External Interference-Aware Channel AssignmentFelix Juraschek Mesut Günes

Institute of Computer ScienceFreie Universität Berlin and Humboldt Universität

Berlin, Germany{felix.juraschek, mesut.guenes}@fu-berlin.de

Abstract—Channel assignment is an efficient method to in-crease the network capacity of multi-radio mesh networks byusing orthogonal channels for otherwise interfering transmissionsin the network. With the massive increase of commercial andpersonal WLAN deployments in urban areas, co-located networksmay cause a performance degradation if operated on the samewireless channels. Thus, in order to fully exploit the benefits ofspectral diversity, channel assignment needs to become awareof external interference. In this paper, we propose EICA, adistributed algorithm for external interference-aware channelassignment. We show how EICA is capable to detect co-locatedsources of interference with spectrum sensing using commodityIEEE 802.11 radios of the network nodes. Thus, no additionalhardware is required. Based on the channel conditions, thealgorithm favors the lesser congested channels for the channelassignment. We evaluate EICA in a large-scale testbed with 126multi-radio network nodes in different scenarios with increasinglevels of interference. The testbed study shows that EICA cansignificantly increase the network capacity in scenarios with highlevels of external interference.

Index Terms—wireless mesh network; channel assignment;external interference; testbed; spectrum sensing

I. INTRODUCTION

Multi-radio mesh routers allow the communication overmultiple wireless network interfaces at the same time. How-ever, this can result in high interference of the wirelesstransmissions leading to a low network performance [1].Channel assignment for multi-radio wireless mesh networks(WMNs) increases the network performance by exploiting theavailability of fully or partially non-overlapping channels sothat otherwise interfering wireless transmissions utilize dif-ferent non-overlapping channels [2]. Thus, the intra-networkinterference can be reduced. This is feasible, since commonlyused wireless network technologies, such as IEEE 802.11,provide multiple non-overlapping channels [3] and the cost ofsuch radios has significantly dropped in the last decade. Forthese reasons, the number of WMN deployments is rising, forinstance to provide last-mile broadband Internet access.

However, the increasing density of WLANs in urban areasand the pervasive use of wireless devices in our everydaylives leads already to significant capacity limitations in theunlicensed 2.4 and 5 GHz bands [4]. Furthermore, with thedawn of the Internet of Things, the number of wireless devicesoperating in the unlicensed frequency bands is dramaticallyincreasing [5]. From the point of view of the network operator,co-deployed WLAN networks can be considered external,

since they are not under control of the network operator.Recent studies have shown that external interference can beextremely harmful to the performance of WLAN networks [6].

Thus, channel assignment approaches have to consider notonly intra-network interference, but also external interferencecaused by co-located devices in order to exploit the benefitsof spectral diversity. This paper presents an algorithm forexternal interference aware channel assignment that is capableto efficiently adapt to environments with co-located WLANs.

The contributions of this paper are the following:

• A distributed algorithm for external interference-awarechannel assignment (EICA).

• A method to measure the channel condition withIEEE 802.11 hardware that is capable of detecting co-located WLAN radios in the communication and inter-ference range.

• A performance evaluation of EICA in a large-scale multi-radio testbed with 126 network nodes.

The remainder of the paper is structured as follows. In Sec-tion II we give an overview of related work. The DES-Testbed,is introduced in Section III. The method to detect interferersis described in Section IV and EICA follows in Section V.The performance evaluation is presented in Section VI andthe paper concludes with an outlook on future work.

II. RELATED WORK

Over the last decade, a wide range of channel assignmentalgorithms have been developed for different scenarios regard-ing the network technology, application scenario, and networkarchitecture [2]. The approaches can be distinguished intocentralized and distributed algorithms. Centralized algorithmsrely on a central entity, which calculates the network-widechannel assignment [7]. In distributed approaches, each meshrouter calculates its channel assignment based on local infor-mation [8], [9]. Distributed approaches usually react faster totopology changes due to node failures or node mobility. Asa result, distributed approaches are considered more suitableonce the network is operational and running [2].

Only a few proposed algorithms, however, consider externalinterference caused by co-located devices in the channelassignment decisions. The main reason is, that these co-locateddevices are not under the control of the network operator andtheir radio activity patterns are thus hard to capture. Therefore,most algorithms aim to reduce only intra-network interference.

978-1-4673-2480-9/13/$31.00 ©2013 IEEE 1302

The first algorithm for channel assignment that considers ex-ternal interference has been proposed in [10]. The centralizedalgorithm incorporates the external interference by periodicallyassessing the interference levels of the available channels. Thisis done by setting a wireless interface into monitor mode,in which frames of co-located IEEE 802.11 stations can bereceived that pass the frame check sequence (FCS). This way,co-located IEEE 802.11 interfaces in the communication rangecan be determined. The congestion of the channel is measuredby analyzing the captured packets for the particular channel.The results are passed to a CAS which then calculates thenetwork wide channel assignment and sends the decisionsback to the mesh nodes. One of the drawbacks is, that onlyIEEE 802.11 radios in the communication range of a nodeare captured. Remote interferers, which reside outside thecommunication range but inside the interference range, arenot captured because their transmissions do not pass the FCS.

The same method to detect external interference with aninterface in monitor mode is also used in [11]. A hybridchannel assignment algorithm is then extended that considersthe measured data of external interference. Another central-ized approach in a small metropolitan testbed with 6 nodesequipped with directional antennas is proposed in [12]. Theauthors use the round-trip delay of a link as a channel selectionmetric, which is influenced by external interference. However,information about the channel conditions is not collected.

The closest to our approach is [10]. However, the au-thors state that their approach is limited to detect co-locatedIEEE 802.11 devices that reside in the communication rangeof the network node. Our approach differs, since it allows usto capture remote interferers, that only reside in the carriersensing range but not in the communication range of a node.

III. DES-TESTBED

All experiments in this paper were carried out on theDistributed, Embedded Systems-Testbed (DES-Testbed) at theFreie Universität Berlin [13]. The DES-Testbed comprises126 multi-radio indoor and outdoor nodes and is deployedin an unshielded environment over the computer sciencefaculty buildings. Each network node is equipped with threeIEEE 802.11 wireless network interface cards (WNICs). Oneis a Ralink RT2501 USB stick and the other two are MiniPCI cards with an Atheros AR5413 chipset. The cards usethe rt73usb [14] and ath5k [15] drivers, which are part of theLinux kernel. While the Ralink WNICs are IEEE 802.11b/gdevices operating in the 2.4 GHz band, the Atheros WNICsadditionally support the IEEE 802.11a standard on 5 GHz.

IV. DETECTING EXTERNAL INTERFERENCE

A crucial prerequisite for external interference-aware chan-nel assignment is the capability of the network nodes to detectthe RF activity of external devices. This is not a trivial task,since the external devices are usually not under control of theoperator of the network in question. Thus, the number andlocations of external interferers are usually not known beforethe network operation. Additionally, the traffic patterns of the

external devices may be dynamic over time. One solutionto estimate the local level of interference is to determinethe congestion levels of the radio channels. Specialized hard-ware systems, such as the commercial spectrum analyzersWiSpy [16] and Spectrum XT [17], allow the detection of RFactivity in the unlicensed 2.4 and 5 GHz bands. Alternatively,software solutions using commodity IEEE 802.11 hardwarecan be used for the same purpose.

In our approach, we measure the channel conditions bydirectly accessing the carrier sensing statistics of the wirelessnetwork interface. Atheros-based WNICs maintain two regis-ters that can be used for this purpose [18]. The cycle time (ortime slot) counter is incremented with every clock tick anddepends on the clock of the particular Atheros chipset. Themedium busy counter holds the information for how manycycles the medium has been sensed busy in regard to theclear channel assessment (CCA) function. Current drivers forthe Atheros-based WNICs, such as ath5k [15], expose thesestatistics. We retrieve the statistics for the channel timeand channel time busy by periodically reading the cor-responding 32-bit registers of the wireless network interfaceswith the AR5413 chipset in the DES-Testbed. With theseinformation we can determine the channel congestion (orchannel load) CGc for channel c:

CGc =channel time busy

channel time(1)

The result for the channel congestion is a value between 0and 1, describing what fraction of the measurement interval themedium was sensed busy on the particular channel. A valueCGc = 0.5 corresponds to a channel congestion of 50% forchannel c in the measurement interval.

The procedure based on this method to assess the channelcondition is the following. One interface per node is configuredinto monitor mode and measures the carrier sensing statisticsfor a short duration on each channel, which is set to 1 secondin our implementation. The results in form of the congestionlevels for each channel are provided to the channel assignmentalgorithm. Since the channel conditions may be dynamic overtime, periodic calls to the procedure are required to maintaincorrect information of the channel congestion.

Our approach for estimating the local interference levelsbased on the carrier sensing statistics has several advantagescompared to the method in [10] and [11]. Remote interferersare captured as well because their transmissions trigger thecarrier sensing at the sensing node. Additionally, the methoddoes not require additional specialized hardware and can thusbe integrated easily into existing wireless mesh networks.

V. INTERFERENCE-AWARE CHANNEL ASSIGNMENT

The external interference-aware channel assignment(EICA) algorithm incorporates the channel congestioninformation to avoid already heavily congested channels inthe local environment of the network nodes. The idea forEICA has been first published in [19] and was refined inthe process of implementing and evaluating a prototype in a

1303

real testbed environment, which is the focus of this paper.EICA results from an extension of the distributed greedyalgorithm (DGA) [8]. With DGA, the least used channelsin the interference set of a node are assigned to its networkinterfaces. The interference set Sn of a node n consists ofall nodes and their channel assignment whose transmissionsaffect sending and receiving at node n. Thus, the originalDGA considers only intra-network interference and notexternal interference. To preserve the network connectivity,one interface on every node is operating on a global commonchannel (or default channel) and is not used for channelassignment. EICA extends DGA with the presented procedureto measure the channel congestion on each network node.Based on these measurement data, a blacklisting mechanismis introduced, that removes channels with a high load fromthe set of the available channels.

The EICA algorithm including the channel congestion mea-surement procedure (lines 1–6) and the channel selectionprocedure (lines 7–15) is given in Figure 1. At the networkinitialization, all nodes measure the local channel congestionon all available channels K. If the channel congestion CGc ata particular node exceeds a threshold TCG for channel c ∈ K,the channel c is removed from the set of the available channelsand not considered for the channel assignment on this node.More formally, the set of available channels K after measuringthe channel congestion is updated as follows

K ← {c ∈ K | CGc < TCG} (2)

In other words, if the local congestion on channel c exceedsthe threshold TCG, an impact on the performance of linksoperating on this channel is expected and thus, the channelshould be avoided. It is important to note, that the channelcongestion is measured in a distributed manner on all nodes.The channel congestion may vary throughout the environmentof the network, especially in large-scale wireless networksthat cover a large area and possibly multiple buildings. Thus,channels should not be avoided globally, but only in theareas in which the network nodes have detected a significantload on the particular channels. The measurement interval perchannel is set to 1 second, which results in a runtime of thechannel congestion measurement procedure of several seconds,depending on the number of available channels.

After the channel congestion measurement, the network isinitialized with setting one interface to the default channel.As next, the local communication graph based on the ETXmetric [20] is retrieved and the interference set Sn of eachnode is determined. With an interference cost function fI asgiven in Equation 3, the spectral overlap of two channels isused to determine the level of interference in the interferenceset.

fI(f1, f2, α) = max(0, α− |f1 − f2|) (3)

The cost function takes the center frequencies f1 and f2of two radio channels and the additional parameter α intoaccount. α denotes the minimum frequency difference of

Require: Sn ← Interference set of node nK ← Set of available channelscj ← channels assigned to each node j ∈ Sn

ci ← current channel of the interface i1: for k in K do2: CGk ← measure channel congestion(k)3: if CGk > TCG then4: K ← K \ k5: end if6: end for7: Fbest ←

∑j∈Sn

fI(α, ci, cj)8: cbest ← ci9: for k in K do

10: Fk ←∑

j∈SnfI(α, k, cj)

11: if Fk < Fbest then12: Fbest ← Fk

13: cbest ← k14: end if15: end for

Figure 1: Channel congestion measurements and channel se-lection running on every node with EICA.

orthogonal channels. A value of fI = 0 states that the channelsare orthogonal, meaning they do not interfere with each other.

Using the interference cost function, each node tries to as-sign the channel to its interfaces (except the default interface)that results in the lowest level of interference in its interferenceset (lines 7–15). In each iteration, the algorithm considersexactly one interface and calculates the interference cost forall available channels k with the current channel assignmentof the interference set. The result is the channel cbest withthe lowest overall interference cost. If cbest is different thanthe currently assigned channel for this interface, a channelswitching procedure is initiated. As an additional constraint,only channels can be assigned that are used by at least oneneighboring node, in order to avoid isolated interfaces.

The channel switching procedure is based on a 3-wayhandshake for synchronization [8]. It is initiated by sending aChannelRequest message to all nodes in the interferenceset Sn. The recipients of a ChannelRequest reply with aChannelAccept message, if the proposed channel switchdoes not conflict with their own intended channel switches.The originator of the request applies the channel switch, if allnodes of the interference set reply with a ChannelAcceptmessage. A following ChannelUpdate message is sent tothe nodes in the interference set to confirm the channel switch.If the ChannelRequest conflicts with another pendingchannel switch, a node replies with a ChannelRejectmessage. When receiving a ChannelReject message, theoriginator of the request aborts the current channel switchingprocedure by sending a ChannelAbort message to all nodesin its interference set. The algorithm terminates, when the localinterference can not be reduced any further. The authors ofDGA prove the convergence of the algorithm by showing thatthe overall network interference decreases monotonically with

1304

Parameter ValueNumber of nodes 126 DES-Nodes each equipped with

one IEEE 802.11b/g radio and twoIEEE 802.11a/b/g radios.

Algorithms DGA, EICANumber of interferers 10, 30, 50Interference traffic pattern 1CBR, 6CBR, BURSTDefault interface / channel wlan0 / 14Available channels 36, 44, 48, 52, 60, 64, 100, 108, 112Threshold TCG 0.1Performance metric EXRInterference model 2-hop modelReplications 35 for each interferer class

Table I: Settings for the performance evaluation.

each channel switch in static networks [8]. The algorithm hasbeen implemented for the DES-Testbed based on the channelassignment framework DES-Chan [21].

VI. EVALUATION

For the evaluation, we compare the performance of theoriginal DGA algorithm with EICA in multiple scenarios withincreasing levels of interference. For this study, we use a subsetof the network nodes of the DES-Testbed to act as co-locatedinterferers. The performance metric is based on the networkcapacity when the network is exposed to external interference.Additionally, we investigate the interference levels of thescenarios and the impact of the blacklisting mechanism onthe set of available channels.

A. Experiment design

The settings for this experimental study are defined in Ta-ble I. We use 126 network nodes of the DES-Testbedfor the experiments. For the channel assignment, all threeIEEE 802.11 interfaces are used, set to ad-hoc mode and6 MBit/s data rate. In all experiments, a subset of these nodesacts as interferers (varying from 10, 30, and 50 nodes), whilethe remaining network nodes run the channel assignment. Anode selected as interferer will generate traffic throughout theexperiment using two of its WNICs tuned to random channelsof the available channel set (see Table I). We use three differenttraffic patterns: with 1CBR the interferers generate 1MBit/sconstant bit rate (CBR) traffic. With 6CBR, the interferersgenerate 6 MBit/s CBR traffic, meaning that they jam theparticular channels since the data rate of the interfaces is setto 6 MBit/s as well. For the last traffic pattern, the interfer-ers replay a bursty traffic schedule (BURST) that has beenrecorded before in the testbed of single hop TCP traffic. Theaverage data rate for BURST is roughly 2 MBit/s, the delaybetween subsequent bursts is chosen randomly and may exceedthe duration of the channel congestion measurement procedurefor a single channel. The 3 groups of interferers combined withthe 3 different traffic patterns result in a total of 9 scenarios.We replicate each scenario 35 times for both algorithms. Ineach replication we choose the interferers randomly from alltestbed nodes. Channel 14 is used as common global channelsince measurements have shown that it is not utilized in theenvironment of the DES-Testbed.

B. Performance metric

In each scenario, we measure the external interference ratio(EXR) performance metric with the following steps. First,the specified number of interferers are randomly selected ofall available network nodes. At the beginning, the interferersstart generating traffic while the other nodes run the channelassignment procedure. After the channel assignment procedurehas finished, the evaluation phase is started to measure thenetwork capacity. For this, we select all wireless links L inthe network with the constraint, that each node is at mostincident to not more than one link in L. This gives us themaximum number of node-disjoint single hop links in thenetwork. Then, we measure the throughput of all links in Lwhen they are activated simultaneously with 6 MBit/s UDPtraffic. The aggregate throughput over all links expresses thenetwork capacity and is denoted as:

Tnoif =∑l∈L

tnoifl (4)

We repeat this measurement, but this time, the interferersare active throughout the measurement. The network capacityof the second measurements is defined as:

Tif =∑l∈L

tifl (5)

The EXR metric is then defined as the ratio between thenetwork capacity while the interferers are active and whenthey are not, which is:

EXR =TifTnoif

(6)

A value for EXR close to 1 means that the aggregatethroughput is not reduced when the interferers are activecompared to when they are not. In this case, the networkcapacity is not affected by the external interference. A lowervalue for EXR means that the transmissions of the interferershave decreased the network capacity.

C. Results

1) EXR: The results for the scenarios with the 1CBR trafficpattern for the EXR metric are depicted in Figure 2a. Asexpected, with the growing number of external interferers,the EXR decreases for DGA since the algorithm does notconsider external interference. With 50 interferer, the networkcapacity drops by about 15% compared to when the externalinterferers are not active. The plot shows that with EICAthe network capacity is hardly affected in this scenario, thenetwork capacity drops only by about 3% percent. The resultsfor the increased interferer activity by using 6CBR trafficare depicted in Figure 2b. As expected, the higher level ofexternal interference has an impact on the network capacityof both algorithms. The impact is again higher on DGA, thenetwork capacity drops by 30% percent compared to whenthe interferers are not active. With EICA, the network capacityonly drops by 5%. The final set of results for the BURST trafficpattern are depicted in Figure 2c. The results for DGA are

1305

(a) Results for 1CBR. (b) Results for 6CBR. (c) Results for BURST.

Figure 2: Results of DGA and EICA for the EXR performance metric. In (a), the results are shown for the 1CBR trafficpattern. The results for the 6CBR traffic pattern are shown in (b), and for the BURST traffic pattern in (c).

similar to the ones for the 1CBR scenario. This is as expected,since the data rate of the interfering nodes is comparable inboth scenarios. However, we can see, that the results for EICAare slightly worse compared to the 1CBR scenario. With 50interferer the network capacity drops by about 6% in themedian. This can be explained by the bursty traffic pattern,which makes it harder to capture the external interference inthe channel congestion measurement phase. It may happen,that while the channel condition is measured, the interferingnode is currently not transmitting, thus it is not captured.

2) Interference level: As next, we quantify the interferencelevel of the environment created by the interfering nodesin regard to the traffic pattern and the amount of externalinterferers. For this, we take a look at the mean channelcongestion of all available channels as measured during thechannel congestion measurement procedure in Figure 3. Asexpected, for all three traffic patterns the measured channelload increases with the growing number of interferers. Also,the more traffic the interferers generate, the higher is themeasured channel congestion. Thus, the channel congestionis highest for 6CBR. The mean channel congestion for 6CBRover all available channels measured at all network nodes isabout 9% for 10 interferers and about 29% for 50 interferers.The short confidence intervals show, that we are able to controlthe interference level quite well with our setup.

3) Set of available channels: As next, we investigate theimpact of the blacklisting mechanism for congested channelswith EICA on the set of available channels. The results aredepicted in Figure 4. For all scenarios, the number of availablechannels decreases with the growing amount of interferers.This is as expected, since with the higher level of externalinterference, the mean channel congestion rises and thus morechannels are blacklisted. Since DGA does not blacklist anychannels, it uses 10 channels for all scenarios. It is interestingto note, that the number of available channels for 1CBR

Figure 3: Mean measured channel congestion.

and 6CBR is almost the same. The reason is that with bothtraffic patterns, the interferer generate traffic with a constantbit rate. Thus, the measuring duration of 1 second is longenough to detect all external interferers. The number of theavailable channels for BURST is higher. Some interferersare not detected, since the randomly chosen delay betweensubsequent packet bursts may be longer than the duration ofmeasuring the channel condition for a single channel. Thisexplains, why the results for EXR for the BURST trafficpattern scenarios are similar to the ones for 6CBR, even thoughmuch more traffic is generated by the interferers using 6CBR.

4) Runtime: In all scenarios, all instances of DGA andEICA finished in less than 10 minutes. The channel conges-tion measurement phase takes only about 12 seconds of thetotal runtime. Thus, the approach is practical for large-scalewireless mesh deployments.

5) Limitations and channel re-assignment: An importantquestion is the frequency of executing the channel assignment

1306

Figure 4: Mean amount of available channels after the channelcongestion measurement procedure for EICA.

algorithm to adapt to possible changing interference levelsin the network environment caused for instance by nodemobility. While this depends on how dynamic the networkenvironment is, one solution is two run the relatively shortchannel congestion measurement phase periodically. Based onthe degree of changes to the last measurement, a new iterationof the channel assignment algorithm can be initiated. We willinvestigate the impact of such a channel re-assignment strategyin future work, since it exceeds the scope of the currentstudy. The selection of the global common channel is also animportant factor for the algorithm. We have chosen channel14 in our setup because it is not used in the environmentof the DES-Testbed. However, this might be different in otherlocations, thus measuring the channel congestion levels to findthe least congested channel for the global common channel isdesirable prior to the network operation.

6) Discussion: In summary, EICA outperforms DGA inall 9 scenarios. For the scenario with the highest level ofexternal interference (50 interferers using 6CBR), with EICAthe network capacity only drops by 6% when the interferersare active compared to when they are not. Thus, EICA iscapable to detect highly congested channels and to avoid thesechannels for the channel assignment. The results show, thatmeasuring the channel congestion is a reliable method to detectexternal interferers.

VII. SUMMARY AND OUTLOOK

In this paper, we presented the external interference-awarechannel assignment algorithm EICA. We showed how the esti-mation of external interference can be achieved with spectrumsensing based on the carrier sensing statistics of the WNICs.The evaluation of EICA in a large-scale wireless testbed with126 multi-radio network nodes showed, that EICA is robusttowards high levels of external interference.

Future work will focus on the performance evaluation ofEICA in scenarios with additional non-IEEE 802.11 devicesthat operate in the unlicensed frequency spectrum. A recent

study has shown, that the method of spectrum sensing withcarrier sensing is capable of capturing non-IEEE 802.11 de-vices [6]. We plan extend the DES-Testbed with additionalnon-IEEE 802.11 devices in order to evaluate EICA in sce-narios with diverse interference patterns.

REFERENCES

[1] P. Gupta and P. R. Kumar, “The capacity of wireless networks,” IEEETransactions on Information Theory, vol. 46, no. 2, pp. 388–404, 2000.

[2] W. Si, S. Selvakennedy, and A. Y. Zomaya, “An overview of channelassignment methods for multi-radio multi-channel wireless mesh net-works,” Journal of Parallel and Distributed Computing, vol. 70, pp.505–524, May 2010.

[3] IEEE Std 802.11-2007, “IEEE standard for information technology —Telecommunications and information exchange between systems —Local and metropolitan area networks-specific requirements — Part 11:Wireless LAN medium access control (MAC) and physical layer (PHY)specifications,” LAN/MAN Standards Committee, New York, NY, USA,pp. C1–1184, June 2007.

[4] P. Pawelczak, K. Nolan, L. Doyle, S. W. Oh, and D. Cabric, “Cognitiveradio ten years of experimentation and development,” CommunicationsSurveys Tutorials, IEEE, vol. 11, pp. 90–100, 2009.

[5] H. Sundmaeker, P. Guillemin, P. Friess, and S. Woelffle, Eds., Vision andchallenges for realising the Internet of Things. Cluster of EuropeanResearch Projects on the Internet of Things, European Commision, 2010.

[6] S. Rayanchu, A. Patro, and S. Banerjee, “Airshark: detecting non-wifirf devices using commodity wifi hardware,” in Proceedings of the 2011ACM SIGCOMM, New York, NY, USA, 2011, pp. 137–154.

[7] M. K. Marina and S. R. Das, “A topology control approach forutilizing multiple channels in multi-radio wireless mesh networks,” inIn Broadnets, 2005, pp. 381–390.

[8] B.-J. Ko, V. Misra, J. Padhye, and D. Rubenstein, “Distributed channelassignment in multi-radio 802.11 mesh networks,” in Wireless Commu-nications and Networking Conference (WCNC), 2007, pp. 3978–3983.

[9] A. P. Subramanian, H. Gupta, S. R. Das, and J. Cao, “Minimuminterference channel assignment in multiradio wireless mesh networks,”IEEE Transactions on Mobile Computing, vol. 7, no. 12, 2008.

[10] K. N. Ramachandran, E. M. Belding, K. C. Almeroth, and M. M.Buddhikot, “Interference-Aware Channel Assignment in Multi-RadioWireless Mesh Networks,” in 25th IEEE International Conference onComputer Communications (INFOCOM), 2006, pp. 1–12.

[11] S. Kim, J. Cha, and J. Ma, “Interference-aware channel assignmentswith seamless multi-channel monitoring in wireless mesh networks,” inProceedings of the 2009 IEEE ICC, Piscataway, NJ, USA, 2009.

[12] V. A. Siris and M. Delakis, “Interference-aware channel assignmentin a metropolitan multi-radio wireless mesh network with directionalantennas,” Comput. Commun., vol. 34, no. 12, Aug. 2011.

[13] B. Blywis, M. Günes, F. Juraschek, and J. Schiller, “Trends, Advances,and Challenges in Testbed-based Wireless Mesh Network Research,”Mobile Networks and Applications, vol. 15, pp. 315–329, 2010.

[14] “Ralink technology, Ralink RT2501U chipset product specification.”[Online]. Available: http://www.ralinktech.com/

[15] Linux wireless, “Ath5k driver homepage,” last checked: 09/2012.[Online]. Available: http://linuxwireless.org/en/users/Drivers/ath5k

[16] WiSpy, “WiSpy homepage,” http://www.wi-spy.co.uk/, last checked:09/2012.

[17] Spectrum XT, “Spectrum XT homepage,”http://www.flukenetworks.com/products/airmagnet-spectrum-xt, lastchecked: 09/2012.

[18] P. A. Acharya, A. Sharma, E. M. Belding, K. C. Almeroth, and K. D.Papagiannaki, “Rate adaptation in congested wireless networks throughreal-time measurements,” IEEE Transactions on Mobile Computing,vol. 9, pp. 1535–1550, 2010.

[19] F. Juraschek, M. Günes, and B. Blywis, “External Interference-AwareDistributed Channel Assignment in Wireless Mesh Networks,” in Pro-ceedings of the 5th UBICOMM, 2011.

[20] D. S. J. De Couto, D. Aguayo, J. Bicket, and R. Morris, “A high-throughput path metric for multi-hop wireless routing,” in MobiCom’03. New York, NY, USA: ACM, 2003, pp. 134–146.

[21] F. Juraschek, M. Günes, M. Philipp, B. Blywis, and O. Hahm, “DES-Chan: A Framework for Distributed Channel Assignment in WirelessMesh Networks,” in Proceedings of ATNAC, November 2011.

1307