The Impact of Adjacent Channel Interference inMulti-Radio Systems using IEEE 802.11

Jens Nachtigall, Anatolij Zubow, Jens-Peter RedlichHumboldt University, Germany.

E-mail: {nachtiga, zubow, jpr}@informatik.hu-berlin.de

AbstractA promising approach for improving the capac-ity of Wireless Mesh Networks is by making use of multi-ple non-overlapping RF channels. Multi-channel protocols havethe advantage that several devices can transmit in parallelwithin a collision domain on distinct channels. When usingIEEE 802.11b/g/a most protocol designers assume 3 and 12non-overlapping channels, respectively. However, this simpliedassumption does not hold. We present results from measure-ments that show that the number of available non-interferingchannels depends on the antenna separation, PHY modulation,RF band, trafc pattern and whether single- or multi-radiosystems are used. The problem is caused by Adjacent ChannelInterference (ACI) where nearby transmitters bleed over toother frequencies and either cause spurious carrier sensing orframe corruption. For nearby transceivers, as in the factorydefaults of multi-radio devices, this results in at most two non-interfering channels, one within 2.4 GHz and the other withinthe 5 GHz band. Only if the distance between the antennas isincreased, non-interfering channels within the bands themselvesbecome available. Moreover, our comparison of single- and multi-radio systems allows us to isolate ACI from board crosstalk andradiation leakage of which only the multi-radio systems seemto suffer. Finally, we show how a packet-level simulator can beimproved to realistically incorporate ACI. With the help of thissimulator more condent statements about the performance ofvarious multi-channel protocols can be made.

Index TermsWireless Networks, Multi-Radio, Multi-Channel, Cross-Channel, Adjacent Channel Interference,Measurements

I. INTRODUCTION

It is known that throughput and latency within a WirelessMesh Network (WMN) dramatically worsen with the numberof hops that need to be traversed for communication [10].This multi-hop nature is the main difference compared totraditional cellular networks. As WMNs usually operate onthe same frequency channel to avoid network partitioning, theyalso share a common collision domain. That is, while a nodetransmits a packet to another node within its communicationrange, all remaining nodes within the much bigger interferencerange need to remain silent. This has a dramatic impact onthe performance of multi-hop networks whose very nature itis to relay other nodes packets. It is widely assumed thatIEEE 802.11b/g offers 3 and 802.11a 12 non-overlapping RFchannels. Channels are referred to as being non-overlapping ororthogonal if they can be used in parallel without any negativeinterference effects. An obvious strategy is to utilize differentnon-overlapping channels within a WMN in order to split upthe collision domain as a way to alleviate the adverse effects

of interference. Neighboring nodes can transmit in parallelas long as they use non-overlapping channels. This increasesthe throughput and reduces the delay signicantly. Declininghardware prices boosted this trend: It is not only worthwhileto increase the number of nodes and hence possibly channelsof a single-radio, multi-channel network, but also to equipdevices with multiple wireless transceivers to establish a so-called multi-radio, multi-channel network.

II. RELATED WORK

A signicant work was done on protocols that exploitmultiple RF channels available in IEEE 802.11 to increasecapacity [13]. Most multi-channel protocol designers assumethe existence of several non-overlapping channels, e.g. 3for 802.11b/g and 12 for 802.11a (FCC domain). Whileimplementing proof-of-concept prototypes of their multi-radioprotocols, some authors accidentally stumbled across the factof Adjacent Channel Interference (ACI) between supposablynon-overlapping channels. For instance, Draves et al. [7] couldnot nd any non-interfering channels within 802.11b/g and802.11a. As a result they had to operate one transceiver in the2.4 GHz and the other in the 5 GHz band. Of the expected 15non-interfering frequency channels only two remained. Theirtheoretical multi-radio routing metric was effectively reducedto a two-radio routing metric. Adya et al. [1] propose a linklayer protocol operating on distinct frequency channels. Onlyby separating the radios by at least 30 cm three non-interferingchannels became available for 802.11b/g. They also noted thatACI is highly hardware dependent. Robinson et al. [16] used amore systematical measurement methodology. However, theyused rather untypical hardware for WMNs Dell workstationswith four PCI slots. They observed that merely plugging anadditional wireless card into a workstation and operating itin passive monitor mode can reduce the throughput. Theyaccounted this to board crosstalk and radiation leakage of thepassive cards. In contrary to our results, they found that eventwo receiving radios were strongly interfering with each other.It should be noted that they sent 802.11 unicast packets thatrequire acknowledgments in the opposite direction, whereaswe sent broadcast packets only. Therefore, Robinson et al.did not isolate the different trafc patterns of receiving andtransmitting as clearly from each other as we did. However,they also observed the need to increase the distance betweenthe radios, in their case 1 m, to get at least two non-overlappingchannels within 802.11b/g. The same was emphasized in [4],

978-1-4244-2202-9/08/$25.00 2008 IEEE 874

[14] and [12], whereas the latter required a distance of at least1 m between their omni-directional 5 dBi antennas for a xedbitrate of 11 Mbps. Cheng et al. [5], [6] presented resultsfor 802.11a. However, they only considered three neighboringchannels within 802.11a. The most recent work on ACI canbe found in [8]. Four Intel laptops (with 2200BG Mini-PCIcards) formed two links, whose two ends were close to eachother. One link was set to channel 3 of 802.11b and the otherto channel 8. Surprisingly, the radio on channel 3 could evendecode packets from channel 8.

III. THEORYIn this chapter we explain why two distinct channels might

interfere with each other. The notion of Adjacent ChannelInterference is most important. An additional problem, so-called Board Crosstalk or Radiation Leakage occurs when us-ing multi-radio systems. Finally, the number of non-interferingchannels also depends on the used radio spectrum.

A. InterferenceDenition 1: Adjacent Channel Interference

Interference that is caused by nearby transmitters on distinctfrequency channels bleeding over to another channel iscalled adjacent channel interference (ACI) [15, p. 74].

The resulting phenomenons are commonly referred to asthe near-far or near-eld effect [11, p. 279]. For 802.11, onecan distinguish between the following consequences (to bediscussed more deeply in Sec. IV): For the case of two nearbytransmitters (referred to as TX-TX) the overlapping ACI ofone transmitter causes a spurious carrier sensing at the other.Remember that 802.11 is a CSMA/CA protocol which followsthe listen-before-talk paradigm. That means that a station isonly allowed to transmit if the medium is idle. ACI may triggerthe carrier sensing mechanisms to report that the mediumis busy. In this case the station will misleadingly defer itstransmission. For the case of a receiver and a transmitteron distinct channels in close proximity (RX-TX) the weakincoming signal at the receiver gets corrupted by the strongoutgoing signal of the nearby transmitter. Both consequencesare also described in the measurement based studies of [8]and [5]. We also observed a problem for two receiving nodesin close vicinity (RX-RX): The ACI causes a variant of thehidden-node problem that cannot be tackled with RTS/CTSsince the two links are on distinct channels and therefore thereceiver is unable to decode the NAV value from the RTS/CTSpackets.

Throughout our evaluation we considered two different de-vice types: single- and multi-radio systems. The latter refers todevices equipped with multiple radio transceivers. In contrast,single-radio systems have only one radio transceiver. Whenusing multi-radio systems an additional problem appears, theso-called board crosstalk or radiation leakage:

Denition 2: Board Crosstalk and Radio LeakageBoard crosstalk is dened as noise caused by the usage ofa common bus by several WiFi cards. On the other hand,radiation leakage refers to over-the-air interference due toimperfect shielding of the WiFi cards [16].

Frequency (MHz)

Re

lative

sig

na

l str

en

gth

(d

B)

2400 2420 2440 2460 24802380

0

-20

-40

-60

-80

Channel 1

Channel 6

Channel 11

Frequency (MHz)

Re

lative

sig

na

l str

en

gth

(d

B)

2400 2420 2440 2460 24802380

0

-20

-40

-60

Channel 1

Channel 6

Channel 11

OFDM

DSSS

Figure 1. Channel separation in 802.11b DSSS (top) and 802.11g OFDM(bottom). Adapted from [9].

30 25 20 15 10 5 fc 5 10 15 20 25 3060

50

40

30

20

10

0

Frequency (MHz)

Po

we

r S

pe

ctr

al D

en

sity (

dB

m)

802.11 OFDM TX spectrum mask

TX Spectrum Mask (802.11 spec)

Typical Spectrum (802.11a 18dBm)

Figure 2. OFDM transmit mask as required by 802.11g/a (outer line)compared with a real transmission at 18 dBm measured with a spectrumanalyser (inner line) [6].

Since we used both single-radio as well as multi-radiosystems, we could isolate the interference effects caused byboard crosstalk or radiation leakage from ACI.

B. Radio Spectrum Usage

a) 802.11b/g (2.4 GHz): The DSSS PHY has at most 13channels in the 2.4 GHz band, each 5 MHz wide. Channel 1 isplaced at 2.412 GHz, channel 2 at 2.417 GHz, and so on up tochannel 13 at 2.472 GHz. DSSS is a single-carrier modulation.Within a channel, most of the signal energy is spread acrossa 22-MHz band. To prevent interference to adjacent channels,the rst side lobe is ltered to 30 dB below the power at thechannel center frequency. Additional lobes are ltered to 50 dBbelow the power at the channel center [9]. ACI inuences thenumber of channels that can be used simultaneously. The IEEE802.11b species that 25 MHz spacing is sufcient. Fig. 1(top) shows the spectral mask of transmissions on the so-callednon-overlapping channels (1, 6, and 11). The multi-carrierOFDM PHY is also available in the 2.4 GHz band. As withthe DSSS PHYs, a transmit mask limits power leakage intothe side bands. Fig. 2 shows the transmission mask as requiredby 802.11g/a and a real OFDM transmission measured witha spectrum analyser. Fig. 1 (bottom) shows the spectral maskof OFDM on the so-called non-overlapping channels.

875

MM0

L R

M1

(a) (b)

d1

M0

R

d1

M1

L

d2

d3d3

d2

Figure 3. Node placement: (a) multi-radio M with interfaces M0 and M1(b) two single radios M0 and M1

b) 802.11a (5 GHz): The OFDM PHY in the 5 GHzband offers eight channels for indoor and eleven for outdooruse (ETSI domain), each 20 MHz wide. In comparison to the2.4 GHz band, the channel spacings are larger 20 instead of5 MHz [9].

IV. MEASUREMENTSThe measurements were taken on the topology depicted in

Fig. 3. Three nodes referred to as left (L), middle (M ) orright (R) node were put indoors in an almost straight line.The left and right node were single-radio devices, whereaswe considered two different setups for the middle node: Atrst a multi-radio system equipped with two radio transceivers(Fig. 3a) was used to evaluate the impact of both ACI as wellas board crosstalk and radiation leakage. To isolate both effectsfrom each other we also replaced the multi-radio device in themiddle by two single-radio devices (Fig. 3b). Five differentdistances were used for d1 (15, 40, 80, 160 and 320 cm),of which a separation of 15 cm corresponds to the factorydefaults. Distances d2 and d3 were kept xed at 140 and1020 cm. We used WRAP.2E boards1 (233 MHz AMD Geodex86 CPU, 128MB RAM) equipped with two Mini-PCI slotstogether with Routerboard R52 wireless IEEE 802.11a/b/gcombo cards (Atheros AR5414 chipset)2. All boards installedin waterproof outdoor metal cases3 were placed 26 cm abovethe ground and had clear line of sight to each other. Omni-directional dual-band antennas with a gain of 5 dBi weremounted with equal vertical polarization. Unless the defaultantenna separation of the WRAP-Board (15 cm) was used,we added an RF cable (HDF-200) of 150 cm length to eachconnector for the multi-radio case in order to increase theantenna separation at the middle node to up to 320 cm. TheRF cable induced an additional attenuation of about 2 dB. Onthe software side we chose OpenWrt4 version 7.09 with Linuxkernel 2.6.22 as operating system and MadWi5 version 0.9.3as WiFi driver. MadWis regdomain setting was changed tothe ETSI and countrycode to Germany in order to obtain13 channels for 802.11b/g. Antenna Diversity was disabledand only the main connector used. The packet generationand capturing was done with the Click Modular Router6

1WRAP.2E boards: http://www.pcengines.ch/wrap2e3.htm2R52 cards: http://www.mikrotik.com/pdf/R52.pdf3Metal cases: http://www.mini-box.com/WRAP-BOX-2A1E4OpenWrt Linux distribution for embedded devices: http://openwrt.org5Multimode Atheros Driver for WiFi on Linux: http://madwi.org6The Click Modular Router: http://www.read.cs.ucla.edu/click/

Parameter ValueSystems Single- and two-radio devicesScenarios TX-TX (M0, M1 transmitting),

RX-RX (M0, M1 receiving),RX-TX (M0 receiving, M1 transmitting)

Antenna separations 15, 40, 80, 160, 320 cmPhysical layer 802.11a/b/gTransmission power 6 (RX-TX only), 16 dBmBitrates 1 Mbps (DSSS), 6 Mbps (OFDM)Radio frequencies 2.4 GHz (ch. 113), 5 GHz (ch. 3664)RTS/CTS DisabledWiFi frame size 1500 bytesTransmission mode BroadcastFlow duration 10 secNumber of runs 10

Table IMEASUREMENTS PARAMETERS

software version 1.5. During transmission and reception wemonitored that the CPU load remained within safe groundsand did not become the bottleneck. The interference due toexternal networks was negligible. Our scripts and all dumples (PCAP) are available on our website for further analysis7.

The parameters we used throughout our measurements aresummarized in Table I. We considered the following threescenarios: i) The middle node transmits on both of its in-terfaces (TX-TX). The left node is set to receiving packetsfrom M0, the right from M1. For non-interfering channels onewould expect that both radios in the middle can transmit inparallel. So the total throughput should be doubled comparedto setups were both radios are on the same channel. ii) Bothinterfaces in the middle receive packets from left and right(RX-RX). Since two ows can be received simultaneously,one would again expect the throughput to double. iii) The leftnode sends packets to M0 while at the same time M1 transmitsto the right node (RX-TX). This is the full-duplex case andfor non-interfering channels one would expect the throughputto remain stable and not halved as seen for single-channelsystems.

We considered both the DSSS and the OFDM physicallayers with a bitrate of 1 and 6 Mbps, respectively. BroadcastWiFi frames with a size of 1500 B were sent out as fast aspossible. Since acknowledgments were not used, we were ableto obtain strictly directive ows. For all experiments the linkbetween L and M0 was xed at channel 1 for 802.11b/g and36 for 802.11a, respectively, while the channel for the linkbetween M1 and R was varied (Fig. 3). For each scenario andchannel assignment the experiment lasted 10 seconds and wasrepeated 10 times. Beforehand, the links were independentlymeasured to ensure that the signal is strong enough and thePacket Error Rate (PER) zero for all links (baseline tests).

A. TX-TX ScenarioIn this section we evaluate channel orthogonality for two

transmitting radios in close vicinity. We setup two ows: fromM0 to L and M1 to R (Fig. 3).

Fig. 4 shows the results for a multi-radio system with DSSSPHY of 1 Mbps for the distance of 15 cm. This is the default

7http://sartrac.informatik.hu-berlin.de/channel-orthogonality

876

1 2 3 4 5 6 7 8 9 10 11 12 130

0.2

0.4

0.6

0.8

1

RF channel at M1

Thro

ughput (M

bps)

RX(L)

TX(M0)

TX(M1)

RX(R)

singleflow capacityalmost reached

high packetdrop

asymmetric flows,M0 defers its TX

Figure 4. TX-TX scenario (multi-radio, DSSS PHY, 15 cm, 2.4 GHz).

antenna separation without additional RF cables. From channel15 both transmitters equally share the medium as would beexpected from a fair CSMA/CA protocol. From channel 68the total throughput remains the same, however, one of thetransmitting nodes (M0) defers its transmission in favor ofthe other node (M1). This may be due to spurious carriersensing at node M0 in this range of low ACI received fromnearby M1, whereas M1 seems to sense the ACI from M0less strong. From channel 913 the total throughput constantlyincreases as the ACI decreases. Hence, the medium is notsensed busy anymore and a maximum of about 0.9 Mbpsis reached at channel 12. For farther antenna separations weachieved a maximum of about 1 Mbps. This suggests thatthe throughput for the setup in Fig. 4 would still increase ifadditional channels were available. Notice that there is alwaysa packet drop of about 10 % on both links which is neitherpresent if only one link is active as for the baseline tests norfor the single-radio setups. Board crosstalk or radiation leakagemay be the reason for that (see Def. 2).

The above situation (15 cm) improves if d1 is increased to160 cm or more. The total throughput reaches its maximum of1 Mbps at channel 6 for a antenna separation of 160 cm andfor 320 cm already at channel 5. The results are summarized inTable II (col. DSSS 2.4 GHZ at TX-TX). The asymmetricow gap of Fig. 4 is shifted to the left for those higherdistances, that is, towards nearer channels since a lower ACIis reached sooner.

Fig. 5 shows the results for the same setup, but using theOFDM PHY of 6 Mbps. Here, the total throughput remainsalmost the same for all channels. This is in contrast withDSSS PHY where at channel 12 for the second radio (M1)the total throughput almost doubled (Fig. 4). By increasing thedistance between the antennas the situation also improves forOFDM. With a separation of 40 cm channels 1 and 8 becomenon-interfering (Table II, col. OFDM 2.4 GHZ at TX-TX).For the multi-radio system the packet drops are smaller withOFDM PHY than with DSSS PHY, but again still present incontrast to the single-radio setups.

The results of the multi-radio setup in the 5 GHz band(OFDM) are depicted in Fig. 6. When the radios are spatiallynearby as with the factory defaults of 15 cm, then only two

Scenario Antenna DSSS OFDM OFDMseparation 2.4 GHz 2.4 GHz 5 GHz

TX-T

X

15 cm 12 (12) n/a (n/a) 60 (48)40 cm 9 8 4880 cm 7 (7) 6 (8) 44 (44)160 cm 6 6 44320 cm 5 (5) 6 (6) 44 (44)

RX-R

X

15 cm 5 (5) 5 (5) 44 (44)40 cm 5 6 4480 cm 5 (5) 5 (6) 44 (44)160 cm 5 6 44320 cm 5 (5) 5 (6) 44 (44)

RX-T

X

15 cm n/a (n/a) n/a (n/a) n/a (48)40 cm 5 n/a 5280 cm 5 (6) n/a (n/a) 44 (48)160 cm 5 9 48320 cm 6 (5) 9 (11) 48 (44)

Table IIMEASUREMENT RESULTS OF ALL SCENARIOS: THE 1ST NUMBER IS THE

NEAREST NON-INTERFERING CHANNEL ON M1 R FOR THEMULTI-RADIO, THE NUMBER IN BRACKETS FOR THE CORRESPONDING

SINGLE-RADIO SETUP. THE LEFT LINK (LM0) WAS FIXED ATCHANNEL 1 FOR 802.11B/G AND 36 FOR 802.11A.

1 2 3 4 5 6 7 8 9 10 11 12 130

1

2

3

4

5

6

RF channel at M1

Thro

ughput (M

bps)

RX(L)

TX(M0)

TX(M1)

RX(R)

asymmetric flows

low packetdrops

hardly any increaseof the data rate

Figure 5. TX-TX scenario (multi-radio, OFDM PHY, 15 cm, 2.4 GHz).

non-interfering channels are available: 36 and at nearest 60.This is surprising as the 5 GHz band for indoor use offersa bandwidth of 140 MHz. As the distance between the twoantennas is increased to 80 cm the channels 36 and 44 becomenon-interfering for the multi-radio system. The results arebetter for the single-radio system. Although the two radiosare still only separated by 15 cm channels 36 and 48 cannow be used simultaneously (Table II, col. OFDM 5 GHZat TX-TX).

B. RX-RX ScenarioNext we evaluate channel orthogonality for two receiving

radios in close vicinity.In Fig. 7 the results are displayed for DSSS PHY where

the antennas separation was set to 15 cm. The maximumthroughput of 1 Mbps per ow is reached at channel 5. Further,when using channels 24 the packet ow is asymmetric. Atchannel 1 and 2 the PER is about 11 %. We noticed a strangeresult at channel 12. At this channel the observed throughputwas very low for some setups. It looks like there is a MadWi

877

36 40 44 48 52 56 60 640

1

2

3

4

5

6

RF channel at M1

Th

rou

gh

pu

t (M

bp

s)

RX(L, d1=15cm)

TX(M0, d1=15cm)

TX(M1, d1=15cm)

RX(R, d1=15cm)

RX(L, d1=80cm)

TX(M0, d1=80cm)

TX(M1, d1=80cm)

RX(R, d1=80cm)

asymmetricflows (d1=15cm)

symmetricflows (d1=80cm)

Figure 6. TX-TX scenario (multi-radio, OFDM PHY, 15 and 80 cm, 5 GHz).

1 2 3 4 5 6 7 8 9 10 11 12 130

0.2

0.4

0.6

0.8

1

RF channel at M1

Thro

ughput (M

bps)

TX(L)

RX(M0)

RX(M1)

TX(R)

channel 12 bug

asymmetricflows

high packetloss

singleflowcapacity reached

Figure 7. RX-RX scenario (multi-radio, DSSS PHY, 15 cm, 2.4 GHz).

driver bug at this channel. We replaced the hardware, butthe problem remained. This only occurs with DSSS PHY.Increasing the distance between the two receiving radios doesnot yield better results channel 5 remains the closest non-interfering channel (Table II, col. DSSS 2.4 GHz at RX-RX).

Fig. 8 shows the results for OFDM PHY in the 2.4 GHzband. As with DSSS PHY we have a high packet loss atchannels 1 and 2. We observed simular PERs for other setupswith very close channel assignments where the carrier sensingshould actually trigger more reliably. This suggests that arather aggressive CCA policy is used by Atheros, since theother link is on the same or very nearby channel. At channel 34 there is no packet loss. For the channels 46 the performancesuffers again due to packet drops. We believe that this is nowa variant of the hidden node problem nodes L and R do notsense each others ACI anymore and go on transmitting whichleads to collisions at the middle node M . This type of hiddennode problem results from ACI and is therefore different fromthe well-known cases due to co-channel interference. EvenRTS/CTS cannot counteract this as an interfering radio onan adjacent channel would not decode it. For higher channelseparations (713) the high packet loss of about 10 % foreach ow remains whereas it only amounts to 2 % for thesame setup with single-radio devices. As with DSSS PHY a

1 2 3 4 5 6 7 8 9 10 11 12 130

1

2

3

4

5

6

RF channel at M1

Th

rou

gh

pu

t (M

bp

s)

TX(L)

RX(M0)

RX(M1)

TX(R)

flow asymmetry,L defers its TX

high packet dropremains

high packetloss

high packet drop(hidden node)

Figure 8. RX-RX scenario (multi-radio, OFDM PHY, 15 cm, 2.4 GHz).

36 40 44 48 52 56 60 640

1

2

3

4

5

6

RF channel at M1

Thro

ughput (M

bps)

TX(L)

RX(M0)

RX(M1)

TX(R)

high packet drops

Figure 9. RX-RX scenario (multi-radio, OFDM PHY, 15 cm, 5 GHz).

further increase in the antenna separations of the middle node(e.g. 320 cm) does not make nearer frequencies than channel5 available for simultaneous use (Table II, col. OFDM 2.4GHz at RX-RX).

At last, we also show the effects of using closely separatedradios in the 5 GHz band (Fig. 9). Channel 36 and 44 are non-overlapping. However, there is always a high PER of about9 % for each ow. This seems to be a board crosstalk orradiation leakage problem as the drops do not occur whenusing two single-radios.

C. RX-TX ScenarioHere we look at channel orthogonality when one radio is

receiving while the other one is transmitting. This is a veryimportant setup as some multi-channel protocols use multi-radio systems to operate in full-duplex mode. We setup thepacket ow from L to M0 and M1 to R. This section isdivided into two parts. At rst we consider the case where thereceived signal is strong in term of signal strength. Here, allradios transmit with 16 dBm. In the second part we decreasedthe transmit power of node L to 6 dBm. The idea here is tosimulate links of real-world mesh networks which are oftenweak [3].

878

1 2 3 4 5 6 7 8 9 10 11 12 130

0.2

0.4

0.6

0.8

1

RF channel at M1

Thro

ughput (M

bps)

TX(L)

RX(M0)

TX(M1)

RX(R)

very high flow assymmetry(L defers its TX)

channel 1112 bug

Figure 10. RX-TX scenario (multi-radio, DSSS PHY, 15 cm, 2.4 GHz).Radio L is transmitting with 16 dBm.

Let us start with the case where all devices transmit with16 dBm. Again, we rst consider the default antenna sep-aration of 15 cm between the two middle radios M0 andM1. Fig. 10 illustrates the results. Channels 1 and 5 are non-interfering. For channels 1 and 2 we again notice a PER ofabout 12% at node M0. At channel 3 the total throughputremains the same, however, the ows are very asymmetric aswe have already seen for other scenarios like TX-TX in Fig. 4.There are signicant packet drops, e.g. at channels 5, 8 and9. The bug at channel 1112 is again present. The results donot change when the distance between the two middle radiosis increased.

The performance of OFDM in the 2.4 GHz band is depictedin Fig. 11. The most interesting fact here is that we cannotnd any two non-interfering channels even with the strongtransmission power at the left node. Even when using channel13 a PER of 41 % remains for the link between L and M0.The situation improves if we increase the distance betweenthe antennas to 320 cm. From channel 7 onwards the negativeconsequences of the near-far effect are mitigated, although theloss on the left link (LM0) remains high with about 8 %compared to the lossless right link (M1 R). The effectsof ACI diminish when switching to the 5 GHz band. Herechannels 36 and 52 are non-interfering for the distance of only15 cm. By increasing the distance between the two radios to320 cm channel 48 can be used.

Next, the transmitting power of node L was reduced to 6dBm to simulate more real-life links. Note that the PER wasstill zero for this weak link in the baseline tests. We start withthe case where the two radios are spatially close to each other(15 cm). Fig. 12 illustrates the results. Between channel 2 and4 CSMA/CA again does not seem to be able to guaranteefor a fair sharing of the medium. The left node defers itstransmission in favor of M1. From channel 5 onwards L doesnot seem to backoff anymore, presumably because the ACIfrom M1 does not trigger its CCA anymore. The results areeven worse than in the TX-TX scenario where the transmissionof packets was only deferred due to spurious CCA at the MAClayer. But here the left node sends packets and therefore seizesthe medium so that all other nodes within its neighborhood

1 2 3 4 5 6 7 8 9 10 11 12 130

1

2

3

4

5

6

RF channel at M1

Th

rou

gh

pu

t (M

bp

s)

TX(L)

RX(M0)

TX(M1)

RX(R)

high packet loss at left link

no loss on right linkfrom channel 2 onwards

Figure 11. RX-TX scenario (multi-radio, OFDM PHY, 15 cm, 2.4 GHz).Radio L is transmitting with 16 dBm.

1 2 3 4 5 6 7 8 9 10 11 12 130

0.2

0.4

0.6

0.8

1

RF channel at M1

Thro

ughput (M

bps)

TX(L)

RX(M0)

TX(M1)

RX(R)

channel 12 bug

packet dropsat weak left link

flow asymmetry,L defers its TX

Figure 12. RX-TX scenario (multi-radio, DSSS PHY, 15 cm, 2.4 GHz).Radio L is transmitting with 6 dBm.

cannot transmit at the same time. Moreover, the packets sentby the left node are almost always corrupted at M0 dueto the strong output power of the transmitting radio M1.This requires retransmission at node L along with exponentialbackoffs. Hence, for RX-TX the problem is at the PHY layer.

Increasing the distance between M0 and M1 helps to coun-teract the near-far problem (Table II, col. DSSS 2.4 GHzat RX-TX). The transmitting radio M1 is now far enoughseparated from the receiving antenna M0. Due to the path-loss attenuation the disrupting ACI of the transmitting radio isnow weak enough at the receiving radio, so that it cannot beharmfull anymore. However, even for such a high separationthe sharing of the medium remains asymmetric at channel 24.

When using OFDM PHY in 2.4 GHz we get slightly worseresults as with DSSS PHY (Table II, col. OFDM 2.4 GHzat RX-TX). An interesting observation can be made in the 5GHz band (Fig. 13). Only the received rate at M0 is illustratednow, while the transmission rates at L and M1 were alwaysat maximum from channel 44 onwards and shared at channel36 and 40. For the nearest distance of 15 cm interference ispresent at all available channels. For a separation of 160 cmthe effects of ACI are mitigated from channel 52 onwards. The

879

36 40 44 48 52 56 60 640

1

2

3

4

5

6

RF channel at M1

Th

rou

gh

pu

t (M

bp

s)

RX(M0, d1=15cm)

RX(M0, d1=40cm)

RX(M0, d1=80cm)

RX(M0, d1=160cm)

RX(M0, d1=320cm)

d1=15cm

d1=80cm

d1=320cm

d1=40cm

d1=160cm

Figure 13. RX-TX scenario (multi-radio, OFDM PHY, 15 and 320 cm,5 GHz). Radio L is transmitting with 6 dBm.

same holds true for the distance of 40 cm. For 320 cm the near-far effect diminishes already from channel 48 onwards and for80 cm even from channel 44 onwards. The order of distancesis not fully logical as additional effects like multi-path mightplay a role.

V. SIMULATIONIn this section we describe the modications we made

to our packet-level simulator in order to support multiplechannels. Our aim was to nd a very realistic model whichalso incorporates ACI. Throughout this section we considersingle-radio systems using 802.11g.

A. ModelJiST/SWANS [2] was used as packet-level simulator. We

extended the radio model to support multiple overlappingchannels. In the initial version of JiST/SWANS the radio ofeach node is modeled as a nite state machine (Fig. 14a). In arst step we changed this by increasing the number of radiosbelonging to a node from 1 to 13 one for each channel incase of the 2.4 GHz band. These 13 radios per node now runin parallel. An incoming RX signal is classied according toits annotated channel and directed to the radio with the samechannel (Fig. 14b). In a second step we implemented the ACI.As an example, consider a packet which was sent on channel3. Our observation from the measurements was that not onlynodes operating on channel 3 may receive or be interfered bythat signal, but also nodes operating on neighboring channels(e.g. 2 and 4) are affected. To correctly model this behavior wehave to take a closer look at a typical transmit spectrum maskof an 802.11g transmission. Fig. 2 shows the theoretical trans-mission mask according to 802.11g/a as well as a transmissionmask from a real 802.11g/a OFDM transmission at 18 dBmmeasured with a spectrum analyzer [6]. Remember that thechannel spacing is 5 MHz. A receiver operating on channel5 instead of 3 will receive that signal with reduced signalstrength of 13 dB. We used this observation to model ACI.Therefore, for each incoming packet we calculate the signalstrength on all neighboring channels according to Table III.

(a)

Single Channel

Radio Model

...

RX Signal

on channel 3

RX

Signal

(b)

Multi Channel

Radio Model

ch #1 ch #2 ch #3

Figure 14. The modied radio model for multiple overlapping RF channels.

Frequency offset (MHz) 5 10 15 20 25 30Spectral attenuation (dB) 0 13 32 42 53 55

Table IIINUMMERICAL VALUES OF THE TYPICAL TRANSMIT MASK OF FIG. 2.

Then these signals are simultaneously processed by the radiosrunning in parallel. As an example, consider Fig. 14b. Here asignal is sent on channel 3. This signal is passed on unmodiedto the radio on channel 3. The signal is also passed unmodiedto the radios on the neighboring channels 2 and 4. However,for channels 1 and 5 the signal is reduced by 13 dB. Theremaining channels are modied according to Table III.

B. ResultsFig. 15 illustrates the simulation results for the TX-TX

scenario. Here, F1 and F2 represent the ows on the linksL M0 and M1 R. That is, we only consider the totalthroughput of a link and do not differentiate between sent andreceived packets anymore. For the short distance of 15 cmonly the channels 1 and 8 are non-interfering. By increasingthe distance to 160 cm we can use channel 6. Channels 1 and5 are non-interfering when the distance becomes greater than500 cm. For most setups of the RX-RX scenario channels 1and 4 are non-interfering (Fig. 16). However, if we increase thedistance between both senders signicantly (20 m), channel1 and 3 can be used at the same time. Finally, considerthe RX-TX scenario depicted in Fig. 17. The ows are veryasymmetric for distances of 15 and 40 cm. In the former

1 2 3 4 5 6 7 8 9 10 11 12 130

1

2

3

4

5

6

RF channel at M1

Thro

ughput (M

bps)

F1(d1=15cm)

F2(d1=15cm)

F1(d1=160cm)

F2(d1=160cm)

F1(d1=500cm)

F2(d1=500cm)

increase in distance reduces effects of ACI

Figure 15. Simulated TX-TX scenario (single-radio, OFDM PHY, 2.4 GHz).

880

1 2 3 4 5 6 7 8 9 10 11 12 130

1

2

3

4

5

6

RF channel at M1

Th

rou

gh

pu

t (M

bp

s)

F1(d1=15cm, d2=2000cm)

F2(d1=15cm, d2=2000cm)

F1(d1=15cm, d2=500cm)

F2(d1=15cm, d2=500cm)

The RXRX scenario isless sensitive to ACI.

Figure 16. Simulated RX-RX scenario (single-radio, OFDM PHY, 2.4 GHz).

1 2 3 4 5 6 7 8 9 10 11 12 130

1

2

3

4

5

6

RF channel at M1

Thro

ughput (M

bps)

F1(d1=15cm)

F2(d1=15cm)

F1(d1=40cm)

F2(d1=40cm)

F1(d1=80cm)

F2(d1=80cm)

symmetric flows(d1=80cm)

asymmetric flows(d1=40cm)

asymmetric flows(d1=80cm)

Figure 17. Simulated RX-TX scenario (single-radio, OFDM PHY, 2.4 GHz).

case, the throughput of the second ow is nearly zero forthe channels 4 and 5. The performance improves when thedistance is increased. Channels 1 and 8 are non-interfering forthe distance of 15 cm. For distances larger than 80 cm channel4 is non-interfering.

VI. CONCLUSIONSOur measurements show that the number of available or-

thogonal channels in IEEE 802.11b/g/a depends on the antennaseparation, PHY modulation, RF band, trafc pattern andwhether single- or multi-radios are used. If two transceiversare in close proximity to each other, as it is the defaultsetup for multi-radio systems, the results are very dramatical.From the frequency ranges of 802.11b/g and 802.11a onlyone channel can be used at the same time. This results in atmost two non-overlapping channels, one within the 2.4 andthe other within the 5 GHz band. Only increasing the antennaseparations of the two transceivers helps. The main reason isACI. Obviously, multi-radio systems were not the main focusduring the standardization process of 802.11b/g/a. The secondissue we noticed is board crosstalk and radiation leakageof the wireless cards. To overcome this problem one mightuse single-radio systems only. Alternatively, Ramachandran etal. [14] connected several single-radio devices via an Ultra

Wide Band backhaul or ethernet cables to form one unit, wherethe WiFi cards and antennas are farther separated from eachother. We also showed that the OFDM PHY with 6 Mbps ismore vulnerable to ACI than the DSSS PHY with 1 Mbps. Thiscan be connected to the higher bitrate in OFDM. Furthermore,OFDM is a multi-carrier modulation whereas DSSS makes useof a single-carrier.

We hope that our detailed measurement results may help todesign better multi-channel protocols which work well withcontemporary 802.11 off-the-shelf hardware. Moreover, withthe help of a simulator that incorporates ACI as presented inthis paper more condent statements about the performanceof various multi-channel protocols can be made.

REFERENCES[1] ADYA, A., BAHL, P., PADHYE, J., WOLMAN, A., AND ZHOU, L. A

multi-radio unication protocol for IEEE 802.11 wireless networks. InBROADNETS04: Proceedings of the First International Conference onBroadband Networks (Washington, DC, USA, 2004), IEEE ComputerSociety, pp. 344354.

[2] BARR, R., HAAS, Z. J., AND VAN RENESSE, R. JiST: an efcientapproach to simulation using virtual machines. Software: Practice andExperience 35, 6 (2005), 539576.

[3] BICKET, J., AGUAYO, D., BISWAS, S., AND MORRIS, R. Architectureand evaluation of an unplanned 802.11b mesh network. In MobiCom05: Proceedings of the 11th annual international conference on Mobilecomputing and networking (New York, NY, USA, 2005), ACM Press,pp. 3142.

[4] CHEBROLU, K., RAMAN, B., AND SEN, S. Long-distance 802.11blinks: performance measurements and experience. In MobiCom06:Proceedings of the 12th annual international conference on Mobilecomputing and networking (New York, NY, USA, 2006), ACM Press,pp. 7485.

[5] CHENG, C.-M., HSIAO, P.-H., KUNG, H. T., AND VLAH, D. Paralleluse of multiple channels in multi-hop 802.11 wireless networks. InMILCOM06: Military Communications Conference (October 2006).

[6] CHENG, C.-M., HSIAO, P.-H., KUNG, H. T., AND VLAH, D. Wsn07-1:Adjacent channel interference in dual-radio 802.11a nodes and its impacton multi-hop networking. In GLOBECOM06: Global Telecommunica-tions Conference. IEEE (San Francisco, CA, USA, 2006), pp. 16.

[7] DRAVES, R., PADHYE, J., AND ZILL, B. Routing in multi-radio, multi-hop wireless mesh networks. In MobiCom04: International Conferenceon Mobile Computing and Networking (2004), pp. 114128.

[8] FUXJAGER, P., VALERIO, D., AND RICCIATO, F. The myth of non-overlapping channels: interference measurements in IEEE 802.11. InWONS07: Fourth Annual Conference on Wireless on Demand NetworkSystems and Service (2007), pp. 18.

[9] GAST, M. 802.11 Wireless Networks: The Denitive Guide, SecondEdition. OReilly Media, Inc., 2005.

[10] GUPTA, P., AND KUMAR, P. R. The capacity of wireless networks.IEEE Transactions on Information Theory 46, 2 (2000), 388404.

[11] HAYKIN, S. Communication Systems. Wiley & Sons, 2001.[12] LIESE, S., WU, D., AND MOHAPATRA, P. Experimental characteriza-

tion of an 802.11b wireless mesh network. In IWCMC06: Proceedingsof the 2006 International Wireless Communication and Mobile Comput-ing Conference (New York, NY, USA, 2006), ACM Press, pp. 587592.

[13] MO, J., SO, H.-S. W., AND WALRAND, J. Comparison of multichannelMAC protocols. IEEE Transactions on Mobile Computing 7 (January2008), 5065.

[14] RAMACHANDRAN, K., SHERIFF, I., BELDING, E. M., ANDALMEROTH, K. A multi-radio 802.11 mesh network architecture.Mobile Networks and Applications (MONET), Special Issue onAdvances in Wireless Mesh Networks (2008).

[15] RAPPAPORT, T. S. Wireless Communications: Principles and Practice,2nd ed. Prentice Hall PTR, December 2002.

[16] ROBINSON, J., PAPAGIANNAKI, K., DIOT, C., GUO, X., AND KRISH-NAMURTHY, L. Experimenting with a Multi-Radio Mesh NetworkingTestbed. In WiNMee05: Proceedings of the First Workshop on WirelessNetwork Measurements (Riva del Garda, Italy, April 2005).

881