Vous êtes sur la page 1sur 5

On Mobility Load Balancing for LTE Systems

Raymond Kwan1, Rob Arnott1, Robert Paterson1, Riccardo Trivisonno1, Mitsuhiro Kubota2
2

Telecom Modus Ltd, UK NEC Corporation, Tokyo, Japan raymond.kwan@eu.nec.com


(f ) ) M j + O (j f ) + Oi(,cn + Oi( cs ) + j > M i + Oi

Abstract In this paper we present simulation results to demonstrate that a simple distributed intra-frequency load balancing algorithm based on automatic adjustment of handover thresholds can significantly reduce the call blocking rate and increase cell-edge throughput in an LTE network. Keywords- LTE, load balancing, handover, self-organising network (SON), radio resource management (RRM)

(1)

where indices i and j correspond to the serving and neighbour cells respectively, M i and M j correspond to the UE measurement result of cells i and
j ,

Oi( f ) and O (j f )
( cs )

correspond to the frequency-specific offset for frequency

I.

INTRODUCTION

f i and f j that are used by cells i and j respectively, Oi


( cn )

is

The term load balancing can be used to describe any mechanism whereby highly loaded cells distribute some of their traffic to less heavily loaded neighbours in order to make the use of the radio resource more efficient across the whole network. In this paper, we are concerned with intra-frequency load balancing mechanisms that have reaction times measured in several minutes or hours, and which can be implemented in a Long Term Evolution (LTE) network with a minimum of additional signalling. There are many ways to re-distribute load between cells. One approach is to adjust the cell coverage by modifying the pilot power [1]. A larger pilot power effectively allows more distant mobiles to access the cell, thereby increasing the coverage area. However, automatic adjustment of cell coverage area runs the risk of creating coverage holes. Another way to re-distribute the load is to modify the handover regions between neighbouring cells. Such an approach is referred to as mobility load balancing (MLB). The principle behind mobility load balancing is to adjust the handover regions by biasing the handover measurements, causing cell-edge users in highly loaded cells to migrate to less heavily loaded neighbour cells, thereby improving the efficiency of resource utilisation [2]. The result is an improvement in call blocking rate and celledge throughput. Since the re-distribution of load is done automatically between cells, this technique is an example of a Self-Organizing Network (SON) algorithm. The organisation of this paper is follows. In section II, a simple distributed load balancing algorithm is described. In section III, a simulation model is described and numerical simulation results are presented. Finally conclusions are given in section IV. II. MOBILITY LOAD BALANCING

the cell-specific offset of the serving cell, and Oi , j is the cellspecific offset of the neighbour cell j with respect to cell i. The quantities and are a hysteresis term and a fixed offset respectively. The measurement M i may be in the form of the Reference Signal Received Power (RSRP) in the unit of dBm, or Reference Signal Received Quality (RSRQ) in the unit of dB [4]. In this paper we consider only intra-frequency handover, in which case (1) becomes
) M j M i > Oi( cs ) Oi(,cn j + + .

(2)
) Oi(,cn j , it

From (2), it can be seen that by increasing the offset

is possible to cause a mobile served by cell i to be handedover to the neighbour cell j , thereby reducing the load in cell i. In this paper we perform load balancing by automatically adjusting the offsets
) Oi(,cn j based on cell load measurements. A

simple way to do this is as follows


) ( cn ) min Oi(,cn j + , Omax , if j i th ) ( cn ) max Oi(,cn j , Omax , if i j th O ( cn) , if i j < th i, j

( cn ) i, j

( (

) )

(3)

where is the offset step-size, i and j are the load of cell i and cell j respectively, and the quantity

th > 0 is
O

a pre( cn ) i, j

defined threshold for triggering load balancing. All

are

According to [3], a handover can be triggered a number of events. In this paper we are concerned with one particular event, known as event A3, in which it is detected that a neighbouring cell offers a better signal quality for a particular UE than its currently serving cell. This condition can be expressed as follows.

initialised to zero. The update (3) is applied periodically, in response to new load measurements made by each cell. Thus this method is suitable for a distributed implementation, in which each cell exchanges load measurements and offset values with its neighbours. One property of (3) is that the offsets are always symmetrical, i.e.
) ( cn ) Oi(,cn j = O j , i . This is important in order to

978-1-4244-3574-6/10/$25.00 2010 IEEE

ensure that a UE handed over from one cell to another is not handed straight back (i.e. to prevent ping-ponging). In order to apply the update in (3) a cell load measurement is required. A simple method of measuring load is to calculate the mean utilisation of physical resource blocks (PRBs) in the cell, as follows.

hours, with the cell offsets

) Oi(,cn j updated every 60 seconds.

For a simulation of this size, fully modeling the MAC scheduler, RLC, HARQ and physical layers would be prohibitively complex. Instead we use a simplified (but representative) model as described below. In each sub-frame, an SIR value k is computed for each user k as follows.

i =
k

nk K T

(4)

k =
where

i, k i , k P
N0 + P j,k
j i

(6)

Here

number of PRBs in the system bandwidth, and the summation is performed over all bearers connected to cell i during the measurement period. Note that i is a value in the range

nk is the total number of PRBs allocated to bearer k over a measurement period of T sub-frames, K is the total

N 0 is the thermal noise spectral density (including UE

receiver noise figure),

j , k is

the mean path gain (including

0 i 1 . The problem with this method is that the PRB


utilisation is not always a good indication of the real load in a cell. If there are non-GBR bearers connected to the cell, the PRB utilisation may always look high because the scheduler will assign any spare resources in the cell to those users. Even in the case of GBR traffic, a high PRB utilisation may not be a reliable indication of high load because it may be possible to achieve more efficient PRB usage at the expense of slightly increased packet delay, whilst still achieving the required QoS for all connected bearers. We can obtain a more useful load measurement by taking into account the QoS requirements of the connected bearers, as follows.

shadowing and antenna sectorisation gain but not fast fading) between UE k and cell j, P is the transmitted power (assumed to be the same for all cells), i is the index of the serving cell for UE k, and i , k is a unit-mean geometrically distributed random variable representing Rayleigh fading between the UE and cell i. The SIR value k is mapped to a modulation and coding scheme (MCS) based on look-up-tables obtained from link level simulation, to give an instantaneous achievable data rate rk . A proportional-fair scheduling metric is then calculated for each user as follows, where

=1.5 is a fairness parameter and


(7)

i = min KC ,
k

nk Rk( req ) K T Tk

(5)

Here

T and Rk(req ) is the required mean data rate for bearer k


derived from its QoS requirements [5]. (In the case of nonGBR services, Rk represents an equivalent GBR, i.e. a minimum throughput that would be considered acceptable for non-GBR users). K C is a positive real constant (typically around 2.0) used to keep the load under a reasonable value in case Tk becomes very low due to exceptionally poor radio channel quality for a particular bearer. III. NUMERICAL RESULTS
( req )

Tk is the achieved throughput of bearer k over the period

Tk is the mean throughput achieved by user k. r k = k Tk

In each cell, the connected users are ranked in descending order of k and each user is assigned as many PRBs as needed to empty its queue until all users are served or all PRBs are used. An admission control mechanism [6] is included in each cell which blocks new calls when the measured load reaches 0.7, and blocks incoming handover attempts when the measured load reaches 0.8. A congestion control mechanism [7] is also included which drops users if the load exceeds 0.9. Call arrival is modeled as a Poisson process, and to simulate the effect of non-uniform geographical traffic distributions we assume that new calls are generated in geographically localised clusters according to the following method. At any given time there is always the same number of clusters, N clust . Each cluster exists for a fixed period, Tclust , and when a cluster expires a new cluster replaces it at a randomly chosen location. The call arrival rate of each cluster increases linearly over time from 0 up to a maximum value at time Tclust 2 , and then decreases linearly to 0 again. The peak call arrival rate is a constant and the same for all clusters. The lifecycles of the clusters are offset in steps of Tclust N clust , which ensures that at any time the total call arrival rate

A. Simulation Model Simulations are used to evaluate the performance of the proposed load balancing scheme in the downlink of an LTE network. The main simulation parameters are given in Table 1. We assume video streaming traffic with a constant date rate of 64kbps per user and a mean call duration of 3 minutes. The relatively slow action of load balancing means that long simulation durations are needed to study its behaviour. We have simulated a network of 21 cells for a period of 4

summed over all clusters is constant (i.e. the average call arrival rate over the whole network remains constant throughout the simulation). New UE positions are generated by adding a random displacement to the centre location of the originating cluster. The displacement is generated from a 2D normal distribution, and wrap-around is taken into account. This model generates an offered traffic distribution which continually and gradually changes, whilst maintaining a constant average call arrival rate over the whole network. The rate of change is governed by the parameter Tclust , which we set to 2 hours. (After a period of 2 hours, all existing clusters will have been replaced, so the traffic distribution will have completely changed). The degree of traffic localization is determined by the cluster radius, i.e. the r.m.s. distance between new UE locations and the centre of the originating cluster. For illustration, Figure 1 shows the locations of the UEs generated in the first 15 minutes of the simulation for N clust = 32 with cluster radii of 25m and 100m. B. Results In a first set of simulations, the cluster radius is fixed to 100m and the simulation is repeated at different mean call arrival rates such that the mean offered load per cell varies between 1 Mbps and 4 Mbps. We compare the cases of no MLB, MLB based on PRB utilisation (equation (4)) and PRB based on load (equation (5)). Figure 2 - Figure 4 show the blocking, dropping, and handover failure probabilities as a function of the offered load. As expected, as the offered load increases, the blocking probability increases. However, loadbased MLB provides significant reduction in blocking probability compared to the other cases. The results for PRB utilisation-based MLB are poor, confirming the fact that PRB utilisation is not a reliable indicator of actual cell load. By biasing the handover decisions users will be handed over when their radio conditions in the target cell are less favourable, increasing the risk of dropping. However Figure 3 shows that MLB causes little change in dropping probability. (The trade-off between improved blocking and worsened dropping can be adjusted by means of the maximum offset ( cn ) parameter Omax ). Figure 4 shows that load-based MLB significantly reduces handover failure rate at medium load, but slightly increases it at high load. Figure 5 shows that at medium load, load-based MLB yields a significant improvement in 5%-tile user throughput, with little effect on mean user throughput. At both very high and very low load, MLB shows no benefit, because in these cases the load variation between cells is generally not large enough to trigger MLB. Figure 6 shows the CDF of cell load aggregated over all cells and over the whole simulation, for the case of mean offered load 2.5Mbps/cell. As expected, MLB results in a narrower CDF (i.e. more evenly distributed load) but a higher mean load. The mean load is increased because MLB causes some cell-edge users to handover earlier, resulting in worse

radio conditions in the target cell, and hence worse average radio conditions over the whole network. This does not result in a loss of throughput because it is more than compensated for by the more even distribution of users between cells. In second experiment, the mean offered load is fixed at 2.5 Mbps per cell and the cluster radius is varied from 25m to 400m. Figure 7 - Figure 9 show the blocking, dropping, and handover failure probabilities as a function of the cluster radius respectively in the case of no MLB and load-based MLB. These results show that, as expected, the gain of MLB reduces as the cluster size increases, since larger clusters give rise to a more uniform traffic distribution which reduces the load variation between cells. On the other hand, small clusters give rise to a highly non-uniform traffic distribution, thereby creating larger load differences between cells. In these results, a large gain from MLB is seen even with very small clusters, even though in such cases the gain of MLB is clearly very dependent on the exact locations of the clusters relative to the cell handover boundaries. IV.
CONCLUSIONS

In this paper, simulations are performed to investigate the performance benefits of MLB in LTE networks. Although more work is needed, the results suggest that significant gains in call blocking and cell-edge user throughput are possible even with a very simple MLB mechanism. However it is also shown that PRB utilisation is not a sufficient indicator of cell load for MLB, even in the case of constant bit rate traffic. ACKNOWLEDGMENT Telecom Modus Ltd is a wholly owned subsidiary of NEC Corporation. The authors would like to thank their colleagues at Telecom Modus and NEC Japan for their constructive comments. REFERENCES
[1] K. A. Ali, H. S. Hassanein, and H. T. Mouftah, "Directional Cell Breathing Based Reactive Congestion Control in WCDMA Cellular Networks", in Proc. of IEEE Symposium on Computers and Communications, July 1-4, 2007. R. Nasri and Z. Altman, Handover Adaptation for Dynamic Load Balancing in 3GPP Long Term Evolution Systems, Proc. of International Conference on Advancnes in Mobile Computing & Multimedia (MoMM), Dec. 2007. TS 36.331 3rd Generation Partnership Project; Technical Specification Group Radio Access Network; Evolved Universal Terrestrial Radio Access (E-UTRA); Radio Resource Control (RRC); Protocol specification (Release 8). TS 36.214, 3rd Generation Partnership Project; Technical Specification Group Radio Access Network; Evolved Universal Terrestrial Radio Access (E-UTRA); Physical layer Measurements (Release 8). TS 36.413, 3rd Generation Partnership Project; Technical Specification Group Radio Access Network; Evolved Universal Terrestrial Radio Access (E-UTRA) S1 Application Protocol (S1AP) (Release 8). R. Kwan, R. Arnott et al., On Radio Admission Control for LTE Systems, Proc. of IEEE Vehicular Technology Conference (VTC), Fall, 2010. R. Kwan, R. Arnott et al., On Pre-emption and Congestion Control for LTE Systems, Proc. of IEEE Vehicular Technology Conference (VTC), Fall, 2010.

[2]

[3]

[4]

[5]

[6]

[7]

[8]

NGMN Alliance, NGMN Radio Access Performance Evaluation Methodology, January, 2008.
14 No MLB 12 PRB utilisation-based MLB Load-based MLB 10 Blocking Prob. (%)

System bandwidth Simulation duration Cell Layout

Table 1 Simulation Parameters 5MHz (25 PRBs) 4 hours Hexagonal grid, 7 cell sites, 3 sectors per site with wrap-around Inter-Site Distance (ISD) 500m Pathloss 128.15 + 37.6log10( max(dkm,0.035) ) Shadowing Log-normal with standard deviation 8 dB Intra-site correlation 1.0 Inter-site correlation 0.5 Antenna pattern (horizontal) 2 ,

A( ) = min 12 , Am 3dB

where eNB Tx Antennas UE receive antennas Load measurement averaging period

3dB = 70 deg, Am = 20 dB

0 1

1.5

O
O

( cn ) i , j update rate ( cn ) max

1 per sector 2 400 ms for admission control and congestion control, further averaged over 60 seconds for MLB 60 seconds 6 dB
Dropping Prob. (%)

2.5 Offered Load Mbps

3.5

Figure 2 Blocking probability as a function of offered load.

0.25 No MLB PRB utilisation-based MLB 0.2 Load-based MLB

0.5 dB 0.25 Proportional Fair Constant bit rate 64 kbps (64 bytes/8ms) Geometric with mean 180 seconds 64 kbps 1 m/s random walk 3 dB 100 ms 0.7 for new users, 0.8 for handover users 0.9 32

0.15

MLB trigger threshold, th Scheduler Traffic model Call duration

0.1

0.05

Ri(req )
UE speed Handover hysteresis Handover measurement period AC load threshold CC trigger load threshold

1.5

2.5 Offered Load Mbps

3.5

Figure 3 Dropping probability as a function of offered load.


5 4.5 4 Handover failure prob. (%) 3.5 3 2.5 2 1.5 1 0.5 0 No MLB PRB utilisation-based MLB Load-based MLB

N clust Tclust
Cluster Radius 25m

2 hours

Cluster Radius 100m

1.5

2.5 Offered Load Mbps

3.5

Figure 4 Handover failure probability as a function of offered load.

Figure 1 UE positions generated in first 15 minutes

350 No MLB PRB utilisation-based MLB 300 5% user throughput in kbps Load-based MLB

0.1 0.09 0.08 0.07 Dropping Prob. (%) Offered load increases 0.06 0.05 0.04 0.03 0.02 0.01 No MLB Load-based MLB

250

200

150

100

50 150

200

250 300 350 mean user throughput in kbps

400

450

0.1

0.2

0.3 0.4 0.5 Cluster radius/ISD

0.6

0.7

0.8

Figure 5 Relationship between 5% user throughput and mean user throughput.


1 0.9 0.8 0.7 Prob(X < abscissa)

Figure 8 Dropping probability as a function of normalized cluster radius.


1 0.9 0.8 Handover failure prob. (%) 0.7 0.6 0.5 0.4 0.3 0.2 0.1 No MLB Load-based MLB

0.6 0.5 0.4 0.3 0.2 0.1 0 No MLB Load-based MLB

0.1

0.2

0.3

0.4 Cell Load

0.5

0.6

0.7

0.8

0.1

0.2

0.3 0.4 0.5 Cluster radius/ISD

0.6

0.7

0.8

Figure 6 CDF of cell load (mean offered load 2.5Mbps/cell).


4.5 4 3.5 3 5% user throughput in kbps 2.5 2 1.5 1 0.5 0 No MLB Load-based MLB

Figure 9 Handover failure probability as a function of normalized cluster radius.


200 190 180 170 Cluster size increases 160 150 140 130 120 0 0.1 0.2 0.3 0.4 0.5 Cluster radius/ISD 0.6 0.7 0.8 110 100 250 No MLB Load-based MLB

Blocking Prob. (%)

260

Figure 7 Blocking probability as a function of normalized cluster radius.

270 280 290 300 mean user throughput in kbps

310

320

Figure 10 Relationship between 5% user throughput and mean user throughput.

Vous aimerez peut-être aussi