Cone of Silence: Adaptively Nulling Interferers in Wireless Networks

Transcription

1 UCL DEPARTMENT OF COMPUTER SCIENCE Research Note RN/1/ Cone of Silence: Adatively Nulling Interferers in Wireless Networks 3 th January 1 Georgios Nikolaidis Astrit Zhushi Kyle Jamieson Brad Kar Abstract Dense 8.11 wireless networks resent a ressing caacity challenge: users in roximity contend for limited unlicensed sectrum. Directional antennas romise increased caacity by imroving the signal-to-interference-lus-noise ratio (SINR) at the receiver, otentially allowing successful decoding of ackets at higher bit-rates. Many uses of directional antennas to date have directed high gain between two eers, thus maximizing the strength of the sender's signal reaching the receiver. But in an interference-rich environment, as in dense 8.11 deloyments, directional antennas only truly come into their own when they exlicitly null interference from cometing concurrent senders. In this aer, we resent Cone of Silence (CoS), a technique that leverages software-steerable directional antennas to imrove the caacity of indoor 8.11 wireless networks by adatively nulling interference. Using in situ signal strength measurements that account for the comlex roagation environment, CoS derives custom antenna radiation atterns that maximize the strength of the signal arriving at an access oint from a sender while nulling interference from one or more concurrent interferers. CoS leverages multile antennas, but requires only a single commodity 8.11 radio, thus avoiding the significant rocessing requirements of decoding multile concurrent ackets. Exeriments in an indoor 8.11 deloyment demonstrate that CoS imroves throughut under interference.

2 Cone of Silence: Adatively Nulling Interferers in Wireless Networks Georgios Nikolaidis Astrit Zhushi Kyle Jamieson Brad Kar University College London UCL CS Research Note RN/1/ fg.nikolaidis, a.zhushi, k.jamieson, ABSTRACT Dense 8.11 wireless networks resent a ressing caacity challenge: users in roximity contend for limited unlicensed sectrum. Directional antennas romise increased caacity by imroving the signal-to-interference-lus-noise ratio (SINR) at the receiver, otentially allowing successful decoding of ackets at higher bit-rates. Many uses of directional antennas to date have directed high gain between two eers, thus maximizing the strength of the sender's signal reaching the receiver. But in an interference-rich environment, as in dense 8.11 deloyments, directional antennas only truly come into their own when they exlicitly null interference from cometing concurrent senders. In this aer, we resent Cone of Silence (CoS), a technique that leverages software-steerable directional antennas to imrove the caacity of indoor 8.11 wireless networks by adatively nulling interference. Using in situ signal strength measurements that account for the comlex roagation environment, CoS derives custom antenna radiation atterns that maximize the strength of the signal arriving at an access oint from a sender while nulling inteference from one or more concurrent interferers. CoS leverages multile antennas, but requires only a single commodity 8.11 radio, thus avoiding the signi cant rocessing requirements of decoding multile concurrent ackets. Exeriments in an indoor 8.11 deloyment demonstrate that CoS imroves throughut under interference. 1. INTRODUCTION Surred by the availability of low-cost commodity radio hardware and freely usable unlicensed sectrum, users have enthusiastically embraced 8.11 wireless networking in home and of ce environments. As these networks roliferate raidly, articularly in oulous urban areas, their deloyment density increases signi cantly. Measurements of 8.11 base station deloyments in major US cities taken in 5 already showed thousands of cases in which four or more 8.11 access oints mutually interfered [1]. As only three non-overlaing channels are available in 8.11b/g, these increasingly dense deloyments ose a wireless caacity challenge hysically roximal networks must share - nite bandwidth. C1 AP1 C AP AP3 C3 Figure 1: Tyical dense 8.11 deloyment in aartments or of- ces. AP1, AP, and AP3 are access oints; C1, C, and C3 are clients. Solid arrows denote desired transmissions; dashed arrows denote unintended interference. Consider a dense deloyment of 8.11 networks on overlaing channels, tyical in today's residential and commercial areas, as shown in Figure 1. The occuant of each of three aartments (or of ces) oerates his own access oint (AP) with a standard omnidirectional antenna, and because of their close roximity, AP1 and AP3 interfere with recetion by AP. In articular, if client C and AP1 are hidden from one another, client C may transmit to its AP, AP, and AP1 may simultaneously transmit to its client C1. Consider C as the sender, AP as the receiver, and AP1 as the interferer. Interference from AP1 will reduce the throughut C achieves at AP. Indeed, more than one interferer may transmit concurrently; e.g., AP3 might transmit to clients of its own, too, further interfering at AP. When multile wireless networks oerated by indeendent, non-cooerating individuals interfere, a receiver in one network derives no bene t from successfully decoding transmissions from an interferer in another network; data from another network are tyically of no interest. Moreover, oerators of these networks do not centrally coordinate transmit schedules, or share decoding information among nodes. Many recent advances in mitigation of interference have targeted environments where one enterrise oerator closely coordinates multile cooerating senders [3, 6], or where concurrent transmissions' contents are all of interest to a single receiver (i.e., where interferers are art of the same network) [1]. In contrast, in this work, we seci cally focus on mitigating interference in the ubiquitous chaotic, noncooerative deloyments described above. Directional antennas hold great romise for imroving throughut on wireless links in such dense, interferencerich deloyments. The throughut achieveable on a link deends on how well the receiver can discern a sender of interest's signal, while distinguishing it from comet- 1

3 ing background noise and interference from other concurrent senders on the signal-to-interference-lus-noise ratio (SINR). The greater the bit-rate at which a acket is transmitted, the greater the SINR with which the receiver must receive the acket in order to decode it successfully. And at a given transmit bit-rate, as SINR increases, bit-error rate (BER) decreases, reducing costly link-layer retransmissions. Extracting the greatest SINR from a receiver's directional antenna entails solving two distinct roblems. First, how can one direct gain toward a sender of interest's signal, thus imroving the strength with which it is received? And second, how can one avoid directing gain toward interfering signals from concurrent transmitters, and thus null interference? A system may indeendently address either or both of these roblems. Solving either increases SINR, and can thus imrove throughut. 1 The relative bene ts to SINR of directing gain toward a sender's signal vs. nulling interference deend heavily on the deloyment scenario. In an interferencerich environment, nulling is vital to achieving the full SINR and throughut imrovements a directional antenna can offer. Lakshmanan et al. offer a technique for maximizing gain toward a sender of interest indoors [5], but this technique does not exlicitly null interferers. Multiath roagation, commonlace indoors, signi cantly comlicates effective use of directional antennas by causing a sender's signal to arrive at a receiver in multile comonents from unredictable bearings, each with a different hase. Figure offers an idealized illustration of this henomenon in a simle toology, where a receiver equied with a directional antenna attemts to receive from sender S while an interferer I transmits concurrently. The solid arrows indicate multile comonents from S while the dashed arrows indicate multile comonents from I. We see that multile comonents arise from re ections of each transmitter's signal (off walls and any of the many other re- ective objects in the indoor setting) that deend not only on the locations of the sender and receiver, but on the timevarying minutiae of the hysical surroundings. In order to maximize SINR successfully, and extract full bene t from directionality, the receiver must con gure its antenna such that high-gain lobes are directed toward the incident bearings of S's signal, while low-gain nulls are directed toward the incident bearings of I's signal. The heavy line undulating about the receiver reresents just such a gain attern for its directional antenna. Radial distance from the center of the attern to this line indicates the gain of the antenna in db in that radial direction, and lobes are oriented toward S's comonents, while nulls are oriented toward I's. Softwaresteerable, hased array antennas, such as the one used in 1 Comlementary arguments aly when a sender transmits with a directional antenna. In this aer, we focus on recetion using directional antennas, though as we discuss in Section 5, we believe the techniques we describe will be useful at senders, too. For simlicity of exosition, we ignore hase in this discussion; we describe the role of hase in detail in Section. S Figure : Examle of multiath roagation between a sender S, inteferer I, and a receiving directional antenna. Solid arrows reresent comonents from S; dashed from I. Boundaries are re ective walls; the triangle reresents a re ective object. this work (described in Section.), allow quick shaing of the gain attern entirely electrically, under software control, without any mechanical motion. While hased array antennas have reviously been used successfully indoors with xed, essentially single-lobe beam shaes [6], such beam shaes cannot maximize SINR as effectively as ones exibly and dynamically customized in accordance with the seci c multile arrival bearings of senders' and interferers' signals. In this aer, we resent Cone of Silence (CoS), a technique for imroving throughut under interference in 8.11 wireless networks. CoS incororates two main techniques: SamlePhase, a method for accurately, robustly, and ef- ciently deriving a custom attern for an AP's antenna that maximizes the signal strength received from a seci c sender or interferer; and Silencer, a method that, given signal-maximizing atterns for a sender and one or more interferers, roduces a single attern that simultaneously nulls the interferers while maximizing signal strength from the sender of interest, thus maximizing SINR and throughut. An evaluation of a rototye of CoS on an indoor 8.11b/g testbed demonstrates that CoS can imrove a sender's throughut under interference over that achieved by an omnidirectional receiver by between 1.6 and 17, and that CoS imroves receive throughut by nulling one or two concurrent interferers. CoS achieves these substantial erformance imrovements while offering the following key roerties: I An AP using CoS can null even uncooerative interferers from which it can receive ackets CoS does not require APs to schedule transmissions collaboratively, as do revious techniques for mitigating interference with directional antennas [6] or multile antennas [3]. Unlike schemes that receive many concurrently transmitted ackets, but require rocessing-intensive full decoding of each one [1, 3], CoS can null multile concurrent interferers using multile antennas connected to only a single commodity 8.11 radio. To our knowledge, CoS's Silencer is the rst such imlementation of a decorrelator [11].

4 . a= mejφ DESIGN a a a7... We begin with a brief overview of the use scenario for CoS, followed by a rimer on the hased array antenna hardware latform on which CoS is built. Thereafter, we next resent SamlePhase, an algorithm for measuring the wireless channel between the CoS AP and other radios. We then resent the design of Silencer, which builds on SamlePhase to simultaneously steer our AP towards associated clients and null one or more interferers. Use Scenario Figure 3: Left: Phocus hased array 8.11b/g antenna with eight elements. Right: simli ed array model. Consider again the toology in Figure. Recall that in order to imrove SINR at a receiving AP equied with a beam-steerable directional antenna, CoS must derive a attern for the antenna that gathers ath comonents from the sender of interest, S, while nulling ath comonents from the interferer, I. CoS goes about that goal in two logical stes: have foreknowledge of when each client will transmit.3 As do others who have roosed interference mitigation techniques for 8.11 networks [3], we envision that an AP would use a TDMA schedule among its own clients to allow it to redict when each client transmits. Because TDMA scheduling among a base station and its clients is a wellunderstood area in the literature the original 8.11 seci cation includes the PCF MAC, a TDMA aroach [4] we assume the availability of a TDMA imlementation in this work, and focus on nulling interference given that a CoSenabled AP can use TDMA to redict which of its clients will send. First, CoS considers the sender of interest and each in- terferer individually. For each such remote transmitter, a CoS AP alies the SamlePhase algorithm, described in Section.3, to derive one receive attern for each remote transmitter that maximizes received signal strength from that remote transmitter alone. Second, for each sender of interest, a CoS AP alies the Silencer algorithm, described in Section.4, whose inut consists of the receive attern that maximizes signal strength at the AP from the sender of interest, as well as one attern for each interferer that does the same for that interferer. Silencer roduces one attern for each sender of interest that nulls all interferers while directing gain to maximize signal strength from the sender of interest.. Hardware Platform We have built the CoS rototye ato the Phocus hased array 8.11b/g antenna [] manufactured by Fidelity Comtech Inc. and shown in Figure 3, left. The array consists of eight elements saced equally on the circumference of a circle, each of which consists of four stacked dioles. A signal rocessing module (Figure 3, right) mixes the signal from or to element k with a comlex gain ak allowing adjustment of its hase fk and magnitude mk during recetion and transmission, resectively. The eight resulting signals are summed and in turn connect to an antenna ort of a standard Atheros AR chiset. The gain and hase alied to each element may be indeendently controlled in software, the hase in single-degree increments between 18 and 18 degrees and the gain in 1% increments between % and 1%. Taken together, the hase shifts and magnitudes con gured for all elements de ne a comlete attern. Changing the array's attern takes aroximately 1 m s. As sulied by the manufacturer, the antenna's software suorts only 17 factory-con gured stock atterns: one omnidirectional, and the remaining 16 high-gain, each of which consists of a single high-gain lobe aroximately 43 degrees in width, ointing toward one of 16 equally saced directions about the antenna's center. The CoS software is CoS must determine the identities of interferers. Doing so for 8.11 transmitters who interfere strongly at the AP is not dif cult; an AP may simly scan the channels that overla its own eriodically, and record the MAC addresses of any senders that occuy the channel heavily. It is recisely the strongest interferers that stand the greatest chance to reduce a sender of interest's throughut to the AP that will be most easily identi ed in this fashion. When receiving from any sender of interest, CoS always nulls toward all interferers that it has identi ed. CoS cannot be certain that an interferer is sending at any given time, and so may needlessly null that interferer. That choice would be roblematic if nulling signi cantly reduced throughut from the sender of interest as a side effect. In Section 3., we resent exerimental evidence to argue that nulling does not do so in effect, that nulling an interferer tends to be safe for the sender of interest. Like any AP, CoS maintains a list of associated clients. CoS stores the Silencer-roduced, throughut-maximizing attern tailored to each associated client, and con gures its directional antenna to the aroriate such attern each time a client transmits to the AP. To do so, however, CoS must 3 It is imortant to note that the AP need only redict transmissions from its own clients as exlained above, CoS assumes that inteferers, who are uncooerative because they are art of other networks, transmit constantly. 3

5 not bound by this restriction; it can con gure each element indeendently to any of the suorted hase and gain values, yielding a vast variety of ossible multi-lobed atterns..3 Measuring the Channel: SamlePhase The rst challenge we face is measurement of the wireless channel from each client to all of the AP's antennas. Re ections off walls and other objects found in the tyical indoor home or of ce mean that indoor wireless networks oerate in the resence of strong multiath re ections, wherein transmissions arrive from multile directions at the AP. In order to roduce the best end-to-end erformance, we argue that a channel measurement algorithm should satisfy the following objectives: Performance: The algorithm should roduce measurements that result in the best throughut. Ef ciency: The overhead of the algorithm should be as low as ossible without sacri cing erformance. Reliability: The algorithm should meet the above two objectives consistently, with low erformance variance, even in challenging wireless environments. In rior work, Lakshmanan et al. [5] have roosed a channel measurement algorithm that we imrove uon here. We exerimentally comare the two algorithms in Section 3.1 and brie y touch uon ef ciency differences between the two algorithms in Section 4. The SamlePhase algorithm. To reduce comlexity, our aroach leverages received signal strength (RSS) readings measured at a client, from ackets sent by the AP. 4 Consider an aroximation of the wireless channel between the k th AP element and the client as a single comlex number of a certain hase q k and magnitude P k. 5 Then based on RSS readings, SamlePhase oututs a set of eight channel measurements: P 1 e jq r1 ; P e jq r ; : : : ; P8 e jq r8, where r is the index of a reference element in the array, and q kl = q l q k is the hase of element k relative to element l. SamlePhase only measures the relative differences between the channel hases q k, since these determine beam shae. SamlePhase measures the individual channel magnitudes from each individual element P k (k = 1 : : :8) directly. Our algorithm transmits 5 contiguous bursts of three robe ackets each from each individual element of the AP to the remote node to which the channel is being measured. The 4 In our rototye, the CoS AP sends measurement robes to all remote nodes, which record RSS measurements and return them out-of-band to the AP. In a roduction deloyment, we envision that the CoS AP would send 8.11 null data frames to its clients to elicit ACKs, and measure RSS on these returning ACKs (in fact, the Phocus array software already imlements this functionality as shied). For interferers, CoS could simly measure RSS on received frames. 5 The 8.11b/g wireless channel is MHz wide and so in fact cannot be comletely characterized by a single comlex value; we touch on this oint in Section 4. bursts are interleaved across elements, so that a eriod of interference imacts just a small number of ackets from any articular element. SamlePhase's hase measurements are based on the following observation about the wireless channel. Suose the hase of the channel from element k to the remote node is some value q k, and the hase of the channel from element l to the remote node is another value q l. Further, suose that the AP transmits data with a hase difference d between elements k and l that we choose and rogram into the AP. Then, by the rincile of suerosition (Lakshmanan et al. rovide a detailed derivation), the ower of the elements' combined transmissions at the client is P kl (d ) = P k + P l + P k P l cos (q l q k + d ) : (1) Rearranging the above, we nd the following: cos (q kl + d ) = P kl (d ) (P k + P l ) P k P l () The above suggests the following way of estimating q kl : using emirically measured values of P k, P l, and P kl (d ), samle the exression on the right-hand side of Equation at one or more values of d. Then, the best estimator of q kl will minimize the sum of squared errors between the emirically measured values from the right-hand side of Equation and comuted values from the left-hand side of the same equation. Samling multile evenly saced values of d removes any hase ambiguity with high likelihood, as exlained in the Aendix A. For simlicity and to overcome ractical limitations on the number of antenna atterns that the Fidelity Comtech array can store at once, SamlePhase uses four evenly saced values of d for its hase measurements. SamlePhase microbenchmarks. Figure 4 shows examle measurements of P kl (d ) for three reresentative element airs on an AP-client link in our testbed (described in Section 3), as we vary d between 18 and 18 degrees. We see the exected sinusoidal relationshi, and note that estimating q kl by using measurements of P kl near the eak or trough of the sinusoid may decrease accuracy, because of the quantization of the samle data and the decreased sloe of the sinusoid near those oints. On the same lots, the vertical lines reresent the eaks of the sinusoid as redicted by the estimators of q kl roduced by the SamlePhase and the Lakshmanan et al. methods. We see that by using multile samle oints, SamlePhase nds the eaks of the sinusoidal data better than the rior method for these reresentative element airs. Mean absolute error Variance of the absolute error SamlePhase L. et al Table 1: Mean and variance of absolute error for SamlePhase and L. et al. methods. Table 1 lists the mean and the variance of the absolute error for both methods. To derive these results, we t the emirical data to the sine wave whose hase minimizes the 4

6 Power (nw) Power (nw) Power (nw) Phase difference (degrees) SamlePhase Emirical L. et al Figure 4: Emirical ower measurements (P kl ) of RSS for three reresentative element airs for an AP-remote node link when airs of AP antenna elements transmit simultaneously with varying hase difference (d ) to a remote node ( Emirical oints). The SamlePhase and Lakshmanan et al. estimates of the eak of the sinusoid aear as vertical lines. Power (nw) Theoretical Phase (degrees) Emirical Figure 5: Emirical data with the theoretical t suerimosed. sum of squared errors (Figure 5). Note that this tting to a sine wave uses many more samles then both SamlePhase and Lakshmanan et al.'s method. We measured the absolute error of each of the two methods by comuting the absolute difference between each method's estimate and the eak of the best- t sine wave. Then looking across all links in our testbed (shown in Figure 8 on. 7), and at all array element airs on every testbed link, we comuted the average absolute error and variance of the absolute error for both SamlePhase and the Lakshmanan et al. method. The results are shown in Table 1. In this table, we see that SamlePhase offers both a lower absolute mean error and a lower variance of absolute error. We conclude that SamlePhase more accurately nds the antenna hase that maximizes the signal strength when two elements send together. In Section 3, we show that SamlePhase furthermore nds better overall beamforming and interference-cancelling atterns than the rior method, as determined by the end-to-end metric of throughut. Beamforming toward a client. Once the AP has measured the channel between itself and a client, it can beamform its transmissions to or recetions from that client by weighting the kth element's inut by the channel measurement to the kth element, P k e jq kr, where r is the reference element chosen during SamlePhase's measurement. This results in cohasing the signals from all antennas so that they align and constructively interfere. The combination of co-hasing and weighting roortional to P k maximizes signal-to-noise ratio (SNR) at the receiver and is known as maximal ratio combining (MRC) in the literature. MRC does not maximize signal to noise lus interference ratio (SINR), however, and so interfering transmissions will imact a beamforming AP's throughut, as we show in Section 3. We therefore seek a way to null interfering clients and maximize SINR..4 Nulling Interferers: Silencer Silencer is an imlementation of a decorrelator [11], a comutational structure that allows distinct signals to be received concurrently. What distinguishes Silencer from other decorrelator imlementations is that Silencer can recover a signal from a sender of interest while nulling other concurrently received signals without decoding these other signals. Using channel measurements from the methods in Section.3, we can reresent the channels to clients as vectors in an eight-dimensional sace (since our AP has eight elements): 3 P1 e jq 1 P e jq h c = 6 7 (3) 4. 5 P8 e jq 8 where the measurements for h c are taken at client c. To null an interferer i (either another AP or an interfering client), the AP measures the channel h i between itself and the interferer, and using the Gram-Schmidt algorithm, comutes a basis in C 8 for the vector sace orthogonal to h i (indicated by V i in Figure 6). Then, Silencer rojects the received signal y onto V i (indicated by Proj Vi (y) in Figure 6). After the interference cancellation ste, Silencer directs gain in the direction of the intended client's channel h c (in the V i vector subsace). If we reresent rojection onto V i with the 8 8 comlex matrix Q i, the overall oeration on the received signal y is therefore (Q i h c ) Q i y. We rogram the AP with the eight-element, comlex-valued vector Q Q i ih c to imlement this oeration. Generalization to multile interferers. Silencer easily generalizes to multile interferers i 1 to i 7, each with a different channel estimate h i1 : : : h i7, by using the Gram-Schmidt rocess to construct a vector subsace orthogonal to the san of all the inteference vectors. In Section 3.4, we resent ex- 5

7 V i h i y Proj V i (y) Figure 6: By rojecting the received signal y onto the vector subsace orthogonal to an interferer's channel h i, Silencer (shown here in R for ease of exosition) nulls the signal from the interfering client. -18 db -1 db -6 db db -18 db -1 db -6 db db h c -18 db -1 db -6 db db Figure 7: Imact of Silencer on antenna gain attern: note that these emirical gures show gain vs. direction, but do not show how the antenna maniulates the hase of the received signals. Left: an MRC gain attern maximizing SNR for a sender. Center: an MRC gain attern maximizing SNR for an interferer. Right: the resulting Silencer gain cancelling the interferer and beamforming toward the sender. erimental results nulling u to two out of three simultaneously transmitting senders. Practical limitations in nulling interferers. Although CoS can theoretically entirely remove interference from u to seven simultaneous interferers, several ractical design issues limit real-world system erformance: Hearing the interferer. In order to comute h i, CoS needs to receive a suf cient number of ackets from interferer i. This recludes nulling the most distant interferers, since our commodity hardware detects ackets only down to 94 dbm. Nonetheless, we show in Section 3. that CoS can even null interferers with PRRs to the AP as low as 9%. Estimating the channel to the interferer. The accuracy of the channel estimation algorithm will imact the degree to which CoS can beamform towards clients and null interferers. For OFDM modulations, since CoS measures and beamforms across all OFDM subcarriers simultaneously, it does not cature inter-subcarrier differences in the MHz WLAN channel. This is a ractical tradeoff: measuring inter-subcarrier differences would require a software-de ned radio, or a PHY interface that returns er-subcarrier RSS readings. CoS sacri ces some erformance for the simlicity of running on a commodity hardware latform and the seed of only requiring three C 8 matrix multilications and a Gram-Schmidt iteration in order to comute a beamforming attern that nulls a new interferer. Adating when an interferer ceases sending. When CoS nulls an interferer that has since ceased transmission, it sacri ces some amount signal ower that would have imroved the overall SINR had it not nulled that interferer. In Section 3. we quantify the throughut imact of nulling towards an interferer desite that interferer's not transmitting. The degree of similarity between the client's channel and the sender's channel. The more orthogonal the sender of interest's channel is to each of the interferers' channels, the less of the sender's signal Silencer will null along with those of the interferers. Fortunately, since CoS uses eight antennas and comlex-valued channel vectors, there are many more degrees of freedom than the two shown in Figure 6. We examine how well CoS can null interferers endto-end in Section 3. The time between channel estimation and interference cancellation. One imortant concern is how much time the beam-steering and inteference-cancelling atterns we derive last, because their longevity, together with the time needed to measure the channel, determine CoS's overhead. In Section 3.3 we measure attern lifetime. 3. EVALUATION There are several key erformance questions surrounding indoor interference nulling. First, by how much does SamlePhase increase received signal ower and throughut on a single link? Next, how well does Silencer null interference and allow that same link to function in the resence of an interferer? For both of the receding questions, how long do the atterns derived last? Finally, how many simultaneous interferers can CoS null? In this section, we answer these questions using exeriments in a tyical indoor of ce environment, on the 13-node testbed shown in Figure 8. Exerimental setu. Our testbed consists of three Phocus hased-array antenna nodes on which we run CoS and 1 Soekris nodes each equied with a single omnidirectional antenna. All nodes use Atheros 5413 WiFi cards and the madwi driver under Linux. Soekris nodes use madwi v.9.4, whereas the hased arrays use madwi v.9..1, including atches from the OenWRT roject (for back-orted bug xes) and Fidelity Comtech (for antenna hase and gain control functionality). To exlore many different toologies, we use Soekris nodes to transmit as either senders or interferers. Notionally, these may be thought of as omnidirectional APs. Methodology. Our exeriments run in channels one (.41 MHz) and six (.437 MHz) of the.4 GHz ISM band. Using a WiSy 6 dbx sectrum analyzer to monitor the entire.4 GHz sectrum, we measured the noise oor of the network at 94 dbm throughout our exeriments. We also veri ed 6 htt://metageek.net 6

8 Cumulative fraction of ackets SamlePhase L. et al Highgain Received Power (dbm) Omni Scaled Highgain Scaled Omni Figure 8: The indoor of ce environment and wireless network toology for the exeriments in this aer. Filled dots reresent nodes with a single omnidirectional antenna and hollow dots reresent nodes with a hased-array antenna. the resence of light background traf c from one other network being received at an average 9 dbm (measured at the middle of the testbed in Figure 8) on channel six and the occasional resence of background traf c on channel one. Senders in our throughut exeriments send 15-byte UDP ackets. For the auto bit-rate exeriments, we enable bit-rate selection at senders using the madwi imlementation of the SamleRate algorithm. Exeriments roceed by measuring the throughut of each of the atterns evaluated over 3 seconds. Unless otherwise stated below, when we comare different antenna atterns, we normalize total antenna gain, running SamlePhase and using its total radiated ower as a reference ower level, and scaling the Silencer, directional, and omnidirectional atterns to emit the same total ower. We label the latter two Scaled Highgain and Scaled Omni, resectively. To ut our erformance into ersective, we also comare against omnidirectional ( Omni ) and stock high gain ( Highgain ) atterns with the highest antenna gain con gurable by the user:.15 dbi for omnidirectional 7, and 15 dbi with a 43 beam width for high gain. Unless otherwise stated, senders and interferers in the exeriments below transmit at full ower (18 dbm). Table gives a roadma for the key exeriments we resent in this section, and the erformance gains they acheive. 3.1 Beamforming with SamlePhase We rst examine how well SamlePhase imroves throughut at the receiver comared to the Lakshmanan et al. method and simle omnidirectional atterns. We also determine whether SamlePhase's measurements of the chan- 7 The vendor does not rovide a gure for gain relative to an isotroic antenna. The gure above is based on the assumtion that the antenna in omnidirectional mode acts as a half-wavelength diole. Figure 1: RSS distributions of ackets drawn from xed bit-rate (54 Mbit/s) exeriments on testbed link A. nel yield any throughut imrovement as comared with the highest-throughut attern among the manufacturer's xed high-gain atterns. We determine which high-gain attern among the 16 ossible such atterns offers the greatest throughut by emirical measurement. To do so, we interleave two iterations through the 16 directional atterns, each saced equally around the 36 circle at :5 increments. We test each attern for 3 seconds, comleting the exeriment in 16 minutes. We ick the stock high gain attern that yields the best overall average throughut, and comare SamlePhase against this attern, both at full array ower (7 dbm) and scaled on a er-link basis to use the same total ower as the SamlePhase attern. In a toology with the AP A receiving, and a single sender S sending, we comare receive throughuts at the AP using a SamlePhase-derived attern steered toward S, the best high-gain attern for S, the scaled best high-gain attern for S, an omni attern, and a scaled omni attern. For many different two-node (S; A) toologies, we measure receive throughut at A for the above atterns. Each measurement we reort in Figure 9 is the mean of 1 oneminute measurements with error bars reresenting 95% con- dence intervals. We re-run the otimization at the start of each one-minute measurement interval. Figure 9 resents the main SamlePhase throughut result, in which we comare atterns generated by the SamlePhase and the Lakshmanan et al. measurement methods, described in Section.3, with the high gain and omnidirectional antenna atterns described above. From the gure, we see that in the absence of interference, SamlePhase offers the greatest throughut. Two factors exlain this throughut imrovement. The rst is that SamlePhase derives atterns that maximize RSS better than other methods. In Section.3, we resent microbenchmarks that show that SamlePhase derives atterns that maximize element-airwise RSS better than cometing methods. In order to show that SamlePhase's atterns im- 7

9 Exeriment Conclusion or erformance imrovement Discussed in SamlePhase throughut Over many links sending one at a time in our testbed, SamlePhase imroves 3.1 throughut over an omnidirectional attern by %, imroves throughut over the best directional attern by %, and imroves throughut over the Lakshmanan et al. method by %. Silencer throughut Over many testbed links with an interferer laced near the receiver of each, 3. CoS imroves throughut over an omnidirectional attern by 4 113% and imroves throughut over the best stock directional attern by 31 %. Silencer can also null traf c on adjacent WiFi channels. Longevity of atterns In a busy indoor of ce environment, CoS atterns work effectively for on the 3.3 order of 1 hours during quiet times and two hours during busy times. Nulling many interferers Over many testbed links, Silencer can null two simultaneous interferers (a total of three concurrent transmissions), acheiving throughut gains of over the best omnidirectional attern with the same interference. Nulling distinct interferers yields additive throughut gains. Table : Summary of exerimental results for the roosed techniques, and the corresonding conclusion or erformance imrovement. Throughut (Mbs) A 1 C 6 C 6 A 5 B 3 C SamlePhase L. et al Highgain Omni Scaled Highgain Scaled Omni Figure 9: Emirically measured throughut for six different antenna gain atterns across six different testbed links (labeled by sender identi er and access oint identi er: cf. Figure 8 on. 7). In the absence of interference, SamlePhase offers the greatest throughut because of the accuracy of its channel measurement method. rove RSS over other antenna atterns, we ran a xed bitrate microbenchmark over a link with high acket delivery rate at that bit-rate. This microbenchmark eliminates a samling bias favoring strongly received ackets that we would introduce if we examined RSS readings from the auto bitrate exeriments in this section. In Figure 1 we see that these airwise gains in RSS translate into imroved RSS for the attern as a whole, which increases the SNR of received ackets, making successful decoding more likely. The second henomenon that exlains the throughut imrovement in Figure 9 is that as a result of incurring fewer bit errors, SamleRate, the bit-rate adatation scheme, chooses to use higher rates when using SamlePhase-derived atterns. For a reresentative link in our testbed, we examined SamleRate's data structures over a 3-second reresentative eriod in the middle of our throughut exeriment. Figure 11 shows the fraction of ackets that SamleRate chooses to send at each bit rate over this link. From Figure 9, we see that with the SamlePhase attern, SamleRate chooses the to bitrate (54 Mbit/s) for slightly more than 4% of all ackets it sends and either of the to two highest (54 and 48 Mbit/s) for a total 8% of all ackets sent. None of the other measurement methods chooses the to two bitrates for more than 4% of all ackets sent. Omnidirectional atterns use 4 Mbit/s and lower bit-rates for half of all ackets. 3. Nulling Interferers with Silencer 8.11 networks oerate at relatively high SNRs, so the cause of oor erformance is often interference. In this exeriment, we test how well Silencer nulls a single interferer, the most common case in a light to moderately loaded network. In a toology with an AP A receiving, a sender S sending, and an interferer I interfering, we measure the imrovement in receive throughut at the AP using a Silencer attern comared to those of a SamlePhase-derived attern steered toward S, the best high-gain attern for S, the scaled best high-gain attern for S, an omni attern, and a scaled omni attern. For several three-node (S; I; A) toologies, we measure receive throughut at A for the above atterns. Each measurement we reort in Figure 1 is the mean of 1 one-minute measurements with error bars reresenting 95% con dence intervals. We re-run the otimization for each one-minute 8

10 Fraction of ackets Bit rate Omni Scaled Omni Highgain Scaled Highgain L. et al. SamlePhase Figure 11: Bit-rates chosen by SamleRate during a reresentative 3-second interval at link 6A (cf. Figure 9). SamlePhase chooses signi cantly higher bit rates than other methods, choosing the to two bit-rates (48 and 54 Mbit/s) 8% of the time. Toology Sender (dbm) Interferer (dbm) 3 A C B B C B A C A 6* 1 18 Table 3: Single-interferer exeriments: ower levels used at sender and interferer. measurement interval. The limited hysical extent of our testbed makes it dif cult to lace sender-interferer airs such that neither senses the other's carrier. Therefore, we use the TX_STOMP register of the Atheros chiset to run exeriments with carrier sense turned off at both S and I. Doing so yields more exibility to try different ower levels (shown in Table 3) at the sender and interferer, to more broadly exlore how Silencer erforms at the AP. Because the ath between sender and interferer is distinct and indeendent of the sender-ap and interferer-ap aths, it is reasonable to turn off carrier sense in order to emulate toologies in which the sender and interferer are mutually hidden terminals. In all exeriments, the interferer sends broadcast ackets at 54 Mbit/s. On the nal link (labeled 3A6 ) in Figure 1, we test the ability of Silencer to null interference on an adjacent channel. For these data oints, we tune the interferer's radio to WiFi channel 7, leaving the sender on channel 6. We allow SamlePhase to send measurement robes on channel 7 to the interferer, and on channel 6 to the sender. We then offer the resulting atterns as inuts to Silencer in the usual way. Once Silencer has generated a attern that nulls the interferer, we tune the AP to channel 6. We note that Silencer still effectively nulls interference and increases the throughut of the link, desite combining atterns generated by SamlePhase on adjacent channels. Penalty associated with nulling interferers. Because CoS always nulls inteferers of which it is aware, but interferers do not send during every acket time, there may be an oortunity cost associated with needless nulling (of an interferer not sending). That is, nulling an interferer may collaterally also reduce the signal strength from the sender of interest. To evaluate whether such an effect noticeably reduces throughut, we erformed an exeriment in which we comared the throughut achieved by a attern derived with SamlePhase with that of a attern derived using Silencer. In the latter case, we used the same sender as in the former, with the addition of an interferer in order to calculate the Silencerderived attern. We then measured the throughut achieved by the same sender in the absence of interference when receiving using these two atterns. The result was a.1% reduction in throughut for Silencer, suggesting that nulling a non-active interferer may not reduce a sender's throughut signi cantly. 3.3 Longevity of Interference-Nulling Patterns In this exeriment, we ask the following question: How long can we reasonably exect to be able to use a throughut-maximizing attern before changes in the channel cause the attern's erformance to degrade signi cantly? In other words, how do received ower and throughut evolve over time when a CoS-enabled AP receives after deriving beam atterns using SamlePhase and Silencer? 9

11 Throughut (Mbs) A 6 7 C 8 5 B 4 B 1 7 C 1 5 B 1 7 A 8 3 A 6 * 3 C 6 Silencer SamlePhase Highgain Omni Scaled Highgain Scaled Omni Figure 1: Emirically measured throughut in the resence of one interferer, for six different antenna gain atterns across six different testbed links (labeled by sender identi er, access oint identi er and interferer identi er: cf. Figure 8 on. 7). By nulling, Silencer yields the best end-to-end throughut in the resence of interference. For one link on which Silencer rovides a throughut gain over omni and high gain atterns, we receive using both the Silencer-derived and high gain directional atterns. In Figure 13 we lot a time series of the thirty-second moving average of throughut. We interleave throughut measurements of Silencer, a SamlePhase attern that does not attemt to null, the high gain directional attern that achieves the highest throughut on the link, the omnidirectional attern and ower-normalized versions of the last two. We can see that the attern derived by Silencer outerforms all the others consistently for about 45 minutes, after which its erformance degrades roughly to theirs. 3.4 Nulling Multile Interferers Since multile simultaneous interferers are common in densely deloyed wireless networks, Silencer's erformance when more than one interferer sends concurrently is aramount. We now answer this question exerimentally. For ve links in our testbed, we set u two interferers and a sender, for a total of three senders, each of which transmits ackets simultaneously, as fast as ossible, with carrier sense disabled as in the single-interferer exeriments in Section 3.. Over all links evaluated in this exeriment, the sender transmits at 18 dbm ower and the interferers transmit at 1 dbm. We comare the throughuts of a CoSenabled AP in omnidirectional mode, Silencer nulling one of the interferers, and Silencer nulling both interferers. The results of this exeriment aear in Figure 14. We see that Silencer nulling one interferer while not taking into account the other offers substantial throughut gains over an omnidirectional attern, doubling throughut on some links. Furthermore, when Silencer measures the channel to both interferers and nulls both, it achieves additional substantial throughut gains over Silencer nulling just one interferer. 4. RELATED WORK Beamforming shaing the transmit or receive atterns of a multi-antenna array to maximize signal strength between a sender and receiver is a well-known communications technique for multiath communication channels, but has only recently been investigated in local area wireless networks. In contrast, base stations in mobile telehone networks erform transmit beamforming on the downlink and receive beamforming on the ulink in order to multilex transmissions. This is made ossible in art by well-lanned cellular toologies where interference is carefully managed. In local-area wireless and mesh networks, recent MIMO 8.11n chisets erform receive beamforming. As these ASIC-based imlementations have direct access to hysicallayer information, they can estimate the channel for each OFDM subcarrier indeedently, and beamform indeendently for each subcarrier. That aroach allows more effective maximization of received signal strength than either SamlePhase or the earlier method of Lakshmanan et al. can achieve, as these latter two techniques only observe channel measurements from a commodity 8.11 card's er-acket RSSI measurements, which are averaged across all OFDM subcarriers. Ruckus Wireless, a startu comany, manufactures Zone- Flex APs that erform transmit beamforming to maximize receive signal strength at clients. While the algorithmic details of the techniques used by these roducts are rorietary, an examination of the marketing literature on the comany's web site [8] strongly suggests they do no exlicit nulling of any kind, neither in the transmit nor receive direction. DIRC [6] increases indoor network caacity by having APs transmit directionally, but always receives using only an omnidirectional antenna attern. DIRC further is intended for use in an enterrise setting where all wireless infrastructure is controlled by one authority, as it centrally schedules all APs' transmissions to avoid causing interference. By contrast, CoS is intended to mitigate interference in dense chaotic deloyments, where APs run by noncooerating users interfere. While based on the same Phocus array antenna latform as CoS, DIRC's APs erform no beamforming of any kind when they transmit neither to maximize signal strength nor to null. Instead, DIRC's APs transmit using 16 xed, manufacturer-sulied high-gain 1

12 Throughut (Mbs) : 13:1 13: 13:3 13:4 13:5 14: 14:1 14: 14:3 Silencer SamlePhase Highgain Omni Scaled Highgain Scaled Omni Figure 13: Longevity of a Silencer interference-nulling attern. We run the SamlePhase measurement method only once, at the beginning of this time series. Then, as the wireless channel changes, we measure throughut in the resence of a continuous interferer. We see that the Silencer attern's throughut does not degrade signi cantly for aroximately 45 minutes. atterns, each with one main lobe and several side lobes. These atterns are identical to the high-gain atterns used in the evaluation of CoS. CoS and DIRC are comlementary: because DIRC identi es APs whose transmissions interfere at receivers and revents those APs from transmitting concurrently, we believe that DIRC could otentially exerience imroved transmit concurrency by nulling toward unintended receivers during transmission, using atterns roduced by CoS's Silencer. The Phocus array antenna has also found use for multicast [9] and vehicular network access [7]. Like DIRC, both these systems use only the manufacturer's 16 xed, highgain atterns; neither does any beamforming or nulling. Sace-Division Multile Access (SDMA) [11], which allows a receiver equied with multile antennas and multile radios to receive multile concurrent ackets successfully, has been a toic of investigation in the communications theory community since the late 199s, and has been alied to mobile telehone base stations in the ast decade. More recently, Tan et al. [1] have combined SDMA with successive interference cancellation (SIC) [11] for use with acketized data, and successfully decoded two concurrent transmissions in a dual-antenna MIMO 8.11 receiver imlemented ato a software radio latform. (They also describe techniques intended to decode more than two concurrent transmissions.) SDMA is a good t in settings where a base station must receive data from many clients. In the dense, chaotic environments CoS targets, where APs run by different users interfere, however, data from an interferer on one network is of no interest to users on another, and the comutational cost of decoding many ackets is unattractive. CoS instead ots for an aroach that nulls multile interferers with eight antennas and a single radio, decoding only the acket from the sender of interest. Gollakota et al. [3] describe a MIMO interference mitigation system in which APs align their concurrent transmissions in time, and aly SIC by exchanging information over a wired Ethernet. These techniques t the case in which all interfering APs are controlled by a single authority, but are ill-suited to today's commmon dense 8.11 deloyments, where many indeendent administrators oerate APs that intefere. Again, CoS nulls interfering APs without requiring coordination among them, and thus is well suited to dense deloyments of uncoordinated 8.11 networks. Finally, Lakshmanan et al. [5] imlement transmit beamforming using MRC for the Phocus hased array antenna, the identical exerimental latform we use for CoS. Their technique insired SamlePhase, but it does not exlicitly null interferers; i.e., their technique includes no functionality analogous to CoS's Silencer. SamlePhase uses a randomly chosen base element in its measurements to avoid consistently using any one base element that cannot successfully transmit to a client. The method of Lakshmanan et al. always uses the same base element, and thus is vulnerable to the aforementioned athology. One subtle but signi cant ractical difference between SamlePhase and Lakshmanan et al.'s method concerns the comlexity of the two algorithms' measurements. Essentially, SamlePhase can statically comute all antenna atterns used during its measurements for deriving an MRC attern, while the method of Lakshmanan et al. cannot. This difference arises because the latter method cannot analytically determine the correct sign of the hase difference between two elements in an MRC attern; it must instead exerimentally comare the received ower levels achieved with hase differences of oosite sign. The Lakshmanan et al. method therefore requires two sequentially deendent measurement stes, the second of which uses antenna atterns known only after the rst ste. Because the Phocus array latform incurs signi cant delay when being con gured with new atterns (on the order of 5 ms for each attern con gured), the extra measurement ste of the Lakshmanan et al. method is a significant cost. 8 As a consequence, SamlePhase will in many cases be able to roduce an MRC receive attern using mea- 8 Note that this delay is only for con guring a newly de ned attern into the array; switching among reviously con gured atterns takes only 1 ms. 11