(WiFi/LTE) Heterogeneous Networks - IEEE Xplore

Globecom 2013 Workshop - Broadband Wireless Access

Characterizing Performance of Load-Aware Network Selection in Multi-Radio (WiFi/LTE) Heterogeneous Networks Mikhail Gerasimenko† , Nageen Himayat? , Shu-ping Yeh? , Shilpa Talwar? , Sergey Andreev† , and Yevgeni Koucheryavy† †

Tampere University of Technology, Tampere, Finland; ? Intel Corporation, Santa Clara, CA, USA Email: {mikhail.gerasimenko, sergey.andreev}@tut.fi, {nageen.himayat, shu-ping.yeh, shilpa.talwar}@intel.com, [email protected] Abstract—In this paper, we consider the problem of network selection between different radio access technologies (RATs) deployed as part of an operator managed multi-RAT heterogeneous network. An urban deployment scenario is studied where WiFi small cells are overlaid on top of the 3GPP LTE network. We assume limited cooperation across the multi-RAT network and emphasize user-centric network selection algorithms to minimize feedback overhead and to better account for user preferences. Specifically, we investigate schemes that rely on network loading information with suitable adaptation of hysteresis mechanisms and compare them with WiFi-preferred schemes that only account for signal strength measurements. We also benchmark the performance of load-aware schemes against conventional cellrange extension methods that use network-wide optimization to offload users to small cells. The results of our systemlevel performance evaluation show that load-aware user-centric schemes can provide improved performance compared to the WiFi-preferred schemes and may even outperform networkbased cell-range extension schemes under some conditions.

I. I NTRODUCTION A. Motivation and scope In recent years, there has been an increasing shift towards tighter interworking across different radio access technologies (RATs). For example, in the past, WLAN (WiFi) and cellular standards evolved largely independently, but today mobile network operators increasingly rely on WiFi offload as a low cost solution to relieve traffic congestion on their networks. As WiFi becomes an integral part of operator’s strategy to meet future traffic demand, cellular operators desire more control on how WiFi capacity is utilized and managed in their networks. For this reason, the Third Generation Partnership Project (3GPP) has recently been very active in specifying interworking solutions between 3GPP cellular technologies such as LTE (UMTS) and the IEEE 802.11/WiFi technology. The areas of investigation range from solutions for trusted access to 3GPP services with WLAN devices, seamless mobility between 3GPP & WLAN technologies, and support for Access Network Discovery and Selection (ANDSF) functions [1], [2], [3]. While the bulk of considered interworking solutions involve functions within the 3GPP core network, interworking options that require cooperation within the Radio Access Network (RAN) are also starting to be addressed [4]. We expect this trend to continue with increased integration of WiFi into the 3GPP cellular RAN. 978-1-4799-2851-4/13/$31.00 ©2013IEEE

LTElink

LTEMacro BS

WiFi link

Wi-Fi onlyAP LTE-OnlyPico-cell LTE-OnlyFemto-cell Integrated LTE/AP BS LTE-OnlyUE Multi-RATUE Wi-Fi only device

Fig. 1.

Example of Het-Net topology

Figure 1 illustrates our vision of an operator’s multi-RAT heterogeneous network (Het-Net) deployment. It shows a hierarchical deployment of wide-area macro cells providing ubiquitous coverage augmented with the overlay tier of smaller cells (picos, femtos, WiFi access points, integrated WiFi-LTE small cells, etc.) to enhance capacity. Reference [5] provides a more detailed view of multi-RAT Het-Nets and the available integration options. In a nutshell, when WiFi is managed as part of an operator’s RAN, increased level of cooperation between WiFi and 3GPP infrastructure may become feasible. In the simplest case, no cooperation between WiFi and cellular RAN is available and the users (UEs) are left to determine how the two RATs are utilized [6]. One may also envisage an architecture where integrated WiFi-LTE small cells may enable full cooperation between the two RATs, allowing for WiFi to simply become a “virtual 3GPP carrier” anchored on the 3GPP radio network [7]. We note that multi-RAT small cells with collocated WiFi and 3GPP interfaces are an emerging industry trend for lowering deployment costs by leveraging common infrastructure across multiple RATs. However, given that such deployments are presently not common, current standardization efforts aim to improve UE-centric interworking architectures while assuming only limited cooperation or assistance across multi-RAT network. In this paper, we focus on the important problem of network selection between WiFi & LTE RATs, assuming WiFi is part of an operator deployed and managed multi-RAT HetNet. We consider simple extensions to improve performance of UE-centric network selection schemes. To be consistent

397


with current network deployments, we consider distributed small cell overlay with standalone WiFi access points (APs), assuming that there is no interface between the WiFi and the 3GPP radio networks. In particular, we investigate schemes that account for network loading information across the 3GPP and the WiFi RATs and compare them with schemes that only rely on signal strength measurements. We also benchmark the performance of UE-centric RAT selection with optimized network-based load balancing mechanisms. Intuitively, network-centric solutions may seem to offer better performance compared to UE-based approaches as network-wide radio link information across users can be employed to develop optimum RAT assignment algorithms [7]. However, with distributed architectures assuming no direct cooperation between 3GPP & WiFi RATs, such solutions may only be developed through extensive UE feedback which could result in significant overheads. UE-centric RAT selection may also be desirable because the UE can better account for user preferences and application QoE. In summary, this paper compares performance of load-aware schemes with conventional/existing UE and network-based schemes used in current systems. In particular, the network selection mechanisms considered are operating at the RAN layer, which resides below the IP layer. More specifically, we focus on uplink performance as it has not been fully addressed (e.g., [5] has focused on downlink performance results). Complete comparison with downlink performance is not included due to limited space and will be addressed in future submissions. Comparison with more advanced networkcontrolled schemes will also become an area of our future investigation. B. Related work and our contribution Recent literature sets extensive focus on cellular Het-Net deployments using small cells. The work in [8] offers a comprehensive view of LTE development trends wherein small cells are envisioned as one of the main approaches for capacity enhancement in cellular networks. Reference [9] compares conventional cellular vs. Het-Net deployments and emphasizes their relative performance with different interference coordination schemes. In [10], small cell discovery mechanisms are studied in detail by explicitly accounting for energy efficiency. Frequency planning analysis for Het-Nets is considered in [11] while the importance of self-organization in a Het-Net context is highlighted by [12]. Whereas most existing research concentrates exclusively on LTE small cells, recent papers now consider operator deployments based on WiFi small cells as well. In [13], the potential for traffic offloading as well as the economic benefits of Het-Nets are evaluated accounting for the use of WiFi small cells. Further, in [7], Het-Net performance inclusive of WiFi APs and integrated WiFi-LTE small cells is considered. The literature on network selection in Het-Net deployments is also represented by multiple papers. Here, a bulk of contributions focuses on cell-range extension methods that offload users from the macro cell to the small cell tiers [9]. Reference [7]

studies network-controlled cell/RAT selection and assignment schemes for Het-Nets based on integrated WiFi-LTE small cells. While conventional cell-range extension methods are used to offload users to multi-RAT small cells, a utility maximization framework is used therein for each small cell to partition users between WiFi and LTE RATs. Reference [14] also follows a distributed per cell approach for cell/RAT assignment using load and signal strength based adaptive learning mechanisms. Fully distributed UE-centric techniques for inter-RAT network selection are considered in [6], where extensive theoretical characterization of loadaware UE-centric network selection is presented. Here, a UE selects a RAT based on greedy maximization of its throughput. The paper uses generic throughput classes representative of WiFi and cellular RATs to characterize the optimality gap and the convergence behavior of UE-centric RAT selection. Further, RATs are assumed to be orthogonal, wherein selection of a particular RAT does not change the interference environment in the network. While suitable hysteresis mechanisms are designed to guarantee stability, it is shown that RAT selection between WiFi and LTE could take several steps to converge and could also be sensitive to measurement errors. Fast convergence is extremely important design parameter for UE-based schemes, as current inter-RAT mobility procedures are expensive in terms of overhead and latency. This paper builds on the work in [6] to evaluate system performance of UE-centric RAT selection with more practical 3GPP-complaint multi-RAT Het-Net models, which also account for detailed modeling of WiFi MAC protocols. Hence, more realistic interference environment is considered across both 3GPP and WiFi RATs. We also focus on important performance metrics such as the aggregate system throughput, distribution of user throughputs, and fairness criteria not captured by Pareto optimality considerations in [6]. Practical load-aware network selection algorithms for throughput maximization are proposed with suitable adaptation of hysteresis mechanisms. Another key contribution of this work is the consideration of uplink performance in which the RAT selection decisions could change the interference environment in 3GPP networks in addition to WiFi networks. The rest of the text is organized as follows. In Section II, we detail the important performance metrics, the system evaluation methodology, and the operation of the LTE/WiFi system-level simulator used in our evaluation. Section III describes the network selection algorithms considered in this paper, including the improved load-aware scheme. The system performance results are summarized in Section IV together with the analysis of proper hysteresis mechanisms. Finally, Section V concludes the paper and gives directions for some future work. II. M ETHODOLOGY AND OBJECTIVES A. Scenario and evaluation parameters In this work, we concentrate on a multi-RAT simulation model representative of an urban deployment, where WiFi small cells are overlaid on top of the 3GPP cellular network. Outdoor deployments are considered and are based

398

Globecom 2013 Workshop - Broadband Wireless Access TABLE I I MPORTANT SIMULATION

PARAMETERS

Parameter

Value

LTE/WiFi configuration

10 MHz FDD / 20 MHz

Macro cell layout

7 cells, 3 sectors each

Inter-site distance (ISD)

500 m

UE to eNodeB/AP pathloss model

ITU UMa/UMi [17]

eNodeB antenna gain

12 dB

eNodeB/AP/UE maximum power

43/20/(23/20 LTE/WiFi) dBm

LTE power control

Fractional (α=1.0) [15]

WiFi power/rate control

Max-power/ARQ

UE/eNodeB/AP antenna height

1.5/25/10 m

UE noise figure/feeder loss

9 dB/0 dB

Traffic model

Full-buffer

Number of UEs/APs

30/4 per Macro cell (3 sectors)

AP/UE deployment type

Uniform/clustered (4b in [15])

AP/UE-eNodeB, AP/UE-UE distance

> 75/35 m, 40/10 m [16]

Modeling time

3s

Number of trials per experiment

30

on recommendations in [15] and [16]. A brief summary of the parameters is provided in Table I. Specifically, we consider a loaded WiFi network with WLAN APs uniformly distributed across the cellular coverage area. Most UEs cluster around the APs, which recreates a hot-spot area (airport, restaurant, shopping mall, or university campus) with many bandwidth-hungry users loading the WiFi network. Moreover, around one third of UEs are still deployed uniformly across the cellular network [15] mimicking regular mobile users. Whereas this scenario may not accurately capture all practical urban conditions, it represents a harmonized 3GPP vision of a characteristic Het-Net deployment. Our additional assumptions are as follows. (i) Users do not move over one simulation run and transmit their full-buffer uplink data into the network. This simplification allows us collect reliable network performance statistics. (ii) The 1-3-3 frequency reuse is assumed where every LTE base station (eNodeB) employs three different frequencies across its three sectors. This corresponds to operator’s desire to facilitate interference coordination thus improving network performance at the cell edge. (iii) LTE network schedules users according to a round-robin discipline. Whereas practical schedulers are typically more elaborate, the round-robin assumption reflects the case of fair resource allocation, which is in a way similar to what WiFi (802.11e/n) does in the saturated scenario. B. LTE/WiFi system-level simulator The system-level simulator (SLS) developed for the purposes of this study is an elaborate event-driven tool that mimics a complete LTE-WiFi system deployment compatible with 3GPP LTE Release-10 and IEEE 802.11-2012 specifications. It comprises several software modules modeling the deployment of wireless infrastructure and user devices, control events related to transmission of signals between several distinct types of transmitters and receivers, abstractions for

wireless channels, mechanisms for collecting measurements and statistics for quantifying system performance, etc. The SLS takes advantage of flexible modular structure which allows easy integration of technology updates and network selection mechanisms. Specifically, our tool employs four programming languages: Python (for MAC and upper layers), C++ (for PHY layer), Matlab (for parsing the collected statistics), and Bash (for shell scripts). Owing to its event-driven nature and efficient PHY-layer implementation (in C++), it allows producing statistically reliable results without hitting high computational intensity. The supported PHY-layer features include pathloss, shadow fading, fast fading and antenna directivity effects, as well as broadcast interference calculations and wrap-around of border-cell signals. MAC layer embraces rate and power control, ARQ/HARQ, as well as scheduling and resource allocation mechanisms. Since the implementation of LTE functionality is defined and detailed in [15], [16], below we only concentrate our attention on WiFi features. The SLS reproduces 20 MHz singlechannel WiFi based on carrier sense multiple access with collision avoidance (CSMA/CA) function. Due to contentionbased channel access, we employ RTS/CTS handshake to mitigate the hidden node problem. The binary exponential backoff (BEB) collision-resolution protocol is modeled in every detail including support for advanced signaling (such as the concepts of transmission opportunity and block acknowledgment). WiFi MAC operation also supports data frame fragmentation and reassembly as well as packet aggregation mechanisms. All the above ensures accurate performance results by contrast to many past works which oversimplify contention-based operation or model legacy versions of the MAC protocol. More detailed description could be found in [18]. C. Metrics of interest The major expected outcome of leveraging WiFi small cells is efficient offloading of cellular user traffic resulting in significant user benefits. For that reason, our primary metric of interest is the UE throughput which, in turn, determines the overall system capacity. However, as quality of service (QoS) may be equally important, we also account for fairness between the users expressed in terms of Jain’s index: !2 N X 1 , J= xi N P 2 i=1 N· xi i=1

where xi is the throughput of user i and N is the total number of UEs in the system. Naturally, fairness index indicates how large is the deviation between actual user throughput and the cell-average performance. Stability of UE-centric schemes is another very important aspect of UE-centric RAT selection, as excessive ping-ponging between RATs is undesirable due to the overhead and latency of mobility protocols as well as due to energy efficiency considerations. Here we observe and report the number of cellular/WLAN re-connections and explore hysteresis mechanisms to improve performance.

399


Broa d load cast mea sure men ts

AP/eNodeB SNR measurements

SNR>40 dB? Yes No

tch Swi P A n o

y

ta to/s

Swit ch on e to/sta y Nod eB

eNodeB

No

rand(0..1)>pmi+1

No

Yes

Switch to new Stay on old WiFi/LTE AP/BS i, WiFi/LTE AP/BS mi+=1

AP/eNodeB RSSI measurements

Et_new>Et_old*

Et_new/Et_old- expected throughput on new/old interface (IF)

Fig. 2.

b

Considered network selection algorithms

III. N ETWORK SELECTION SCHEMES A. SNR-threshold algorithm The simplest SNR-threshold algorithm serves as our baseline user-centric network selection scheme [19]. With this solution, a UE is continuously monitoring the signaling messages from the neighboring WiFi APs to obtain timely SNR information. Ideal SNR measurements are assumed in this paper. In practice, such estimates may be generated by measurements on the beacon signals. When a particular SNR value exceeds a predefined threshold (which we chose to be 40 dB as per [19]), the user starts steering its traffic to the respective WiFi AP. Otherwise, it keeps transmitting on LTE network. Naturally, such behavior is an automatized version of what human user would do: whenever a hot-spot with reliable signal is available, UEs switch to WiFi to enjoy higher data rates and reduce expenses associated with paid cellular traffic. The operation of the SNR-threshold scheme is shown in Figure 2(a). B. Load-aware algorithm Due to its simplicity, the baseline SNR-threshold scheme may experience limitations in dense interference-limited scenarios which are typical for urban deployments. For instance, a hot-spot AP may experience overload conditions when a significant number of users try to steer their traffic through it. Moreover, nomadic WiFi users, such as those with laptops, could consume most of WLAN capacity. To make matters worse, the WiFi medium access is contention-based which results in non-linear degradation of the throughput performance with increasing number of users. Therefore, the load-agnostic SNR-threshold scheme is not expected to stay effective in environments with varying load. In such situations, UEs may attempt to combine SNR knowledge with additional knowledge of the loading information from the network infrastructure (eNodeB/AP). While accounting for WiFi load would certainly improve performance beyond the SNR-threshold scheme, it is easy to construct scenarios where accounting for WiFi load only will not be sufficient. Hence, we focus our investigation on schemes that account for both 3GPP and WiFi load and compare them with existing network-based schemes which have been standardized in 3GPP for small cell offload. The improved load-aware scheme may work as follows.

eNodeB e valu Bias

me

Yes

UE

WiFi AP

st dca Broa load ents rem asu

Et_new>Et_old?*

AP/eNodeB SNR measurements

a

UE

WiFi AP

Measurement window

Measurement window

eNodeB

Measurement window

UE

WiFi AP

RSSI_LTE< RSSI_WiFi+bias?*

Yes No

t to/s tch Swi P on A

Swit ch on e to/sta y Nod eB

ay

AP/eNodeB RSSI measurements * hysteresis is taken into account

c

Throughput estimation: User attempts to listen to both interfaces in order to monitor the SNR information in its neighborhood and estimates its expected throughput. For WiFi, such estimation is conducted based on predicted network capacity divided by number of UEs connected to a particular AP (as advertised by AP through the load indicators in the beacon frames) as well as accounting for several weighting factors. First, a coefficient α is taken into account, which equals 0 if SNR of the AP candidate is lower than 40 dB and equals 1 otherwise (modified by UE if SNR situation changes). The motivation behind such SNR weighting is to exclude APs with low signal quality, whereas its value is chosen to benchmark against the SNR-threshold scheme fairly. Second, a coefficient β accounts for the contention-based nature of WiFi channel access and includes signaling overheads as well as collision losses. The value of β can be derived based on characteristic dependencies of saturation throughput on the number of users [20] and has been taken constant (around 0.7) in our evaluations. Importantly, a user may adjust its α and β parameters to improve the associated performance. For LTE, throughput prediction may be simply based on the scheduler advertisements and the used power control. Randomization: User selects the network with the highest expected throughput value probabilistically [6] rand(0..1) < pmi +1 , where mi is the number of recent connections to this AP/eNodeB and p is the number in (0, 1), which is representing the re-connection probability. Initially, p is equal to 0.5, so that a user would be likely to attempt other RAT. The proper use of p reduces the number of concurrent re-connections to the same AP/eNodeB, which will prevent uncontrollable hopping from one interface to another. For simplicity, we assume the number of concurrent connections to be ideally known by a user. In practice, this information may be obtained by e.g. listening to the broadcast transmissions from its neighbors and/or using load indications by APs/eNodeBs. If a network re-selection occurs, mi is incremented for AP/eNodeB i. Other users are taking into account this information by dividing their expected throughput value for this eNodeB/AP by mi + 1. This allows to control dynamic re-selections on both networks. Otherwise, the expected throughput information may not be very useful due to frequent re-selections (up to 12 per 2 s).

400


90

80

80

70

70

60

60

CDF, %

CDF, %

90

50 40

30 20 SNR-threshold scheme, Total Load-aware scheme, Total

0 0

2

4

6 8 Throughput, Mbps

a

Load-aware scheme, LTE Load-aware scheme, WiFi SNR-threshold scheme, WiFi SNR-threshold scheme, LTE

40

20

70

12

Fig. 3.

0 0

LTE WiFi

50 40 30 20 10

10 10

Average time spent by user on each interface, %

60

50

30

10

Per user throughput, Mbps

100

time if was active/total time, %

Per user throughput, Mbps

100

2

4

6 8 Throughput, Mbps

b

10

12

0

SNR-threshold and load-aware schemes comparison

Hysteresis: To decrease the number of cell-border switchings, a hysteresis value is added to the current expected throughput value. We have chosen the value of 1.76 dB for our scenario. The choice of hysteresis parameters will be discussed in the following section. Filtering of throughput estimation: Further improvement in throughput estimates is obtained through averaging. After each measurement window of T = 100 ms (chosen empirically), we measure the actual throughput obtained over this period and filter our expected value with a moving average filter. The resultant value, which combines the measured and the predicted throughput, is then used as the expected throughput value for this AP/eNodeB. This averaging is made to achieve more reliability, which could suffer due to contention-based channel access. The simplified time-diagram of load-aware network selection scheme is shown in Figure 2(b). C. Cell-range extension based on RSSI bias The above two network selection schemes are user-centric in nature. Hence, they may not result in the optimal system wide performance, which may otherwise be achievable through network-based centralized mechanisms. Here we consider the conventional cell-range extension schemes applied to steer users to small cells employing a network optimized RSSI bias value. The RSSI bias serves to increase/decrease the effective WiFi AP coverage area depending on the network capacity expectations. One limitation of this method is that optimal bias value needs to be adapted based on network-wide knowledge of user distribution. For example, our results show that the optimal bias depends on user deployment model as well as the interference levels in the network, which may not always be feasible as typically WiFi cells may not have a direct interface to eNodeB. In this paper, we evaluate RSSI-based cell-range extension with bias values optimized for the target scenario. We also use hysteresis for the RSSI-based algorithm. The timediagram of this method is shown in Figure 2(c). IV. N UMERICAL RESULTS A. Comparing load-aware and SNR-threshold schemes The cumulative distribution function (CDF) of user throughput comparing performance between SNR-threshold and loadaware scheme is shown in Figure 3(a). The results indicate that the load-aware scheme gives visible benefits at the cell edge (e.g., over 44% of improvement is observed in 5% quantile),

Load −aware

SNR threshold scheme type

c

as well as some improvement in the average throughput. In more detail, Figure 3(b) clarifies the CDF of user throughput across LTE and WiFi separately and Figure 3(c) highlights the average time spent by users on each interface. It may be seen from Figures 3(b) and 3(c) that the load-aware scheme is effective in utilizing the capacity available from the LTE eNodeB, whereas SNR-threshold scheme continues to keep the users that locate near the WLAN APs on the WiFi interface. This is because the SNR of these users is high, while the real throughput they could obtain may be very low because of intensive neighboring interference. In terms of fairness, the Jain’s index of the load-aware scheme (0.7) is also higher than that for the SNR-threshold scheme (0.4). B. Comparison with cell-range extension schemes We also account for the performance of optimized cell-range extension scheme based on RSSI bias, where the network-wide optimization is expected to result in improved performance. However, from Figure 4(a) we learn that even with a network controlled bias value, the mean value of throughput per node is still lower than that in the load-aware case. This is due to the fact that our optimization, based on RSSI value, may not be the best solution in the considered scenario. The main reason is that the cell-range extension scheme just increases the effective WiFi cell radius with respect to the bias level. This could work well in the scenario with uniformly-deployed UEs, but in the clustered case the interference between WiFi users needs to be considered as well. Similar effects are observed for the 5% quantile performance. C. Analysis of re-connection behavior and hysteresis In Figures 4(b) and 4(c), we characterize the re-connection behavior of RSSI-based and load-aware schemes. Figure 4(b) shows that the load-aware scheme on average may have increased number of re-connections when compared to the RSSI-based scheme. Higher number of re-connections, caused by the so-called ping-pong effect (continuous cell-edge reconnections), could significantly decrease the benefits of this scheme, from both capacity and QoS points of view. Hence, suitable hysteresis mechanism should be used to control this behavior. With high enough values of hysteresis, the difference between the number of re-connections becomes vanishingly small. Nevertheless, too high hysteresis may decrease the celledge (Figure 4(c)) and mean system throughputs. Therefore,

401

Globecom 2013 Workshop - Broadband Wireless Access Mean per user throughput, Mbps

4

Mean Number of reconnections per node

1.5

Mean per user throughput, 5% quantile, Mbps

1.2

3.5

2 1.5 1

R S S I-based scheme L oad-aware scheme 1

Throughput, Mbps

2.5

N umber of reconnections

Throughput, Mbps

1.1

Total, Load-aware scheme Wifi interface, RSSI-based scheme Lte interface, RSSI-based scheme Total, RSSI-based scheme 5% quantile, Load-aware scheme 5% quantile, RSSI-based scheme

3

0.5

−5

0.9 R S S I-based scheme Load-aware scheme 0.8

0.7

0.5 0

1

0

5 10 Bias level, dBm

a

15

20

25

0

0

1.76

3

Fig. 4.

4

6 Hysteresis, dB

b

RSSI-based scheme analysis

there is an evident trade-off between system performance and lowering the number of re-connections. To achieve good balance, in our simulations we choose a value of 1.76 dB as the optimal point for the RSSI-based scheme. On the other hand, Figure 4(c) indicates that the hysteresis value of 3 dB is compensating the fading effects better and hence improves the cell edge comparatively to the value of 1.76. However, it also decreases the importance of load prediction and thus degrades the overall system capacity. Summarizing, increased margin may lower the number of re-connections, but could also adversely impact system performance as shown by Figure 4(b) and 4(c). Hence, a suitable balance must be achieved. Further, the hysteresis margin may also be adapted based on expected estimation error in throughput estimates (e.g., higher unreliability in throughput may warrant a higher hysteresis margin). V. C ONCLUSIONS AND FUTURE WORK In this paper, we consider several UE-based network selection algorithms for multi-RAT Het-Net deployments. Our results show that the load-aware user-centric scheme, which augments SNR measurements with additional information about network load, could improve the performance of WiFipreferred scheme based on minimum SNR threshold. We observed a 44% improvement in 5% cell-edge throughput for the clustered user deployment scenario considered in this paper. Under some conditions, load-aware scheme may even outperform network-based cell-range extension schemes. It was also observed that despite capacity and cell-edge advantages of this algorithm, it exhibits unwanted inter-RAT switching for cell-edge users. We showed that such switching may be controlled through application of suitable hysteresis mechanisms. The main advantages of load-aware schemes stem from the fact that the SNR-threshold scheme, as well as the network-centric cell-range extension scheme, do not account for the loading and interference on the WiFi APs typically encountered in clustered UE deployments. Next steps include further investigation of UE-based schemes while explicitly considering load variation in the network and accounting for application-layer statistics. System behavior in the presence of uncoordinated (rogue) WiFi interference must also be accounted for and hysteresis mechanism may further be improved to combat the uncertainty in estimating user throughput. The effect of user mobility must also be addressed.

8

0

1.76

3

4

6 Hysteresis, dB

8

c

ACKNOWLEDGMENTS This work is supported by Intel Corporation and the IoT SRA program of Digile, funded by Tekes. R EFERENCES [1] “Access to the 3GPP Evolved Packet Core (EPC) via non-3GPP access networks,” 3GPP Technical specification (TS) 24.302, 2013. [2] “Architecture enhancements for non-3GPP accesses,” 3GPP Technical specification (TS) 23.402, 2013. [3] “Network based IP flow mobility,” 3GPP Technical Report (TR) 23.861, 2012. [4] “Study on WLAN/3GPP Radio Interworking,” 3GPP Technical Report (TR) 37.834, 2013. [5] S.-P. Yeh, S. Talwar, G. Wu, N. Himayat, and K. Johnsson, “Capacity and coverage enhancement in heterogeneous networks,” IEEE Wireless Communications, vol. 18, no. 3, pp. 32–38, 2011. [6] E. Aryafar, A. Keshavarz-Haddad, M. Wang, and M. Chiang, “RAT selection games in HetNets,” in Proc. of IEEE INFOCOM, pp. 1–11, 2013. [7] A. Y. Panah, S.-P. Yeh, N. Himayat, and S. Talwar, “Utility-based radio link assignment in multi-radio heterogeneous networks,” in Proc. of International Workshop on Emerging Technologies for LTE-Advanced and Beyond-4G on IEEE Globecom, pp. 618–623, 2012. [8] B. A. Bjerke, “LTE-Advanced and the evolution of LTE deployments,” IEEE Wireless Communications, vol. 18, pp. 4–5, Oct. 2011. [9] J. G. Andrews, “Seven ways that HetNets are a cellular paradigm shift,” IEEE Communications Magazine, vol. 51, no. 3, pp. 136–144, 2013. [10] A. Prasad, O. Tirkkonen, P. Lund´en, O. N. C. Yilmaz, L. Dalsgaard, and C. Wijting, “Energy-efficient inter-frequency small cell discovery techniques for LTE-Advanced heterogeneous network deployments,” IEEE Communications Magazine, vol. 51, no. 5, pp. 72–81, 2013. [11] W. Lei, W. Hai, Y. Yinghui, and Z. Fei, “Heterogeneous network in LTE-advanced system,” in Proc. of IEEE International Conference on Communication Systems (ICCS), pp. 156–160, 2010. [12] M. Peng, D. Liang, Y. Wei, J. Li, and H.-H. Chen, “Self-configuration and self-optimization in LTE-Advanced heterogeneous networks,” IEEE Communications Magazine, vol. 51, no. 5, pp. 36–45, 2013. [13] M. El-Sayed, M. Hill, and P. Gagen, “Mobile data explosion and planning of heterogeneous networks,” in Proc. of XVth International Telecommunications Network Strategy and Planning Symposium (NETWORKS), 2012. [14] M. Bennis, M. Simsek, W. Saad, S. Valentin, and M. Debbah, “When cellular meets WiFi in wireless small cell networks,” IEEE Communications Magazine, vol. 51, no. 6, pp. 44–50, 2013. [15] “Further advancements for E-UTRA physical layer aspects,” 3GPP Technical Report (TR) 36.814, 2010. [16] “Coordinated multi-point operation for LTE physical layer aspects,” 3GPP Technical Report (TR) 36.819, 2011. [17] “Guidelines for evaluation of radio interface technologies for IMTAdvanced,” ITU, 2009. [18] “WINTERsim system-level simulator description, 2013,” http://www.cs.tut.fi/˜andreev/data/wintersim.pdf. [19] Intel Corporation, “R2-131348. WLAN/3GPP access network selection based on maximum achievable rate metric,” 2013. [20] C.-E. Weng, C.-H. Chen, C.-H. Chen, and J.-H. Wen, “The Performances Study of EDCF with Block Ack in WLANs,” in Network and Parallel Computing, pp. 328–335, 2012.

402