered load ( HS ) kbps 5120 MB ( 80 MT x 64 kbps)
Number of Handovers 4516 4411 5087
Number of LB Handovers 0 0 681 (13%)
No of started calls 12246 12282 12287
Unhappy users, UL 2.0% 1.0% 1.0%
Power limited UEs 0.3% 0.1% 0.1%
a) b) Fig. 2. Springwald with a traffic hotspot. a) 128 kbps CBR. b) 64 kbps CBR. 166