On Resource Allocation in Dense Femto-deployments

On Resource Allocation in Dense Femto-deployments Jonathan Ling, Dmitry Chizhik, and Reinaldo Valenzuela Wireless Communications Research Alcatel Lucent, Holmdel NJ Abstract — Femtocells offer a promising way of extending macrocellular network coverage to indoor residential environments. If femtocells are spaced one per home, each on it's own lot, interference low. For dense deployments, i.e. multi-unit construction, SIRS can become very low. Thus we propose a greedy frequency planning algorithm which only uses measurements at the femto-access point. Simulation results show an improvement in SIR due to using the new frequency plan. Index Terms — radio spectrum management, propagation, land mobile radio cellular systems

radio

for all UE locations. Similar algorithms are investigated in the context of 802.11 wi-fi [2], where user assignment and channel selection are optimized for minimum message delay. The remainder of this paper is organized as follows. Section II describes the suburban and urban interference scenarios. For random placement of a user in the home, cumulative density function of interference are presented, motivating the need for interference mitigation. Section III describes the frequency allocation algorithms, and their performance.

I. INTRODUCTION

II. INTERFERENCE ASSESSMENT

Femtocells offer a promising way of extending macrocellular network coverage to indoor residential environments [1]. A femtocell may either share or be on a separate carrier from the macro network. Expected transmit powers are typically 20 dBm or less, to cover homes. A user on the macrocellular network entering the vicinity of its own femtocell will handoff and enjoy higher rates along with improved coverage. Proper operation of femto networks requires careful planning of power and frequency resources, and many of these issues are being discussed in 3GPP and the Femto-Forum. In this study we assume the femto network has a separate channel. We examine femto-femto interference for two different deployment densities. For dense deployment, a distributed frequency planning algorithm is evaluated which improves the rates of users that would otherwise be in outage. This paper studies frequency resource allocation, as 4G systems employ OFDM and easily lend themselves to this kind of orthogonal access. Dividing allocated spectrum further into smaller channels, reduces the spectrum assigned to each user, as well as the interference. Centralized schemes collect all the signal and interference data from all the UEs and APs, and able to plan frequency and power optimally, but are impractical given the large number of femtos. We study the improvement in SIR and the effect on data rate when R multiple non-interfering channels are available for the AP to choose. Two techniques are explored, random where the AP makes a choice of channel without any other information. This could be considered a baseline. Second, a measurement based autonoumous distributed algorithm where the AP makes a measurement of other AP’s pilots, and tries to avoid their interference. Choice of channels that the AP makes, is not dependent of traffic or the location of the UE, and so the UE’s SIR could be low, i.e. the AP’s choice is good, but not optimal

Using ray-tracing, radio propagation is computed for uplink and downlink, over random femto UE placement. We simulated the scenario where femtos are deployed relatively spasely, i.e. one per suburban home with its own lot. The cumulative density function of SIR shows that femto-femto interference is minimal. For denser deployments, i.e. high density dwellings, i.e. apartments, femto-femto interference becomes an issue. We study interference in a garden apartment complex, whose structures consist primarily of wood and sheetrock. Note, such construction is relatively transparent to RF, as compared to concrete, and can lead to high levels of interference. The femto may have either a closed subscriber group policy (CSG) when the femto user equipment (UE) must be connected to it’s own femto access point (AP), or open subscriber group policy (OSG), where the femto UE may be connect to the strongest femto AP. Clearly for CSG, outages may occur when the femto UE moves too close to a neighbor, which we quantify in the simulations. A single-story suburban environment is considered, which consists of regularly placed homes with the interior partitions, Figure 1. Statistics are taken from a center house in the neighborhood. Neighbors may also have femtocells and therefore contribute to interference levels. The spacing between houses is 30 m in x and y axes, and each house is each 20 m by 10 m. Receiver points are specified in the center house every ½ wavelength. Average building penetration loss is related to building construction type, and was found to vary between 3 to 16 dB for suburban homes depending on the construction of the exterior wall [3][4][5]. If the outer walls are concrete block without windows, the loss can be much higher, up to 20 dB around 2 GHz. We have assigned the transmission loss of a wall LBP as 10 dB corresponding a typical exterior wall with foil backed insulation. LBP = 5dB was also used to model

⎛ σ ⎞ γ = ω μ0ε 0ε R ⎜ ⎟ Np/m , ⎝ ωε 0ε R ⎠

Using the suburban scenario, we have simulated the downlink SIRs that a user would experience inside the center house with all neighbors also having a femtocell. All femtocells are transmitting at full power, which is worse case “busy hour”. The results are shown in the CDF of Figure 2. For the partitioned home, 90% of the SIRS are greater than 0 dB for LBP=5 dB and 10 dB for LBP=10 dB. Note, the threshold 0 dB SIR is a typical of outdoor macro systems employing frequency reuse of 1. However, exceeding 10% outage is generally considered undesirable, and it may be concluded that full deployment in areas with more transparently construction homes (i.e. 5 dB exterior wall loss) would lead to unacceptable levels of femto-femto interference. 1 0.9 0.8 0.7 Prob(SIR < abscissa)

more transparent walls, that is exterior walls constructed primarily wood with no metal siding or foil. Vegetation has not been modeled, which would tend to increase over-all building to building losses, or improve in-building SIRs. The interior partitions are sheetrock over wood frame, and modeled as a multilayer dielectric, with additional small empirical reflection and transmission losses to account for furnishings. Where the primary residential construction material is wood, RF propagation losses besides spherical spreading in a home are mostly due to the energy required to polarize molecules in wood which generates heat. The dipole moment is a measure of the total polarisability in the dielectric. Measurements were taken of wood at 1, 5, and 10 GHz at different levels of humidity [7]. These results are reprinted in [6], Figures 4 and 10, and at 2.5 GHz the relative dielectric constant εR is approximately 4, with the loss tangent of approximately 0.25 for typical indoor relative humidities. As an alternative to LBP an exterior wall could be modeled as a multilayer dielectric. It typically consists of wood studs on 16 in centers, covered with ½ inch plywood on the exterior, and sheetrock on the interior. Let the volume of the studs be added to the thickness of the exterior wood sheet. Thus the equivalent wood thickness comes to 0.72 inches. The real part of the propagation exponent is

0.6

LBP=5dB

0.5 0.4

LBP=10dB

0.3 0.2

(1)

and for the parameters assigned to wood γ=26 Np/m. Thus a plane wave will lose half its power traveling through 1/2 inch of wood. Transmission loss at normal incident angle is about 6 dB.

0.1 0 -10

-5

0

5

10

15 20 SIR (dB)

25

30

35

40

Figure 2: Downlink femto-femto interference CDF for suburban homes. A floor plan, Figure 3, was used to create a garden style apartment building, typical of densely populated areas in the United States. Both the interior an exterior walls are woodframed. Covering the frame on the exterior is plywood with plastic vapor barrier, and plastic siding. Interior partitions are all sheetrock, even between units. Floors are also wood framed cover with carpet, and ceilings are sheetrock. A neighborhood of 3-story apartment buildings was simulated using WiSE [8] ray tracing. The neighborhood consisted of a regular grid of 3-story buildings with 4 apartments on each floor, as illustrated in Figure 4.

Figure 1: Layout and signal strength of a suburban house with partitions. Figure 3: Floor plan for urban apartment.

1

0.9

Home Femto Only Strongest Signal Femto

0.8

Prob(SIR < abscissa)

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 -20

-15

-10

-5

0 SINR (dB)

5

10

15

Figure 4: Neighborhood of 3 story apt. buildings Figure 5: CDF of femto-femto interference in urban apts. The CDF of SIR in an urban apartment in the presence of co-channel interfering femtocells in all neighboring apartments, floors and buildings is plotted in Figure 5. It may be observed that here with wooden and sheetrock wall providing little RF isolation and a re-use of 1, the interference levels provide for very poor SIR, with 45% of all locations suffering SIR of less than 0 dB. In this environment, the situation may be improved through greater frequency re-use. To determine the sensitivity of the SIR to exterior wall properties, the exterior wall assigned angle independent 10 dB transmission loss. This setting results in a figure nearly identical to Figure 5, and may be explained by the fact that the properties of the interior partitions remained the same. Thus the power of eleven interferers remain unaffected. Even with the mild exterior transmission loss, a ray must leave and then enter, and thus the total transmission loss is doubled in dB, making signals from neighboring apartments weak in general. If however the reuse factor is large, even weak interferers are significant. A femto user is configured to connect only its home femto AP regardless of the signal strengths of neighboring femto APs. The radio propagation may be such that a UE is faded to home but has a strong signal to neighbor. In macro-cellular systems this problem does not occur as a UE will eventually handoff to the base with the strongest signal. To compare with the policy of always staying on home AP rather than handing off (closed subscriber group) , the urban apartment scenario is simulated where the UE policy is to always select the strongest femto AP (open subscriber group). The greendashed curve in Figure 5, shows that choosing the strongest femto gives about 5 dB SIR gain at the 10-percentile. Interestingly the dashed curve is very similar to the reference outdoor macro SIR curve.

III. INTERFERENCE MITIGATION The frequency selection the AP does, resemble the dynamic channel allocation (DCA) problem that has been researched extensively. For canonical geometries, i.e. cells on a line and hexagonal cells, a quantiy of interest is maximum packing density. A constraint that the next user to be a distance D slots away, so D=1, which means every other cell can be reused. For linear geometry, Cmax=50%, and in general Cmax=(D+1)-1 [9]. The goal of DCA research has been to construct algorithms, with good reuse properties, along with low amount of channel reconfiguration, i.e. channel switching of other users. Katzel and Naghshineh provide survey of various techniques [10]. A way to improve peak rates, is to allocate the full bandwidth during gaps in the neighbors traffic. Usually this requires sensing, as in 801.11, however a way in femto networks to determine the traffic of neighbors is to request the length of the queues from the neighbors from the femto controller. The algorithms investigated here generate static frequency plans which are appropriate as the network becomes heavily loaded. The inclusion of traffic awareness would further improve performance at the expense of signaling to a centralized network element, and this refinement is a topic for future study. The channel allocation problem may more generally be considered an optimization of power over the R subchannels. One expression of the problem may be stated as a maximization of the rate with QOS constraints: Maximize: R Nusers

∑ ∑ log (1 + SINR ) j

i

j

Subject to:

i

(1)

20

R ⎧ ∑j Pi j ≤ Pi,max ∀i ⎪ ⎪ ⎨R ⎪ log (1 + SINR j ) ≥ r ∀i i i ,min ⎪⎩∑ j

(2)

Variables:

∀ i =1...N users j =1...R

(3)

where

SINRi = j

Pi j Gi ,i

∑P G j

k ≠i

k ,i

k

+ ni

(4)

and Gk,i is the pathgain from the k-th AP to the i-th user, and ni is the noise power. The first constraint is on maximum transmit power, and the second specifies a minimum rate. If solved centrally, maximization of the network throughput, requires the identification and measurement of every interferer, reporting the powers back to a central location. Thus the controller will know, if a UE is put on a certain channel with a certain power, what the signal will be to its own AP, and how much interference it will cause others on the same channel. This problem is non-convex, which means that a solver may get stuck in local maxima. Even if centrally solved, the reporting requirements appear to be too high to be practical. The distributed frequency planning algorithm we label Adaptive Autonomous (AA), and it is different from the previous problem of optimizing the network throughput. Rather at each AP we seek to choose a channel that minimizes the interference at the AP itself, assuming this will give good performance over the AP’s coverage area. This may be cast as a Gibbs Sampler [11] or as a classical optimization problem. Given: partitions Ω indexed from 1…R, and APs indexed from 1…NAP

1 Minimize: f (Ω) = N AP Where: wk =

∑∑ G j

i≠ j

ij

In the static case, where the number of nodes and the pathgains are fixed, the algorithm always converges [2]. At each AP, the pathgain Gk,i is known, as well as Ω. Say, AP k leaves channel rn to r’n+1 resulting in Ω’. The level of inteference observed on a channel is also the interference caused to all other APs on this channel. Thus choosing the minimally interfered channel results in f(Ω’) < f(Ω). Since there are a finite number of states, the algorithm will find a (local) minima. To compare different algorithms we define the ratio:

Θ=

k =1

k

0.9

Exhaustive

(5)

i, j ∈ Ω( k )

w is the weight of each partition and is simply the sum of the interference seen at each AP belonging to that partition, and the objective function is the average interference seen at the AP. (5) can be solved exhaustively by computer for small N R and NAP with search space sized R AP . Due to symmetry the search space actually smaller, i.e. for NAP =3, R=2 [ 1 2 1] and [2 1 2] are equivalent, and computing the smaller set efficiently is an open problem. Simulated annealing also obtains the solution, at the expense of slow convergence Greedy approaches converge quickly, but typically do not find

(6)

1

R

∑w

f ( Ω) ’ f (Ω1 )

where Ω1 denotes a set of partitions with one non-empty partition, re-use R=1, i.e. all transmitters on the same channel. This metric quantifies the relative improvement in SINR at the AP achieved through frequency selection. We compared three algorithms: centralized exhaustive search, random channel selection, and Adaptive Autonomous algorithm. In random selection, APs initialize to one of R channels, without regard for their neighbors, and the channel stays fixed thereafter. A numerical experiment was performed for NAP=12, R=3, and with G from a single 12 unit apartment. Figure 6 contains a plot of the cumulative probability distribution (cdf) of Θ defined in (6). It may be observed that random selection does poorly, while AA is within a factor of 2 of the solution found by exhaustive centralized search.

0.8

Pr ( Θ < Abscissa)

Pi j

the optimal solution. Adaptive Autonomous may be stated as follows: 1. APs initialize to some random order. 2. When an AP initializes it takes measurements over all R channels (where APs are broadcasting pilots), and chooses the one with the lowest power, if the interference is 1 dB below the interference in its current channel. It then emits a beacon/pilot on this channel. 3. Wait a random time delay. Goto Step 2, until no AP changes channels.

AA

0.7

Random

0.6 0.5 0.4 0.3 0.2 0.1 0 -2 10

10

-1

Interference Ratio: Θ

10

Figure 6: Performance of different channel selection algorithms.

0

18 16 14 Percentage of time

Adaptive Autonomous attempts to improve SINRS for UEs in its vicinity, but does not try change the frequency plan for UE’s specific propagation condition. So for example, if a UE moves closer to a neighboring AP and its SINR becomes worse, nothing is done. When the measurement location and interference source (to other) are at the same, and this improves the stability of the selection. That is, if the AP measures a channel with weak interference, this means that pathloss to those interferers is high, and if the AP broadcasts on this channel it will affect those other AP’s minimally as compared to broadcasting on any other channel. The algorithms were run over the full urban scenario of 9 buildings, and 36 apartments. In Figure 7, we observe that random performs worse than R=1. However for AA, we observe doubling in rate, with R=5 and 10, for UEs with poor SINR, i.e. around 10th to 20th percentile. Both R=5 and R=10 perform similarly in terms of throughput over most of the cell, although lower values of R provide some benefit when the UE is close to the AP.

8 6

2 0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Number of iterations

Figure 8: Histogram of number of interations for AA to complete. 100 Percentage of UEs Swapping Channels

0.9 0.8 Prob(Rate < abscissa)

10

4

1

0.7 0.6 0.5 0.4 R=1 R=5 Random R=10 Random R=5 AA R=10 AA

0.3 0.2 0.1 0 0

12

1

2 3 4 5 6 7 8 Shannon Rate (Mbits) with 3.8 MHz bandwidth

9

10

Figure 7: CDF of Shannon rates for different R and algorithm. The speed (or number of iterations) the algorithm requires is an important quality. Another desirable feature is that the number of nodes switching decay as the algorithm is run. These qualities are investigated numerically. The algorithm AA was run for several values of R, and since the results are similar, R=5 is taken as representative. Figure 8 shows the percentage of mobiles switching their channel each iteration, if the algorithm did indeed need to continue to this iteration. The percentage is 100% for the first iteration, and drops quickly to less than 10% on the second, and falls slowly in for Figure 9 gives a histogram of the following iterations. number of iterations necessary to settle on a frequency plan. The maximum number of iterations observed is 16, but 95% of the time the algorithm completes in only 8 iterations. We conclude that AA requires only a small percentage of APs switching channels, and requires only a few iterations.

90 80 70 60 50 40 30 20 10 0 0

2

4

6

8 10 12 Iteration Number

14

16

18

Figure 9: Percentage of UE's switching channels. CONCLUSIONS Femto-femto interference is assessed in suburban and urban areas, where each home is assumed to have a single femto broadcasting at full power. For suburban areas, if the building construction results in relatively transparent outer walls (5 dB avg. loss) , SIRs are 0 dB or less 10% of the locations. For urban areas, specifically garden style apartments with wood construction, SIRs are 0 dB or less 45% of the locations. Here, the open subscriber group model improves SIRs by 5 dB at 10%. A greedy channel selection algorithm was tested on a garden apartment complex, with 12 apartments in each building. The neighborhood size is a total of 9 apartments. In a small example, the performance of the algorithm, is compared to the exhaustive search for small R and NAP. The greedy algorithm gets within a factor of two of the optimal solution found by exhaustive search. The actual performance,

is quantified using the new frequency plan, for different vales of R. With UEs placed randomly over the apartment area, the SIR cumulative density function steepens and shifts to the right, demonstrating the usefulness of the approach. REFERENCES [1] L. T.W. Ho, and H. Claussen, “Effects of user-deployed, co-channel femtocells on the call drop probability in a residential scenario,” Proc of IEEE PIMRC 2007. [2] B. Kauffmann, F. Baccelli, A. Chaintreau, V. Mhatre, K. Papagiannaki, C. Diot, “Meaurement-based self organization of 802.11 wireless access networks,” Proc. of IEEE Infocom, 2007, pp. 1451-1459. [3] D. C. Cox, R. R. Murray, A. W. Norris, “800 MHz attenuation measured in and around suburban houses”, Bell Labs Technical Journal, Vol. 63, No. 6, July-August 1984, pp. 921-954. [4] J. H. Loew, Y. Lo, M. G. Laflin, E. E. Pol, “Building penetration measurements from low-height base stations at 912, 1920, and 5990 MHz,” NTIA-95-325, Sept. 1995. G.

[5] Durgin, T. Rappaport, H. Xu, “Measurements and models for radio path loss and penetration loss in and around homes and trees at 5.85 GHz”, IEEE Trans. On Comm., Vol. 46, No. 11, Nov. 1998. [6] W. James, “Dielectric properties of wood and hardboard: variation with temperature, frequency, moisture content, and grain orientation,” USDA Forest Research Paper FPL-245, 1975. [7] K. Kroner, L. Pungs, “Dielectric loss factors in natural wood at high frequencies,” Holzforsch, 7(1), 1953, pp. 12-18. [8] S. J. Fortune, D.M. Gay, B.W. Kernighan,O. Landron, R.A. Valenzuela, M.H Wright, “WISE design of indoor wireless systems: practical computation and optimization,” IEEE Computational Science & Engineering, Spring 1995, pp. 58 – 68. [9] L. Cimini, G. J. Foschini, L. A. Shepp, “Single-channel user-capacity calculations for self-organizing cellular systems,” IEEE Trans. On Comm., Vol. 42, No. 12, Dec. 1994, pp. 31373143. [10] I. Katzela, M. Naghshineh, “Channel assignment schemes for cellular mobile telecommunications systems: a comprehensive survey,” IEEE Personal Communications, June 1996, pp. 10-31. [11] P. Bremaud, “Markove chains, Gibbs field, Monte Carlo simulation and queues,” Springer-Verlag, 1999.