Available online at www.sciencedirect.com

www.elsevier.com/locate/comcom

Computer Communications 31 (2008) 1564–1578

Can the current generation of wireless mesh networkscompete with cellular voice?

Andres Arjona a,*, Cedric Westphal b, Jukka Manner c, Antti Yla-Jaaski c, Sami Takala a

a Nokia Siemens Networks, Linnoitustie 6, 02600 Espoo, Finlandb DoCoMo USA Labs, 3240 Hillview ave., Palo Alto, CA94304, USA

c Helsinki University of Technology, P.O. Box 5400, FIN-02015 HUT, Finland

Available online 31 January 2008

Abstract

Wireless mesh networks are being deployed to provide broadband wireless connectivity to city-wide hotspots. The typical architecturein these deployments thus far is a single-radio architecture: mesh nodes carry only one radio, which is used both to receive the traffic fromthe clients and to relay this traffic through the mesh to the wired Internet gateway.In this paper, we study the performance of a repre-sentative single-radio mesh network both in a live setup and in a laboratory environment. We characterize the performance of differentapplications (e.g. VoIP), and study some key challenges of mesh networks such as the fairness in bandwidth allocation and hidden nodeterminal. Finally, we compare the results of the study with traditional cellular networks, and discuss various options to enhance the per-formance of wireless mesh networks in the future.� 2008 Elsevier B.V. All rights reserved.

Keywords: Wireless mesh; WiFi; VoIP; Performance; Multihop; Measurements

1. Introduction

Many cities in the US, and quite a few outside, are eithercommitting to, or studying the possibility of deploying acity-wide WiFi coverage using wireless mesh networks.San Francisco, New York, Philadelphia are in the planningstages, while smaller cities have already deployed networkswhich attempt to provide broadband wireless accessubiquitously. The goals in setting up these wireless meshnetworks are multiple, and include providing broadbandaccess to underserved communities or supporting emer-gency services. However, one of the main reasons is toreduce the cost per bit of the wireless access, in order to sup-port applications which reduce the expenditures of a city.

The hope is that a lower cost per bit would provide theincentive to use applications on the go, thereby increasingthe productivity of city employees. Alternatively, if the net-

0140-3664/$ - see front matter � 2008 Elsevier B.V. All rights reserved.

doi:10.1016/j.comcom.2008.01.043

* Corresponding author. Tel.: +358 50 487 2500.E-mail address: andres.arjona@nsn.com (A. Arjona).

work is operated by a provider, the lower cost per bitwould provide the margin to compete for mobile applica-tions with cellular operators. Metropolitan wireless meshnetworks are seen by some investors as a potential disrup-tive technology for legacy cellular operators. A wirelessmesh network combined with a VoIP handheld devicecould become an alternative to the cellular handset.

Many vendors are proposing infrastructure productsand solutions to provide wireless broadband connectivityover extended areas [9,38,39,43]. It is however, a testamentto the uncertainty of the business model to see such a bal-kanized market with no clear technology consensus. Themarket is pulling in different technological directions, hop-ing that one will prove profitable.

The current generation of mesh networks is based on asingle-radio architecture. This architecture is the only tech-nology that can provide the required coverage at the cur-rent budget envelope. The market leader in volume forthis architecture is Tropos Networks. In this paper, westudy the performance of such single-radio mesh networks

A. Arjona et al. / Computer Communications 31 (2008) 1564–1578 1565

and compare the cost benefit of using such architecturewith that of a cellular network.

We study the VoIP performance in an actual, fullydeployed network, and show that, at the price point cur-rently supported by the business model rules of thumb, asingle-radio wireless mesh network is unable to provide sat-isfying VoIP service, let alone compete with a cellularinfrastructure.

Subsequently, we study in the lab the performance of amesh network of a single-radio platform against some keyperformance metrics. We will show that issues of unfairnessarise quickly, and that, even in the lab, the performanceproves disappointing.

Finally, we compare and evaluate the deployment feasi-bility of a single-radio mesh solution against other cellulartechnologies in regards to data and voice capacity, cover-age, and cost of deployment. Our evaluation shows thatthis generation of single-radio mesh networks cannot cur-rently rival with cellular networks.

The remainder of the paper is organized as follows. Sec-tion 2 introduces the single-radio WiFi mesh architecture.Section 3 describes our research approach. Section 4 pre-sents the results and analysis and in Section 5 we comparemesh networks to the current cellular offerings. In Section 6we describe the related work, and finally, in Section 7 wedraw conclusions.

2. Deployed single-radio WiFi mesh architectures

The wireless mesh network is composed of three basicelements: an 802.11 WiFi client, an access point (AP) and

Fig. 1. WiFi mesh in an

mesh routers. The users connect to the network AP withthe 802.11 WiFi client, such as a laptop, PDA or a VoIPphone. The mesh routers forward the traffic over poten-tially multiple wireless hops in between the AP and thewired Internet gateway. The mesh routers form the back-haul of the mesh network (see Fig. 1).

The most common commercial solutions deploy a sin-gle-radio mesh node, where both AP and backhaul func-tionality are merged on the same platform. The single-radio architecture uses the same-radio frequency for wire-less access and backhaul. The same node functions both asan access point to the end-user clients and as a relay forthe traffic to the gateway. Single-radio architectures werechosen mostly due to their lower price point [47], and totheir implementation and deployment simplicity in anemerging market.

Many metropolitan wireless mesh deployments are cur-rently based on this architecture, usually with Tropos hard-ware, such as those from Earthlink in Philadelphia orGoogle in Mountain View [24]. The single-radio architec-ture is substantially cheaper than dual or multi-radio archi-tectures. On the other hand, multi-radio architectures offerbetter performance, as the backhaul traffic typically is in adifferent frequency band as the access traffic, and does notinterfere with it. For a performance benchmark of multi-radio mesh equipment, please refer to [28].

In a mesh network the nodes might be several hopsaway from the wired gateway, as depicted in Fig. 1, andthe nodes might need to use other nodes as relays. There-fore, not only does the backhaul traffic interfere with theaccess traffic, it also interferes with itself. As more mesh

outdoor scenario.

1566 A. Arjona et al. / Computer Communications 31 (2008) 1564–1578

access points are added, a higher percentage of the wire-less bandwidth is dedicated to packet forwarding. Thus,less capacity is available for the users’ traffic [10].

As an example of a live state-of-the-art single-radioWiFi Mesh network, we consider the Google WiFi Net-work in Mountain View. The network has been deployedto serve the community for free, but aside from the sub-scription model, it is expected to be similar to the futurePhiladelphia WiFi mesh networks. The main reasons forchoosing this network are the following: (a) it was astate-of-the-art deployment at the time of our measure-ments; it has operated for the community since August2006; (b) Tropos, which provided the access points, isone of the market leaders and provides equipment for mostof the high profile networks.

Google’s Muni WiFi mesh deployment cost is estimatedat roughly one million US dollars [30]. It consists of a net-work providing coverage to the city of Mountain View, inwhich Google’s headquarters are located (see Fig. 2). Thenetwork is free of cost and has the objective to be a proofof concept for Muni WiFi mesh deployments. The popula-tion in Mountain View is about 72,000 inhabitants. Thenetwork provides coverage to about 12 square miles (about30 sq. km) and consisted of the following elements [24] atthe time of our study:

Fig. 2. Google’s WiF

� 380 Tropos access points mounted on street lamps� 1 Alvarion gateway per every six access points� 3 bandwidth aggregation points connected to Google’s

headquarters using GigaBeam equipment.

The network is based on a single-radio architecture inthe 2.4 GHz band. The deployment consists of small clus-ters of 6–8 access points connected to a gateway, fromwhich most nodes are one hop away, i.e., roughly five gate-ways per square mile. The gateways are connected to anaggregation point using a point-to-point Gigabit wirelesslinks. Such aggregation points support about 20 gateways.The set of access points connected to the same gatewayoperate in the same frequency channel. Additionally, thebandwidth is restricted to 1 Mbps for both downlink anduplink for each user.

3. Methodology and test environments

In this section we describe the setup and methodology ofour experiments. We made two studies; in the first we eval-uated the performance of an existing wireless mesh net-work in situ, and in the second we analyzed theperformance of single-radio mesh technologies in a labora-tory setup.

i mesh network.

Fig. 3. VoIP testing configuration.

A. Arjona et al. / Computer Communications 31 (2008) 1564–1578 1567

3.1. Case study: evaluation of a live WiFi mesh network in

Mountain View

In the first case study we performed an empirical perfor-mance analysis of the Google’s Muni WiFi mesh networkin Mountain View [6]. The main goal of the tests was todetermine the performance level that can be achieved indifferent common VoIP calls (e.g. Skype) and data applica-tions. VoIP calls are a especially relevant in order to assesswhether WiFi mesh networks are a serious alternative tocellular networks. All the scenarios and tests were carriedout in outdoor conditions. The performance indoors isexpected to be worst than indoors very poor due to signalattenuation and, since we will show a rather poor perfor-mance outdoors in the first place, the indoor case was omit-ted in this study.

The Google’s mesh network quality at a spatial locationwas measured based on the signal-to-noise ratio (SNR)which gives a good indication of the expected performance.The SNR is measured based on the noise floor and thereceived signal strength from the access point (AP) to themobile station (MS). The measured SNR is a relativelygood estimation of the WiFi link condition. Naturally thisis not an absolute measure since there is a differencebetween uplink and downlink transmission power fromthe MS (�15–17 dBm) and AP (�20 dBm) which can causeunbalance in the bandwidth.

The transmission rate between the AP and the MS isdetermined by the MS receiver sensitivity. Based on thetransmission rate, two scenarios are depicted:

� Excellent signal conditions: SNR > 25 dB, providingtransmission rates up to 54 Mbps� Medium signal conditions: SNR 18–24 dB; providing

transmission rates up to 36 Mbps

Signal conditions with SNR lower than 18 dBm are con-sidered poor and due to packet loss and constant retrans-missions they cannot support VoIP services. However, itmight still be possible to achieve connectivity with theAP. The minimum SNR required in order to associate withan AP is roughly 12 dB.

3.1.1. Voice over IP

Voice over IP was measured objectively based onPESQ Mean Opinion Score (MOS) as depicted by Mal-den Digital Speech Level Analyzer [29]. The tests werecarried in static locations selected based on SNR (excel-lent and medium). Fig. 3 shows the VoIP testing config-uration.

The testing methodology consisted of originating callswith Skype software to (a) another Skype user and (b) acellular subscriber. Multiple iterations were carried outwith different voice samples. In addition, the unlicensedspectrum was monitored with a spectrum analyzer [17]to look for fluctuations in signal conditions during thetests.

3.1.2. Data performance

Data performance was measured based on PING roundtrip time and using FTP to download and upload 10 MBfiles. The tests were carried out in static locations selectedbased on SNR levels (excellent and medium).

3.1.3. Network coverage

Wireless links perform better when there is clear line ofsight between the nodes. Any physical obstacles like build-ings, cars, trees, etc., have impact on the link performance.The coverage survey of the Mountain View network wasperformed in three different environments within the 12square mile network.

� Residential area: Locations with dense foliage, manyhomes one next to each other, cars parked in front ofhomes, few people and little traffic. The streets are fairlynarrow and most homes are single story housing.� Corporate campuses area: Locations consisting mainly

of enterprise buildings, wide streets, parking lots, regu-lar traffic, few trees and people.� Downtown area: Location in the downtown area, con-

stant traffic and pedestrians, mixture of moderate trees,bushes and two story commerce buildings.

Mountain View does not contain any urban canyon withtall high rise buildings; therefore, we could not study ultra-dense urban environments. Based on the above character-izations, several surveys were carried out in the three sce-narios and coverage maps were generated withAirMagnet Surveyor tool [2]. AirMagnet interpolates the

1568 A. Arjona et al. / Computer Communications 31 (2008) 1564–1578

discrete measurements along the path of the survey torecreate the SNR in a coverage area.

3.2. Single-radio mesh evaluation in laboratory conditions

In the second case study we analyzed the performance ofa mesh network in a laboratory environment with state-of-the-art single-radio technology [7]. In the laboratory setup,we were interested in assessing the upper bound on themesh network capacity using single-radio nodes. As such,we make three assumptions:

Assumption 1: line topology. Actual mesh networks fol-low tree topologies, where the different branches of the treeinteract with each other. Considering a line networkreduces the amount of interference by eliminating otherbranches potentially competing for radio resources, andthus gives us an upper bound on performance vs. the num-ber of hops.

Assumption 2: interference limited only to neighbors.

Again, different topologies and node placement will havedifferent results in terms of achieving a transmission sche-dule. However, removing interference from nodes morethan one hop away gives an easily reproducible upperbound.

Assumption 3: network with size limited to three hops.

While this could seem small, it is actually the deploymentrule followed in real life networks: nodes are never furtherthan three hops away from a gateway. Typically, the gate-way is not a wired gateway, but a point-to-multipoint linkwith a higher capacity, in an orthogonal frequency band,such as an Alvarion [4] or Motorola Canopy [32] link.

Under these assumptions, only one wireless link can beup at any given time, and consequently, the expected

Fig. 4. Single-radio mesh

throughput of a single-radio mesh network should coincidewith its theoretic behavior and be proportional to theinverse of the number of hops. Assumptions 1–3 enablereproducible measurements with results that are still usefulperformance measures of the mesh network.

In our laboratory measurements we evaluate a single-radio node from the latest generation, the Meraki mini

node [31]. The Meraki mini exists in indoor and outdoorconfigurations, and was introduced commercially inAugust 2006 as a fully functional beta version. The testingtook place in November 2006, at which point the firmwarehad been upgraded to the latest version. Meraki is a start-up founded by students from the Roofnet project [35] atMIT, and the product incorporates the state-of-the-artresult of this research group. It should be noted that theexact numerical measurement results are always dependenton the products themselves since there are variationsbetween the different implementations. For instance, theuse of proprietary Medium Access Control (MAC) mecha-nisms in the backhaul, and specialized routing algorithmscan result in differences in performance [40]. However,the given results are indicative for single-radio mesh net-works in general.

The evaluation methodology consists of a variety of testcases with multiple combinations of number of hops, num-ber of clients and network traffic. With such variables weare able to evaluate the single-radio solution from the fol-lowing test objectives:

� Trade-off between client bandwidth and number of hopsfrom the gateway.� Trade-off between bandwidth share for clients near the

gateway over clients far from the gateway.

testing configuration.

Table 1Objective voice quality tests summary

High SNR Medium SNR

Signal-to-noise ratio (SNR) 29 20Noise �92 �89Signal strength �63 �69Transmit rate 54 36Receive rate 54 24Channel 3 9Skype – Skype MOS 3.3 2.8Skype – Mobile MOS 2.5 2.4

A. Arjona et al. / Computer Communications 31 (2008) 1564–1578 1569

� Effect and capacity for simultaneous clients connected todifferent network nodes.� Effect of rate limiting the users at 1 Mbps, which is a

typical value for subscriber service in commercial meshnetworks.� Hidden node problem effect in single-radio equipment.� VoIP quality with multiple clients and number of hops

from the gateway.

The testbed configuration consisted of single-radio Mer-aki mesh access points located individually in RF enclo-sures. The access points had default configuration andmaximum modulation enabled. Each of the RF enclosureswas connected to each other in such a way that multiplehops are possible. No attenuators were placed betweenthe access points in order to allow the backhaul to operateat a maximum rate. One of the access points at the edge ofthe wireless network was connected to a switch in the wirednetwork, where the management consoles and a web serverwere located (see Fig. 4). Laptops were connected directlyto the RF enclosure that contained the respective accesspoint to which the client was to associate. The number oflaptop clients varied between one and three. The RF enclo-sures assured that the laptop clients were only able toreceive a high WiFi signal from the access point locatedin the same RF enclosure. The signal-to-noise ratio foreach of the laptop clients in the enclosures was largeenough to support the highest transmission rates(54 Mbps).

The tests executed consisted of multiple iterations of testcases for each of the mentioned test objectives. Throughputwas measured by means of 25 MB FTP file downloads anduploads. Delay was characterized with PING. HTTPbrowsing was measured as download times for webpageswith different sizes and layouts located in a web server.Background traffic was generated with continuous FTP filedownloads and uploads. Finally VoIP quality and capacityis measured with IxChariot testing software. Chariot’s

Fig. 5. Objective voice qu

mean opinion score values are based on ITU-T E-Model[26,25,18].

4. Results

4.1. Live WiFi mesh network results

In the first case study we measured the performance ofthe Google’s Muni WiFi mesh network in Mountain View[6].

4.1.1. Voice over IP performance

The performance of VoIP services is important sincemunicipal mesh networks intend to compete with cellularnetworks. The subjective voice over IP quality tests resultsare shown in Fig. 5 and Tables 1 and 2.

The results show that good voice quality is possible inthe case of Skype to Skype calls. However, the quality isvery poor or literally at an unusable level in the case of callsmade to cellular phones in the mobile network (see Fig. 5and Table 1). A possible reason for the reduced quality isdue to transcoding of incoming Skype VoIP packets intocircuit switched based voice audio. Likewise due to the nat-ure of the unlicensed spectrum, signal conditions might notbe constant and decrease from time to time due to interfer-ers. That is a possibility for the drops in voice quality dur-

ality (PESQ MOS).

Table 2Interferers in the 2.4 GHz band during voice quality tests with high SNR

Interferer Avg. pulseduration (ms)

Avg. power(dBm)

Channelsaffected

DECT BS 0.317 �60 1–14DECT BS 0.282 �85 1–14DECT BS 0.369 �72 1–14DECT BS 0.256 �64 1–14DECT BS 0.251 �82 NoneDECT BS 0.217 �74 1–14DECT handset 0.839 �83 1–7; 14DECT BS 0.274 �86 1–2; 5–14DECT BS 0.245 �82 NoneDECT BS 0.299 �77 1–14DECT BS 0.244 �81 1–14

1570 A. Arjona et al. / Computer Communications 31 (2008) 1564–1578

ing the tests in high SNR where interference from multipledevices was constant (see Table 2). DECT devices operat-ing in the 2.4 GHz band are widely implemented and avail-able in the United States and therefore a commonadditional source of interference [20].

4.1.2. Data performance

The data performance results are summarized in Table3. The measurements show that delay (round trip time) isin average short and does not fluctuate too much. Also,

Table 3Data performance

High SNR Medium SNR

Signal-to-noise ratio (SNR) 29 20Noise �92 �89Signal strength �63 �69Transmit rate 54 36Receive rate 54 24Channel 3 9FTP DL (kbps) 912 944FTP UL (kbps) 725 691

PING (ms) Min Max Avg. Min Max Avg.

32 bytes 19 41 25 23 52 33256 bytes 19 75 26 22 194 351460 bytes 23 138 35 30 322 53

Fig. 6. Coverage maps f

the average throughput for downloads is fairly stable andclose to the 1 Mbps limit imposed by Google. However,the uplink performance is roughly 30% lower. This is dueto the different multiple access contention mechanism; inthe downlink, the access point schedules packets with acentralized medium access, whereas on the uplink, the cli-ents compete with each other for the channel.

4.1.3. Network coverage

Fig. 6 show the network coverage under the three differ-ent environment scenarios, residential, corporate cam-puses, and downtown areas. The islands of yellow depictthe areas were the signal is good.

The measurement results from the residential area showthat SNR levels between 15 and 25 dB are available. How-ever, coverage with excellent signal conditions(SNR > 25 dB) is only available in very few areas. Someof the reasons for this are the distance between each accesspoint and also the dense foliage from the trees in the side-walks. The access points do not seem to be in optimal posi-tion and in some occasions the tree branches could directlyobstruct the access points.

The measurement results from the corporate campusesarea show that only some areas with excellent signal condi-tions (SNR > 25 dB) are available. This difference, whencompared to the residential area, is likely to be due tothe wider streets, less trees in the area and parking lots thatprovide the opportunity to create a line of sight betweenthe mobile station and the access point.

The measurement results from the downtown area showthat with an increased number of access points (comparedto the previous locations), and the lack of trees and foliage,the coverage in the area is much better. Likewise, areaswith excellent signal conditions (SNR > 25 dB) are avail-able in most of the downtown area.

4.2. Single-radio mesh in laboratory results

In the second case study we measured a mesh network ina laboratory environment with single-radio technology. Ina single-radio mesh network the same radio is used for the

or tested locations.

A. Arjona et al. / Computer Communications 31 (2008) 1564–1578 1571

access and wireless backhaul. This means that the samescarce radio resource is used both for the user access trafficand packet forwarding traffic between mesh nodes resultingin complex contention situations for the channel access.The larger the hop count to the Internet gateway, the lowerthe available user traffic capacity [7].

4.2.1. Available bandwidth vs. number of hops

The performance tests carried out show that the maxi-mum available bandwidth per cell is determined at thegateway; 5.2 Mbps and 4 Mbps for downlink and uplink,respectively, when there is no background traffic. Regard-less of supporting 54 Mbps transmission rates, the Merakinodes limit client bandwidth to 5 Mbps because they arealso mesh nodes and dedicate a fraction of their time listen-ing to potential traffic from other mesh routers. The samethroughput is achieved regardless of whether there wasonly one client associated to the access point or two. Themaximum available bandwidth is shared among the twoclients associated to the access point.

We first measure how the channel capacity is divided fortwo clients and how the total user traffic drops with respectto the number of hops. Fig. 7 shows that the bandwidthshare between the users is not equal. One of the users isable to get a considerably higher throughput, especiallywhen both users are associated to the gateway or the firstrelay. In contrast, Fig. 8 shows that the share for the uplinkdirection is almost equal, yet the maximum throughputachievable is lower than for downlink.

Fig. 7. Bandwidth vs. number of hops [DL].

Fig. 8. Bandwidth vs. number of hops [UL].

The capacity decrease in our test bed approached theexpected (1/number of hops) rate. However, we must notethat in our test environment, access points and clients areisolated via the use of RF enclosures (see Fig. 4). There-fore, access points can only hear adjacent neighbors. In areal outdoor configuration, the gateway node is likely tobe located in a middle location to reduce the number ofhops to its neighbor nodes (see Fig. 9). As a result, the gate-way will probably hear more than one or two relay nodes.Since all clients and access points must operate in the samefrequency channel, and contend for access based on the802.11 standard, the capacity will possibly be lower in suchscenarios, and our measurements give an upper bound onthe achievable throughput.

In regards to latency, the average delay was very low(10–15 ms), despite of the number of hops from the gate-way. The browsing experience was good in average, withdownload times between 2 and 6 s for 275 KB pages.

4.2.2. Near client vs. far client

The bandwidth allocation already exhibits unfairnesswhen clients are connected to the same access point. Thefollowing test results show that connecting the clients todifferent APs only exacerbate the unfairness. To investigatethis, the setup consisted of one client associated with thegateway and a second client associated to the first, secondand third relaying AP, respectively, (see Figs. 10 and 11).

Fig. 10. Near client vs. far client [DL].

Fig. 9. Single-radio mesh cluster.

Fig. 11. Near client vs. far client [UL].

1572 A. Arjona et al. / Computer Communications 31 (2008) 1564–1578

The results show that the maximum throughput of thefurthermost client, that is, the one that is far from the gate-way, is directly affected by the clients utilizing resourcesfrom the gateway. When clients are located at differentnodes, the total bandwidth available for all nodes isreduced to roughly 3.6 Mbps. Additionally, the client inthe gateway increases its share of the bandwidth allocationthe further away from the gateway the other client connectsto the network. The approximate bandwidth gain is around14%. Moreover, browsing experience is affected very muchwhen a client is more than one hop away (second relay orfurther) and simultaneous traffic from the other clientexists. In such cases, the webpage download times increasedrastically to 10–30 s. Therefore, in a real deployment, thenumber of hops should be limited to one or two at the max-imum in order not to compromise the browsing applicationperformance.

4.2.3. Client traffic mixes

The following tests aimed at understanding the effect ofmultiple users in different nodes and number of hops fromthe gateway. The test setup consisted of associating a clientand running FTP file download and uploads on the gate-way and the first relay, and subsequently increasing thenumber of clients by adding another client to the secondrelay or further. Fig. 12 shows the test configuration forthree clients, one in the gateway, one in the first relayand one in the third relay.

The results show that the limit for simultaneous clientsin different nodes appears to be three. Already with three

Fig. 12. Test example: simultaneous clients in gateway, first relay andthird relay.

clients, most of the time, the client that is furthermost fromthe gateway does not receive any service at all in the down-link direction, and sometimes it does not either have anyservice in the uplink. This issue occurs both when connect-ing to the second relay and to the third relay.

Client that is connected to second relay or furtherdegrades the mesh network performance even if itself doesnot receive service. This is due to the fact that as long as ittries to access the medium it will affect the performance ofclients in the gateway and first relay considerably. Theavailable bandwidth can drop as much as 75% as long asthe third client tries to gain access to the medium. Fig. 13shows the bandwidth performance view for the client con-nected to the gateway when a third simultaneous client fur-ther apart intends to access the medium.

The starvation at nodes at distances of more than onehop is due to TCP behavior in relation to the window size[21,16]. TCP senders are aggressively probing for morecapacity, which results in excess packets in flight; the win-dows size of the senders grows beyond the optimal value,and results in packets being dropped.

4.2.4. Effect of maximum bandwidth limit of 1 Mbps per user

In order to investigate the problem with multiple usersfurther, we limited the maximum available bandwidth peruser to 1 Mbps, using a management feature of the Merakinodes. Several of the metropolitan wireless mesh networkswith single-radio architecture enforce this limit, includingthe Google network in Mountain View [24]. With this lim-itation, the furthermost client is able to receive service (seeFig. 14). However, latency increased to an average of200 ms. This illustrates the need for rate limiting, trafficmonitoring and potentially connection admission controlin the mesh network.

4.2.5. Susceptibility to hidden node problems

In order to increase the throughput of the mesh net-work, single-radio nodes customarily disable the RTS/CTS exchange. This exposes the nodes to the hidden termi-nal problem [3,27,42]. To test the susceptibility of the nodes

Fig. 13. Third simultaneous client effect on performance.

Fig. 14. Bandwidth share for three simultaneous users in different nodes.

A. Arjona et al. / Computer Communications 31 (2008) 1564–1578 1573

to the hidden node, clients placed in different RF enclo-sures were forced to be unaware of each other while stillbeing able to transmit data to a common access point.The results show that the existence of even one hidden cli-ent is very harmful to the network performance. This prob-lem affects uplink transmissions much more than downlink.The reason for this is that even though clients are notaware of each other, the access point is aware of themand synchronizes transmissions to each of them (seeFig. 15).

Furthermore, while the scenario is symmetric for bothclients, the bandwidth allocation is actually unfair. Thethroughput for one client is very low, while the throughputfor the other client fluctuates a lot. The problem is lessdamaging in cases where the majority of users synchronize

Fig. 16. VoIP capacity without

Fig. 15. Hidden node problem

transmissions with each other. However, the active hiddenclient still hinders performance by forcing heavy servicefluctuations in the rest of the clients.

The trade-off for the single-radio node is thus to maxi-mize the throughput by disabling the mechanisms to han-dle the hidden node problem. This adversely affects theoverall performance when the hidden node situation actu-ally appears. While the choice to not enable RTS/CTSmight prove better overall for the network, it is very detri-mental to the nodes placed in the adverse situation.

4.2.6. VoIP quality and capacity

VoIP mean opinion score quality was tested with IxCha-riot [26] and consisted of client mixes with and without simul-taneous background traffic. Traffic load was generated with

simultaneous data traffic.

effect on performance.

Fig. 17. VoIP capacity with simultaneous data traffic.

1574 A. Arjona et al. / Computer Communications 31 (2008) 1564–1578

a continuous 2 Mbps UDP data stream on the downlinkdirection. The resulting maximum number of calls usingthe G.711 codec with toll call quality was 6–7 for VoIP onlyscenarios, which is lower than the normal 802.11 g capacity[13]. In scenarios with background traffic load, the maximumnumber of users was reduced to 6. Studies in [41] describesimilar mesh results. We evaluated capacity with G.711codec because it is the most widely supported VoIP codec.However, the maximum capacity can differ depending onthe codec bitrate and packet size. The 802.11 technologyhas an upper limit on the number of packets transmittedper second, regardless of the packet payload size, thus, thecapacity in bps drops as the packet size goes down. Figs.16 and 17 summarize the VoIP results.

5. Mesh networks vs. cellular technologies

In this section we evaluate the current mesh networksagainst the GSM cellular technologies. We are interestedin answering the question: given the performance parame-ters that we have assessed in the previous sections, both ina live deployed network and in the lab, can the current gen-eration of wireless mesh networks be a viable alternative tocellular networks?

Table 4WLAN cards receiver sensitivity

Manufacturer Receiver sensitivity

Orinoco PMCI silver/gold �82 dBm at 11 Mbps�87 dBm at 5.5 Mbps�91 dBm at 2 Mbps�94 dBm at 1 Mbps

Cisco Aironet 350 �85 dBm at 11 Mbps�89 dBm at 5.5 Mbps�91 dBm at 2 Mbps�94 dBm at 1 Mbps

5.1. Coverage and cost

The measurement results prove that the Google networkdoes not provide an adequate service across the city.Except for the downtown area, the access points are notdensely deployed and there are several coverage holes.Our research is based on the assumption that all accesspoints cover a circle of equal size, and are placed in theplane according to a triangular paving (hexagonal cell cov-erage). As a consequence of the current average of 30access points per square mile, each coverage disk is about166 m, and access points are 287 m apart. Using a freespace propagation model (a = 2) and the Friis formula inthe 2.4 GHz frequency band, the link budget at 166 m is�84 dBm. However, if we use a typical path-loss exponentfor mid-density urban areas (a = 3.3) [14], the link budgetat 166 m is �113 dBm. The receiver sensitivity of an Ather-os chipset, which is one of the main WLAN card manufac-

turers, is �71 dBm for the 54 Mbps modulation rate for an802.11 g card, �90 dBm for the 11 Mbps modulation rate,and �95 dBm for the 1 Mbps modulation rate. The trans-mission power for the same chipset is 18 dBm ± 2 dBm.

With only 18 dBm transmission power for the client, it isevident that transmission is possible for free space propa-gation, but not for an exponent observed in real setups.This can be demonstrated with the following calculation:

18 dBm� 113 dBm ¼ �95 dBm

Furthermore, handheld devices have even less transmis-sion power (usually 15 dBm). Therefore, the current net-work architecture in Google cannot provide goodconnectivity. This is also a likely outcome in many otherdeployments based on 30 access points per mile. Table 4shows other WLAN receiver sensitivity values as given bydifferent manufacturers.

With a transmission power of 20 dBm and a receiver sen-sitivity of �90 dBm, the resulting link budget equals110 dBm. Consequently, using the same typical path-lossexponent for mid-density urban areas of a = 3.3, the cellradius is 130 m. For 18 dBm, it is 115 m, and for 16 dBm itis 100 m.

In order to address the low transmission power of hand-held devices, we assume a radius of 100 m. With that cellradius size, 81 access points would be required to coverone square mile without overlapping in the covered areas.Therefore, even a higher number of access points wouldbe required for efficient coverage. Fig. 18 show the esti-mated costs of cellular and Muni WiFi deployments with30 and 81 access point density per square mile.

Fig. 18. Evaluation of cost of deployment between Muni WiFi and cellular.

A. Arjona et al. / Computer Communications 31 (2008) 1564–1578 1575

In contrast, the cost of building a cellular network basestation site costs around $30,000 USD [15]. However,macro sites and installation costs can push the cost to$100,000 USD. This shows that the cost of current MuniWiFi deployments is already similar to cellular. The costof Muni WiFi deployments is not an advantage over cellu-lar deployments. Moreover, the cost of a Muni WiFi net-work with good coverage would raise the cost up to threetimes the current cost. Cellular deployments also providemuch better coverage (e.g. indoors) where WiFi is heavilyhandicapped due to outdoor-to-indoor wall attenuation.Nevertheless, it is important to note that cellular deploy-ments require an expensive licensed spectrum and relativelyexpensive sites. Likewise, site acquisition of cellular sites isnot always easy or cheap. But in such cases, very smallnano base stations similar to WiFi access points can beinstalled. However, in the case of an existing cellular oper-ator, spectrum licenses and sites might already be availablefrom previous deployments. Even though, the spectrumcost is not included in our comparison, the market dynam-ics support our conclusion. A clear example is Earthlinkcancelling future deployments such as San Francisco, sincethe current mesh architecture cannot offer a cheap andgood performing solution to the customer [37].

The current Muni WiFi rule of thumb of 30 accesspoints per square mile is possibly a consequence of the$100,000 USD cost per mile budget. With an average costof $2,000 USD per access point, and the cost of a fewaggregation nodes, network hardware in the backend andinstallation, the whole budget is consumed easily. A budgetof $100,000 USD per square mile with 81 access pointswould impose a cost per access point of less than $740USD.

5.2. Data and voice capacity

The measurements show that single-radio mesh net-works are able to provide equivalent or higher throughputthan 3G networks. However, in live setups, due to fairbandwidth allocation and starvation issues, the maximum

user throughput might be restricted to 1 Mbps as in theGoogle network [24]. Therefore, the gain is not necessarilysignificant. In addition, since the single-radio architecturecapacity is reduced depending on how many hops the traf-fic has to go through before reaching a gateway, the overallmaximum users can be fairly limited compared with cellu-lar technologies. Cellular technologies are able to providedata rates of 7–10 Mbps using Release 5 products(HSDPA). The maximum number of data users rangesfrom 16 to 25 for Release 5, and larger numbers withRelease 6 products (HSUPA).

In regards to voice capacity, our measurement resultsand previous work show that the maximum number ofVoIP users with toll quality in single-radio mesh networksis around 7 per cluster under excellent signal conditions[33,41,45]. Current deployments have about 5 clusters persquare mile, thus, around 35 maximum theoretical VoIPusers. In case multi-radio mesh hardware is used, the maindifference is that the capacity per each node is not sharedamong the other nodes in the cluster. Therefore, the capac-ity is dependant on the number of nodes. In a multi-radiodeployment with 30 nodes per square mile, the theoreticalnumber of VoIP users would be around 210. However, inpractice, with the current cluster configurations, providinggood signal coverage is very challenging, and such capacitylimits are hardly reached. In contrast, the voice capacity incellular networks is around 40 users for a cell covering asquare mile with very reliable indoor penetration and qual-ity of service assured. Moreover, already standardizedRelease 7 features will increase capacity considerably toaround 135 by using circuit switched voice over HSPA[34] and still leave space for data users.

With such capacity limits, the limitations of the currentrule of thumb of 30 APs per square mile are even more evi-dent. Using the Google network as an example, each of the380 APs deployed covers 190 citizens (out of 72,000). Con-sidering the average household size of 2.25 (2000 census),this means there is only one AP per every 84 homes. Fur-thermore, the single-radio capacity is shared within thecluster. Therefore, the current cluster density of 6 APs

1576 A. Arjona et al. / Computer Communications 31 (2008) 1564–1578

attempts to cover �500 homes (1134 citizens). These num-bers are much higher than the maximum capacity the tech-nology can provide, especially for voice services. Therefore,it is clear that the single-radio architecture is not a realisticsolution for providing a competitive service vs. well devel-oped cellular networks.

Furthermore, the above calculations are for externalcoverage only. The current AP density provides a very poorindoor coverage if any. Google attempts to take care of theindoor coverage issues by promoting the use of externaladaptors (e.g. Ruckus [36]) and subsequently adding selfconfigurable access points (e.g. Meraki Mini) to extendcoverage [24]. Since additional hops will affect the wirelesslink further we believe this solution is neither realistic norfeasible. Especially considering that the additional extender(�130 USD) and AP (�50 USD), result in roughly 180USD (below 1 Mbps link) opposed to a traditional ADSLlink (1.5 Mbps link) for 10–30 USD per month.

5.3. Mesh networking in the future

Mesh networking as it is currently deployed has variousshortcomings, which were discussed in the previous sec-tions. In this section, we highlight some of the importanttechnological advances that are needed to make wirelessmesh networks a more feasible technology.

First, a multi-radio based architecture will likelyimprove the current performance. A multi-radio architec-ture uses a different frequency for backhaul communica-tion between mesh access points (e.g. 802.11a at 5 GHz)and the wireless access with the WiFi clients (e.g.802.11b/g at 2.4 GHz). This results in a capacity increasesince multiple channels are used within each cluster. Addi-tionally, unfairness in bandwidth allocation has a lessereffect due to the different frequency used between meshaccess points. However, currently, multi-radio mesh accesspoints are almost double the price of single-radio nodes. Ifdemand for multi-radio mesh access points increase, themanufacturing costs would reduce and make WiFi deploy-ments more feasible.

Second, even though a multi-radio architecture couldimprove performance, many issues will still remain, suchas the required node density to provide continuous cover-age, VoIP capacity, and the hidden node problem. In casecontinuous coverage is not available, there are other solu-tions that could fill in the gaps. One possibility is Unli-censed Mobile Access (UMA). UMA technologysupports seamless handovers between WiFi and 2G and3G for both voice and packet data. Therefore, wheneverWiFi coverage would weaken, users could roam to cellular[8,44]. Another possible approach is by supporting MobileIP-based inter-system mobility between WiFi and 3G net-works. Roaming to 3G when necessary is interesting sincesome 3G networks such as HSDPA and cdma EV-DO arealready able to provide user throughputs above 1 Mbps.Likewise, a Mobile IP does not require an agreement

between network providers as long as the Mobile IP HomeAgent is available via the public Internet.

Multi-radio mesh networks could also use technologiessuch as WiMAX for the backhaul and FLASH-OFDMfor the client link. However, these technologies use licensedspectrum, which limits the number of providers, whileWiFi can be used by anyone. Still, with unlicensed commu-nications, we easily end up with multiple providers usingthe same channels and disturbing each others communica-tions. This can be easily seen in urban areas already todaywith tens of individual 802.11b/g access points setup to usethe same limited number of channels.

6. Related work

There is prior work available investigating several of thelimitations of single-radio mesh networks. For instance,several papers [5,16,21,46] investigate the effects of starva-tion with TCP over wireless networks; which is the reasonwhy the available bandwidth is reduced considerably withmultiple users at different hops. However, these works donot demonstrate the effect of starvation via measurementsor simulations. Additionally, another group of papers[16,22,23] investigate options (e.g. rate limiting) to countermeasure unfairness and possible starvation. Further, manystudies [3,27,42] have been made on the effect of the hiddennode problem and they propose several solutions to it.However, most of the solutions require changes in theaccess nodes and also in the clients. In regards to the num-ber of VoIP calls supported in mesh networks, severalpapers [33,41,45] show similar results; other papers [45]also analyzes optimizations that could increase the VoIPcall capacity.

Finally, there are a few number of measurement studieson the MIT Roofnet mesh network [1,11,12,19]. Thesemeasurements differ significantly from ours since they arebased on a single-tier architecture, while most commercialdeployments are based on a two-tier architecture (e.g. Goo-gle network). Measurements on live commercial networkssuch as the one in this study are not currently available.In addition, the available research does not focus on theperformance of real-time applications such as VoIP, thenetwork capability to carry voice as a primary service,nor measures the coverage area of such deployments.

7. Summary

In this paper we evaluated the single-radio mesh archi-tecture and characterized its performance in a live setupand laboratory environment. We showed that even state-of-the-art network deployments such as the Google net-work in Mountain View are far from optimal and only ableto provide very limited coverage and services. In regards toVoIP service, the probability of finding a spot with a feasi-ble coverage for good voice quality is very low; and themore users make the same choice, this spot eventuallybecomes congested.

A. Arjona et al. / Computer Communications 31 (2008) 1564–1578 1577

Furthermore, we studied limitations of the mesh archi-tecture in a laboratory environment and described severalof its issues, such as the fairness in bandwidth allocationand capacity at different hops, which is an issue relatedto starvation. Likewise, we studied the VoIP user capacity,which plays a major role if these networks are to competewith traditional cellular networks. Our results show thatthe maximum number of users with toll quality is 7. Inaddition, our results show that the hidden node problemis still an existing issue in mesh networks and should beconsidered for future deployments.

Finally, we evaluated the feasibility of singe-radio meshtechnologies and its competitiveness with cellular net-works. Our evaluation shows that despite the popularityof WiFi mesh, the technology is far from optimal andnot really a contender with well developed cellular net-works. Furthermore, the cost of deployment is already ata similar level. Therefore, improving the performance byincreasing the AP density could result in a deploymentup to three times more expensive than current cellular tech-nologies. However, with the use of e.g. Mobile IP or UMA,interworking between WiFi and cellular networks is beingdeployed as a commercial product already. Interworkingbetween networks could potentially reduce the need fordeploying continuous WiFi coverage and therefore, resultin decreased costs. Finally, if demand for multi-radio meshaccess points increase, the manufacturing costs woulddecrease and make WiFi mesh deployments more feasible.

References

[1] D. Aguayo, J. Bicket, S. Biswas, G. Judd, R. Morris, Link levelmeasurements from an 802.11 mesh network, in: Proc. ACMSIGCOMM, Portland, OR, 2005.

[2] AirMagnet Surveyor. Available from: <http://www.airmagnet.com/>.

[3] F. Akyldiz, X. Wang, W. Wang, Wireless mesh networks: A survey,Elsevier Computer Networks 47 (4) (2005) 445–487.

[4] Alvarion. Available from: <http://www.alvarion.com/>.[5] G. Anastasi, E. Borgia, M. Conti, E. Gregori, Rate control in

communication networks: shadow prices proportional fairness andstability, J. Cluster Comput. 8 (2–3) (2005) 135–145.

[6] A. Arjona, S. Takala, The Google Muni WiFi network – can itcompete with cellular voice?, in: Proc. Advanced InternationalCommunications Conference AICT’07, Morne Mauritius, May 13–19, 2007.

[7] A. Arjona, C. Westphal, S. Takala, Empirical analysis of the single-radio mesh architecture performance, limitations and challenges, in:Proc. ACM International Wireless Communications and MobileComputing IWCMC’07, Hawaii USA, August 12–16, 2007.

[8] A. Arjona, H. Verkasalo, Unlicensed moblie access (UMA) –handover and packet data performance analysis, in: Proc. Interna-tional Conference on Digital Telecommunications (ICDT’07), SiliconValley USA, July 1–6, 2007.

[9] BelAir Networks. Available from: <http://www.belairnetworks.com/>.

[10] BelAir Networks, Capacity of Wireless Mesh Networks – Under-standing Single Radio, Dual Radio and Multi Radio Wireless MeshNetworks, BelAir Networks, 2006.

[11] J. Bicket, D. Aguayo, S. Biswas, R. Morris, Architecture andevaluation of the MIT Roofnet mesh network, in: Proc. ACMMOBICOM, Cologne Germany, September 2005.

[12] S. Biswas, R. Morris, Opportunistic routing in multi-hopwireless networks, in: Proc. HotNets-II, Cambridge MA,November 2003.

[13] L. Cai, X. Shen, J. Mark, L. Cai, Y. Xiao, Voice capacity of WLANwith unbalanced traffic, QShine’05, Orlando USA, August 22–24,2005.

[14] J. Camp, J. Robinson, C. Steger, E. Knightly, Measurement drivendeployment of a two-tier urban mesh access network, in: Proc.MobiSys 2006, Uppsala, Sweden, June 2006.

[15] K. Casier, S. Verbrugge, R. Meersman, J. Van Ooteghem, D. Colle,M. Pickavet, P. Demeester, A fair cost allocation scheme for CaPexand OpEx for a network service provider, in: Proc. CTTE 2006,Athens, Greece, June 2006.

[16] M. Chen, A. Zakhor, Flow control over wireless network andapplication layer implementation, in: Proc. IEEE INFOCOM,Barcelona Spain 2006.

[17] Cognio Spectrum Analyzer. Available from: <http://www.cognio.com/>.

[18] R. Cole, J. Rosenbluth, Voice over IP performance monitoring, in:Proc. ACM SIGCOMM’01.

[19] D. De Couto, D. Aguayo, J. Bicket, R. Morris, A high-throughputpath metric for multi-hop wireless routing, in: Proc. ACM MOBI-COM, September 2003.

[20] DECT Forum, Positioning of DECT in Relation to other RadioAccess Technologies, DECT Forum, June 2002. Available from:<http://www.dect.org/pdf/dect_positioning_version1.pdf/>.

[21] Z. Fu, P. Zerfos, H. Luo, S. Lu, M. Gerla, The impact of multihopwireless channel on TCP throughput and loss, in: Proc. IEEEINFOCOM, San Francisco CA, April 2003.

[22] V. Gambiroza, B. Sadelghi, E. Knightly, End-to-end performanceand fairness in multihop wireless backhaul networks, in: Proc. ACMMOBICOM, Philadelphia PA, 2004.

[23] M. Garetto, T. Salonidis, E. Knightly, Modeling perf-flow through-put and capturing starvation in CSMA multi-hop wireless networks,in: Proc. IEEE INFOCOM, Barcelona Spain, 2006.

[24] Google WiFi Mountain View. Available from: <https://wifi.google.com/support/>.

[25] ITU-T, Rec. G.107 – A Computational Model for Use in Transmis-sion Planning ITU-T 2003.

[26] Ixia IxChariot. Available from: <http://www.ixiacom.com/products/display?skey=ixchariot/>.

[27] Y. Kim, J. Yu, S. Choi, K. Jang, A novel hidden detectionmechanism in IEEE 802.11 WLAN, IEEE Comm. Lett. 10 (8)(2006) 608–610.

[28] B. Mandeville, C. Mandeville, A. Bouard, Wireless mesh: Ready!,light reading, June 20, 2006. Available from: <http://www.lightread-ing.com/document.asp?doc_id=96200/>.

[29] Malden, PESQ – Perceptual Evaluation of Speech Quality, Malden.Available from: http://www.malden.co.uk/pesq.htm.

[30] O. Malik, Google Launches WiFi Network in Mountain View,GigaOM, August 2006. Available from: http://gigaom.com/2006/08/15/google-launches-wifi-network-in-mountain-view/.

[31] Meraki. Available from: <http://www.meraki.net/>.[32] Motorola Canopy. Available from: <http://motorola.canopy.com/>.[33] D. Nicolescus, S. Gamguly, K. Kim, R. Aizmailov, Performance of

VoIP in a 802.11 Wireless Mesh Network, in: Proceedings of IEEEINFOCOM 2006.

[34] Nokia Siemens Networks and Nokia, Supporting CS over HSPA,3GPP R2-073487, August 2007.

[35] Roofnet. Available from: <http://pdos.csail.mit.edu/roofnet/doku.php/>.

[36] Ruckus Wireless. Available from: <http://www.ruckuswireless.com/>.[37] R. Selna, S.F. Citywide Wi-Fi Plan Fizzles as Provider Backs Off, San

Francisco Chronicle, August 30, 2007. Available from: <http://www.sfgate.com/cgi-bin/article.cgi?f=/c/a/2007/08/30/MNE-JRRO70.DTL/>.

[38] SkyPilot. Available from: http://www.skypilot.com.[39] Strix Systems. Available from: <http://www.strixsystems.com/>.

1578 A. Arjona et al. / Computer Communications 31 (2008) 1564–1578

[40] W. Terrill, Wireless LAN Mesh: Wireless Wireless, Network andTelecom Strategies In-Depth Research Report, Burton Group, June02, 2005.

[41] K. Ting, Y. Ko, M. Sim, Voice performance study on single radiomultihop IEEE 802.11b systems with chain topology, in: Proc.ICN’05.

[42] F.A. Tobagi, L. Kleinrock, Packet switching in radio channels: Part II– the hidden terminal problem in carrier sense multiple-access modesand the busy-tone solution, IEEE Trans. Commun. COM-23 (12)(1975) 1417–1433.

[43] Tropos Networks. Available from: <http://www.tropos.com/>.[44] Unlicensed Mobile Access. Available from: <http://www.umatech-

nology.com/>.[45] H. Wei, K. Kim, A. Kashyap, S. Ganguly, On admission of VoIP

calls over wireless mesh network, in: Proc. IEEE ICC, 2006.[46] K. Xu, M. Gerla, L. Qi, Y. Shu, Enhancing TCP fairness in adhoc

networks using neighborhood RED, in: Proc. ACM MOBICOM, SanDiego CA, 2003.

[47] Unstrung, Wireless Mesh: From Enterprise to Metro, UnstrungInsider, February 2006.