IEEE Communications Magazine • July 2012128 0163-6804/12/$25.00 © 2012 IEEE

1 http://www.crc.net.nz/crc-net.php

INTRODUCTIONWireless multiradio multichannel mesh networkshave the potential to provide ubiquitous andhigh-speed broadband access in urban and ruralareas, to both fixed and mobile users, with lowoperation and management costs. To investigateissues related to the management and perfor-mance of a multiradio mesh network in an actu-al metropolitan environment, we have deployedan experimental multiradio mesh network thatcovers an area of approximately 60 km2 in thecity of Heraklion, Crete, Greece. Our objectiveis to use the network as a metropolitan-scaletestbed to:• Investigate the performance of a multiradio

mesh network, built from commodity com-ponents and containing 1 to 5 km links withdirectional antennas

• Evaluate channel assignment procedures forefficient wireless spectrum utilization

• Investigate medium access control (MAC)/network layer mechanisms and routing met-rics for supporting performance guaranteesin multiradio multichannel multirate meshnetworks

• Investigate innovative applications thatrequire pervasive, wide-area, and high-speed data transferThe metropolitan testbed is built from com-

modity IEEE 802.11 components, which leads tosignificantly lower costs compared to other tech-nologies, such as 802.16. Several other meshand/or long-distance 802.11 networks have beendeveloped worldwide. The 802.11b-based DigitalGangetic Plains rural area testbed contains 1–23km links [1] with directional antennas. TheWiLDNet network has 50–100 km links [2],which use a modified IEEE 802.11 MAC proto-col to operate over such very long distances. TheRoofnet network [3] considers only single-radiomesh nodes, in contrast to the multiradio nodesin our mesh testbed. The Quail Ridge wirelessmesh network [4] contains 34 mesh nodes and802.11g (2.4 GHz) links, most with distancesmaller than 1 km. Rice University has deployeda metropolitan mesh network in East-End Hous-ton with 21 mesh nodes in an area of approxi-mately 3 km2 [5], using uni- or omnidirectionalantennas and 802.11g. Another mesh networkdeployed by Waikato University in New Zealandhas 17 links with distances from 300 m to 17km,1 and has the goal of offering broadbandaccess to remote schools. All the above wirelessmesh networks operate in the 2.4 GHz band anduse point-to-point links, whereas our metropoli-tan mesh network uses 802.11a (5 GHz) andcontains a point-to-multipoint link. Moreover,an objective in this article is to use the wirelessmesh testbed to investigate the interference, rateand power adaptation, and channel assignmentfor metropolitan wireless links with distances 1.6to 5 km.

An important goal in the design of a multira-

ABSTRACT

Wireless mesh networks comprise nodes withmultiple radio interfaces, and can provide low-cost high-speed Internet access or connectivityfor data transfer. In this article we report ourexperiences and investigations with an experi-mental metropolitan multiradio mesh networkthat covers an area of approximately 60 km2 inthe city of Heraklion, Crete. We present thedesign and deployment of the network, experi-ments to quantify the network’s performance,and an application that runs on top of it andexploits it’s low-cost wide-area connectivity. Themetropolitan network consists of 16 nodes,among which six are core nodes, each with up tofour 802.11a wireless interfaces and an addition-al wireless interface for management and moni-toring. The distance between core mesh nodesvaries from 1.6 to 5 km, and the mesh networkcontains two gateways that connect it to a wirednetwork. Our performance experiments involverate, power, and channel control for long-dis-tance metropolitan links, and include investiga-tions of the timescales for the operation forthese mechanisms. Finally, we present a systemfor continuous online electromagnetic field mon-itoring and spectrum sensing, which utilizes themetropolitan mesh network for collecting wide-area measurements from low-cost EMF mea-surement devices.

TOPICS IN AD HOC AND SENSOR NETWORKS

Vasilios A. Siris, Foundation for Research and Technology — Hellas (FORTH) and

Athens University of Economics and Business

Elias Z. Tragos and Nikolaos E. Petroulakis, FORTH

Experiences with a Metropolitan Multiradio Wireless Mesh Network:Design, Performance, and Application

SIRIS LAYOUT_Layout 1 6/21/12 3:16 PM Page 128

IEEE Communications Magazine • July 2012 129

dio wireless mesh network is the efficient utiliza-tion of the limited wireless resources and theradio network infrastructure, by controlling keyparameters of a wireless communications systemsuch as the assigned channels, and the transmis-sion rate and power. Interference is a key factorthat can lead to reduced capacity and perfor-mance of wireless networks operating in unli-censed spectrum bands. Indeed, interference canexist between links belonging to the same net-work, or originate from external sources employ-ing the same or different wireless technologies.

One method of reducing the interference isto appropriately select the channels of wirelessinterfaces, which also affects the connectivity ofwireless mesh networks since two interfaces withomnidirectional antennas that are within trans-mission range of each other can communicateonly if they operate on the same channel. Theobjective of rate control is to adapt the transmis-sion rate to the channel characteristics in orderto improve performance in terms of throughputand packet transmission delay.

Throughput depends on both the transmis-sion rate and the packet loss ratio, and the maxi-mum throughput is not necessarily achieved bythe maximum transmission rate or the transmis-sion rate with the lowest packet loss ratio. Thepower control problem in wireless networks isthat of selecting the transmission power at eachradio interface in the network, in order to bal-ance energy consumption and performance. Wehave investigated the operation and performanceof the above mechanisms in the metropolitantest-bed, in addition to the time-scales in whichthey should operate.

A key question we address is whether theaforementioned mechanisms should be per-formed on a small timescale (on the order ofpackets or hours) or a much larger timescale (onthe order of days or weeks).

In this article, we also present an innovativesystem that utilizes the low-cost wide-area con-nectivity provided by the metropolitan wirelessmesh network for collecting, processing, andpresenting electromagnetic field (EMF) mea-surements. The rapid growth of wireless technol-ogy has brought to the forefront of publicinterest and concern the issue of increasingEMF radiation, especially from mobile telepho-ny systems; thus, monitoring of EMF radiation isbecoming increasingly important, as is the neces-sity to verify conformance to national and inter-national thresholds. Additionally, the system isable to collect measurements for a wide range offrequencies, and hence can be used for low-costmetropolitan area spectrum sensing in cognitiveradio networks. Such a system can create fre-quency usage maps, which can help cognitiveusers reduce spectrum sensing by avoiding fre-quencies that are already used.

The rest of the article is organized as follows.First, we describe the design and deployment ofthe metropolitan mesh network, and the inter-ference and performance monitoring. Then wediscuss experiments on channel assignment, rateand power adaptation, and investigation of thetimescales of operation for these mechanisms.Next, we describe a system for real-time collec-tion of EMF measurements and spectrum sens-

ing that utilizes the low-cost wide-area connec-tivity provided by the metropolitan mesh net-work. Finally, we conclude the article.

METROPOLITAN MULTIRADIOMESH NETWORK DESIGN

TOPOLOGYThe metropolitan mesh network covers an areaof approximately 60 km2 and currently contains16 nodes, Fig. 1, among which six are coremulti-radio mesh nodes (labeled with a leading“K” in the figure). The other ten mesh nodes(labeled with a leading “M” in Fig. 1) are man-agement nodes, that are used to remotely moni-tor the state and the performance of thenetwork and ensure its normal operation. Man-agement nodes are connected with 802.11a/glinks that operate in parallel to the experimen-tal mesh network links. Core multi-radio meshnodes are connected to the management net-work through a radio interface that is indepen-dent of the interfaces used for the experimentalmultiradio mesh network connectivity. Theabove design ensures that we can remotely andcontinuously monitor the multi-radio meshnodes. The management network is assigned astatic channel, which is different from the chan-nels used in the experimental wireless meshnetwork, and uses directional antennas that areplaced at a distance from the antennas used inthe experimental network.

The network contains both point-to-point andpoint-to-multi-point links, using directional (dishand panel) antennas with gains ranging from 19to 36 dBi. Each wireless interface is assigned astatic IP address, and the Optimized Link StateRouting (OLSR)2 protocol is used for routingtraffic in the network, whereas communicationof the management nodes use static multihoprouting. The mesh network is connected to afixed network via two nodes (M1 — FORTHand K4 — University of Crete/UoC).

Figure 1. Topology of the Heraklion metropolitan wireless mesh network.

2 http://www.olsr.org/

SIRIS LAYOUT_Layout 1 6/21/12 3:16 PM Page 129

MULTIRADIO MESH NODE

Each multiradio mesh node consists of a mini-ITX board (EPIA SP 13000, 1.3 GHz C3 CPU,512 Mbytes DDR400 memory) and a 40-Gbyte2.5 HDD. A four-slot mini PCI-to-PCI adapter(MikroTik RouterBOARD 14) holds four802.11a/g mini PCI adapters (NL-5354 MPPLUS Aries 2, Atheros-based High Power SuperA/G Dual Band 802.11a/b/g). The mini-ITX runsGentoo 2006 i686 Linux (2.6.18 kernel) with theMadWiFi driver version 0.9.2. Finally, the nodesrun OLSR daemon version 0.4.10 (by olsr.org),which implements the OLSR protocol.

One of our design requirements was to allowremote management, monitoring, and recoveryof the mesh nodes, even in situations when anode’s mini-ITX board crashes or its wirelessinterfaces are down. To address this requirementwe added to each mesh node an additional802.11a/g radio interface, which connects thenode to the management and monitoring net-work that operates in parallel to the experimen-tal mesh network. The independent managementand monitoring network has proved to be indis-pensable, allowing us to continuously monitorthe performance of the network, and conductremote experiments with channel, power, andrate control mechanisms without loosing connec-tivity to the mesh nodes. Additionally, each

mesh node contains an intelligent remote powerswitch (Dataprobe iBoot), which supports on/offpower switching through a web interface, andtimed power reboots based on the results whenthe power switch pings other devices, such as themini-ITX board or some remote device to verifythe wireless connectivity. Together with theindependent management and monitoring net-work, the remote power switch helps to providecontinuous remote connectivity to all meshnodes, enabling remote recovery from crashesand deadlocks, which are very common whenexperimenting with wireless devices.

INTERFERENCE ANDPERFORMANCE MONITORING

Interference Measurements — Most priormeasurement studies of outdoor 802.11 linksfocused on measuring the path loss and the timecorrelation of losses, and how loss is affected byfactors such as received signal strength, link dis-tance, interference, weather conditions, and tech-nology (802.11b/g) [1, 6]. Work on the impact ofadjacent channel interference for 802.11b/g is con-tained in [1, 7] and for 802.11a in [8]; the latterfocuses on measuring the impact on the signal-to-noise ratio, and considers link distances of 60 m.The unique feature of our testbed and the resultspresented in this section is that they considermetropolitan 802.11a links with longer distances(1.6–5 km), and focus on the impact that interfer-ence has on the throughput above the MAC layer.

Next we investigate the interference betweenmetropolitan links, when one of the two inter-faces of each link under investigation is locatedin the same mesh node (K2). In particular, weconsider the link pair K2–K3 and K2–K4, andthe link pair K2–K3 and K2–K1. For the firstpair, the two interfaces in node K2 are connectedto two 21 dBi panel antennas, which are both onthe same mast with a distance of approximately0.75 m, and have a relative angle of approximate-ly 150˚. For the second pair, the two interfaces innode K2 are again connected to two 21 dBi panelantennas, which, however, are on a differentmast with a distance of approximately 2.5 m, andhave a relative angle of approximately 90˚. Eachexperiment we present below shows the averagefrom 10 runs, each run lasting for 100 s. For allresults, the 95 percent confidence interval is lessthan 6 percent. Finally, the experiments involvedUDP traffic with rate 3 Mb/s, generated usingthe iperf tool. Although the maximum linkthroughput in 802.11a is higher than 3 Mb/s, ourgoal was to measure the effect of adjacent chan-nel interference, rather than the maximum linkthroughput. For this reason, we selected a lowtraffic rate to keep the CPU utilization low andavoid any impact it might have on the results.

Tx/Rx in the same node: We first considerthe interference between links when a receiveand transmit interface exists in the same meshnode. Two 3 Mb/s UDP streams are transmittedover the links K2 → K3 (2 km distance) and K4→ K2 (1.6 km distance). Note that the twostreams are independent, and the iperf senderfor the first stream is located in a workstationconnected to our internal laboratory network.Table 1 — Experiment 1 shows the throughput

IEEE Communications Magazine • July 2012130

Table 1. Throughput (Mb/s) measurements of twoflows in three different experiments.

Experiment 1 — receiver and transmitter is insame mesh node (K2) and antennas on samemast at distance ≈ 0.75 m

Channel distance TX K2→K3 RX K2→K4

0 (36–36) 2.970 2.358

1 (40–36) 2.995 2.976

2 (44–36) 2.997 2.997

Experiment 2 — receiver and transmitter is insame mesh node (K2) and on different mast atdistance ≈ 2.5 m

Channel distance TX K2→K1 RX K2→K3

0 (36–36) 3 2.75

1 (40–36) 3 3

2 (44–36) 3 3

Experiment 3 — two receivers are in same meshnode (K2), and antennas on same mast at dis-tance ≈ 0.75 m

Channel distance RX K2→K4 RX K2→K4

0 (36–36) 2.996 2.996

1 (40–36) 2.996 3

2 (44–36) 3 3

Together with the

independent man-

agement and moni-

toring network, the

remote power switch

helps to provide con-

tinuous remote con-

nectivity to all mesh

nodes, enabling

remote recovery

from crashes and

deadlocks, which are

very common when

experimenting with

wireless devices.

SIRIS LAYOUT_Layout 1 6/21/12 3:16 PM Page 130

IEEE Communications Magazine • July 2012 131

achieved by each UDP flow for three differentchannel assignments. Observe that when bothlinks are assigned the same channel (channeldistance = 0, channels 36 – 36 in Table 1, thetransmitter significantly affects the receiver (seethroughput for Rx K2 → K4), as they are bothlocated in the same mesh node (K2). When thetwo links are assigned neighboring channels (40-36), the interference is significantly reduced, butstill appears to exist. On the other hand, whenthere is a one channel separation (44–36), thereis no interference and the throughput is essen-tially equal to the UDP sending rate.

We performed the same experiment as theone described above, but with two 3 Mb/s UDPstream transmitted over the links K2 → K1 (5.1km distance) and K3 → K2 (2 km distance).Table 1 — Experiment 2 shows that the interfer-ence in this case is lower than in the previousexperiment. This is due to the larger distance(approximately 2.5 m) between the antennas fornode K2 that correspond to the above two links.

Rx/Rx in the same node: Next we investigatethe interference between links when two receiveinterfaces are located in the same mesh node.Two 3 Mb/s UDP streams are transmitted overlinks K3 → K2 and K4 → K2. In this experimentthe achieved throughput for both flows is identi-cal, and equal to the UDP sending rate (asshown in Table 1 — Experiment 3), even whenthe links are assigned the same channel. Hence,when the receive interfaces are located in thesame node, the interference between the twolinks is not significant.

The above experiments show that, dependingon the distance between antennas, there can besignificant interference between metropolitanlinks, when the transmitting and receiving ends ofthe two links are located in the same node, andeven when the two links are assigned different butadjacent 802.11a channels. The interference canbe avoided if the links are assigned channels witha one channel separation, or by placing antennasat some distance, which for our network hardwareand configuration was typically more than 2 m.

Continuous Online Performance Monitor-ing — Continuous monitoring of core mesh net-work links allows quick detection andidentification of anomalous link behavior.Indeed, as we will see in the following sections,for metropolitan links to achieve high perfor-mance under normal operation, it is sufficient tofix the transmission rate, power, and channel;the performance should be continuously moni-tored, and only when anomalies are observed therate, power, or channel needs to be adapted. Forthis reason, we have developed a set of perl andshell scripts that continuously monitor importantperformance metrics for all links between corenodes. The metrics include the signal-to-noiseratio (SNR), transmission rate, MAC and physi-cal layer errors, two-way delay, and throughput.The scripts are executed every five minutes,except the scripts for measuring the throughput,which are executed every 30 min. The collecteddata is stored in an Round Robin Database(RRD), and the corresponding daily and weeklygraphs are made available through an httpserver3 using the RRD Tool.

PERFORMANCE OF RATE, POWER,AND CHANNEL CONTROL OVER

METROPOLITAN LINKS

Next we investigate the performance of rate,power, and channel control in the metropolitanmesh network, including the time-scales for theoperation of these mechanisms. Note that althoughthe numerical results reported depend on the spe-cific network hardware and configuration of ourtest-bed, and the degree of external interference,the more general/qualitative conclusions are appli-cable to any metropolitan-scale network.

RATE CONTROLWe compare the throughput that is achievedusing a fixed transmission rate scheme, with thethroughput achieved with MadWifi’s SampleRatealgorithm, which we refer to as an auto-ratescheme, that adjusts the transmission rate on aper packet basis. The traffic was generated usingthe iperf tool. The graphs in this section reportvalues averaged over intervals of 2 min. Figure2a shows the throughput achieved on a specificlink (K2–K4), with the auto-rate scheme (hori-zontal straight lines) and fixed transmission ratefor different transmission powers. Observe that afixed transmission rate scheme can achieve high-er throughput than an auto-rate scheme if thetransmission rate is appropriately selected. More-over, the figure shows that higher improvement isachieved in the case of lower SNR values: whenthe transmission power is 15 dBm the improve-ment is approximately 30 percent, whereas whenthe transmission power is 12 dBm the improve-ment increases to approximately 42 percent.

Figure 2b shows that the throughput improve-ments of a fixed transmission rate scheme overan auto-rate scheme remain relatively the sameduring the course of a day.

The conclusions drawn from the above exper-imental results are the following:• For high-quality (high signal-to-noise ratio

[SNR]) links, fixing the rate to the highesttransmission rate achieves similar through-put as SampleRate.

• For low-quality (low SNR) links, an appro-priately selected fixed transmission rate canachieve significantly higher throughput thanSampleRate.

• Finally, a fixed transmission rate schemeshows similar performance over longtimescales (day). Hence, for a metropolitanlink with directional antennas, the transmis-sion rate does not necessarily need to beadapted over smaller timescales.The above suggests that the adaptation of the

transmission rate in small timescales (e.g., on theorder of packet arrivals), which is typically fol-lowed by all auto-rate algorithms, is not neces-sary, and can even reduce performance for longdistance metropolitan wireless links, which dueto attenuation links with low SNR would bemore commonplace.

This conclusion highlights the differencebetween long distance metropolitan links withdirectional antennas, compared to short distancelinks where throughput degradation of tradition-al rate control schemes for 802.11 networks can

3 Online access to some ofthe measurements isavailable athttp://www.ics.forth.gr/HMESH/

The interference can

be avoided if the

links are assigned

channels with a one-

channel separation,

or by placing

antennas at some

distance, which for

our network

hardware and

configuration was

typically more than

2 m.

SIRIS LAYOUT_Layout 1 6/21/12 3:16 PM Page 131

IEEE Communications Magazine • July 2012132

be solved only using sophisticated per-packetrate control mechanisms [11].

JOINT POWER AND RATE CONTROLNext we investigate power control in metropolitanwireless links. The advantage of reducing thetransmission power is the consequent reduction ofthe interference that is produced. Our objective isto identify the minimum transmission power thatachieves close to maximum performance, andinvestigate how this minimum transmission powervaries with time. Figure 3a shows the throughputfor different transmission powers for both auto-rate and fixed transmission rates. The key obser-vation from this figure is that the same highthroughput is achieved with transmission powerhigher than 8 dBm with the auto-rate scheme anda 36 Mb/s fixed transmission rate, and higher than11 dBm for a 48 Mb/s fixed transmission rate.These transmission power values are significantlylower than the maximum transmission power of15 dBm and achieve the same throughput as seenin the figure. Figure 3b shows, for different timesof day, the transmission power required to achievethroughput above 94 percent of the maximumthroughput. This figure shows that the necessaryminimum transmission power to achieve highthroughput does not change significantly through-out the day. The conclusions drawn from thepower control experiments for metropolitan linkswith directional antennas are the following:• The transmission power can be significantly

reduced, without a large impact on theachieved throughput.

• The minimum transmission power toachieve some minimum performance doesnot significantly change throughout thecourse of a day; hence, adjustment of thetransmission power can occur on longertimescales.Of course, the above results and the appro-

priate timescales for power control depend onthe level of external interference and thetimescales over which this interference changes.

CHANNEL ASSIGNMENTNext we present experiments related to channelassignment in a metropolitan mesh network withdirectional antennas. A key requirement forchannel assignment, when links operate in anunlicensed band, is to account for both internal(or intra-network) interference between linksbelonging to the mesh network, and interferencefrom sources external to the network (externalinterference). In this section we present resultsof an approach that captures both types of inter-ference, and show the influence on the overallperformance when internal or external interfer-ence is not taken into account.

One approach for capturing intra-networkinterference is the Multi-Point Link ConflictGraph (MPLCG) presented in [9]. A vertex in theMPLCG represents a multipoint communicationlink, which is a set of interfaces that communicatewith each other; all interfaces belonging to thesame multipoint link should be assigned the samechannel. An edge between two vertices in theMPLCG indicates that the two corresponding linksinterfere, and hence cannot be assigned the sameor neighboring channels; the latter is enforcedbecause interference between adjacent channelscan exist, even in the case of IEEE 802.11a. Inaddition to capturing interference, another impor-tant component of channel assignment is the actualselection of channels. Possible metrics for selectinga channel are the one-way SNR, two-way SNR(which is taken to be the average SNR at the twointerfaces belonging to the same link), or round-trip delay. These metrics can be measured online,and can capture the level of interference fromexternal sources, both 802.11 and non-802.11;other approaches to channel assignment captureonly interference between internal links, or exter-nal interference solely from 802.11 sources. Thetwo SNR metrics capture adjacent channel inter-ference, but do not capture the contention at theMAC layer between interfaces operating on thesame channel. On the other hand, the round-tripdelay metric can capture interference due to bothadjacent and co-channel interference, since it isinfluenced by MAC layer contention.

Figure 4a compares the average packet delayachieved with a channel assignment procedureusing the MPLCG for capturing intra-networkinterference and the packet delay metric for chan-nel selection [9], with the Measurement-basedDirectional Channel Assignment (M-DCA)scheme presented in [10], and channel selectionbased only on the packet delay metric and chan-nel selection that accounts only for internal inter-ference. The M-DCA scheme relies on anantenna overhearing transmissions from otherantennas inside its neighborhood, to identifywhich channels are already used by its neighbors,

Figure 2. Rate control measurements: a) throughput for auto-rate and fixedtransmission rate; link K2–K4, distance 1.6 km; b) throughput for auto-ratefor different periods of the day; link K2–K4, distance 1.6 km.

Time of day(a)

1

Thro

ughp

ut (

Mb/

s)

0

2

3

4

5

6

7

8

15:00 19:00 23:00 3:00 7:00 11:00

Transmission rate (Mb/s)(b)

2

Thro

ughp

ut (

Mb/

s)

0

4

6

8

10

12

14

16

18 24 36 48 54

24 Mb/s36 Mb/sAuto rate

15 dBm12 dBm15 dBm auto rate12 dBm auto rate

SIRIS LAYOUT_Layout 1 6/21/12 3:16 PM Page 132

IEEE Communications Magazine • July 2012 133

and avoid selecting them. Figure 4a shows thatthe M-DCA scheme achieves an average delaywhich is approximately 19 percent higher thanaverage delay achieved by the MPLCG scheme(channel assignment using the multi-link conflictgraph for interference modeling, and the delaymetric for channel selection). The above resultshows that considering only interfering 802.11networks in the channel assignment process is notenough, and that it is important to also considerthe quality of the links, in terms of the packetdelay or the SNR; this is especially important innetworks with long distance links (typicall, above1 km as in our metropolitan testbed), comparedto networks with smaller distances (as in the net-work of [10] that contains links up to 60 m).

Figure 4a also shows the average packet delaywhen only the link quality (in terms of packetdelay) is used, without modeling intranetworkinterference. Observe that the average delay is 29percent higher than that achieved when the mul-tilink conflict graph is used, and 10 percent high-er than the case when interference from 802.11networks is considered. This shows that, in addi-tion to the external interference and link quality(captured using the packet delay metric), it isimportant to consider the intranetwork interfer-ence in order to achieve high performance. Final-ly, the fourth column in Fig. 4a shows the averagedelay when only intranetwork interference istaken into account. Observe that the averagedelay is 46 percent higher than when channelselection takes into account both external inter-ference and link quality, in addition to intranet-work interference; hence, many of the schemesappearing in the literature, which focus exclusive-ly on capturing interference among links insidethe network, would yield very low performance.

Figure 4b compares, for an interval of 22days, the average packet delay when the channelsare selected once at the beginning of the 22-dayperiod, with an adaptive approach where newchannels are selected every day. The MPLCG-based procedure previously proposed in [9] wasused in both cases. The depicted delay is theaverage delay per link for all the links of themesh network. The results show that the fixedapproach achieves an average packet delay within11.5 percent of the adaptive approach. This sug-gests that, for the environment where the experi-ment was performed, there are no significantgains in performing channel assignment on atimescale smaller than 1–2 weeks. This conclu-sion depends on the fact that in our network thetraffic remained constant, and the external inter-ference remained relatively the same, since ournetwork used directional antennas and links with802.11a, which is not as widespread as 802.11b/g.

EMF MONITORING AND SPECTRUMSENSING USING A METROPOLITAN

MESH NETWORK

Metropolitan mesh networks have many poten-tial uses, such as providing wireless broadbandconnectivity in a large-scale environment fortraffic monitoring, acting as sensors, and so on.In this section we present a novel application we

have developed on top of our metropolitan wire-less mesh network, which uses mesh nodes forperforming measurements of EMF radiation.There are several potential uses for such anapplication, such as measuring the radiation atbuildings when there are multiple antennas near-by, something that is always a major concern ofcitizens, especially when the measurements aretaken near schools or hospitals. Another novelapplication of EMF monitoring is related to cog-nitive radio networks (CRNs). In CRNs, sec-ondary users have the ability to access the freespectrum holes via a function called “spectrumsensing,” which aims to avoid creating interfer-ence to licensed users. A longer spectrum sens-ing period results in less available time foraccessing the spectrum and hence lower networkthroughput. Using EMF monitoring deviceslocated at the mesh nodes in a metropolitan net-work, we can perform several measurements ofthe spectrum utilization at multiple bands, andfrom these measurements we can extract amap/database of the available spectrum holesthat can be used by cognitive users in order toavoid long periods of spectrum sensing. Depend-ing on the time and frequency granularity of themeasurements, their collection in some centralserver can be demanding in terms of boththroughput and delay.

The key idea of the monitoring system is touse a low-cost EMF measurement device locatedin a node connected to the metropolitan meshnetwork. The EMF monitoring node consists ofthe following: a low-cost EMF measurementdevice, a mini PC for controlling the EMF mea-

Figure 3. Joint power and rate control measurements: a) Throughput for differ-ent transmission powers. Link K2-K3, distance 2 km; b) Transmission powerfor achieving throughput above 94 percent of the maximum. Link K2-K3, dis-tance 2 km.

Transmission power (dBm)

(a)

1

10

Thro

ughp

ut (

Mb/

s)

5

0

15

20

25

2 3 4 5 6 7 8 9 10 11 12 13 14 15

Time of day

(b)

10:00 11:00Tr

ansm

issi

on p

ower

(dB

m)

11

10

12

13

12:00 13:00 14:00 15:00 16:00 17:00 18:00 19:00 20:00

48 Mb/s36 Mb/sAuto rate

94% of max

SIRIS LAYOUT_Layout 1 6/21/12 3:16 PM Page 133

IEEE Communications Magazine • July 2012134

surement device, and software modules for col-lecting, processing, and presenting the measure-ments.

The EMF measurement device is a low-costAaronia Spectran Analyzer. Our current systemuses the HF 6060, which has a measurementrange of 10–6000 MHz. Additional monitoringnodes we are currently deploying will use the HF6080, which has a range of 10–7000 MHz. TheHF60X0 spectrum analyzer can be connected totwo types of antennas: a directional antenna(HyperLOG 6080) and an omnidirectional dipoleantenna (BicoLOG 20300). The HyperLOG6080 directional antenna has a range of 700MHz–7 GHz. The BicoLOG 20300 omnidirec-tional dipole antenna has a range of 40–3000MHz. The BicoLOG antenna is more costly thanthe HyperLOG antenna, but can measure lowerfrequencies, which include the FM and TVbands. On the other hand, the lower-cost Hyper-LOG directional antenna can measure higherfrequencies, up to 7 GHz, but requires pointingit toward the area we are interested in measur-ing. The currently deployed EMF monitoringnode includes the omnidirectional dipole anten-na BicoLOG 20300.

The EMF monitoring node contains a smallPC, which is based on a mini-ITX board (EPIASP 13000, 1.3 GHz C3, CPU) with 512 MbytesDDR400 memory, and an 80 GB 2.5” HDD.The mini-ITX PC runs the Aaronia software forcontrolling and collecting measurements fromthe HF60X0 analyzer, which also allows externalconfiguration of the spectrum analyzer forparameters that include frequency range, sampletime, and resolution bandwidth. An Apacheserver running on the mini-ITX allows remoteaccess of the measurement graphs through a webinterface. There are four Perl scripts executed

every 5 min, which take the data from theHF60X0 analyzer and create graphs presentingthe EMF measurements in a different manner—per band, per operator, and time series — andcan also be executed on demand through a webinterface.

ADVANTAGESThe advantages of using an EMF measurementdevice connected to a metropolitan mesh net-work for collecting EMF measurements includethe following:• Higher range: An EMF measurement device

can be used to monitor frequencies up to 7GHz (with the appropriate antenna), whichis higher than the capabilities from special-ized standalone EMF monitors, whoserange is typically limited to 3 GHz.

• Real-time remote measurement collection:EMF measurement devices with real-timemonitoring capabilities together with ametropolitan coverage mesh network allowreal-time remote collection of EMF mea-surement data.

• Low cost and high speed: Small (handheld)EMF measurement devices with advancedspectrum analyzer capabilities are signifi-cantly cheaper than standalone EMF moni-toring devices with remote communication(Global System for Mobile Communica-tions [GSM], third generation [3G], etc.)capabilities, and wireless mesh networksprovide higher speeds at lower costs com-pared to mobile technologies.

• Advanced flexibility: Together with the miniPC, the EMF monitor can be controlledremotely to collect measurements in differ-ent frequency ranges (bands) and differenttime windows.

MONITORING CAPABILITIESThe collected measurement data is in dBm units.In addition to the default units, the measure-ments can be presented in dBmV, mV/m,mA/m2, and finally mW/m2. We note here thatelectromagnetic radiation limits are usually rep-resented in mV/m or mW/m2. Figure 5 showsexamples of EMF per-band and spectrum sens-ing measurements.

Per-band and Per-Operator Monitoring —This option allows the presentation of EMF lev-els in frequencies of various well-known bands.Figure 5a shows an example of per-band moni-toring. The displayed values are the average ofmeasurements taken at time intervals of approxi-mately 6 min. The graph is refreshed periodical-ly, based on the last stored values in thedatabase. Each column corresponds to a differ-ent band. Note that the values for the GSM and3G/Universal Mobile Telecommunications Sys-tem (UMTS) bands are higher than approxi-mately –80 dBm, which is due to the mobiletelephony antennas located opposite the EMFmonitor node. In addition to per-band monitor-ing, our system also supports EMF measure-ments, per-operator frequencies, and time seriesmeasurements, which allows the daily, weekly,monthly, and yearly presentation of EMF mea-surements.

Figure 4. Channel assignment measurements: a) comparison of channelassignment algorithms; b) comparison of performance when channels areassigned once at the beginning of a 22-day period with performance whenchannels are adjusted each day.

(a)

(b)

0.2

0.0

Day1

0.2Pack

et d

elay

(m

s)

Pack

et d

elay

(m

s)

0

0.4

0.6

0.8

1.0

1.2

0.4

0.6

0.8

1

1.2

3 5 7 9 11 13 15 17 19 21

Same channels for all daysDifferent channels each day

0.770.90 0.97

1.1

MPLCG + pkt delay metricM-DCAPkt delay metric onlyOnly internal interference (MPLCG)

SIRIS LAYOUT_Layout 1 6/21/12 3:16 PM Page 134

IEEE Communications Magazine • July 2012 135

Spectrum Sensing — This option displays, inaddition to the aggregate transmission power indifferent bands, the spectrum holes as a functionof time. Specifically, the top of Figs. 5b and 5cshows the received aggregate power in differentfrequencies in the 2.4 GHz and TV radio bands

at a specific time. The bottom part of Figs. 5band 5c show the received power in the frequen-cies as a function of time, where time is depictedon the vertical axis. As expected, in the TV radioband only a small portion of the spectrum isbeing utilized, whereas in the 2.4 GHz band

Figure 5. Presentation of EMF and spectrum sensing measurements: a) per band EMF monitoring; b) spec-trum sensing in the 2.4 GHz band; c) spectrum sensing in the TV radio band.

(a)

(b)

(c)

FM

Radio TV GSM 900 GSM 1800 UMTS UMTS-FDD WiFi

-100

-120

dBm

-80

-60

-40

-20

0

VHF UHF UL DL UL

Per band monitoring from 2010-04-13 14:31:14 - 2010-04-13 16:07:53

DL TDD UL DL 802.11b/g

The EMF monitoring

node consists of the

following: a low-cost

EMF measurement

device, a mini PC for

controlling the EMF

measurement device,

and software

modules for collect-

ing, processing, and

presenting the

measurements.

SIRIS LAYOUT_Layout 1 6/21/12 3:16 PM Page 135

IEEE Communications Magazine • July 2012136

almost all the channels of the (unlicensed) spec-trum are utilized most of the time. This showsthe necessity of using cognitive radios in meshnetworks, as they are capable of accessing thefree spectrum and achieving much higher perfor-mance than operation in the crowded WiFibands. From these figures, we can observe theutilization of the spectrum in a long time periodand extract the spectrum white spaces in eacharea. This information can be used to createdatabases of white spaces in order to assist cog-nitive radio devices in their spectrum sensingfunction. By installing the EMF devices in themesh nodes of our metropolitan mesh network,we can extract white space databases for a verylarge geographical area. The cognitive deviceswill consult these databases in order to avoidsensing the overcrowded frequencies and focusonly on the long-term unused frequencies. Thiswill help minimizing the time dedicated to spec-trum sensing and, as a result, maximize the timefor accessing the spectrum, which leads toincreased network performance.

CONCLUSIONSWe present some of our experiences and investi-gations with an experimental metropolitan multi-radio mesh network. In particular, we investigatehow interference between metropolitan 802.11alinks operating on adjacent channels affectsthroughput. For rate control, our results showthat the adaptation of the transmission rate onvery small timescales (on a per packet basis as iscommonly the case with widely used auto-ratealgorithms) not only does not improve perfor-mance, but can also result in significantly lowerperformance compared to an appropriately select-ed fixed transmission rate scheme. For channelassignment, we show results demonstrating theimportance of taking into account both internaland external interference. Our experiments forthe timescales of rate, power, and channel controlin metropolitan links with directional antennasshow that there are no significant improvementsfor performing adaptation on short timescales onthe order of minutes or hours. Hence, a moreappropriate approach would be to select a staticrate, power, and channel, and continuously moni-tor the network to detect any anomalies or per-formance drops; once such events are detected,the rate, power, and channel can be adjusted.Finally, we present a system for continuous onlineEMF monitoring and spectrum sensing, whichutilizes the low-cost wide-area connectivity pro-vided by the metropolitan mesh network for col-lecting, processing, and presenting data fromlow-cost EMF measurement devices.

ACKNOWLEDGMENTThis work was supported in part by the Euro-pean Commission in the 7th Framework Pro-gram through project Enhanced, Ubiquitous,and Dependable Broadband Access using MESHNetworks (EU-MESH), ICT-215320, http://www.eu-mesh.eu. The authors would like to thankVangelis Angelakis, Stefanos Papadakis, KostasMathioudakis, and Manolis Delakis for con-tributing to the design and deployment of themetropolitan wireless mesh network.

REFERENCES[1] K. Chebrolu, B. Raman, and S. Sen, “Long-Distance

802.11b Links: Performance Measurements and Experi-ence,” Proc. ACM MOBICOM, 2006.

[2] R. Patra et al., “WiLDNet: Design and Implementationof High Performance WiFi Based Long Distance Net-works,” Proc. USENIX Symp. Networked Sys. Designand Implementation, 2006.

[3] K. N. Ramachandran et al., “Interference-Aware ChannelAssignment in Wulti-Radio Wireless Mesh Networks,”Proc. IEEE INFOCOM, 2006.

[4] J. Bicket et al., “Architecture and Evaluation of anUnplanned 802.11b Mesh Network,” Proc. ACM MOBI-COM, 2005.

[5] D. Wu, D. Gupta, and P. Mohapatra, “Quail RidgeReserve Wireless Mesh Network: Experiences, Challengesand Findings,” Proc. IEEE TRIDENTCOM, 2007.

[6] J. Camp et al.,” Measurement Driven Deployment of aTwo-Tier Urban Mesh Access Network,” Proc. ACMMobiSys, 2006.

[7] G. Bianchi, F. Formisano, and D. Giustiniano, “802.11b/gLink Level Measurements for an Outdoor Wireless Cam-pus Network,” Proc. IEEE WoWMoM, 2006.

[8] T. Ireland et al., “The Impact of Directional AntennaOrientation, Spacing, and Channel Separation on Long-distance Multi-hop 802.11g Networks: A MeasurementStudy,” Proc. 3rd Int’l. Wksp. Wireless Network Mea-surement, 2007.

[9] C.-M. Cheng et al., “Adjacent Channel Interference inDual-Radio 802.11a Nodes and Its Impact on Multi-hopNetworking,” Proc. IEEE GLOBECOM, 2006.

[10] M. Delakis and V. A. Siris, “Channel Assignment in aMetropolitan Wireless Multi-Radio Mesh Network,”Proc. 5th Int’l. Conf. Broadband Commun., Networksand Sys., London, Sept. 2008.

[11] S. M. Das et al., “DMesh: Incorporating Practical Direc-tional Antennas in Multichannel Wireless Mesh Net-works,” IEEE JSAC, vol. 24, no. 11, 2006, pp. 2028–39.

[12] E. Ancillotti, R. Bruno, and M. Conti, “Design and Per-formance Evaluation of Throughput-Aware Rate Adap-tation Protocols for IEEE 802.11 Wireless Networks,”Perf. Eval., vol. 66, issue 12, Dec. 2009, pp. 811–25.

BIOGRAPHIESVASILIOS A. SIRIS [M’98] (vsiris@aueb.gr) is an assistant pro-fessor at the Department of Informatics of Athens Universityof Economics and Business, and a research associate at theInstitute of Computer Science of FORTH. He received adegree in physics (1990) from the National and KapodistrianUniversity of Athens, Greece, his M.S. (1992) in computerscience from Northeastern University, Boston, Massachusetts,and his Ph.D. (1998) in computer science from the Universityof Crete, Greece. In Spring 2001 he was a visiting researcherat the Statistical Laboratory of the University of Cambridge,and in Summer 2001 and 2006 he was a research fellow atthe research laboratories of British Telecommunications (BT),UK. His research interests include resource management inwired and wireless networks, future Internet architectures,and traffic measurement and analysis.

ELIAS Z. TRAGOS (etragos@ics.forth.gr) is a research associ-ate in the Telecommunications and Networks Laboratory ofthe Institute of Computer Science at FORTH. He receivedhis diploma in electrical and computer engineering (2003),his M.B.A. in technoeconomics (2006), and his Ph.D. (2008)from the School of Electrical and Computer Engineering ofthe National Technical University of Athens, Greece. Hisresearch interests are in the area of mobile and wirelessnetworks, mesh and ad hoc networks, radio resource man-agement, mobility and policy based management, P2P net-works, and cognitive networks. He is a member of theTechnical Chamber of Greece.

NIKOLAOS E. PETROULAKIS [M’09] (npetro@ics.forth.gr) hasbeen a research scientist in the Telecommunications andNetworks Laboratory of the Institute of Computer Scienceat FORTH since 2007. He received his degree in mathemat-ics (2004) from the National and Kapodistrian University ofAthens, Greece, and his M.Sc. (2006) in digital communica-tions from the Department of Engineering and Design ofthe University of Sussex, United Kingdom (2006). He is alsoa part-time member of the Computer Emergency ResponseTeam of FORTH (FORTHcert) since 2009. His research inter-ests include wireless networks, cognitive radios, and net-work security.

A more appropriate

approach would be

to select a static rate,

power, and channel,

and continuously

monitor the network

to detect any

anomalies or

performance drops;

once such events are

detected, the rate,

power, and channel

can be adjusted.

SIRIS LAYOUT_Layout 1 6/21/12 3:16 PM Page 136