cognitive frequency hopping based on interference...

Cognitive Frequency Hopping Based on InterferencePrediction: Theory and Experimental Results ∗

Stefan Geirhofera John Z. Sunb Lang Tonga Brian M. Sadlerc

[email protected] [email protected] [email protected] [email protected] of Electrical and Computer Engineering, Cornell University, Ithaca, NY 14853

bResearch Laboratory of Electronics, Massachusetts Institute of Technology, Cambridge, MA 02139cU.S. Army Research Laboratory, Adelphi, MD 20783

Wireless services in the unlicensed bands are proliferating but frequently face high interfer-ence from other devices due to a lack of coordination among heterogeneous technologies.In this paper we study how cognitive radio concepts enable systems to sense and predictinterference patterns and adapt their spectrum access accordingly. This leads to a newcognitive coexistence paradigm, in which cognitive radio implicitly coordinates the spec-trum access of heterogeneous systems. Within this framework, we investigate coexistencewith a set of parallel WLAN bands: based on predicting WLAN activity, the cognitive radiodynamically hops between the bands to avoid collisions and reduce interference. The de-velopment of a real-time test bed is presented, and used to corroborate theoretical resultsand model assumptions. Numerical results show a good fit between theory and experimentand demonstrate that sensing and prediction can mitigate interference effectively.

I. Introduction

Wireless devices and services are growing rapidly andare becoming ubiquitous. The unlicensed bands haveplayed a fundamental role in this trend as illustratedby the rapid proliferation of WiFi and Bluetooth de-vices. As new technologies gain momentum, how-ever, the unlicensed bands are becoming increasinglycrowded, and consequently interference managementneeds to be addressed.

The lack of coordination in the unlicensed bandsmay lead to high interference among heterogeneoustechnologies that operate in close proximity. At thesame time, however, average spectrum utilization istypically low because of traffic burstiness and the inef-ficiencies associated with distributed medium accessprotocols. This contrast motivates a study on how bet-ter coordination may improve spectral efficiency andlead to reduced interference.

Throughout this paper we assume that there is no

∗Manuscript submitted October 31, 2008; accepted March 17,2009. This work is supported in part by the U.S. Army Re-search Laboratory under the Collaborative Technology AllianceProgram, Cooperative Agreement DAAD19-01-2-0011. The U.S.Government is authorized to reproduce and distribute reprintsfor Government purposes notwithstanding any copyright notationthereon. This work has been presented in part at the AsilomarConference on Signals, Systems and Computers, Monterey, CA,November 2007 and the IEEE Military Communications Confer-ence, Orlando, FL, October 2007. The present manuscript extendsprevious work by including new results on the open- and closed-loop behavior of the test bed.

direct information exchange among interfering sys-tems. Instead we study how a cognitive radio mayreduce collisions by predicting the WLAN’s activ-ity and adapting its medium access accordingly. Thecognitive concept of sensing and adaptation conse-quently serves as an implicit means for coordinatingthe medium access of the interfering systems. Ul-timately, this leads to a new coexistence frameworkwhich we refer to ascognitive coexistence.

Cognitive coexistence is based on the same prin-ciples as dynamic spectrum access (DSA) but facesdifferent challenges and tradeoffs in typical deploy-ments (see [1] for a review of DSA). In many coex-istence scenarios radios are collocated and separatingtransmissions in the temporal domain therefore seemsa natural design choice. Compared to DSA, sensing ismuch easier accomplished in such setups since radiosare in close proximity of each other. Furthermore, inunlicensed bands CR concepts can be used in a sim-ilar fashion to orthogonalize interfering systems dy-namically. While the notion of primary and secondaryusers does not arise in this scenario because all sys-tems have equal access rights to the medium, a hierar-chical setup is typically a natural design approach.

I.A. Main Contribution

This paper presents an experimental study of coexis-tence between a cognitive radio (CR) and a set of par-allel WLAN channels; seeFig. 1. Assuming that the

Mobile Computing and Communications Review, Volume 13, Number 2 49

Ch. 1

Ch. i

Ch. M

1 2 3 4 5 6 7 8 9 10 11 12 13 14time slots

WLAN Cognitive radio

Figure 1: System setup. A frequency hopping cogni-tive radio coexists with a set of parallel WLAN chan-nels.

CR operates in a frequency range that overlaps witha set of WLAN channels, we study a cognitive fre-quency hopping (CFH) protocol, which avoids packetcollisions by periodic spectrum sensing and predictingtemporal WLAN activity. By hopping to temporarilyinactive channels, packet collisions are reduced andcoexistence improved. Theoretically, CFH is based onan ON/OFF continuous-time Markov chain (CTMC)model, which approximates WLAN’s decentralizedmedium access while being tractable enough to derivethe optimal hopping pattern.

This paper focuses on comparing theoretical andexperimental performance results. The previous work[2] is taken as a basis for CFH but the protocol isclosely related to other contributions on time domainOSA [3,4]. Theoretical assumptions are corroboratedby a real-time test bed, which is used for validatingmodel assumptions and evaluating the dynamic inter-action between the CR and WLAN. For example, ourtheoretical analysis is based on assuming that colli-sions do not impact the validity of the temporal pre-diction model as long as the collision rate is kept be-low some interference constraint. In an actual system,however, does the CTMC prediction model remainapproximately valid or are there significant changesin the WLAN’s temporal activity? Similarly, is theWLAN’s sensing-based medium access affected? Theexperimental approach helps quantifying the impactof retransmissions, a result that would be hard to ob-tain analytically. Our findings show a good matchbetween theory and practice and demonstrate that theCR can reduce interference effectively. To the best ofour knowledge this paper presents the first test bed tovalidate the above model assumptions experimentally.

Limitations. The experimental test bed has beentailored to the above objectives. As such, it is neces-sary to clarify some limitations due to hardware and

complexity constraints.First, the test bed only supports a single WLAN

channel due to bandwidth limitations of the down-conversion module. Consequently, the test bed’smedium access reduces to transmitting or not trans-mitting at the beginning of each slot. Nevertheless,even though no frequency hopping takes place, the ex-perimental results for the single-channel case can begeneralized to multi-channel setups. Specifically, wewill demonstrate that by characterizing the behaviorof a single WLAN channel for arbitrary traffic inten-sities, it is possible to extrapolate performance metricsto multi-channel scenarios as long as the behavior ofparallel channels is statistically independent.

The test bed is further limited to a single cognitivetransmitter; no receiver has been implemented. Thisremoves the need to maintain slot synchronization butdoes not limit our ability to measure CFH’s through-put and interference through monitoring channel ac-tivity. While methods such as collaborative sensing[5] or acknowledgement feedback [6] may be used forsynchronization, their implementation in real-time isdifficult and goes beyond the scope of this work. Sim-ilarly, multi-user aspects of the cognitive system arenot addressed. However, well-studied concepts couldbe applied, such as Bluetooth’s piconet structure [7],where a master node acts as a central controller.

Finally, mutual interference among collocated ra-dio systems crucially depends on the underlying prop-agation conditions. In this paper, we focus on theworst-case collision model in which any overlap infrequency and time results in a packet drop. This ap-proximates a setup in which WLAN devices and CRnodes operate in close proximity.

Organization and notation. The paper is struc-tured as follows.Sec. IIintroduces CFH analytically,and Sec. III describes the test bed implementation.The measurement methodology and performance re-sults are presented inSec. IV and Sec. V, respec-tively. Performance trends and tradeoffs are discussedin Sec. V.C.

I.B. Related Work

While there is a growing body of literature on cog-nitive radio and dynamic spectrum access, see,e.g.,[1,8], reports on experimental and implementation as-pects of CR and DSA have been limited. This workis most closely related to contributions on temporalDSA. Among the first to address this problem, Zhaoetal. investigate the problem of accessing slots that areleft idle by primary users [9]. The optimal medium ac-cess is derived within a Markovian decision-theoretic

50 Mobile Computing and Communications Review, Volume 13, Number 2

framework and a separation principle between sens-ing and medium access is shown [10]. Similar setupshave recently received increasing interest [4, 11, 12].An experimental test bed which heuristically accesseswhite space in WLAN is presented in [13]. Recent ex-perimental studies and related work include [14–17].

Coexistence between WLAN and Bluetooth de-vices has conceptual similarities with our contributionbecause of a similar physical-layer setup. Differentcoexistence methods have been proposed for this sce-nario and range from interference cancelation at thephysical layer [18] to changes in MAC layer schedul-ing at the WLAN stations [19] and adaptive frequencyhopping (AFH) [20].

AFH techniques are most closely related to CFH:both schemes adapt their hopping sequence to reduceinterference. The major difference lies in how inter-ference is detected and modeled. AFH typically clas-sifies channels as being either “good” or “bad” ac-cording to the empirical error rates of its own trans-mission attempts. However, the WLAN’s tempo-ral activity is not modeled explicitly and no spec-trum sensing is performed. Bad channels are simplyavoided by reducing the hopping set to good chan-nels, if possible. Naturally, this approach is wellsuited to suppress interference that is static or slowlytime varying with respect to the Bluetooth slot length.In many cases, however, WLAN packets are onlyslightly longer than the slot length, reducing the ben-efit of this modeling technique. In contrast, CFH’ssensing and prediction framework is well suited to ac-count for the time variant behavior of WLAN traffic.

Some aspects of CFH have been investigated inprior work by the authors of this paper. The model-ing of WLAN traffic has been proposed in [21], andmedium access schemes based on these models wereintroduced in [2]. This paper goes beyond these con-tributions in presenting an experimental test bed andcomparing theoretical with experimental results. Ourfindings give rise to system-level insights on the per-formance of this method.

II. CFH Protocol Design

To improve coexistence between the WLAN channelsand the frequency-hopping CR, CFH adapts the CR’shopping sequence such that it preferably transmits inbands that will likely remain idle for the duration ofthe transmission.

As depicted inFig. 1, the CR transmission is slot-ted in time with durationT . At the beginning of eachslot the activity (idle or busy) of the WLAN chan-

nels is detected and prediction is based on a two-statecontinuous-time Markov chain model. We assumeperfect sensing in our theoretical work.

II.A. Modeling Spectrum Opportunities

Since the WLAN channels do not overlap in fre-quency it is reasonable to assume that they evolve in-dependently in time. Hence, we can focus on a singlechannel, and deal withN independent models (pos-sibly with different parameters) rather than a singlemodel encompassing all channels.

The activity of each WLAN band is modeled bya two-state CTMC{X(t), t ≥ 0} with idle stateX(t) = 0 and busy stateX(t) = 1. The sojourntimes are exponentially distributed with parameterλin the idle state andµ in the busy state.

This simple model approximates the statistics of theidle periods of real WLAN systems. As shown inFig. 2, by comparing with empirical data gathered bymeasurement, we see that the overall behavior of theempirical cumulative distribution function is approxi-mated by this exponential fit. The measurement setupand assumptions will be described in more detail inSec. III.

We have shown in previous work [2] that a semi-Markov model (SMM) is able to achieve a better fitwith the empirical data by approximating idle peri-ods based on a mixture of two generalized Paretodistributions (one corresponding to the WLAN con-tention behavior, the other modeling random packetarrivals). This leads to the cumulative distributionfunction (CDF)

Fm(t) = p1F1(t; k1, σ1)+(1−p1)F2(t; k2, σ2), (1)

wherep1 denotes the mixture coefficient and

Fi(t; ki, σi) = 1 −

(

1 + kit

σi

)−1/k

(2)

represents the CDF of the generalized Pareto distribu-tion with shape parameterki (measuring the deviationfrom an exponential distribution) and scale parameterσi (quantifying the decay rate). This model shows anexcellent fit with the empirical data as can be seen inFig. 2and has also been validated statistically throughthe Kolmogorov-Smirnov test [21].

While a mixture distribution leads to an accurate fit,it is less tractable analytically. Our theoretical anal-ysis is therefore based on the more tractable CTMCmodel. While its deviation from the empirical data issignificant inFig. 2, our results will show that the ex-ponential model predicts overall system performancequite accurately and is a useful theoretical tool.


trans-

mitidle

Ft(t)

Fi(t)

(a) State transition model

0 5 10 15 200

0.2

0.4

0.6

0.8

1

idle period [ms]

cum

ula

tiv

e d

istr

ibu

tio

n f

un

ctio

n

Empirical CDF

Exponential model

SMM model

σ = 0.2

σ = 0.6

(b) Sojourn time distribution in the idle state

Figure 2: Temporary prediction model of WLAN ac-tivity. The semi-Markov model approximates the em-pirical sojourn time in the idle state with good accu-racy (shown for two WLAN traffic intensitiesσ).

The busy periods are directly related to the packetlength and transmission rate chosen by the schedulerof each WLAN station. In our work, we focus ontraffic scenarios with constant payload packets andtransmission rate, and can therefore treat the sojourntime as deterministic. For the CTMC approximation,which requires exponential sojourn times, we choosethe distribution’s mean to equal the deterministic so-journ time.

II.B. Designing the Optimal Control

Based on the CTMC model we discuss the optimalCFH behavior for a single WLAN channel (due to sin-gle channel operation of the test bed). The generalcase ofM channels has been addressed in [2].

The objective of CFH is to maximize CR through-put while constraining the rate of packet collisions.The CR throughput is measured in terms of successfulpacket transmissions per unit time and the interferenceconstraint is characterized by the WLAN packet errorrate (that is, the number of dropped divided by thenumber of transmitted WLAN packets). This normal-izes the interference constraint by the WLAN’s traffic

intensity and ensures that no more than a certain frac-tion of packets gets dropped.

In the single channel case, the optimal medium ac-cess reduces to finding transmission probabilities foridle and busy sensing outcomes, respectively. When-ever the channel is observed idle, we transmit with aprobabilityp, which is designed such that the interfer-ence constraint is met with equality [2]. Whenever abusy sensing outcome is observed, it is optimal not totransmit because this would only cause higher inter-ference while not increasing throughput.

In the multi-channel case we need to select one outof M bands for transmission. This case can be ad-dressed through decision-theoretic analysis by formu-lating the problem as a constrained Markov decisionprocess. Due to space limitations and because our ex-perimental work focuses on the single channel case,we refer to [2] for details.

III. Test Bed and Experiment Design

Having introduced CFH, we describe the test bed im-plementation, compare the experimental design withour analytical setup, and discuss fundamental designobjectives.

The test bed has been developed with the objectiveof validating some of the implicit assumptions madein our analytical work. This specifically includes dy-namic effects between both systems that are difficultto characterize analytically. Our analytical work fo-cused on designing transmission probabilities, givena stochastic model for the WLAN, such that interfer-ence constraints remained met. The idea behind thismodeling approach is that as long as the interferenceconstraints are sufficiently tight, the residual interfer-ence caused to the WLAN will not impact its temporalbehavior. While analytical approaches for justifyingthis assumption are difficult, our experimental resultsenable us to confirm its validity.

A block diagram of the test bed is shown inFig. 3.The implementation is based on a Sundance soft-ware defined radio (SDR) development kit, consistingof processing and data acquisition modules. Radio-frequency (RF) signals are down-converted using acommercial WLAN transceiver and up-converted us-ing an Agilent vector signal generator. The basebandprocessing is done entirely on the SDR board. TheCR’s slot structure is implemented using an accuratetimer, which triggers periodic interrupts with a periodof T = 625µs. The analog-digital converter (ADC)is triggered at the beginning of each slot. A1µsblock at a rate of 72 MHz is captured and passed on


Energy

detection

RF

Inp

ut

Est.

, !

Find

p( ,!)

Prob.

Tx

Transmit

Data

Generate

waveform

Down-

converter

Find

thres.

Up-

converter

RF

Ou

tpu

t

CFH Control

Ta Tb Tc Td

Sensing time: Ta 10!s

Processing time: Tb 100!s

Retuning time: Tc 100!s

Transmission time Td 500!s

Spectrum Sensing

CR Transmission

Figure 3: Block diagram of the cognitive radio (left). The slot operation and timing is shown on the right. Picturesof the cognitive radio test bed are available online at http://acsp.ece.cornell.edu/mc2r.pdf.

to the energy detector, which computes the signal’senergy and compares it with a threshold. This resultsin a binary sensing result (idle/busy), which is usedby the CFH controller to determine the CR’s mediumaccess. Depending on the outcome of the stochasticcontrol a transmission may be initiated by using a pro-grammable signal generator. Upon digital-to-analog(DAC) conversion, the signal is up-converted by theRF front-end.

The above operations introduce processing delaysthat reduce the actual transmission time per slot. Typ-ical values for these delays are shown inFig. 3. Spec-trum sensing relies on blocks of less than10µs, and isalmost negligible compared toT . The processing timefor the sensing result and the CFH controller togetheramount to roughly100µs in our implementation. Thetime it takes to re-tune the transmitter to a differentchannel amounts to approximately100µs (this delaydoes not occur in our setup, however, as we only dealwith a single WLAN channel). The remainder of theslot can be used for the CR’s transmission.

The baseband processing can be categorized intothree parts, namely (i) the spectrum sensor, (ii) theCFH controller, and (iii) the CFH transmitter. In thefollowing each component is discussed in more detail.

III.A. Spectrum Sensor

Spectrum sensing plays a key role in CR systems andthe challenges associated with reliably detecting sig-nals at very low SNR have been the subject of muchinvestigation. Compared to some DSA setups, how-ever, the burden of reliable spectrum sensing is re-duced in this cognitive coexistence setup. In contrastto DSA schemes that orthogonalize systems by suffi-cient spatial separation (and therefore require the abil-

ity to detect weak primary signals), typical SNR val-ues faced in this cognitive coexistence setup will besubstantially larger as both systems operate in closeproximity of each other.

Thanks to the fairly high SNR conditions, energydetection can be used efficiently with very little com-plexity. Energy detection is mathematically formu-lated as a binary hypothesis testing problem on a setof N samples that either represent just noise, or a sig-nal in noise, respectively. This leads to

H0 : Yi = Vi, i = 1, . . . ,N (3)

H1 : Yi = Si + Vi, i = 1, . . . ,N,

whereYi denotes the complex baseband samples,Vi

are noise samples,Vi ∼ CN (0, σ2

0), andSi denotes

the signal samples drawn from a complex Gaussian,Si ∼ CN (0, σ2

1). This hypothesis testing problem is

standard [22]. The optimal Neyman-Pearson detectoris given by

T (y) =N

∑

i=1

|yi|2H1

≷H0

γ, (4)

where the thresholdγ needs to be chosen such thatthe probability of false alarm (i.e., erroneously declar-ing a busy channel) is no larger than a specific value.The Neyman-Pearson detector then yields the optimalprobability of detection.

In theory, when (3) holds exactly, the thresholdγcan be determined analytically by finding closed-formexpressions for the probability of false-alarm and de-tection. In the experimental domain, however, numer-ous other factors need to be taken into account. Afundamental problem in this work is the fact that hy-pothesesH0 andH1 cannot be observed isolated from


each other because the test bed and the WLAN packettransmission are not synchronized. Therefore it is notpossible to sample exclusively during idle periods orexclusively during busy periods. This results in a mix-ture of the distributions under either hypothesis wherethe mixture parameter depends on the long term prob-ability of observing idle and busy slots. The hypothe-ses can however be separated by statistical analysis.

Empirical observations of the sufficient statistic (4)are plotted inFig. 4. We expect to observe a mixtureof chi-square distributions because bothT (y|H0) andT (y|H1) are chi-square distributed. More than twomixture components may be necessary, however, dueto slightly different power levels of the WLAN termi-nals. Indeed, the empirical CDF can be well approxi-mated by a mixture of three chi-square distributions,

f(x) =3

∑

i=1

pifi(x;αi, βi), (5)

wherepi ≥ 0 for all i,∑

3

i=1pi = 1, and

fi(x;αi, βi) = xαi−1βαi

i e−βix

Γ(αi)(6)

represents the PDF of a Gamma distribution withshape parameterαi and rate parameterβi. AnExpectation-Maximization algorithm [23] is used tofind the model parameters of (5) and the fitting resultis shown inFig. 4. A good match with the empiricaldata is observed. We also found that busy and idlemixture components have very different rate parame-ters. This illustrates the very different energy levelspresent in idle and busy sensing slots.

Numerical performance analysis demonstrates thatthe test bed’s spectrum sensor works reliably. Bychoosing the decision threshold appropriately a de-tection probability of 98.5% can be realized with lessthan 1% false alarms. The good performance of en-ergy detection is, of course, due to the moderate tohigh SNR conditions, which make spectrum sensingsimilar to the carrier-sensing employed in systemssuch as IEEE 802.11.

III.B. Cognitive Controller

The role of the cognitive controller is to initiate atransmission probabilistically, whenever an idle sens-ing result is observed. Transmissions are never ini-tiated after a busy sensing result because this wouldlead to a collision with high probability. A full imple-mentation of CFH encompasses the tracking of traffic

0 0.1 0.2 0.3 0.4 0.50

0.2

0.4

0.6

0.8

1

signal power (normalized)

cum

ula

tiv

e d

istr

ibu

tio

n f

un

ctio

n

Empirical CDF

Gamma mixture

Figure 4: Decision-statistic for the energy detector.A mixture of three gamma distributed components isobserved.

variations and estimation of the CTMC model param-etersλ andµ, the computation of the optimal trans-mission probabilitiesp(λ, µ), and the randomized se-lection of the control action based on these probabil-ities. Due to the complexity associated with estimat-ing λ and µ, the experimental test bed uses a fixedtransmission probabilityp. Despite this limitation, itis possible to infer the throughput and collision ratefor the case in which the transmission probability isadjusted with the traffic intensity. Since transmissionsare never initiated in busy slots, any choice ofp corre-sponds to a certain throughput and collision rate. Fur-thermore, increasingp leads to an increased through-put but also to a larger collision rate. For the single-channel case in which the medium access is describedby a single parameter it is therefore possible to mea-sure throughput and collision rate experimentally andthen normalize accordingly such that the interferenceconstraint is met. Thus, the performance of CFH canbe evaluated without much loss of generality.

III.C. Cognitive Transmitter

If a CR transmission is initiated, it lasts for the remain-der of the slot duration. For the CFH operation it is notrelevant what specific transmission scheme is used.For optimal usage of the white space between con-secutive packets the CR could use the same frequencybands as the WLAN. For simplicity and motivated bythe conceptual similarity with Bluetooth/WLAN co-existence, however, we designed the transmitter to re-semble that of Bluetooth. It therefore transmits innarrowband channels of 1 MHz and similar modula-


tion parameters [7]. Gaussian Frequency Shift Key-ing (GFSK) with a time-bandwidth product of 0.3 wasused at a symbol rate of 1 MSps. We should notethat the bandwidth of the sensing frontend, however,amounted to 36 MHz as discussed before.

The test bed’s transmitter is implemented based ona programmable signal generator which is integratedinto the acquisition module, and can be triggered insoftware. Data contained in an internal buffer is thentransferred to the DAC and played back in an infiniteloop (for further implementation details we refer to[24] due to space limitations).

IV. Measurement Methodology

The previous section described the test bed’s imple-mentation. As CFH relies on sensing and predictingpacket collisions, its performance naturally dependson the specifics of the coexistence setup, such as prop-agation conditions, traffic intensity, system parame-ters, etc. This section is devoted to describing themeasurement methodology that underlies the perfor-mance assessment.

IV.A. Hardware Setup

The experimental coexistence setup is depicted inFig. 5 and consists of the WLAN system (composedof an access point and three workstations), the CR,as well as a vector signal analyzer which was used tomonitor the operation of the system.

Two fundamentally different configurations wereconsidered. In theopen-loopsetup the CR’s outputis not fed back to the WLAN system but only used todetermine the packet error rate. While this does not re-flect what would occur in practice, this setup enablesus to draw a direct comparison with our analytical re-sults. In theclosed-loopsetup, on the other hand, in-terference impacts the WLAN and leads to frequentretransmissions. We analyze the CR’s impact and re-late the results to the open-loop setup.

WLAN configuration. The WLAN consists ofcommercial off-the-shelf IEEE 802.11 devices and in-cludes an access point and three workstations withadapter cards. The access point is connected to afourth workstation using a wired ethernet connection.All wireless devices are configured identically to op-erate in channel 6 (corresponding to a center fre-quency of 2.437 GHz) and use a transmission rate of5.5 Mbps.

The wireless devices’ RF outputs are all connectedto a resistive power divider using coaxial cables. Thisisolates the transmissions from the environment and

reduces interference that could otherwise result fromunrelated transmissions in the unlicensed bands (mea-surements were taken in an office building with anumber of unrelated WLAN access points). In addi-tion, this configuration removed any propagation ef-fects and allowed for repeatable results. While thepropagation conditions encountered in practice willdeviate from this idealized setup, our results corre-spond to a worst-case scenario in which packet over-laps inevitably result in collisions.

Open-loop setup. The open-loop configuration isshown inFig. 5(a). While the combined WLAN sig-nals serve as the input to the CR, its output is not fedback to the WLAN system. Instead it is combinedwith the WLAN signal and detected by a separateworkstation computer with a WLAN adapter card.The two circulators isolate the output of the test bedand prevent it from impacting the WLAN.

The workstation capturing the combined signal ofthe WLAN and the test bed uses an adapter card inpromiscuous mode together with commercial WLANanalysis software to detect WLAN packets. The out-put power level of the CR is set large enough suchthat a collision between both systems will prevent theadapter card from successfully receiving the packet(either a packet error will be displayed or the packetwill be missed altogether, depending on whether thesynchronization preamble or only the payload is af-fected). The rate of successful packet reception is thusmeasured and, by comparison with settings of the traf-fic generators, the rate of packet losses can be inferred.

Closed-loop setup. The closed-loop setup is de-picted in Fig. 5(b). The test bed again receives thecombined WLAN signal, but its output is now con-nected to the same power-divider as the WLAN de-vices. Therefore, packet collisions lead to packetdrops at the WLAN devices themselves (and not at areference station as in the open-loop setup). Packetloss will therefore initiate WLAN retransmission,which may in turn affect the test bed’s performance.The closed-loop setup therefore allows for dynamicinteraction between both devices.

In the closed-loop setup, PC1, which is connectedto the WLAN access point by ethernet, runs trafficanalysis software. The traffic generators are config-ured such that PC1 is the intended receiver, and hencethe successful portion of the WLAN traffic (includingany retransmissions that may occur) is measured. Bycomparing with the case of no interference, the packetloss rate is inferred.


PC3

PC2

Ag

ilent 89640A

VSA

RT CMAP

Power Divider

Primary system Cognitive System

RT...RouterVSA...Vector Signal Analyzer

PC1

PC4

(a) Open-loop setup

PC3

PC2

Ag

ilent 89640A

VSA

RT CMAP

Power Divider

Primary system Cognitive System

PC4

PC1

(b) Closed-loop setup

Figure 5: Measurement methodology. In the open-loop setup (left) the CR’s transmission are not fed back to theWLAN. The closed-loop setup (right) is used to quantify the impact of retransmissions.

IV.B. Traffic Characteristics

WLAN transmissions are generated by using a traf-fic generator on each workstation. The DistributedInternet Traffic Generator (D-ITG) [25] is used andallows to specify packet lengths, the distribution ofinter-arrival times, and transmission rates. The trafficproperties form an important component of the mea-surement methodology.

This work focuses on UDP traffic with constantpayload of 1024 bytes and plots performance with re-spect to traffic intensity. Specifically, we define theWLAN traffic load σ, which is normalized such thatσ = 0 corresponds to an inactive WLAN andσ = 1represents a WLAN operating at full capacity.

Experimentally, the WLAN traffic loadσ is mea-sured as follows. First, the rates of all traffic genera-tors are set to a value that is too large to be supportedby the WLAN (even in the case of no interference)and by measuring the actual rate, the WLAN capac-ity is found. Then, the settings of the traffic gener-ator are normalized by this value, leading to values0 ≤ σ ≤ 1. For example, given the traffic and prop-agation settings in our setup, a maximum of approxi-mately 450 packets per second could be supported bythe WLAN. Configuring each of the traffic generatorsto transmit at a rate of 50 packets per second thereforecorresponds toσ = 1/3.

The measurements focus on the average throughputand interference for stationary traffic scenarios withdifferent intensitiesσ. Measurement results are com-pared with simulations using the model parameters inTab. 1. These parameters were obtained based on sta-tistical analysis, as discussed inSec. II. The resultscan be extended to non-stationary traffic setups, pro-vided that parameters of the traffic model are trackedover time [21].

IV.C. Measurement Process

The measurements are performed in the followingmanner. First, with the CR portion of the test bedturned off, the WLAN traffic generators are adjustedsuch that a specific traffic loadσ is realized. The CRis then enabled and the successfully received WLANpackets are counted. By comparing with the nominalpacket rate in the interference-free case, the packet er-ror rate can be obtained. At the same time, the CRkeeps statistics of the number of initiated and suc-cessful slot transmissions. Lacking a CFH receiver,a successful CFH slot transmission is defined by initi-ating slot transmissionand observing an idle channelat the beginning of the next slot. Due to the reliablespectrum sensing performance and the fact that, in oursetup, WLAN packets are always longer than the slotdurationT , this is a valid metric.

V. Performance Results

This section presents experimental performance re-sults and compares them to simulation results ob-tained using the SMM and CTMC models discussedin Sec. II. For the open-loop setup, where theoreti-cal and experimental results should coincide, we ob-serve an excellent match. In the closed-loop setup,where mandatory re-transmissions affect the exper-imental performance results, we introduce a simpleheuristic formulation that allows us to approximatethe WLAN’s behavior based on open-loop results.

The performance assessment in this section focuseson the CR’s throughput and the interference it causes(in terms of WLAN packet errors induced by the in-terference). Clearly, both performance metrics are in-terrelated: by increasing the CR’s transmission prob-ability p we can increase throughput at the expense ofa larger number of collisions and vice versa. By mea-suring both metrics, we can get a better insight on how


aggressive the CR transmission policy can be withoutsignificantly affecting WLAN performance.

We compare the performance of CFH with a blindreference scheme that neither detects nor predictsWLAN activity but obliviously initiates transmissionswith probability pr regardless of the state of themedium. Our results demonstrate that CFH intro-duces a significant performance gain and increases CRthroughput while reducing WLAN interference.

V.A. Open-Loop Measurement Result

The open-loop measurement result is shown inFig. 6.The left panel shows the CR throughput in terms ofthe expected number of successful slot transmissionsper unit time and the right panel shows the packet er-ror rate of the WLAN. We stress that the experimentalcurves have been obtained by counting the rate of suc-cessfully received WLAN packets and the packet errorrate is computed by comparing with the average num-ber of packets that should have been received duringthat time period. The results are compared with sim-ulations based on the SMM and CTMC model. TheSMM-based simulation also incorporates the fact thatthe CR does not use the entire slot for transmissionand should therefore approximate the measurementresults with good accuracy.

The SMM-based simulation results indeed show anexcellent match with the experimental results. Thethroughput curves shown in the left panel ofFig. 6al-most coincide (the largest aberration amounts to lessthan two percentage points) and for the packet errorrate (shown on the right) we observe a maximum ab-beration of two percentage points. By comparison, theCTMC model is less accurate. While the throughputresults match fairly well, the predicted packet errorrate deviates significantly. From a modeling stand-point, this is not surprising because the exponentialapproximation does not capture the WLAN’s con-tention behavior.

Similar performance trends are observed for the ref-erence scheme without spectrum sensing. As shownin Fig. 6, measurement and SMM-based simulationagain match very well while the CTMC approxima-tion shows noticeable deviation in terms of predictedpacket error rate.

V.B. Closed-Loop Measurement Result

The results for the closed-loop setup are shown inFig. 7. Here, the WLAN terminals are strongly in-terfered with by the CR and therefore initiate retrans-missions whenever a packet is dropped. If a retrans-

mission attempt is ultimately successful, the packet iscounted as successfully received (the retransmissionpackets themselves are not counted toward the WLANtraffic load).

Compared to the open-loop setup, the throughputof the cognitive system stays approximately the same,but the interference to the WLAN changes drasticallydue to the retransmission behavior. Up to approx-imately σ = 0.8, no packet loss is observed. Atσ = 1.0, it increases to roughly 5%.

The non-cognitive reference scheme exhibits a sim-ilar behavior. No packet errors occur forσ = 0.5or below, and an approximately linear increase is ob-served for higher values ofσ. The packet error ratereaches a maximum of roughly 35% atσ = 1.0.

The reason for the reduced packet errors lies in theWLAN’s retransmission behavior. If a packet trans-mission fails, the standard mandates that a retransmis-sion be initiated. At low WLAN rates, it is likely thatthese retransmissions will be successful because themedium is predominantly idle. At high rates, how-ever, it may no longer be possible to accommodatesuch retransmissions, leading to an increasing packeterror rate.

In order to compare experimental and theoreticalresults, we consider a simple approximation that in-corporates the retransmission behavior. For simplic-ity, we assume that packets are retransmitted indefi-nitely until they are received successfully. A trafficloadσ encountering a collision probabilityq leads tothe cumulative traffic load

σ′ = σ(1 + q + q2 + · · · ) =σ

1 − q. (7)

We assume that as long as the traffic loadσ′ is be-low the capacity of the WLAN channel̄σ, no packetloss occurs because retransmissions can be accommo-dated. However, forσ′ > σ̄ this is no longer possibleand there will be a packet error rate of approximately

σ′ − σ̄

σ′. (8)

This simple heuristic approximation is used to com-pare closed and open-loop results. Clearly, in or-der to compute (8) only open-loop performance result(obtained by measurement or simulation) are needed.Based on these it is possible to approximate the packeterror rate in the closed-loop setup.

The performance curves gathered in this way areshown inFig. 7 and again show CR throughput (left)and packet error rate (right). The measurement curvesare obtained directly by measurement from the closed-loop setup. The simulation curves are obtained based


0 0.2 0.4 0.6 0.8 10

0.05

0.1

0.15

0.2

0.25

0.3

0.35

WLAN Traffic Load σ

Av

g R

ate

of

Su

cces

sfu

l P

ack

et T

ran

smis

sio

ns

Cognitive Radio Throughput

0 0.2 0.4 0.6 0.8 10

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5


WL

AN

Pac

ket

Err

or

Rat

e

WLAN Packet Error Rate

Measurement

Sim. (SMM)

Sim. (CTMC)

CFH

No-sensing

No-sensing

CFH

Figure 6: Open-loop performance result. The CR throughput (left) and WLAN packet error rate (right) areshown. The measurement results match closely with SMM-based simulations. The results obtained under theCTMC model show a larger deviation.

on open-loop simulations with the SMM and CTMCmodels and applying (8). We observe a good matchwith the measurement results for the closed-loop casesuggesting the simple heuristic predicts the closed-loop behavior quite accurately.

Lastly, the results show that even at high trafficload, there exists a residual throughput of the CR sys-tem. This appears to be the result of WLAN’s re-transmission behavior, which due to frequent colli-sions enlarges some contention windows to accommo-date other stations. As can be seen in the left panel ofFig. 7, this results in a residual throughput of approx-imately 0.03 even when the WLAN is fully loaded.

V.C. Performance Trends

Based on the performance results presented in the pre-vious section, we discuss some tradeoffs and chal-lenges which may arise in a complete implementationof such a system.

The choice of system parameters in this paper ismotivated by facilitating a comparison between the-ory and practice. Consequently, we focus on a worst-case propagation scenario in which any time overlapbetween transmissions results in packet errors. Thisapproximates the case in which the devices are locatedin close proximity.

An important design parameter in trading off

throughput versus interference is the transmissionprobability p. Clearly, CR throughput increases withp, but so does the interference that is inflicted uponthe WLAN. A natural design approach is to choosep based on the traffic parameters such that a specificconstraint on the WLAN packet error rate is met withequality. In the multi-channel case the optimal vec-tor of transmission probabilities can be found throughdecision-theoretic analysis, which leads to a linearprogramming solution with fairly low implementationcomplexity. Further, transmission probabilities onlyneed to be recomputed whenever the prediction pa-rameters change significantly; on a slot level, the ran-dom access can be implemented by storing the vectorof transmission probability in a lookup table.

A tradeoff arises in selecting the slot length. Ifthere was no overhead associated with sensing andre-tuning the CR, reducing the slot duration wouldenable us to more efficiently “fill up” the idle peri-ods between packets. Due to this overhead, however,choosing a small slot duration leads to small CR pay-load and consequently reduced performance. Becauseof conceptual similarity we have chosen a slot lengthof 625µs, which equals the value used in Bluetooth.Further, we have found that around this value, thethroughput of the CR changes very little.

Throughout this paper we have focused on station-ary traffic scenarios with varying WLAN traffic inten-


0 0.2 0.4 0.6 0.8 10

0.05

0.1

0.15

0.2

0.25

0.3

0.35


Av

g R

ate

of

Su

cces

sfu

l P

ack

et T

ran

smis

sio

ns

Cognitive Radio Throughput

0 0.2 0.4 0.6 0.8 10

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4


WL

AN

Pac

ket

Err

or

Rat

e

WLAN Packet Error Rate

Measurement

Sim. (SMM)

Sim. (CTMC)

CFH

No-sensing

No-sensing

CFH

Figure 7: Closed-loop performance result. CR throughput (left) and WLAN packet error rate (right) are shown.The packet error rate includes retransmissions and is therefore smaller than in the open-loop case.

sity. This enabled us to measure throughput and inter-ference by recording long packet traces and comput-ing time averages. Nevertheless, we believe that ourresults will extend to the non-stationary traffic sce-narios observed in practice. In fact, the time scaleof our prediction model is in the order of tens ofmilliseconds, which is much smaller than the typicaltime scale associated with changes in usage patterns.Tracking non-stationarities by adapting model param-eters is therefore a viable approach. We have verifiedthat this leads to an accurate fit in previous work [21].

While our comparison between theory and experi-ment shed some light on implementation aspects, ourwork represents the first steps toward developing afully functional prototype. Synchronization is one as-pect that goes beyond the scope of this paper. Due tothe dependence on local sensing results, synchroniza-tion is more difficult to achieve than in related AFHsetups. Collaborative sensing concepts are a possi-ble solution approach. By exchanging sensing met-rics, the detection process could be coordinated, andthe stochastic control actions can be synchronized byusing identical random seeds. Methods such as ac-knowledgement feedback are also applicable [6].

VI. Conclusion

In conclusion, this paper presented an experimentalcognitive radio test bed which uses sensing and pre-

diction to exploit temporal white space between pri-mary WLAN transmissions. By comparing measure-ment results with simulations based on an idealizedmathematical model, we were able to validate modelassumptions that have frequently been used in otherworks.

The experimental aspect of our work resulted insome test bed limitations which were due to hard-ware and complexity constraints. However, while thetest bed is limited to single-channel operation, it cap-tures fundamental deviations from the idealized math-ematical model and incorporates carrier sensing andretransmission effects. We have further shown thatgeneralizations to the multi-channel case are possibleby extrapolating performance measurements for thesingle-channel case.

References

[1] Q. Zhao and B. M. Sadler, “Dynamic Spec-trum Access: Signal Processing, Networking,and Regulatory Policy,”IEEE Signal Process.Mag., vol. 55, no. 5, pp. 2294–2309, May 2007.

[2] S. Geirhofer, L. Tong, and B. M. Sadler, “Cogni-tive Medium Access: Constraining InterferenceBased on Experimental Models,”IEEE J. Sel.Areas Commun., vol. 26, no. 1, pp. 95–105, Jan.2008.


Table 1: Model parameters for SMM and CTMC formulations. The parameters were obtained by experiment(see [21] for details) and used to obtain the simulation results presented in the previous section.

WLAN Traffic Load0.05 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

CTMCλ [ms] 23.3 11.6 7.89 5.42 3.32 2.34 1.63 1.01 0.68 0.43 0.24µ [ms] 2.0 2.0 2.0 2.0 2.0 2.0 2.0 2.0 2.0 2.0 2.0

SMM

p1 0.11 0.33 0.27 0.34 0.41 0.54 0.62 0.76 0.82 0.88 0.95k1 0.75 -0.39 -0.33 -0.40 -0.52 -0.3 -0.09 0.02 0.12 0.14 0.07

σ1 [ms] 29.6 18.6 10.7 8.14 5.48 4.81 3.43 3.63 2.59 2.42 2.17k2 0.75 -0.39 -0.33 -0.40 -0.52 -0.3 -0.09 0.02 0.12 0.14 0.07

σ2 [ms] 0.07 0.34 0.32 0.34 0.41 0.29 0.22 0.18 0.15 0.12 0.11

[3] Q. Zhao, L. Tong, A. Swami, and Y. Chen,“Decentralized Cognitive MAC for Opportunis-tic Spectrum Access in Ad Hoc Networks: APOMDP Framework,”IEEE J. Sel. Areas Com-mun., vol. 25, no. 3, pp. 589–600, Apr. 2007.

[4] S. Huang, X. Liu, and Z. Ding, “Opportunisticspectrum access in cognitive radio networks,” inProc. IEEE INFOCOM, Apr. 2008, pp. 1427–1435.

[5] A. Sahai, N. Hoven, S. Mishra, and R. Tandra,“Fundamental tradeoffs in robust spectrum sens-ing for opportunistic frequency reuse,” Univer-sity of California, Berkeley, Tech. Rep., 2006.

[6] Q. Zhao, S. Geirhofer, L. Tong, and B. M.Sadler, “Opportunistic Spectrum Access via Pe-riodic Channel Sensing,”IEEE Trans. SignalProcess., vol. 56, no. 2, pp. 785–796, Feb. 2008.

[7] Bluetooth Special Interest Group, “Specificationof the Bluetooth System,” Nov. 2004.

[8] I. Akyildiz, W. Lee, M. Vuran, and S. Mo-hanty, “NeXt generation/dynamic spectrum ac-cess/cognitive radio wireless networks: A sur-vey,” Computer Networks, vol. 50, no. 13, pp.2127–2159, Sep. 2006.

[9] Q. Zhao, L. Tong, and A. Swami, “Decentral-ized Cognitive MAC for Dynamic Spectrum Ac-cess,” inProc. First IEEE Internatial Symposiumon New Frontiers in Dynamic Spectrum AccessNetworks, Nov. 2005, pp. 224–232.

[10] Y. Chen, Q. Zhao, and A. Swami, “Joint De-sign and Separation Principle for OpportunisticSpectrum Access in the Presence of Sensing Er-rors,” IEEE Trans. Inf. Theory, vol. 54, no. 5, pp.2053–2071, May 2008.

[11] A. N. Mody, S. R. Blatt, D. G. Mills, T. P.McElwain, McElwain, N. B. Thammakhoune,J. D. Niedzwiecki, M. J. Sherman, C. S. My-ers, and P. D. Fiore, “Recent Advances in Cog-nitive Communications,”IEEE Commun. Mag.,vol. 45, no. 10, pp. 54–61, Oct. 2007.

[12] A. Motamedi and A. Bahai, “MAC ProtocolDesign for Spectrum-agile Wireless Networks:Stochastic Control Approach,” inProc. IEEEInternational Symposium on New Frontiers inDynamic Spectrum Access Networks (DySPAN),Apr. 2007, pp. 448–451.

[13] S. D. Jones, N. Merheb, and I.-J. Wang, “An ex-periment for sensing-based opportunistic spec-trum access in CSMA/CA networks,” inFirstIEEE Internatial Symposium on New Frontiersin Dynamic Spectrum Access Networks, Nov.2005, pp. 593–596.

[14] P. Bahl, R. Chandra, and J. Dunagan, “SSCH:Slotted Seeded Channel Hopping for CapacityImprovement in IEEE 802.11 Ad-Hoc WirelessNetworks,” inProc. ACM MobiCom, 2004.

[15] K. G. Shin, H. Kim, C. Cordeiro, and K. Chal-lapali, “An Experimental Approach to Spec-trum Sensing in Cognitive Radio Networks withOff-the-Shelf IEEE 802.11 Devices,” inProc.IEEE Consumer Communications and Network-ing Conference (CCNC), Jan. 2007, pp. 1154–1158.

[16] H. Kim and K. G. Shin, “In-band spectrum sens-ing in cognitive radio networks: energy detec-tion or feature detection?” inProc. ACM Mobi-Com, 2008, pp. 14–25.

[17] F. Perich, R. Foster, P. Tenhula, andM. McHenry, “Experimental Field TestResults on Feasibility of Declarative Spectrum


Management,” inProc. IEEE Symposium onNew Frontiers in Dynamic Spectrum AccessNetworks (DySPAN), Oct. 2008, pp. 1–10.

[18] M. Ghosh and V. Gaddam, “Bluetooth interfer-ence cancellation for 802.11g WLAN receivers,”in Proc. IEEE International CommunicationsConference (ICC), May 2003, pp. 1169–1173.

[19] N. Golmie, “Bluetooth Dynamic Scheduling andInterference Mitigation,”Mobile Networks andApplications, vol. 9, pp. 21–31, 2004.

[20] N. Golmie, N. Chevrollier, and O. Rebala,“Bluetooth and WLAN Coexistence: Challengesand Solutions,”IEEE Trans. Wireless Commun.,vol. 10, no. 6, pp. 22–29, Dec. 2003.

[21] S. Geirhofer, L. Tong, and B. M. Sadler, “Dy-namic Spectrum Access in the Time Domain:Modeling and Exploiting Whitespace,”IEEECommun. Mag., vol. 45, no. 5, pp. 66–72, May2007.

[22] H. V. Poor,An Introduction to Signal Detectionand Estimation, 2nd ed. Springer-Verlag, 1994.

[23] Z. Liu, J. Almhana, V. Choulakian, and R. Mc-Gorman, “Traffic Modeling with Gamma Mix-tures and Dynamical Bandwidth Provisioning,”in Proc. IEEE Communication Networks andServices Research Conference (CNSR), May2006.

[24] S. Geirhofer, J. Z. Sun, L. Tong, and B. M.Sadler, “CMAP: A Real-Time Prototype forCognitive Medium Access,” inProc. IEEE Mil-itary Communications Conference (MILCOM),Oct. 2007.

[25] S. Avallone, A. Botta, D. Emma, S. Guadagno,and A. Pescape, “D-ITG V.2.4 Manual,” Uni-versity of Napoli “Federio II”, Tech. Rep., Dec.2004.


cognitive frequency hopping based on interference...

Documents