HAL Id: hal-01759108 https://hal.archives-ouvertes.fr/hal-01759108 Submitted on 5 Jun 2018 HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. Handover Triggering in IEEE 802.11 Networks Nicolas Montavont, Alberto Blanc, Renzo Efrain Navas, Tanguy Kerdoncuff, German Castignani To cite this version: Nicolas Montavont, Alberto Blanc, Renzo Efrain Navas, Tanguy Kerdoncuff, German Castignani. Handover Triggering in IEEE 802.11 Networks. IEEE 16th International Symposium on A World of Wireless, Mobile and Multimedia Networks (WoWMoM), Jun 2015, Boston, United States. 10.1109/WoWMoM.2015.7158126. hal-01759108
Transcript
HAL Id: hal-01759108https://hal.archives-ouvertes.fr/hal-01759108
Submitted on 5 Jun 2018
HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.
Handover Triggering in IEEE 802.11 NetworksNicolas Montavont, Alberto Blanc, Renzo Efrain Navas, Tanguy Kerdoncuff,
German Castignani
To cite this version:Nicolas Montavont, Alberto Blanc, Renzo Efrain Navas, Tanguy Kerdoncuff, German Castignani.Handover Triggering in IEEE 802.11 Networks. IEEE 16th International Symposium on A Worldof Wireless, Mobile and Multimedia Networks (WoWMoM), Jun 2015, Boston, United States.�10.1109/WoWMoM.2015.7158126�. �hal-01759108�
Abstract—The current and future IEEE 802.11 de-ployment could potentially offer wireless Internet con-nectivity to mobile users. The limited AP radio coverageforces mobile devices to perform frequent handovers whilecurrent operating systems lack efficient mechanisms tomanage AP transition. Thus we propose an anticipation-based handover solution that uses a Kalman filter topredict the short term evolution of the received power.This mechanism allows a mobile device to proactively startscanning and executing a handover as soon as better APsare available. We implement our mechanism in Androidand we show that our solution provides a better wirelessconnection.
I. INTRODUCTION
Due to the proliferation of Wifi hot-spots and com-munity networks, we have recently observed a greatevolution of IEEE 802.11 networks especially in urbanscenarios. These 802.11-based networks allow mobileusers to get connected to the Internet, providing a highthroughput but a limited mobility due to the short cover-age area of access points (APs). In our previous work [1]we have shown that community networks appear to behighly dense in urban areas, generally providing severalAPs (15 in median) per scanning spot. Under this condi-tion, a mobile user may be able to connect to communitynetworks and compensate the low AP coverage areaby transiting between APs. We call such AP transitiona handover. However, two main issues currently limitmobile users from using community networks in sucha mobility-aware scenario. First, operators have notdeployed the necessary infrastructure to allow mobileusers to perform handovers without being disconnectedat the application layer, i.e., after a handover on-goingapplication flows are interrupted. This limitation may beaddressed by deploying a Mobile IP [2] infrastructure, inwhich the application flows may be tunnelled througha Home Agent that belongs to the operator. Second,independently from the first issue, there is still a lack ofmechanism to intelligently manage a layer 2 handoverbetween two APs. In current mobile devices, when ahandover occurs, we observe a degradation of on-going
flows corresponding to a dramatic reduction of the TCPcongestion window (CWND) and of the throughput. Inthis paper, we focus on this latter issue by analyzingthe impact of layer 2 handovers on mobile users. Wepropose Kalman-filter-based HAndover Trigger algo-rithm (KHAT) that succeeds in intelligently triggeringhandovers and reducing the scanning impact on themobile device. We propose a complete implementationof our handover mechanism in Android ICS (4.0) andshow a comparative study to show that our approachoutperforms the handover mechanism that is currentlyimplemented on these devices.
The paper is organized as follows. Section IIpresents the litterature on handover optimization andSection III analyzes the handover impact on on-goingcommunications. Section IV introduces KHAT which isevaluated indoor and outdoor in Section V. Section VIconcludes the paper.
II. HANDOVER PROCESS AND RELATED WORK
The IEEE 802.11 standard defines a handover as athree steps process: scanning, authentication and asso-ciation. The standard proposes two different scanningalgorithms namely passive and active scanning. In pas-sive scanning, the mobile station (MS) simply tunes itsradio on each channel and listens for periodic beaconssent by the APs. In active scanning, the MS proactivelysends requests in each channel and waits for responsesduring a pre-defined timer.
In 2012, the IEEE has published new amendmentsfor IEEE 802.11 handover optimization, aimed at re-ducing its duration and its impact on higher layers. TheIEEE 802.11k amendment proposes mechanisms forradio resource measurement for seamless AP transition,including measurement reports of signal strength (RSSI)and load of nearby APs. Additionally, the IEEE 802.11ramendment contains a Fast Basic Service Set Transition(FT), which avoids exchanging 802.1X authenticationsignaling under special conditions by caching authenti-cation data.
While these features may enhance the handover per-formance, they heavily rely on a cooperation betweenAPs, which might not always be a viable solution. Inaddition, users may access various networks operatedby different providers. In that case, operators shouldshare network information and performance amongthem, which is quite an unlikely scenario. In this paper,we focus on MS-based solutions, where the MS itselfhandles the handover without the help from the network.Several works have been proposed in the literature sofar. In general, those studies cover different aspects ofthe handover mechanism. We may group them into threemain categories:
• Handover triggering: when to decide that adisconnection with the current AP will occur.
• AP discovery: how to search for APs on differ-ent channels by minimizing the impact on thehigher layers.
• Best AP selection: with which AP to associate,among the discovered ones.
The simplest mechanism to trigger a handover is tomonitor the RSSI as an estimation of the link qualityand start the handover process if the current RSSI islower than a pre-established threshold (commonly setat −80 dBm). Fig. 1a shows the relationship betweenthe RSSI measured on an MS and the TCP throughputthat we have gathered during more than 600 connectionsto community networks in a urban area in Rennes,France [1]. We observe that the TCP throughput is ex-tremely variable for high RSSI, but starts degrading forRSSI lower than −70 dBm, and it becomes significantlylow around −80, dBm.
Some works focus on the anticipation of the han-dover triggering in order to minimize the impact onongoing communications. Mhatre et al. [3] proposea set of handover algorithms based on continuouslymonitoring the wireless link, i.e., listening to beaconsfrom the current and neighboring channels. These ap-proaches give handover latencies varying between 150and 800 ms. However, since these approaches need tolisten to beacons from neighboring channels, it is neces-sary to modify the firmware of the wireless card, which
may not always be possible. Yoo et al. [4] propose anumber of handover triggering mechanisms based onpredicting RSSI samples at a given future time usingLeast Mean Square (LMS) linear estimation. In thisalgorithm, the device continuously monitors the RSSIand computes the LMS prediction if the RSSI is belowa certain threshold (PPred). Then, if the predicted RSSIvalue is lower than a second threshold, PMin, the MSstarts a handover.
Wu et al. [5] propose a handover mechanism aimingat decoupling the AP discovery phase from the APselection and reconnection phase. The MS alternatesbetween scanning phases and a (normal) data modewhere the MS is connected to its current AP. Thetime interval between two scanning phases is adapteddepending on the current signal level and varies between100 and 300 ms. In each scanning phase, the sequence ofchannels to scan is selected based on a priority list thatis built based on the results of a periodic full scanning(i.e., here all channels are scanned).
As far as Android devices are concerned, Silva etal. [6] present a mobility management solution based onIEEE 802.21. They propose a mapping of IEEE 802.21primitives for handover initiation, preparation, executionand completion to existent Android OS methods andfunctions.
III. HANDOVER IMPACT
During an L2 handover, the MS is not able to sendor receive application flows. This is because, usually,when a MS triggers a handover, the link quality doesnot allow exchanging frames anymore, and becausethe MS is often switching operating channel. In thissection we evaluate the handover and scanning impacton application flows, and determine which parametersinfluence the scanning latency and success rate.
This testbed consists of nine Cisco Aironet 1040APs installed in the roof of our building at the locationsgiven in Fig. 2. All APs are connected to a dedicatedwired LAN. APs broadcast a single SSID, correspond-ing to an open-system authentication network belongingto a single IP subnet. We also use a dedicated (fixed)server for traffic generation and tracing. iPerf is usedto generate TCP downlink traffic to the MS. For eachexperiment, we walk from AP1 to AP6 and then backagain to AP1.
A. Operating Systems Benchmark
To illustrate how the handover is currently impactingdata flows, we have performed a set of experimentsto evaluate the degradation of TCP performance fordifferent devices and Operating Systems (OS). Table Ishows the number of handovers and the average TCP
2
-20
0
20
40
60
80
100
120
-90 -80 -70 -60 -50 -40 -30
Thro
ughput / K
B/s
Signal Strength / dBm
Throughput Std. Dev. Throughput Average
(a) RSSI and TCP Throughput Relation (b) Downloaded data for different OS
0 1 2 3 4
0
2
4
6
Time /s
Dow
nloa
ded
Dat
a /M
Byt
es
(c) Scanning impact on TCP download
Fig. 1: Various TCP performance
Fig. 2: Campus AP Deployment
throughput we have observed for the same path andsame MS speed using different devices and operatingsystems. As a baseline, we also show the maximumachieved throughput for each device remaining staticand connected to a single AP. Using Windows, weobserve the best result, since the MS performs upto four handovers, reaching an average throughput of0.875MB/s. Additionally we observe that for Win-dows, the time in which no data is downloaded (i.e., thedisconnected time) is relatively short compared to theother OSs. The netbook running Ubuntu reacts slowlyto changing channel conditions: in this case the MS isdisconnected for more than 20 s and executes only twohandovers, indicating that the MS waits until the qualityof the radio link is significantly degraded. Fig.1b showsthe evolution of the downloaded data for each case.Additionally, we have observed that for the Windowsdevice, the average round-trip time (RTT) is the lowestone (103 ms) having also a low standard deviation. Thisdiffers from the other devices that reach larger RTTvalues.
B. Scanning Interactions with Data Traffic
We focus on active scanning where an MS sendsProbe Requests on each channel to discover potentialAPs, instead of just waiting for periodic beacons (pas-sive scanning). We chose active scanning because it
allows spending less time in each channel to determinethe AP availability. If the handover phases are doneone after the other, all packets that arrive during thehandover process will be lost. In order to reduce theimpact of handovers on applications flows, it is possibleto introduce a gap between the scanning phase andthe other handover steps, i.e., the decision to handover,the authentication and the association, as presentedin [5]. An MS may use the power saving mode definedin IEEE 802.11 to request its current AP to bufferincoming packets during the time the MS scans otherchannels. This way, instead of loosing packets duringthe scanning phase, an MS can receive the packets afterthe scanning phase, albeit with an extra delay. Thisbehavior is illustrated in Fig. 1c, where we plot thesequence number of the received packets of a TCP flowwhen an MS is performing one scan of the 13 channelswith an active timer set at 50ms. We can see that thescan is starting just before the time 1 s, at which no moredata packets are received from the server. Once the scanis finished, around 850ms after, the MS comes back toits current AP, and starts receiving TCP packets again.
This technique can also be used to split a scanningphase into several sub-phases where only a subset ofchannels are scanned. For example, to scan the 13channels, an MS could sequentially scan three timesa subset of 4 (or 5) channels each time, interleavingthese sub-phases with the data mode with the current APto retrieve data packets. The impact of the number ofscanned channels, and the timers used in each channelis given in the next subsection.
C. Scanning Parameters
We analyze the scanning performance under differ-ent values of timers used to wait for Probe Responses(from 5 ms to 100 ms) and different number of scannedchannels during a sub-phase (between 1 and 13). Inthe standard IEEE 802.11 scanning algorithm, the MSis supposed to scan each channel using two timers
3
Device OS Version Chipset Static Avg. Mob. Avg. Number of Mean RTT σRTTThr. (MB/s) Thr. (MB/s) handovers (ms) (ms)
Asus N10J Win XP SP2 AR5006 5.19 0.878 4 103 43Asus N10J Ubuntu 10.04 AR5006 4.88 0.601 2 161 360Nexus S Android 4.0.3 BCM4329 3.80 0.568 5 129 114MacBook MAC OS 10.7.4 BCM4322 8.44 0.613 3 167 276
TABLE II: Percentage of discovered APs for differentvalues of AT and number of scanned channels
namely MinCT and MaxCT (see section II). However,the IEEE 802.11 Android driver uses a single timer,namely Active Timer (AT) for scanning. AT is defined asthe time an MS waits for Probe Responses on a channel.
We ran 60 scanning sub-phases for each AT andsubset of scanned channels and measured the averagenumber of discovered APs, the RSSI distribution of thediscovered APs and the average duration of the scanning(i.e., the scanning latency). Results are presented inTable II. As a baseline, we consider that all the availableAPs are discovered when scanning the full channelsequence (i.e., 13 channels) using AT=100 ms. In theother cases, the MS discovers only a fraction of theAPs, since it either does not wait long enough to receiveall AP Probe Responses, or because only a subset ofchannels are scanned.
We have also observed that when using a shortAT, even if the MS discovers a low number of APs,those APs have a high RSSI. On the other hand, whenusing higher AT values, the MS discovers more APsbut a large part of them have a low RSSI. This can beobserved in Fig. 3a, where we see that for AT=5 ms theaverage RSSI of candidate APs is −67 dBm, while forAT=20 ms, this decreases up to −76 dBm.
IV. KHAT: PROACTIVE HANDOVER ALGORITHM
We propose a handover algorithm called KalmanFilter-based HAndover Triggering (KHAT for short) thatprovides link going down detection, optmized scanningstrategy, and new AP selection. An MS monitors itslink quality with its current AP, and when the signalstrength is degrading, it starts alternating between scanperiods and data communication with the current AP.The scan periodicity and the timer values are determinedaccording to the current link quality and whether a can-didate AP has already been found. Once the candidate
AP becomes better than the current AP, the handover istriggered.
A. RSSI modelling
One way of keeping track of the changing radiocondition is to track the RSSI on the MS. While farfrom being perfect, the RSSI has the advantage of beingalways available, whether the MS is exchanging data ornot, as it is updated not only whenever the MS receivesdata frames but also when it receives beacon frames,which are typically sent every 100 ms by most APs.As the RSSI can fluctuate rapidly, especially when auser is moving, its instantaneous value is not necessarilyrepresentative. At the same time, its local average andtrend are more useful in deciding whether the radiochannel conditions are improving or not and whetherthey are reaching the point where communication is nolonger possible. Using the well known Kalman filter,it is possible to extract this information from the RSSImeasurements. Many authors have already use Kalmanfilter and other time series techniques in order to modelradio channels and the received signal strength, see, forexample, the works by Jiang et al. [7], by Baddour etal. [8] and references therein.
More formally, let X(ti) , Xi be the receivedsignal strength at time ti. In our case, we samplethe RSSI roughly every 100 ms; but, as we rely onsoftware timers, there are no guarantees that the ti’swill be equally spaced. Figure 3b shows the empiricaldistribution of ∆ti = ti− ti−1 for a subset of the traceswe collected. The average is 96 ms and the standarddeviation is 8.2 ms. Given that roughly 90% of thesamples are within less than 100 ms of each other, itseems reasonable to “re-sample” the time series with atime-step of 100 ms.
In all the traces we have collected, it is often thecase that several consecutive samples have the samevalue, indicating that the received signal strength isoften constant during periods that are longer than theaverage distance between samples. The presence of sev-eral samples with exactly the same value is an obstaclewhen one is trying to estimate the local trend of a signalas, in this case, the estimated slope would be exactly0. The Kalman filter does not perform well in thesecircumstances. As we rely on the values reported by the802.11 driver, we wondered whether these consecutivesamples with the same values were caused by the driver
4
●
●
●
●●
●●●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●●
●
●●
●●●●
●●
●●
●
●●
●●●
●
●
●
●
−100 −90 −80 −70 −60 −50 −40
1
5
10
50
100
500
1000
Power /dBm
Num
ber
of d
isco
vere
d A
Ps
● AT= 5 msAT= 10 msAT= 20 msAT= 50 msAT= 100 ms
(a) Number and RSSI of discovered APsfor different AT
80 100 120 140 160 180
0.0
0.2
0.4
0.6
0.8
1.0
∆t /ms
CD
F
●● ●●●●
●●●●●
●●●●
●
●
●
●
●●●
●●●●●●●●
●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ● ●
(b) Sampling interval CDF for the originaltime series
(c) The CDF of the length of the periodswhere all the received power samples areequal for the community networks
Fig. 3: RSSI analysis
not updating the values often enough. Figure 3c showsthe distribution of the length of the periods where thesignal strength was constant for several traces collectedusing a static MS which was not sending or receivingdata (note that the distribution was the same whetherthe test was performed using a laptop running Linux ora smartphone running Android). The median is 305 msand the standard deviation is 306 ms. In the case ofmobile MSs and/or data traffic the median values aresmaller (around 110 ms for MS with data traffic, butthe standard deviation is always larger). In order tomitigate the effect of these periods, we pre-processthe RSSI samples, using a time-varying exponentialaverage, before applying the Kalman filter. In order tofurther reduce the lag of the smoothed signal, we use atime-varying weight in the exponential smoothing.
Let Yi be the re-sampled RSSI time series. Weconstruct the smoothed series Zi as:
Zi = αiYi + (1− αi)Zi−1
where Z1 = Y1 and αi = αup if the RSSI is increasing,instead whenever the RSSI starts decreasing αi = α1.Whenever the RSSI is constant the value of αi isdetermined by the last change before the beginning ofthe constant samples. If the last change was an increaseαi = αup, otherwise αi = min(0.8 · αi−1, αmin). Thiscorresponds to the pseudo-code in Algorithm 1.
We have used αup = 0.5, so that the smoothed timeseries will react quickly to upward changes, α1 = 0.4and αmin = 0.01 so that it will, instead, react muchmore slowly to downward changes. The reason of thisasymmetric behavior is that we are interested in havingan accurate estimate of the level and, above all, of thetrend, only when the signal is decreasing. By using alarger αi when the signal is increasing we assure thatZ will quickly reach the value of Y reducing the lagbetween Y and Z.
Algorithm 1 The algorithm used to compute αi1: increasing ← FALSE2: lastV alue← Y1
3: α1 ← α1
4: i← 15: while i ≤ length(Y ) do6: if Yi 6= lastV alue then7: if YI > lastV alue then8: increasing ← TRUE9: αi ← αup
10: else11: increasing ← FALSE12: αi ← α1
13: end if14: else15: if increasing =FALSE then16: if αi−1 > αmin then17: αi ← 0.8 · αi−1
18: else19: αi ← αi−1
20: end if21: else22: αi ← αup23: end if24: end if25: end while
We have verified that the received power times seriesof our sample are indeed non-stationary by computingtheir autocorrelation functions, which were all slowlydecreasing (as a function of the lag). We have thendecided to use a state-space based model to representthe evolution of the power over time. In particular wehave used the local linear trend model (see, for example,Durbin and Koopman [9]):
Zi = µi + εi εi ∼ N (0, σ2ε)
µi+1 = µi + νi + ξi ξi ∼ N (0, σ2ξ ) (1)
νi+1 = νi + ζi ζi ∼ N (0, σ2ζ )
5
where Zi is the time series under scrutiny, µi is thelevel at time i, νi is the slope at time i and εi, ξi, ζiare independent identically distributed (i.i.d.) Gaussianrandom variables with 0 mean and variance σ2
ε , σ2ξ , σ
2ζ
respectively. These variances can be obtained by Max-imum Likelihood Estimation from sample realizationsof Z.
Once the values for the variances are specified, andgiven a realization of Zi (i = 0, . . . , n), one can usethe well known Kalman filter algorithm to computeµi and νi for any value of i (again, see, for example,Durbin and Koopman [9]). To be more precise, one cansolve the “filtering” problem, where the values of µiand νi are computed using the samples Z0, Z1, . . . , Zi.At first, we have used the dlm [10] R [11] package tosolve the filtering problem. Note that, as the filteringproblem uses only the samples between 0 and i, it canbe implemented in real time as it depends only on pastvalues of the time series. The Kalman filter can alsobe used to predict future values. In the case of thelocal linear trend model (1), the prediction algorithm isextremely simple: one can just model the future valuesof the time series using a straight line with slope νi,starting at the value µi at time i.
We have also implemented the Kalman filter on aSamsung Nexus S smartphone, in the WiFiStateMachinemodule of the Android Java framework. For the sake ofsimplicity we have used a straightforward implementa-tion of the Kalman recursion. The general form of theKalman filter is:
Zi = FΘi + v vi ∼ Nn(0, Vi)
Θi = GΘi−1 + wi wi ∼ Nm(0,Wi)
For the local linear trend model Z is a scalar, and so isv, while:
Θi =
(µiνi
), G =
(1 10 1
),W =
(σ2ξ 0
0 σ2ζ
), F = (1 0) .
We are interested in computing the 2× 2 vector mi =(E[µi] E[βi])
T , containing the expected values of thelevel (µi) and slope (νi). It is known [9] that one cancompute these values using the following equations:
mi = ai +RiFTi Q−1i ei fi = Fiai
Ci = Ri −RiFTi Q−1i FiRi Qt = FiRiFTi + Vi
ai = Gimi−1
Ri = GiCi−1GTi +Wi
where ei = Yi − fi, and the following initial values:C0 =
( σ21 0
0 σ22
), m0 = (Z0 0)T are used for C and
m. The values of σ21 and σ2
2 have almost no influenceon the computations as the matrices R,F, T, C quickly
converge to steady state values which are independentfrom the initial values.
The local linear trend model 1 is characterizedby three parameters: σ2
ε , σ2ξ , σ
2ζ . It is possible to use
Maximum Likelihood Estimation (MLE) methods toestimate the values of these parameters from samplerealizations of Z. We have used the MLE functions ofthe dlm package to this end, but in some cases theoptimization algorithm used to compute the MLE didnot converge. When it converged its estimations for σ2
εand σ2
ζ where not always consistent over all the samplesbut the order of magnitude was fairly consistent, with σ2
εusually smaller than σ2
ζ and with σ2ε often fairly close
to 0. It should be stressed that, in this case, there areno guarantees about the convexity of the optimizationproblem solved by the MLE procedure, which canvery well converge to a local minimum instead of aglobal one. Also it is not uncommon to tune the modelparameters in order to improve its performances. In ourcase we have observed that using σ2
ε = 0.5, σ2ξ = 1
and σ2ζ = 2.5, we obtain fairly smooth level and slope
values, which can be effectively used by the KHATalgorithm.
B. Algorithm design
KHAT adapts the scanning strategy, the scanningperiod and the handover trigger by comparing an es-timate of the link quality and the quality of candidateAP as presented in Fig. 4. The main process consistsin continuously monitoring the RSSI of the current linkand detect a link going down event. To achieve this,we use the Kalman filter to obtain the current valueof the RSSI (µ) and the slope (ν). After analyzing alarge number of RSSI time series, we have estimatedthat the link going down trigger can be declared ifµ < −70 dBm and ν < −0.2 dBm/s. If the linkgoing down condition is satisfied, the MS check on itscandidate AP list. If there is not a valid candidate AP,the MS will attempt scanning only if there has not beenanother scanning instance for the last TScan seconds. Onthe other hand, if after triggering a link going downcondition, the MS has a valid candidate it will attempta handover only if the difference between the candidateAP RSSI and the current exponentially smoothed RSSIsample (µ) is greater than ∆, where ∆ is defined asfollows:
∆ =
8 , if µ > −70
5 , if − 70 ≤ µ < −75
3 , if − 75 ≤ µ < −80
2 , if µ ≤ −80.
(2)
After a scan completes, an existing candidate APwould be updated with a new RSSI, or a new candi-date may be selected. Additionally, in order to avoid
6
Monitor
RSSI going down?
Candidate AP ?
Handover Condition ?
TScan expired ?
Scan
New Scan Results
Update TScan
Handover Condition ?
Candidate AP ?
Handover
Y
Y N
YYY
N
NN
Y
Update Candidate
Fig. 4: Algorithm Flow Chart
µ Number of Active Time Scanning duration(dBm) channels (ms) (ms)
scanning at a very high frequency, we adapt the valueof TScan depending on the scanning results. Each timethe MS triggers a link going down, if a candidate APexists, we double the current value of TScan (up to 1 s)since it is not necessary to scan at a high frequency ifthere is at least one candidate AP. On the other hand,if no candidate exists at that time, we set TScan to itsminimum value (250 ms).
The scanning strategy itself is also adapted depend-ing on the current link condition. Each scanning strategyconsists in determining a number of channels to scanand the time to wait on each channel (AT in Androidsystem). Based on results presented in section III-C wefixed AT as presented in Table III: the better the currentlink quality is, the less time the MS will spend scanning,because it still has time to find APs before it disconnectsfrom its current AP. When the signal quality with thecurrent AP is low, we set aside more time for the MS toscan, in order to maximize the probability to find an AP.In order to contain the scanning duration, we propose touse AT in {5 ms, 10 ms, 20 ms}. The reason is that forsmaller scan times, we only find APs with high RSSI(as shown in section III-C) and as we are in fairly goodcondition, the MS would only be interested in AP withhigh RSSI.
V. EXPERIMENTATION
A. Methodology and implementation
We have implemented our solution on the AndroidICS 4.0.3 system working on a Samsung Nexus S(GT-I9023) smartphone. It involves modifications inthe Android Java Framework, the WPA Supplicant
KHAT Stock AndroidIndoor Average Power (dBm) -63 -68
Average Throughput (MB/s) 2.11 0.80Outdoor Average Power (dBm) -71 -77
Average Throughput (MB/s) 0.22 0.12
TABLE IV: Performance comparison
0 50 100 150 200 250 300Scan Occurences
AP 1
AP 2
AP 3
AP 4
AP 5
AP 6
AP 7
AP 8
AP 9
AP 10
AP 11
AP 12
AP 13
AP 14
AP 15
AP 16
AP 17 90
85
80
75
70
65
60
55
50
Fig. 9: Heatmap of the selected AP in the outdoorenvironment
and the Linux driver (BCM4329). We moved at thesame time two identical smartphones, one with KATHimplementation and the other with stock Android, intwo different environments. The first set of experimentswas performed in our building as described in Fig 2.The mobile user walks at a roughly constant speedand each connection lasts for 120 s. The second set ofexperiments was performed in the city of Luxembourg,using the HOTCITY Wifi deployment (see [12] for moredetails). In all cases, we use iperf to generate the TCPtraffic for both MSs and generate several connectionsfor more than one hour.
B. General results
Fig. 5, 6, 7 and 8 show the RSSI and the receivedTCP data for the two considered environments, over oneconnection duration, while Table IV shows the averageover all connections that we made. We can see thatKHAT provides a better RSSI (-69 dBm in average)along the connections and allows the smartphone tohave a better throughput than stock Android (222 kB/sversus 146 kB/s). We can also see that KHAT triggersthe handover systematically before stock Android toavoid suffering from a poor quality with its current AP.Sometimes, as at Time=400s of the outdoor connection,KHAT manages to find an intermediate AP betweenthose chosen by Stock Android which allows to signig-icantly increase both the RSSI and the TCP download.Looking at the zoom of Fig. 8 between Time=200s and
7
0 50 100 150
−90
−80
−70
−60
−50
Time (s.)
RS
SI (
dBm
)
● ● ● ● ● ● ● ● ●
●
KHAT RSSILegacy RSSIKATH HandoverLegacy Handover
Fig. 5: RSSI for Legacy and KHAT smartphones indoor
0 50 100 150
050
100
150
200
Time (s.)
Dow
nloa
ded
Dat
a (M
B)
KHAT RxLegacyRxKHAT HandoverLegacy Handover
Fig. 6: Received data for Legacy and KHAT smartphones indoor
0 200 400 600 800 1000
−90
−80
−70
−60
−50
Time (s.)
RS
SI (
dBm
)
● ● ● ● ● ●● ● ● ●●●● ●●●●● ●●● ● ● ● ● ● ● ●● ●
●
KHAT RSSILegacy RSSIKATH HandoverLegacy Handover
Fig. 7: RSSI for Legacy and KHAT smartphones outdoor
0 200 400 600 800 1000
050
100
150
Time (s.)
Dow
nloa
ded
Dat
a (M
B)
KHAT RxLegacyRxKHAT HandoverLegacy Handover
200 220 240 260 280 300
2030
4050
Fig. 8: Received data for Legacy and KHAT smartphones outdoor
300s, we can see a quite large period of time (around50s) where the Legacy smartphone is not able to receiveany data from the TCP server. On the other hand,
KHAT is performing two handovers. The first handoverat Time=233s is made prior to the Legacy handover, butstill a stagnation in the received data is observed before
8
and after the handover. However the second handoverat Time=275s is smooth and does not impact the datareception.
Fig. 9 shows the list of selected APs from the chosenconnection in the outdoor environment captured by anMS running Wi2me[1]. The APs are shown in theirapparence order along the path. We can see that forthe six first APs, the KHAT AP selection is judicious:when the signal strength of one AP is degrading, anotherAP becomes available. However, between scanning oc-curence 140 to 200, there is not ideal choice for anyAP. Observing that the RSSI is low, KHAT is tryingto handover to different AP at this period of time,osciatilling between AP7, AP8, AP9 and AP10. Theseare the handovers we can see around Time=625s in Fig7. While KHAT avoids to trigger handover when it isnot for a significantly better AP, in areas where all APsoffer low RSSI, KHAT may trigger several handovers. Itfinally finds AP11 at Time=721s which was providing agood coverage area. During this period, we can see thatthe Legacy phone did not trigger any handover and wasunable to receive any data for a long period of time.
VI. CONCLUSION
IEEE 802.11 is one of the most popular wirelessstandard to offer high data rate Internet connection.With the vast number of hot-spots and communitynetworks that are deployed today, there is a potentialfor users to use Wifi network in mobility scenarios.However, as the AP coverage area is usually limitedto few tenths of meters, there is a strong need foroptimized mobility support, when users move from oneAP to another one. We have shown in this paper thatcurrent devices are able to transit between APs, but thehandover performance is quite low.
We proposed a handover algorithm called KHATthat anticipates the signal loss from the current AP topreemptively scan for potential APs. The prediction ofthe link going down is achieved with a Kalman filterwhich estimates the slope of the RSSI to determinethe link condition. If the estimate is below a giventhreshold (smooth RSSI lower than −70 dBm and slopelower than 0.2), we launch a scanning. Data packets canbe buffered (and retrieved later on) by the AP duringthe MS scanning by exploiting the power saving modedefined in 802.11. Depending on the scanning results,the MS will either handover to a new (better) candidateAP that has been found, or it will loop on the linkquality prediction. The scanning period and strategy areadapted depending on the current link condition.
We have implemented KHAT on Android ICS 4.0.3system working on a Samsung Nexus S (GT-I9023). Toaddress the tradeoff between the scanning latency andthe AP discovery, the MS is scanning with AT=20 ms if
a handover is imminent, AT=10 ms when the link qualityis medium and AT=5 ms when the link quality is good.In two different environments (indoor and outdoor), wecompared a Stock Android with a KHAT smartphone.We have shown that KHAT outperforms Stock Androidby anticipating handovers, and using more APs availableon the path. The average RSSI is 6dBm either inthe outdoor environment, and the TCP throughout is0.22MB/s compared to 0.12MB/s for Stock Android.The perspective of this work is to apply the link qualityprediction on candidate APs in order to better choosethe target AP when a handover is needed.
VII. ACKNOWLEDGMENTS
This work has received a French government supportunder reference ANR-10-LABX-07-01 (Cominlabs).
REFERENCES
[1] G. Castignani, A. Lampropulos, A. Blanc, and N. Montavont,“Wi2Me: A Mobile Sensing Platform for Wireless Hetero-geneous Networks,” in International Workshop on Sensing,Networking and Computing with Smartphones (Phonecom),IEEE ICDCS, june 2012.
[2] C. Perkins, “IP Mobility Support for IPv4,” Internet Requestsfor Comments, IETF, RFC 3220, february 2002. [Online].Available: http://www.ietf.org/rfc/rfc3220.txt
[3] V. Mhatre and K. Papagiannaki, “Using smart triggers forimproved user performance in 802.11 wireless networks,” inProc. of the 4th int. conf. on Mobile systems, applications andservices, 2006, pp. 246–259.
[4] S. Yoo, D. Cypher, and N. Golmie, “LMS predictive link trig-gering for seamless handovers in heterogeneous wireless net-works,” in IEEE Military Communications Conference, 2007.MILCOM 2007. IEEE, Oct. 2007, pp. 1–7.
[5] H. Wu, K. Tan, Y. Zhang, and Q. Zhang, “Proactive scan:Fast handoff with smart triggers for 802.11 wireless LAN,”in INFOCOM 2007. 26th IEEE International Conference onComputer Communications. IEEE, May 2007, pp. 749–757.
[6] R. Silva, P. Carvalho, P. Sousa, and P. Neves, “Enablingheterogeneous mobility in android devices,” Mob. Netw. Appl.,vol. 16, no. 4, pp. 518–528, Aug. 2011. [Online]. Available:http://dx.doi.org/10.1007/s11036-011-0322-6
[7] T. Jiang, N. Sidiropoulos, and G. Giannakis, “Kalman filteringfor power estimation in mobile communications,” WirelessComm., IEEE Trans. on, Jan 2003.
[8] K. Baddour and N. Beaulieu, “Autoregressive modeling forfading channel simulation,” Wireless Comm., IEEE Trans. on,july 2005.
[9] J. Durbin and S. Koopman, Time series analysis by state spacemethods, ser. Oxford statistical science series. Oxford Univ.Press, 2001.
[10] G. Petris, “An R package for dynamic linear models,”Journal of Statistical Software, 2010. [Online]. Available:http://www.jstatsoft.org/v36/i12/
[11] R Development Core Team, R: A Language and Environmentfor Statistical Computing. [Online]. Available: http://www.R-project.org
[12] G. Castignani, J. Monetti, N. Montavont, A. Arcia-Moret,R. Frank, and T. Engel, “A study of urban ieee 802.11 hotspotnetworks: towards a community access network,” in WirelessDays (WD), 2013 IFIP, Nov 2013, pp. 1–8.
