This regular paper was presented as part of the main technical program at IFIP WMNC'2013

978-1-4673-5616-9/13/$31.00 ©2013 IEEE

Study on Performance-Centric Offload StrategiesFor LTE Networks

Desta Haileselassie Hagos, Rudiger KapitzaTU Braunschweig

Email: {desta, rrkapitz}@ibr.cs.tu-bs.de

Abstract—Currently, cellular networks are overloaded withmobile data traffic due to the rapid growth of mobile broadbandsubscriptions and the increasing popularity of applications forsmartphones. One possible solution to alleviate this problem isthe offloading of mobile data traffic from the primary accesstechnology to the WiFi infrastructure to gain extra capacity andimprove the overall network performance. As the strategy whatand when to offload data is non-trivial, it is of vital importanceto develop novel algorithms to guide this process.

This paper addresses solutions for WiFi offloading in LongTerm Evolution (LTE) cellular networks when performance needsexceed the capability of the LTE access. It then compares theperformance of each access technology using different networkperformance metrics. In detail, an optimized Signal-to-noise ratio(SNR)-threshold based handover solution and extension to the3rd Generation Partnership Project (3GPP) standard for AccessNetwork Discovery and Selection Function (ANDSF) frameworkfor WiFi offloading is proposed. Our simulation results haveshown that ANDSF discovery can be used to control the amountof offloading.

Keywords—LTE, IEEE 802.11, ANDSF, Access Selection, WiFiAccess Points, Offloading, multi-access, Seamless Handover.

I. INTRODUCTION

The adoption of mobile communication technology hasevolved rapidly in recent years, due to the increasing demandsfor higher data rates and higher quality services. The combi-nation of smartphones, tablets, netbooks and cellular mobilenetworks are rapidly growing in very large numbers and as aresult, this has created an exceptional demand for ubiquitousconnectivity and quality of rich digital content and applica-tions. Mobile broadband traffic communicated over cellularnetworks has seen exponential increases, and a recent reportfrom Ericsson which is a representative base for calculatingworld total data traffic in cellular networks predicts mobiledata traffic to grow 10-fold between 2011 and 2016 [16].

The dramatic increase of mobile data traffic is a majorconcern for network operators with limited radio spectrum.Fortunately, there is a real opportunity of offloading the datatraffic to other networks. Here offloading refers to usingalternative network technologies for delivering data originallytargeted for, e.g., a cellular network when it becomes saturated.It helps to ensure optimal usage of available radio resourcesand load-balancing among available radio accesses. To meetthe requirements of future data-rich applications with improvedmultimedia, future wireless networks are expected to combinemultiple access technologies. Recently, some cellular broad-band network operators, including AT&T, T-Mobile, Vodafone,and Orange, are utilizing WiFi networks as an alternative

access network technology worldwide[19]. Projection datareleased in the Cisco VNI (Visual Networking Index) [15]white paper, estimates that over 800 million terabytes of mobiledata traffic can be offloaded which would be 72% of the totaldata traffic in 2015.

At the technical level, Evolved Packet Core (EPC) providesinterworking functionality between 3GPP and non-3GPP (bothtrusted and non-trusted) accesses according to the 3GPP stan-dards [7], [8]. To allow mobile devices to know when, how andwhere to select an access network discovery support functionsand usage scenarios can be found in [8]. The main drivers forthe interworking of the accesses are to reduce the load on thecellular network (i.e. offload LTE network) and supplement 3Gaccess coverage. When several radio accesses are available, theassignment and the handover of users among these accesses,session management with the flows and much more becomethe fundamental problems. While there has been only a limitednumber of studies so far to mitigate this critical problems, theperformance issue has been even more neglected.

In this paper, our main focus is to address how to overcomethe mobile network congestion by offloading a portion of thedata traffic to complementary access networks, i.e. by usingWiFi whenever there is data congestion in the LTE cellularnetworks. Our offloading strategies are compared to steer WiFioffloading to increase the combined network performance ofLTE and WiFi access connected to the core network with atleast the baseline case of having all the traffic in LTE. Inthis paper, we have also motivate the reasons why ANDSFis a 3GPP choice for network selection in heterogeneousnetworks. Accordingly, we devise and implemented threeANDSF offloading algorithms as per the specified standard[11]. In the 3GPP standard, it is stated that the ANDSFcan trigger the radio cell of a User Equipment (UE) forits discovery information using the geographical location oflatitude, longitude, and/or altitude (radius) coordinates.

We performed extensive simulations to evaluate the algo-rithms considered. An optimized SNR-threshold based han-dover solution for WiFi offloading is proposed. The rest of thepaper is structured as follows. Section II presents a target sce-nario on how and when offloading can be applied. Section IIIpresents the IP flow mobility use cases. Section IV presentsthe network access selection methods used. Section V providesan in-depth explanation and summary of the models andassumptions used. Section VI presents the numerical results.Section VII addresses related work. Section VIII proposes howthe 3GPP standard could be improved and finally, we drawconclusions.

II. TARGET SCENARIO

We assume a 3GPP network operator controlled scenarioas it is shown in Figure 1. Our overall assumption is thatwe have a network operator in charge of both the LTE andWiFi networks. This helps the network operator to have controlover WiFi traffic and to ensure a better customer experienceacross the available networks. This means that, for example,UE’s IP address allocation, access to general IP services aswell as network features like security, charging, Quality ofService (QoS) and policy control can be made independent ofthe access technology. And therefore migration between LTEand WiFi is easily possible. Data session managements areseamlessly handed off by the Packet Data Network Gateway(PDN GW) 1.

Fig. 1: Target Scenario of the work

Figure 1 shows the baseline scenario when a user movesfrom LTE network to a WiFi network coverage. We can thinkof this scenario as an IP flow mobility service: we carry a UEwhich can access among other technologies, LTE and WiFi.We are connected to LTE network and move indoors, intocoffee shops or big malls. There we have a fixed broadbandconnection connected to a IEEE 802.11-capable Access Point(AP). Depending on preferences, the UE in this situation mayswitch access from LTE to WLAN which is the main objectiveof this paper. Whenever WiFi access network is availablenearby, the operator can choose to offload some or all of itstraffic depending on the QoS requirement through a WiFi AP.Selectively offloading traffic to approved WiFi networks givesmobile operators an opportunity to increase their total networkcapacity to meet rising traffic demands and a way to extendnetwork coverage and capacity to WiFi networks.

To enable such an approach ANDSF is a suitable basisas it is a 3GPP approach for controlling handover operationbetween 3GPP and non-3GPP access networks. Release-8 of3GPP [8] has specified the ANDSF framework through whichthe network operator can provide a list of preferred access

1PDN GW is the point of interconnect between the EPC and the externalIP networks. These networks are called Packet Data Network (PDN). ThePDN GW routes packets to and from the PDNs and it also performs variousfunctions such as IP address or policy control and charging.

networks with inter-system mobility policies. As a UE movesacross a heterogeneous network environment, it has to discoverother radio accesses available in its vicinity. For example, a UEusing 3GPP radio access needs to discover when WiFi accessbecomes available and possibly trigger a handover based ona predefined operator policies, or when the radio signal fromits serving 3GPP cell starts to get weak. Hence, the purposeof the ANDSF is to assist user devices to discover accessnetworks in their vicinity and to provide rules so as to prioritizeand manage connections to all networks. Based on operator’spolicies, the IP flows are routed differently when either thecore network is congested or when the current serving cellis overloaded. Different services with different characteristicsin terms of QoS requirements and bandwidth, for example aweb browsing session and non-conversational video streamingsession will be offloaded to the non-3GPP WiFi network. Thishelps to balance the load and relieve the traffic issue of the3G access network usage and to guarantee optimal usage ofthe available radio accesses. It also increases the end-userthroughput for IP flows with high throughput requirements.

III. IP FLOW MOBILITY - USE CASES AND POSSIBLESCENARIOS

In this section, we will describe detailed multi-accessscenarios where the mobile terminal is connected to the EPCvia different access networks simultaneously with service con-tinuity over multiple accesses, sending and receiving differentIP flows when the UE is under the coverage of both 3GPPand non-3GPP access networks, and redistribution of IP flowswhen the non-3GPP access is no longer available. Theseaccesses are interconnected to the EPC via the PDN GW.

We considered a scenario which represents an IP flowmobility services. For example, on the way home from office, auser might only have 3GPP access. Let us assume that the useris simultaneously accessing different services with differentcharacteristics in terms of QoS requirements and bandwidth,for example a web browsing session and a video telephonycall consisting of conversational voice and non-conversationalvideo streaming session. When the user reaches home, his/herdevice selects non-3GPP access (e.g. domestic WiFi hotspot)and based on his/her personal preferences, requirements ofapplications, etc., some of his/her currently running serviceswill be switched over to the non-3GPP interface so as to loadthe balance, to guarantee optimal usage of the available radioaccess and increase the end-user throughput for IP flows withhigh throughput requirement. Some of the flows which theuser is using may be from the same application. Based onoperator’s policies, the user’s preferences, the characteristicsof the application and the accesses, the IP flows are routeddifferently; as an example, the audio media (conversationalvoice which is the hard real-time) of the Video Telephonycall and the video streaming are routed via 3GPP access,while the soft real-time conversational video (live streaming)of the Video Telephony, the P2P download (best effort) andmedia file synchronization are routed through the non-3GPPaccess [2]. In the middle of the IP sessions, the user’s devicemight automatically start a non-real time streaming of FTPfile synchronization with a backup server (best effort) viathe WiFi access system. Due to the huge amount of traffic,WiFi becomes congested and therefore the non-conversationalvideo streaming session doesn’t get the required level of QoS

treatment. This initiates the IP flow to move back to the 3GPPaccess. Later on, when the FTP file synchronization is done,the non-conversational video streaming session will be movedback to the WiFi network.

Based on the IP flow use cases discussed previously, wecan distinguish scenarios where the UE is capable of routingdifferent simultaneously active PDN connections connected tothe Evolved Packet System (EPS) through different accessessystems the UE can stay simultaneously connected with.Possible scenarios could be, adding an IP flow to an existingone via different accesses, removing an IP flow established viadifferent accesses, IP flow mobility between accesses (whenboth interfaces are active simultaneously or when only oneinterface is active) and switchover of all IP flows from oneaccess network to another. The UE selects the access systemwhere to route a specific PDN connection based on the userpreferences and the inter-system mobility policies staticallypre-configured by the network operator on the UE or providedby the ANDSF entity of the EPS [2]. A more thoroughdiscussion of this is available in a technical report [17].

IV. NETWORK ACCESS SELECTIONMETHODS

Next generation wireless heterogeneous networks are char-acterized by the co-existence of multi-access wireless networksutilizing different access networks which complement eachother in terms of offered actual throughput and operationalcosts, e.g., LTE and WiFi. In such networks access selectionis a fundamental problem. Access selection principle does notonly affect the performance of the overall network, but throughits required input parameter set, potentially originating fromdifferent access network technologies, also determines whatnetwork architecture solutions are called for.

Network access selection may be done on different timescales, based on different parameters, and with different ob-jectives in mind, e.g. to maximize capacity or some form ofuser quality measure. But in this paper, the main goal is forWiFi offloading when the LTE network is congested or whenthe current serving cell is overloaded, and access selectionprinciples based on different input parameters are evaluated.The access selection is instantaneous, i.e., the time-scale aspectis neglected and the input parameter set is varied. Generally,an access selection principle can be defined as a function f,which is based on its set of input parameters Pi selects an (orset of) access network(s)/decision variable(s) accessi for eachuser index i:

accessi = f(Pi) (1)

It is of great interest to limit the set of required inputparameters Pi. This is mainly because this set must be madeavailable across different access networks, and potentiallyacross different network operators, which is a challenging taskfrom protocol and an architecture point of view. In this paper,access selection algorithms given in equations 2, 4 and 5are studied for the evaluation of the proposed strategies. Thefirst algorithm considers only the SNR information for accessselection decision and therefore it is traffic load-independent:

accessi =

{WLAN if(SNRWLANi ≥ SNRmin)LTE otherwise. (2)

A mobile device i thus selects WLAN if the SNR fromthe best WLAN AP equals or exceeds the threshold SNRmin.Algorithm 1, the so-called ”WiFi if coverage”, is realized bysetting SNRmin=0dB. A second alternative (Algorithm 2) isto set SNRmin to a higher value so as to balance relativeloads, i.e. so that Un/Cn=Um/Cm, for all n, m, where Un

and Cn are the number of users and capacity of radio accesstechnologies n respectively. A further alternative, is to use aWLAN system load dependent SNRmin. Here, SNRmin is setto 0dB for low WLAN traffic loads, and increased with systemload to maintain a maximum system load just below the APcapacity. These algorithms will be evaluated in Section VI.

V. MODELS AND ASSUMPTIONS

This section discusses a summary of the general modelsand assumptions used in the evaluation, of which a subset ofdefault simulation parameters are listed in Table I.

A. Radio Access Network Simulator

Due to the complexity of the problems and scenariosstudied, a Monte-Carlo static MATLAB-based multi-cell radioaccess network simulator model which separately derivesDownlink (DL) and Uplink (UL) has been used for theevaluations. The simulated system consists of 21 cells and awrap-around technique is used to avoid border effects and toenable faster simulation run times.

B. Traffic Model

A traffic model named Equal Buffer is used in this paper.This traffic model assumes Poisson processes for calculatinguser arrivals that different arrival intensities (probability ofarrival per time unit) may be given for different services. Thismodel is based on these assumptions; the same amount ofinformation has to be transmitted for every active user2. Butdue to different transmission bitrates (channel conditions), andthus users consume different amount of radio resources. Thetransmission time will vary by users and slow users will holdradio resources for a longer time and thus will have a largeimpact on cell performance. For example, cell-edge users3 maythen consume most of the resources.

C. User Behavior Models

We assume a heterogeneous user behavior with spatialtraffic hotspots, covered by WLAN APs. Users are positionedin dense hotspot areas and are modeled by the parameter witha probability (for a UE to be in a hotspot) of Photspot andin the remaining cellular area with probability of 1-Photspot.It is also assumed that all hotspot areas have the same valueof Photspot. A scenario with Photspot=80% and uniform userdistribution where Photspot=0 which puts fewer numbers ofusers to WiFi networks are studied. The hotspots have a radiusof 50m, within which 95% of the hotspot users are locatedand it is also assumed that all hotspot areas have the samevalue of Photspot. The hotspots together cover about 5% of the

2An active user is a terminal that is registered with a cell and is using orseeking to use air link resources to receive and/or transmit data within a shorttime interval (e.g., within 100 ms).

3Cell-edge user : is a user with a bitrate close to a pre-defined percentileof the bitrate distribution

TABLE I: A summary of the default simulation parameters

Traffic ModelsUser Distribution Random and UniformUser position probability (Photspot) 80% (positioned in the hotspot area)

1-Photspot (probability to be in the remaining cellular areas).Terminal Speed 3 km/hData Generation Equal Buffer (EB)Fraction of traffic generated in DL 75%Minimum Data rates [Mbps] in the DL & UL 1, 0.5 in [Mbps]

Radio Network ModelsDistance attenuation L=35.3+37.6*log(d), d = distance in metersShadow fading Log-normal, 8dB standard deviation, 100m correlation distancePath-loss Exponential (rα), α=3.52Propagation model (Path-loss as a funciton) L(d)=β+10*α* log10 (d), where d is distance in metersMultipath fading LTE SCM, Suburban MacroCell layout (System Size) Hexagonal grid, 3-sector sites, 21 sectors in total with wrap aroundISD (Inter Site Distance) 500m or 800mCell radius ∼ 166m or ∼ 266mSubscriber Density/km2 5400Number of hotspots per km2 100Hotspot radius 50

System ModelsSpectrum allocation 10MHz (50 resource blocks) 180KHz (1 resource block)Maximum UE antenna gain 15dBiMax UE output power 250mW into antennaModulation and coding schemes supported 64QAM, 16QAM, QPSK and 3GPP turbo codesNumber of Base station Antennas 2 per cell with 10-wavelength spacingNumber of mobile station antennas for reception 2 per UE with half-wavelength spacingNumber of mobile station for transmission 1Scheduling algorithm Round-robinMiscellaneous Downlink and Uplink

system area, which results in log-scale user density standarddeviations of 0.4 and 0.6 for Photspot=50% and Photspot=80%respectively. We assumed that there are 100 hotspots persquare kilometer (km2), and a subscriber density/km2 of 5400.Hotspot areas, in which WLAN APs are placed half radiusaway from the cellular base station of its cellular cell. It isof course possible to study different sizes and positions ofhotspots. The total number of users in the system can bevaried but as a reference, the value of 1000 active users percellular cell which corresponds to about 110 users per km2

moving at a speed of 3 km/h is assumed. Users are assumedto generate an average traffic of 100kbps each, regardless ofposition4. Furthermore, the output power is set sufficiently highso as to avoid coverage problems, rendering the radio linkquality in the system limited by interference rather than noise.As an underlying assumption for the mobility issues, dual-mode mobile terminals capable of processing LTE networksand WLAN air interface is presumed.

Our scenarios are more appropriate for deployment inplaces in which a network operator is required to supporthighly dense populations in a given location like universi-ties/colleges, malls/shops, airports etc. where a network opera-tor can deploy large-scale WiFi APs as it is depicted in Figure 2so that users can have instant access to the available broadbandaccess, and at the same time a user can be offloaded to WiFias per the predefined operator policies. Figure 2 depicts whenwe have a multi-access network deployed with less number of

4With the system models used it is mainly the aggregate traffic per AP thataffects the performance. Therefore, a hotspot probability of X% with equaltraffic generation per user could also roughly be said to represent a hotspotprobability of Y% but with k times the traffic generation per user in the hotspotfulfilling the relationship kY/(1-Y)=X. Similarly, if there are N hotspots percellular cell rather than 1, this would on average divide the traffic per hotspotwith N, which could be approximated with a correspondingly reduced hotspotprobability.

WiFi APs as compared to hotspots per km2. This simulatedlayout is from the default parameters considering deploymentof less APs by a network operator than hotspots.

Fig. 2: Multi-access network layout with less WiFi AccessPoints per km2

Figure 3 shows when we have a multi-access networkdeployed with almost the same number of WiFi APs ashotspots per cell. In this case, we assumed that there are7 Hotspots and 6 base stations per cell are deployed. Andthe layout shown in Figure 3 reflects the whole evaluationassumption of our paper.

D. Radio Network Models

Radio propagation models are selected to model macroscenarios of an urban environment. The radio network model

Fig. 3: Multi-access network layout with the same number ofWiFi Access Points and Hotspots per km2

and scenarios follows the Heterogeneous Network (HetNet)scenario as described in 3GPP for LTE-Advanced [1]. Simpleradio network models for both cellular networks and WLANsare used to realize the multi-access networks. A regular hexag-onal grid of 7 3GPP urban macro base station sites 5 withthree sectors (cells) per site and carrier bandwidth of 10 MHzwhich consist of 50 resource blocks is considered. Cellularbase stations with Omni directional antennas are deployed atthe center of the cell. For the WLAN APs, it is assumed thatthey are deployed at the center of the hotspots and the defaultpositions of the hotspots are half radius away from the basestation of the host cell.

Cellular propagation is exponentially modeled by a dis-tance dependent path-loss with a constant exponent of 3.52 andlognormal shadow fading with standard deviation of 8dB at acorrelation distance of 100m is assumed. Whereas for WLAN,a ”Keenan-Motley” propagation model assuming line-of-sightup to 60m, followed by a constant attenuation of 0.3dB/m isused. Log-normal shadow fading with a standard deviation of3dB is also assumed for WLAN. Multi-path fading is implicitlymodeled through the utilized link-level performance. The basestation’s antenna gain is 15dBi and the maximum transmitpower from a base station in a macro cell is 46dBm. Inthis evaluation 64-QAM, 16-QAM and QPSK modulation andcoding schemes (MCS) are supported [12].

E. Assumed Scheduling Algorithm

A Round-Robin (RR) scheduling algorithm which takesplace in the eNodeB (LTE base station, i.e where mobiledevices communicate through) is modeled. It assumes andhandles all active users with some priority. With RR schedulingevery user is allocated full bandwidth in the DL and the sched-uler iterates over users. However, in the UL the bandwidthallocation per TTI (Time Transmission Interval) is limited bythe maximum transmit power.

F. LTE and IEEE 802.11a/802.11b

Relatively simple models of LTE and WLAN are used.To ensure reasonable accuracy, benchmarking of the single

5A cell site or simply site is a base station or the geographical location ofa base station which is equipped with transmission and reception equipment.

access results achieved with these models, summarized inFigure 2 and Figure 3 to previously established results havebeen made. LTE detailed simulation uses average Signal toInterference plus Noise Ratio (SINR), where no fast fadingis included and Interference Rejection Combining(IRC) gainis added to average SINR. More about IRC and its algorithmimplementation can be found in [20]. For IEEE 802.11a/11b,one AP with a power of 100mW is deployed at the centerof each hotspot. And therefore, co-channel interference is notincluded. The average packet size is assumed to be 1000bytesin order to model one uplink TCP acknowledgment for everysecond downlink frame of 1500bytes. Each WLAN user isassigned a link rate, which yields the shortest expected time oftransmission, assuming geometrical distribution for the numberof trials. Based on users’ SNR values, corresponding ’radiolink bitrates’ and desired link utilization levels are calculated.Then, based on these and assuming a round-robin scheduler,resulting link utilization levels are derived. The difference withthe LTE model is that active radio link bitrate of each user isscaled by a mapped value of the utilization factor of the APthat the user is connected to. The utilization factor of an APis obtained by summing up the link utilization of the usersconnected to that AP:

ηj =∑APj

ρi (3)

Where ηj is the utilization factor for the users of the AP j.Users are selected in terms of their link quality and the numberof selected users are adjusted to the capacity limit of the AP.The utility factor is done using a table that is based on thesimulation results found in [17].

G. Evaluation Methodology

Simulation methodology has been employed throughout theevaluations, and results for both UL and DL are independentlyderived. To reach satisfactory accuracy and realistic resultsa number of simulations by increasing the number of pointsand different parameter settings (for example LTE Inter SiteDistance (ISD), WiFi Access Points, distance thresholds, load,hotspots, etc) are run for each studied hypothetical scenarios.Users are randomly placed in the network area accordingto the user behavior modeling over the system area as itis described earlier and access selection is done afterwardswith the identified access selection methods. The simulationmethodology is based on the following iterative simulationloop.

• Generate a network. It generated a regular hexagonallayout as it is previously explained and it optionallymitigates the network border effects by wrap-around(default).

• Distribute UEs randomly with a uniform distribution.• Schedule users randomly. For a given load, generate

interferers for each user.• Calculate Signal to Interference plus Noise Ratio

(SINR) for each user (SINR per antenna and perstream after combining).

• Calculate the bitrate for each user and apply SINR-to-bitrate mapping using the mutual information modelof [14]. This gets the combined bitrate per user whenthe user is scheduled.

• Calculate the achievable user throughput, cellthroughput and other performance metrics.

Terminology. The terms definition used in the next section ofthe paper are given here. In these plots we use terminologies11a to denote for IEEE 802.11a, LPN to denote Low PowerNode (WiFi), Macro to denote for LTE and Combined denotesLTE and WiFi.

VI. NUMERICAL RESULTS

This section presents the detailed simulated numerical per-formance results based on the offloading algorithms consideredfor evaluations and hypothetical scenarios analyzed in thispaper. Simulation graphs for scenarios with combined LTEand IEEE 802.11 systems are included .

A. Algorithm 1: WiFi if Coverage

In this algorithm, the user connects to the best WiFi AP ifhis/her SNR is greater than SNRmin, i.e. select WiFi wheneverthere is WLAN coverage. As a result, in the simulation thisalgorithm is realized by setting the SNRmin = 0dB and settingthe SNR threshold to 0dB could be interpreted as ”WiFi ifCoverage” principle.

Fig. 4: Average Number of Users connected to LTE and WiFiusing Algorithm 1

Using this algorithm, an average of 31.5% of users canbe offloaded to the available WiFi access network while theremaining 68.5% average numbers of users can use the LTEnetwork on the available macro cell ID based on the topologydepicted in Figure 3.

B. Algorithm 2: Fixed SNR Threshold

With a fixed SNR threshold from WLAN, a user selectsthe best WLAN if the SNR from the best WLAN AP equalsor exceeds the threshold SNRmin; otherwise the user will bedirected to the cellular network. This algorithm is the sameas the ’WiFi if coverage’ algorithm at the low traffic loadsand hence the SNRmin is a function of the WLAN load.Since only one user is served at a time in WLAN technology,there is no interference between users. And therefore, SNR

instead of SINR is used in our evaluation for a better signalquality between base station and mobile terminal. The onlyparameter needed from this algorithm is the SNR informationfrom WLAN and then the network access selection decisionwill be made according to the following expression:

accessi = f({SNR1i , load

1i }) (4)

and then this could again be expressed as in the followingwhere the SNRmin is a function of the WLAN load.

accessi ={

WLAN SNR1i ≥ SNR1

min(load1i )

LTE otherwise (5)

Where SNR1i is the SNR of user i from WLAN and it is

a function of the WLAN load. The load and the SNRmin aredirectly proportional at each WLAN access point. When theload increases the SNRmin is also increases at each individualWLAN access points so as not to have overloaded WLANs.When load is high, the algorithm sorts the users in terms oftheir link quality and selects the best ones. At the same timethe number of selected users is adjusted to the capacity limit ofWLANs. In real time, users’ access network selection can beinitiated by changing the load or the SINR. Another approachdeployed in this work due to the use of a static simulator isaccess network reselection after fixed time intervals.

C. Algorithm 3: ANDSF Models

ANDSF is a new EPC entity defined in the 3GPP standard[8] that contains data management and control functionalityto provide necessary access network discovery and selectionassistance to the UE as per the operator’s policy. It allowsmobile device users to discover and select from a list ofnearby access networks, determine AP traffic load and theircapabilities, and to connect to the available networks based onthe user preferences and predefined network operator policies.

a) ANDSF Model based on Cell-ID: I)Discovery-

WiFi APs are discovered if they are located in the geo-graphical area corresponding to the current macro cell ID. Thegeographical area of a macro cell is a circle circumscribing thehexagon in the regular macro cell plan where the radius, ( r =ISD3 ) as it is shown in Figure 5.

Fig. 5: Macro-cell with Tri-sector Antenna

II)Connection/Offloading/Handover: In this case of theANDSF model, the mobile connects to the strongest coverageof the discovered WiFi APs provided it has SNR value ofgreater than zero (i.e. SNR > 0).

Fig. 6: Average Number of LTE and WiFi Users, using theANDSF rule-based on Cell ID

Using this offloading algorithm, the simulation result as itis depicted in Figure 6 shows that an average of 75.7% of userscan use the LTE network on the available macro cell ID whilethe remaining 24.3% average of users can be offloaded to theavailable WiFi access network based on the network operator’spredefined policies.

b) ANDSF Model based on Position: I)Discovery- Inthis case of the ANDSF model, WiFi Access Points arediscovered if they are physically located close enough to themobile user, which is the UE.

II)Connection/Offloading/Handover: The user connects tothe strongest coverage of the discovered WiFi APs nearby pro-vided that the mobile is located within the distance thresholdpredefined by the network operator (i.e. if it is in the rangeof the radius specified which in this case the radius is for e.g.200m.

Fig. 7: Average Number of LTE and WiFi Users when theDthr for the discovery=200, SNR=0.

As it is depicted in Figure 7, this offloading algorithmallows more users to be connected to LTE than being offloadedto WiFi when we set the distance threshold for the network

discovery very much less than the radius of the hotspot. Thisalgorithm of course varies on the value of the network discov-ery distance threshold, i.e. the larger the network discoverydistance threshold the more users will be offloaded to WiFi.In this case a load (monthly data volume per user, i.e. vectorsallowed) of 3GB/Month. If the network operator wants to blockall users from being offloaded to the available WiFi accessnetworks and to stay connected to LTE, we only need thedistance threshold for the discovery to set it to 0 (zero).

Fig. 8: Random User Distribution with smaller ISD and thesame number of Hotspots and WiFi Acess Points per cell

Figure 8 shows the 10th percentile bitrate for the number ofusers connected to WiFi versus User Throughput [Mbps] in arandom user distribution (i.e. Photspot=80%). In this scenario,an LTE with smaller ISD of 500m is assumed. Keeping theDthr constant (200m) and varying the SNR value (e.g. 30dB,15dB, 10dB, 6dB, 4dB and 2dB) gives better user throughputthan varying the network discovery distance threshold (Dthr)while setting the SNR value to a threshold of 0dB. This holdstrue where users are randomly placed in a network with thesame number of hotspots and WiFi access points per cell. AndFigure 9 depicts the 10th percentile bitrate when we deploymore base stations/km2 than WiFi APs/km2 in a random userdistribution.

Figure 10 depicts when we have a scenario with larger LTEISD of 800m), for random distribution of users, varying theDthr (0m, 10m, 25m, 50m, 100m and 200m) while keeping theSNR to a fixed value of 0dB which gives better user throughputand it is valid for the majority of our simulation results. It isquite obvious that the larger the LTE ISD, the lesser the userbitrate will be.

c) ANDSF Model based on Cell-ID &Position: ThisANDSF model combines both the ANDSF models describedabove. The following figure shows the average number ofusers connected to LTE and offloaded to WiFi by applyingthis offloading algorithm when the distance threshold for thenetwork discovery is set to 200m while the SNR is set to athreshold value of 0.

Fig. 9: Random User Distribution with smaller ISD and morenumber of Hotspots per cell

Fig. 10: Random User Distribution with larger ISD and thesame number of Hotspots and WiFi Access Points per cell

Fig. 11: Average Number of Users Connected to LTE and WiFiwhen the Dthr=200, SNR=0

D. Average Cell Discovery CostThis algorithm measures the average number of APs dis-

covered when the UE scans the available WiFi access networksas it is shown in Table II. Access network discovery receivedby the ANDSF is used in order to scan for certain accesses onlyin specific area; in this way the energy consumption of the UEcan be reduced and lead to a long battery lifetime. But this isonly part of the problem, because to blindly scan for APs takesa lot of time and as result of this the moible device may notdiscover it in time. This means, handover would be beneficialuntil the blind scanning discovers the available access.

TABLE II: Average Cell Discovery cost

Algorithms Average Cell Discovery CostISD=500, Dthr=200 ISD=800, Dthr=200

Algorithm 1 140 APs 140 APsAlgorithm 2 140 APs 140 APsAlgorithm 3 (ANDSF Rules)Model-1 7.464 APs 7.775 APsModel-2 10.486 APs 14.645 APsModel-3 5.203 APs 2.936 APs

As an extreme scenario, when we set the Dthr to a value of0 or higher SNR (e.g. 30dB) for the ANDSF algorithms, thecell discovery cost will be 0. This means no user is allowed tobe offloaded to the available WiFi networks. In this scenariothe network operator can block users from being offloaded toand stay connected to LTE for some reason.

E. Summary of Simulation results

In this study the networks are assumed to support data-like services. Other types of services, such as voice and alsoa combination of different services could also be studied. Theperformance of the network depends on the value of LTE ISDand it is assumed that the fraction of traffic generated in the DLis 75% and the remaining 25% traffic is generated in the ULwith a monthly data volume load per user of 3GB/Month. Ourdetailed simulation results of LTE and IEEE 802.11a/11b haveshown that, when we have a smaller LTE ISD it is always bestto stay connected to LTE cellular network. Whereas when weset the ISD to a larger value, it is seen that the bitrate offeredby WiFi is typically higher than that of LTE and hence, it isalways best to connect to WiFi. In this case users can benefitfrom being offloaded to the available WiFi access networks.A simple ’WiFi if coverage’ access selection principle maythus be expected to perform reasonably well. With smallerISD, it is also seen that variable SNR-based access networkselection yields better user throughput than keeping the Net-work Discovery distance threshold fixed (Dthr = 200m) in bothuniform and random user distributions. This also applies forscenarios where we have either the same number of WiFi APsas base stations deployed in a network or more number of basestations than APs. As a reference the single-access capacitiesare also shown. Clearly these depend on both system loadand signal quality. And therefore, in order to select the accesstechnology that maximizes user bitrate, both individual signalquality and system load need to be known and we should alsonote that the spectrum allocation differs between the accesstechnologies.More simulation results in terms of mean, 10thand 90th percentiles for user throughput versus traffic loadcharacteristics per user in the UL and DL for LTE cellularnetwork and WLAN subsystems are briefly discussed in [17].

With the assumed hotspot probability of 80% , it is seenthat all algorithms result in the same traffic distribution. Thisis because when Photspot <CWLAN /(CWLAN + CCellular),the cellular network reaches its capacity limit before WLANeven if all hotspot users are allocated to WLAN. In thiscase no bitrate gain is achieved, and therefore as with thelow WLAN load, WLAN is typically the best alternative.For high loads, better mean values are however achievedwith ’WiFi if Coverage’. It is also worth noting that withalgorithm 3(ANDSF model) acceptable bitrates are achievedfor subsystem traffic loads beyond the single-access capacitylimits. This multiaccess related capacity gain is due to thefact that the access selection principle allocates users to thesubsystem in which they consume the least radio resources.It is also seen that ANDSF model based on Cell ID andposition dramatically overloads WLAN, already for relativelylow traffic loads which results in poor bitrates. DL and ULutilization of LTE/WiFi with monthly data volume per user of3GB/Month which can be represented as scalar or vector forcapacity evaluation of 1, 2 is assumed. The higher the load thehigher is the resource utilization and it is directly proportionalwith the given value of LTE ISD. We have only one share ofthe access networks for the given loads both in the DL and ULdue to the fact that the same number of users is assumed inour simulator for both utilizations. In addition to the simulationresults discussed above, a more thorough discussion of severaladditional combined scenarios have been evaluated and arebriefly presented in [17].

Concerning access combinations, scenarios including IEEE802.11a/b as well as LTE indicate that it is the relationshipbetween the cellular and WLAN capabilities that matters. The’better’ the WLAN system the better the ’WiFi if coverage’algorithm. Similarly, the better the cellular system, the higherthe user throughput and the higher the SNR value the lesserthe WiFi share will be. Higher SNR value directs users to thecellular network and makes the WiFi network less loaded. Asa result, fewer users connect to WiFi where they experiencehigh bitrates. Varying hotspot probability and propagationconditions also affects the SNRmin that yields balanced trafficloads. In summary, signal strength can be considered as abasis access selection input, and may also be sufficient inmany scenarios. System load is motivated as an additionalinput in scenarios with high bitrates offered by the LTEnetwork and strongly heterogeneous traffic load, and so thatthe traffic load per WLAN AP is similar to that of the cellularaccess points. Finally, favorable mean bitrate results have beenderived through an extensive simulation.

VII. RELATED WORK

We have also studied the existing cellular traffic offloadsolutions related to efforts in the current 3GPP standardsand mobile industry initiatives. Among the existing solutionsare traffic offloading via deployment of Femtocels for indoorenvironments whereby facing the technical challenges foundin [22] and via WiFi APs. There are also delay-tolerantapproaches to traffic offloading where the delay is usuallycaused by intermittent connectivity [21] and delay the deliveryof information over cellular networks and exploit opportunisticcommunications to offload mobile data traffic [18]. There isalso another approach which has proposed a scheme calledWiffler [13] to augment mobile 3G networks using WiFi for

delay-tolerant applications. The 3GPP standard [5] provides anarchitecture which allows easy interworking between cellularsystems and WiFi hostspots. An enhanced offload solutionis also introduced in 3GPP work item in [6], which makesuse of a client-server based protocol Dual Stack Mobile IP(DSMIP) [3] to achieve seamless handover solution betweencellular networks and WiFi has recently been standardized. Itsupports IP session continuity for IPv4 and IPv6 sessions andthe signaling needed to carry routing filters when the UE isconnected to multiple accesses simultaneously are specified in[4]. As a best offload solution, the 3GPP [9] has produced awork item function called SIPTO (Selected IP Traffic Offload)which enables IP Flow Mobility for an operator to seamlesslyoffload selective IP traffic while supporting simultaneous cel-lular and WiFi access networks.

In all these studies, the performance aspect is neglected;and the strategies of the big vendors that do offloading arenot publicly available. We have considered the performanceas a key issue in mind and to make our offloading algorithmspublicly available. Our intent in this paper is to propose a novelhandover solution as an extension to the 3GPP standard.

VIII. PROPOSAL

The challenge with ANDSF access selection in the cur-rent 3GPP standard is that it only allows prioritized accessselection, that is an access network is only selected if noother access network having a higher priority is available. TheANDSF selection rules have a number of conditions where oneor more access networks to be used whenever a selection ruleis active. When the ANDSF policy selection rules identify anavailable network, the highest priority rule becomes an activeselection rule and network reselection is performed. But thisdoesn’t take network performance into account and it doesn’tseem to be applicable in some situations say for examplein trees/forest environments where radio signal attenuationeasily occurs or when there is damage on buildings andother obstructions which result in loss of radio signal duringmobility. And our system evaluation has demonstrated that thisis non-optimal for heterogeneous wireless environments.

One possible solution to the prioritized access networkselection problem is, if we have a conforming UE we maybe able to tell it ”Connect to this AP only if the signalstrength is greater than SNR-threshold”. We could say, chooseany AP with the strongest SNR which is not possible in thecurrent 3GPP standard. When we have WiFi APs deployed ina distributed wireless environment, for some reason the userthroughput could be good for some of the users and it couldalso be terribly bad for other users and this problem can beeffectively solved by adding SNR.

Based on our evaluation results we propose that the SNR-threshold should be included in the current ANDSF databaseorganization which the ANDSF features for storing both theAccess Network Discovery information and the Inter-systemclient mobility policy parameters [10]. Accordingly, the SNR-threshold could be set with the coverage mapping informationfor the available non-3GPP access networks based on the3GPP cell Identifiers after the ANDSF network discovery isperformed as it is shown in Table III. Because we believethat for a network access selection to perform well, it requires

both signal quality and of course system load information asan input parameters and this in return helps for an efficientoffloading decision.

TABLE III: Proposed ANDSF database organization

UE Location AccessType=WiMAX AccessType=WiFi-3GPP(cell ID)- Other6

Locn 1 NSP-ID=NSP 1 SSID=WiFi1, BSSID=BS1, SNR1Cell Id = Cell 1 -NAP ID=NAP 1 SSID=WiFi2, BSSID=BS2, SNR2

-NAP ID=NAP 2NSP-ID=NSP 2

-NAP ID=NAP 2-NAP ID=NAP 3

Locn 2 NSP-ID=NSP 2 Not availableCell Id = Cell 2 -NAP ID=NSP 3

Locn 3 Not available SSID=WiFi1, BSSID=BS3, SNR3Cell Id = Cell 3 SSID=WiFi4, BSSID=BS4, SNR4

.................... ....... ............Locn n NSP ID=NAP n SSID=WiFi6, BSSID=BS5, SNR5

Cell Id = Cell n NAP ID=NAP 2

IX. CONCLUSION

With the proposed models and assumptions, our systemevaluations have demonstrated that offloading users from LTEto WiFi networks reduces demand on the LTE network with-out affecting user performance. When the hotspot probabilityincreases beyond CWLAN /(CWLAN+CCellular)7, the ’WiFi ifcoverage’ algorithm may yield overloaded WLAN and pooruser bitrates at high traffic loads. This can be avoided byusing a load-based SNR-threshold gradually balancing theload between the access networks. Our simulation results haveshown that it is possible to control the network by changingthe network discovery distance threshold to control the amountof WiFi offloading. Since WiFi is greatly available throughvarious hotspots and at home, and is also in a number ofLTE/3G devices, it offers the capacity to become a seamlessoffloading solution of LTE.

In scenarios with moderate hotspot probability, yielding ahotspot load within the WLAN capacity, the ’WiFi if coverage’algorithm is sufficient. But, in some scenarios with high trafficload concentrated to the LTE networks in the same order asWLANs or hotspots, different conclusions may however bereached with more complex principles taking traffic load intoaccount. In network scenarios with very high hotspot trafficload, the LTE network may offer higher user throughputs.This can be taken into account by introducing some form ofload analysis and admission control for the WLAN networkto prevent it from being overloaded. For a combined LTE andIEEE 802.11, higher user throughputs are offered in the LTEnetwork for more combinations of traffic load. It is thus notobvious that WLAN is always the best alternative. It is alsoseen that signal strength-based access selection yields quitesatisfactory results in many scenarios.

ACKNOWLEDGMENT

We would like to thank Tobias Poegel for his valuablefeedback and proof reading of our paper.

6Geospatial co-ordinates (latitude, longitude)7Where Cn is the capacity of the nework n

REFERENCES

[1] 3GPP TR 36.814 Evolved Universal Terrestrial Radio Access (E-UTRA); Further advancements for E-UTRA Physical Layer Aspects,Rel.9, February 2009.

[2] 3GPP TS 23.861 V1.3.0: Multiple access PDN connectivity and IP flowmobility (Rel. 9), September 2009.

[3] IETF RFC 5555: Network Working Group, Mobile ipv6 support forDual Stack Hosts and Routers, June 2009.

[4] IETF RFC 5648: Multiple Care-of Addresses Registration, October2009.

[5] 3GPP TS 23.234: System to Wireless Local Area Network (WLAN)Interworking; System Description (Rel. 10), March 2011.

[6] 3GPP TS 23.327: Mobility between 3GPP-Wireless Local Area Net-work (WLAN) Interworking and 3GPP Systems (Rel. 10), March 2011.

[7] 3GPP TS 23.401: GPRS Enhancements for E-UTRAN Access (Rel.10), June 2011.

[8] 3GPP TS 23.402 V10.3.0, Architecture Enhancements for Non-3GPPAccesses (Rel. 10), March 2011.

[9] 3GPP TS 23.829: Local IP Access and Selected IP Traffic Offload(LIPA-SIPTO (Rel. 10), March 2011.

[10] 3GPP TS 24.302: Access to the 3gpp Evolved Packet Core (EPC) vianon-3GPP access networks (Rel. 10), April 2011.

[11] 3GPP TS 24.312: Access Network Discovery and Selection Function(ANDSF) Management Object (MO) (Rel. 10), June 2011.

[12] 3GPP TS 36.213 Evolved Universal Terrestrial Radio Access (E-UTRA), (Rel. 10), June 2012.

[13] R. M. A. Balasubramanian and A. Venkataramani. Augmenting Mobile3G Using WiFi. In Proceedings of MobiSys 2010, pages 209–222, June2010.

[14] K. Brueninghaus, D. Astely, T. Salzer, S. Visuri, A. Alexiou, S. Karger,and G.-A. Seraji. Link performance models for system level simulationsof broadband radio access systems. 4:2306 –2311, September 2005.

[15] Cisco. Cisco Visual Networking Index: Global Mobile Data TrafficForecast Update, 2010-2015. White paper, Cisco, February 2011.

[16] Ericsson. Ericsson Research, Annual Report 2011, Traffic and MarketData Report on the Pulse of the Networked Society. Technical report,Ericsson, November 2011.

[17] D. H. Hagos. The Performance of WiFi Offload in LTE Networks.Master’s thesis, Lulea University of Technology, Sweden, June 2012.

[18] B. Han, P. Hui, V. Kumar, M. Marathe, J. Shao, and A. Srinivasan. Mo-bile Data Offloading through Opportunistic Communications and SocialParticipation. Mobile Computing on IEEE Transactions, 11(no.5,,):821–834, May 2012.

[19] A. Handa. Mobile Data Offload for 3G Networks. white paper.Technical report, IntelliNet Technologies, Inc., October 2009.

[20] J. Karlsson, J.; Heinegard. Interference Rejection Combining for GSM,Universal Personal Communications. volume vol.1,, pages pp.433–437.IEEE International Conference, October 1996.

[21] K. roland Fall. A Delay-Tolerant Network Architecture for ChallengedInternets. pages 27–34, August 2003.

[22] J. G. A. V. Chandrasekhar and A. Gatherer. Femtocell Networks:A Survey. IEEE Communications Magazine. IEEE, 46(9):59–67,September 2008.