AirSync: Enabling Distributed Multiuser MIMO with Full ...

AirSync: Enabling Distributed Multiuser MIMO with FullSpatial Multiplexing

Horia Vlad Balan Ryan Rogalin Antonios Michaloliakos Giuseppe Caire Konstantinos Psounis

USC USC USC USC USC

[email protected] [email protected] [email protected] [email protected] [email protected]

ABSTRACTThe enormous success of advanced wireless devices is push-ing the demand for higher wireless data rates. Denser spec-trum reuse through the deployment of more access pointsper square mile has the potential to successfully meet the in-creasing demand for more bandwidth. In theory, the best ap-proach to density increase is via distributed multiuser MIMO,where several access points are connected to a central serverand operate as a large distributed multi-antenna access point,ensuring that all transmitted signal power serves the purposeof data transmission, rather than creating “interference.” Inpractice, while enterprise networks offer a natural setup inwhich distributed MIMO might be possible, there are seriousimplementation difficulties, the primary one being the needto eliminate phase and timing offsets between the jointly co-ordinated access points.

In this paper we propose AirSync, a novel scheme whichprovides not only time but also phase synchronization, thusenabling distributed MIMO with full spatial multiplexinggains. AirSync locks the phase of all access points usinga common reference broadcasted over the air in conjunctionwith a Kalman filter which closely tracks the phase drift. Wehave implemented AirSync as a digital circuit in the FPGAof the WARP radio platform. Our experimental testbed, com-prised of two access points and two clients, shows that AirSyncis able to achieve phase synchronization within a few de-grees, and allows the system to nearly achieve the theoret-ical optimal multiplexing gain. We also discuss MAC andhigher layer aspects of a practical deployment. To the bestof our knowledge, AirSync offers the first ever realization ofthe full multiuser MIMO gain, namely the ability to increasethe number of wireless clients linearly with the number ofjointly coordinated access points, without reducing the perclient rate.

Categories and Subject DescriptorsC.2.2 [Computer System Organization]: ComputerCommunication Networks

General TermsDesign, Experimentation, Performance

KeywordsWireless, Virtual MIMO, Software Radios, Synchroniza-tion

1. INTRODUCTIONThe enormous success of advanced wireless devices

such as tablets and smartphones is pushing the demandfor higher and higher wireless data rates and is causingsignificant stress to existing networks. While new stan-dards (e.g., 802.11n and 4G) are developed almost everycouple of years, novel and more radical approaches tothis problem are yet to be tested. The fundamentalbottleneck is that wireless bandwidth is simply upperbounded by physical laws, in contrast to wired band-width, where putting new fiber on the ground has beenthe de-facto solution for decades. While advances innetwork protocols and modulation and coding schemeshave managed relatively modest improvements, denserspectrum reuse, that is placing more access points persquare mile, has the potential to successfully meet theincreasing demand for more bandwidth. However, verydense infrastructure deployments cannot be carefullyplanned and managed for reasons pertaining to scaleand cost. Therefore, the denser the deployment, thelarger the interference among different access points.Eventually the system becomes interference-limited andwe are back to square one.

In theory, the ultimate answer to this problem isdistributed multiuser MIMO (also known as “virtualMIMO”), where several (possibly multi-antenna) accesspoints are connected to central servers and operate as alarge distributed multi-antenna base station. When us-ing joint decoding in the uplink and joint precoding inthe downlink, all transmitted signal power is useful, asopposed to conventional random access scenarios (e.g.,carrier-sense) which waste power through interference.This approach is particularly suited to the case of an en-terprise network (e.g., a WLAN covering a conferencecenter, an airport terminal or a university), or to thecase of clusters of closely spaced home networks con-nected to the Internet infrastructure through the samecable bundle.

1

Figure 1: Enterprise Wifi and DistributedMIMO. Multiple access points connected to a cen-tral server through Ethernet (red lines) coordinate theirtransmissions to several clients by using distributedMIMO.

However, distributed multiuser MIMO is regarded to-day mostly as a theoretical solution because of someserious implementation hurdles, such as the ability toeliminate phase and timing offsets between jointly coor-dinated access points and the ability to perform efficientjoint encoding at a central server linked to the accesspoints through wired links of limited capacity.

We consider a typical enterprise network as illustratedin Figure 1. Since in such networks the wired links con-necting the access points are fast enough to allow forefficient joint processing at a server, the major obstacleto achieve the full potential of distributed MIMO gainsis eliminating the phase offsets between the different ac-cess points. The perceived difficulty of this task has ledsome researchers to believe that it is practically impos-sible to achieve full spatial multiplexing in the contextof distributed MIMO. In this paper, we present the first(to the best of our knowledge) real-world testbed imple-mentation which achieves the theoretical optimal gainby removing, in real time, the phase offsets between ge-ographically separated access points. We achieve thisvia AirSync.

AirSync is a novel scheme which provides not onlytime but also phase synchronization between access points.In a nutshell, AirSync locks the phase of all accesspoints using a common reference broadcasted over theair in conjunction with a Kalman filter which closelytracks the phase drift between the different oscillators.We have implemented AirSync as a digital circuit in theFPGA of the WARP radio platform. We have also im-plemented Zero-Forcing Beamforming, a physical layerprecoding scheme for multiuser downlink transmission,and investigated the practical requirements of optimalMAC-layer schemes. Finally, we have shown in a testbedconsisting of four WARP radios, two acting as accesspoints connected to a central server and two acting asclients, that the theoretical optimal gain of multiuserMIMO is achievable in practice. We argue later in the

paper that this result will extend to an increasing num-ber of access points as long as there is enough spatialdiversity in the propagation environment. This is some-thing that depends entirely on the richness of the physi-cal channels and has nothing to do with the distributednature of our MIMO system and AirSync. While re-cently there have been a number of very interesting andimportant works in which some of the gains of multiuserMIMO have been shown (see Section 2 for more details)none of these has managed to achieve phase synchro-nization between remote transmitters and thus they allfall short of the optimal gains in the distributed senderscenario.

In summary, the contributions that we make in thispaper are the following:

• We introduce AirSync, the first (to the best of ourknowledge) scheme which achieves phase synchro-nization in a distributed multiuser MIMO setting.

• We implement AirSync as a digital circuit in theFPGA of the WARP platform.

• We showcase in a testbed consisting of 4 WARPradios that, thanks to AirSync, the theoreticallyoptimal spatial multiplexing gain is achievable inpractice.

• We discuss practical implementation aspects of thetheoretically optimal MAC schemes to be used inconjunction with our distributed MIMO system.

We conclude this introduction by providing a briefoutline for the rest of the paper. In Section 2 we dis-cuss in detail related work both on the theoretical side(information theory) and on the practical side (softwareradio implementations). In Section 3 we use a theo-retical approach to show why phase synchronization isneeded to achieve the promised gain, and describe, ingeneral terms, AirSync. In Section 4 we present thehardware implementation of AirSync in detail. In Sec-tion 5 we present a number of results obtained usingour testbed implementation with two access points andtwo clients. We show results regarding the synchroniza-tion accuracy, the beamforming gain, the Zero-Forcingprecision and the multiuser multiplexing gain of the sys-tem. The following section mentions theoretically opti-mal MAC schemes and efficient approximations as wellas their practical realizations. Finally, Section 7 dis-cusses a number of challenging yet promising topics thatwe plan to explore in the future, namely the use of rate-less codes for flexible dynamic scheduling and implicitrate allocation, as well as possible alternative non-linearmultiuser precoding schemes.

2. RELATED WORKThe pioneering papers by Foschini [13] and Telatar

[31] have shown that adding multiple antennas both to

2

the transmitter and to the receiver increases the capac-ity of a point-to-point communication channel. At prac-tical medium-to-high Signal to Noise Ratios (SNRs),this gain manifests as a multiplicative factor equal tothe rank of the matrix representing the transfer func-tion between the transmit and the receive antennas. Forsufficiently rich propagation scattering, with probability1 this factor is equal to min{Nt, Nr}, where Nt and Nrdenote the number of transmit and receive antennas, re-spectively. The MIMO capacity gain can be interpretedas the implicit ability to create min{Nt, Nr} “parallel”non-interfering channels corresponding to the channelmatrix eigenmodes, and it is referred to in the litera-ture as multiplexing gain, or as the degrees of freedom ofthe channel. Subsequently, Caire and Shamai [4] haveshown that the MIMO broadcast channel, where thetransmitter has Nt antennas and serves K clients withNr antennas each, exhibits an analogous capacity fac-tor increase of min{Nt,KNr}, suggesting that a trans-mitter with multiple antennas could transmit simulta-neously on the same frequency to independent users.Such multiuser communication has two additional re-quirements. First, precoding of the transmitted datais needed to prevent the different spatial streams frommutually interfering. Second, the transmitter requiresaccurate knowledge of the channel matrix (channel stateinformation) in order to realize this precoding.

The idea of precoding has spurred research beyondthe scope of this paper. Dirty Paper Coding (DPC) [8]with a Gaussian coding ensemble achieves the capac-ity of the MIMO broadcast channel [36], but is diffi-cult to implement in practice. The well-known linearZero-Forcing Beamforming (ZFBF) achieves the samehigh-SNR capacity factor increase, with some fixed gapfrom optimal that can be reduced when the number ofclients is large and the transmitter can dynamically se-lect the clients to be served depending on their channelstate information [19, 38]. A number of other precod-ing strategies, e.g., lattice reduction, regularized vec-tor perturbation and generalized Tomlinson Harashimaprecoding, have been studied and the interested readeris referred to [29] and references therein. For the pur-poses of this paper ZFBF will be the primary methodof interest because of its conceptual simplicity and goodcomplexity/performance tradeoff.

Generalizing the idea of the MIMO broadcast chan-nel, a vast literature has investigated distributed mul-tiuser MIMO systems where several access points areconnected to a common central processor through somebackbone wired network and coordinate their signals tojointly serve a number of clients. If the backbone net-work has a sufficiently high bandwidth, the problem isconceptually identical to that of a single (distributed)multiple antenna terminal and therefore the same tech-niques of the MIMO broadcast channel can be applied.

However, distributed multiuser MIMO presents severaladditional and non-trivial practical implementation prob-lems related to the synchronization (phase and timingstability) of the separate coordinated access points, thatneed to maintain a very tight synchronization in or-der to be able to coherently precode (e.g., beamform)the signals to the clients without creating unacceptablemultiuser self-interference.

A number of recent system implementations have madeforays into the topics of multiuser MIMO transmissionand distributed, frame aligned OFDM transmission. Thebenefits of using ZFBF as a precoding scheme have beenexamined in [2], in a system which consists of a sin-gle access point with multiple antennas hosted on thesame radio board. The use of interference alignmentand cancellation as a precoding technique, which doesnot require frame alignment or phase synchronization,has been illustrated in [15]. While this solution achievesa part of the potential spatial multiplexing gain, in or-der to realize the full spatial multiplexing with standardprecoding techniques it is required to have tight phasesynchronization [20,34]

Simultaneous OFDM signal transmissions which arenot separated in the spatial domain require precise framealignment to maintain their frequency orthogonality.Two signals whose frame boundaries misalign by morethan a cyclic prefix length cannot be reliably decodeddue to interference leakage over the frequency domainduring the decoding process. Frame alignment was usedin SourceSync [25] in conjunction with space-time blockcoding in order to provide a diversity gain in a dis-tributed MIMO downlink system. In Fine-Grained Chan-nel Access [30], a similar technique allows for multipleindependent clients to share the frequency band in fineincrements, without a need for guard bands, resultingin a flexible OFDMA (OFDM with orthogonal multipleaccess) uplink implementation.

3. SYNCHRONIZATION IN DISTRIBUTEDMIMO SYSTEMS

OFDM and Zero-Forcing Beamforming. OFDMhas become the preferred digital signaling format inmost modern broadband wireless networks, includingWLANs IEEE 802.11a/g/n and 4G cellular systems. Itsmain characteristic is that it decomposes a frequencyselective channel into a set of N parallel narrowbandfrequency-flat channels, where the number of frequencysubcarriers N is a system design parameter. In a mul-tiuser environment it has also a significant side advan-tage: as long as the different users’ signals align intime with an offset smaller than a guard time inter-val called the cyclic prefix (CP), their symbols afterOFDM demodulation will remain perfectly aligned onthe time-frequency grid. In other words, the timingmisalignment problem between user signals, which in

3

single-carrier systems creates significant complicationsfor joint processing of overlapping signals (e.g., mul-tiuser detection [35], successive interference cancella-tion [32], Zig-Zag decoding [14]), completely disappearsin the case of OFDM, provided that all users achieve arather coarse timing alignment within the CP. 1

In a point-to-point MIMO link with Nr receive anten-nas and Nt transmit antennas, the time-domain chan-nel is represented by an Nr ×Nt matrix of channel im-pulse responses. Thanks to OFDM, we can think of thechannel in the frequency domain, such that the chan-nel transfer function is described by a set of channelmatrices of dimension Nr × Nt, one for each of the NOFDM subcarriers. Because signals add linearly overthe shared medium, the signal received at each client an-tenna is a linear combination of the signals sent from theaccess point’s antennas. The receiver, having knowledgeof the channel coefficients, is tasked with solving a linearsystem of Nr equations with Nt unknowns from whichit can generally recover up to min(Nr, Nt) transmittedsymbols (this is multiplexing gain, or degrees of free-dom). Since the OFDM modulation breaks the spec-trum into narrow subcarriers, this process is repeatedon each independent subcarrier.

In contrast to point-to-point MIMO, in multiuser MIMOthe receiver antennas are spatially separated and re-ceivers are not generally able to communicate with oneanother. While before the receiver could find the sentsymbols by just solving a set of linear equations, noweach receiver has only one equation with several un-knowns. In order to be able to solve for the variableof interest (the symbol intended for that receiver), wearrange that the contributions of all other unknownscancel each other out in its particular equation. Oneof the techniques to achieve this is linear Zero-ForcingBeamforming (ZFBF).

In ZFBF, the transmitter multiplies the outgoing sym-bols by beamforming vectors such that the receivers seeonly their intended signals. For instance, let the re-ceived signal on a given subcarrier at user k be givenby

yk = hk,1x1 + hk,2x2 + · · ·+ hk,NtxNt + zk (1)

where hk,j is the channel coefficient from transmit an-tenna j to user k and zk is additive white Gaussiannoise. Then, the vector of all received signals can bewritten in matrix form as

y = Hx + z (2)

where H has dimension K×Nt, K denoting the numberof single-antenna clients. Assuming K ≤ Nt, we wish to1Notice that typical CP length is between 16 to 64 timeslonger than the duration of an equivalent single-carrier sym-bol. For example, for a 20 MHz signal, as in standard802.11g, the time-domain symbol interval is 50 ns, so that atypical CP length ranges between 0.7 and 3.2 µs.

find a matrix V such that HV is zero for all elements ex-cept the main diagonal, that is HV = diag(λ1, . . . , λK).When this occurs, then

y = HVx + z = diag(λ1, . . . , λK)x + z, (3)

assuring that each receiver k will see yk = λkxk + zk,which is an independent channel with no interference.

When H has rank K (which is true with probability1 for sufficiently rich propagation scattering environ-ments typical of WLANs and for K ≤ Nt) a column-normalized version of the Moore-Penrose Pseudoinversegenerally yields the ZFBF matrix. This takes on theform

V = HH(HHH)−1Λ,

where Λ = diag(λ1, . . . , λK) ensures that the norm ofeach column of V is equal to 1, thus setting the totaltransmit power equal to tr(Cov(Vx)) = E[‖x‖2], i.e.,equal to the power of the transmitted data vector x.Since, by construction, HV = Λ is a diagonal matrix, itfollows that left multiplying the vector of user symbolswith the beamforming matrix cancels out the symbolinterference at the receivers.

Why Synchronization Is Needed. Time and phasesynchronization are needed between transmitters in or-der for such precoding to work. Clearly, time synchro-nization is needed to coordinate transmissions, but be-cause OFDM gives some leeway due to the cyclic pre-fix, this is a relatively coarse synchronization. Phasesynchronization, however, is required since ZFBF relieson being able to precisely tune the phase of a signalarriving at a receiver. While a classic MIMO transmit-ter has all of its RF chains running on a single clocksource, each access point in a distributed MIMO sys-tem has its own clock and thus the signal it producesdrifts in phase with respect to the signals of the otheraccess points. We will show that sufficiently accuratephase synchronization is necessary to make distributedmultiuser MIMO a reality.

Why is distributed multiuser MIMO challeng-ing? For simplicity of exposition, let us now consider adistributed multiuser MIMO scenario with two clientsand two access points, each one with a single antenna.All of our considerations will apply equally to a moregeneral scenario. For nomadic users, typical of WLANscenarios, the channel changes quite slowly with time, sothat we may assume that the channel impulse responseis locally invariant with respect to time. In order touse ZFBF, we must estimate the channel matrix coef-ficients at each subcarrier for each transmitter/receiverantenna combination. A number of methods for es-timating channel coefficients have been proposed, in-cluding feedback schemes ( see [3] and the referencestherein) and exploiting uplink/downlink reciprocity inTime-Division Duplex (TDD) systems [18]. For simplic-ity of exposition, we will assume here that the channel

4

estimates correspond perfectly to the real channel.After the channels have been estimated, all of the

access points send their channel estimates to a centralserver, which computes the precoding matrix. For eachsubcarrier n = 1, . . . , N , let

H(n) =

[H11(n) H12(n)H21(n) H22(n)

](4)

denote the 2× 2 downlink channel matrix between thetwo clients and the two access point antennas.

Let the precoding matrix V(n) of subcarrier n besuch that H(n)V(n) = Λ(n) = diag(λ1(n), λ2(n)), asexplained before. If timing and carrier phase refer-ence remain unchanged from when the channel was esti-mated until the signal is transmitted, the received signalat the clients, on each subcarrier, can be written as

y(n) = H(n)V(n)x(n) + z(n) = Λ(n)x(n) + z(n) (5)

Since the overall channel matrix Λ(n) is diagonal, wehave achieved complete user separation, so that the ac-cess point can serve the two clients on the same down-

link slot without interference. The spatial multiplexinggain in this case is 2, as two users are being served si-multaneously on the same time-frequency resource.

Suppose now that the timing reference and carrierphase reference between the estimation and transmis-sion slots of the two access point is not ideal. Withperfect timing, the downlink channel from access pointi to client j would have impulse response hij(τ). In-stead, due to misalignment, the impulse response ishij(τ − τi − δj)ej(φi+θj) where τi, φi are the timing andcarrier phase shifts at access point i and δj , θj are thetiming and carrier phase shifts at client j. For sim-plicity, assume that the timing shifts are integer multi-ples of the time-domain symbol interval Ts (otherwisethe derivation is more complicated, involving the foldedspectrum of the channel frequency response, but the endresult is analogous). From the well-known rules of lin-earity and time-shift of the discrete Fourier transform,we arrive at the following expression for the effectivechannel matrix:

H̃(n) =

[ej( 2πNTs

δ1n+θ1) 0

0 ej( 2πNTs

δ2n+θ2)

]︸ ︷︷ ︸

Θ(n)

[H11(n) H12(n)H21(n) H22(n)

][ej( 2πNTs

τ1n+φ1) 0

0 ej( 2πNTs

τ2n+φ2)

]︸ ︷︷ ︸

Φ(n)

(6)

We notice that the diagonal matrix of phasors Θ(n)multiplying the nominal channel matrix from the leftposes no problems, since these phase shifts can be re-covered individually by each client as in standard co-herent communication [24]. In contrast, the diagonalmatrix Φ(n) multiplying from the right poses a signifi-cant problem, since in each receiver’s equation each un-known will be further multiplied by a different randomfactor. In fact, since the server computes the MIMOprecoding matrix V(n) based on H(n), it follows that

when applied to the effective channel H̃(n) in (6) the

matrix multiplication H̃(n)V(n) is no longer necessar-ily diagonal. We conclude that the presence of timingand carrier phase misalignment between the estimationand transmission slots, at each individual access point,yields residual multiuser interference which may com-pletely destroy the performance of a distributed mul-tiuser MIMO system. To stress the importance of thisaspect, we would like to make clear that the resultingsignal mixing takes place over the actual transmissionchannel, making it impossible for the receivers to elim-inate it.

Why Synchronization Is Possible. Any discus-sion on phase synchronization of distributed wirelesstransmitters must necessarily start with the mechanisms

through which phase errors occur. Digital wireless trans-mission systems are constructed using a number of clocksources, among which the two most important ones arethe sampling clock and the carrier clock. In a typ-ical system, signals are created in a digital form inbaseband at a sampling rate on the order of mega-hertz, then passed through a digital-to-analog converter(DAC). Through the use of interpolators and filters, theDAC creates a smooth analog waveform signal which isthen multiplied by a sinusoidal carrier produced by thecarrier clock. The result is a passband signal which isthen sent over the antenna.

Wireless receivers, in turn, use a chain of signal mul-tiplications and filters to create a baseband version ofthe passband signal received over the antenna. Some de-signs, such as the common superheterodyne receivers,use multiple high frequency clocks and convert a sig-nal first to an intermediate frequency before bringing itback to baseband. Other designs simply use a carrierclock operating at the same nominal frequency as thecarrier clock of the transmitter and perform the passagefrom passband to baseband in a single step. We will befocusing on such designs in the ensuing discussion. Af-ter baseband conversion, the signal is sampled and theresulting digital waveform is decoded.

5

5 10 15 20 25 30 35−180

−120

−60

0

60

120

180

Time (OFDM Symbols)

Ph

ase (

Deg

ree

s)

Figure 2: Pilot phases

There are four clocks in the signal path: the transmit-ter’s sampling clock and carrier clock and the receiver’scarrier clock and sampling clock. All four clocks mani-fest phase drift and jitter. The drift effect, when linearin time and happening at a relatively stable rate, can beassimilated to the presence of a carrier frequency offset.

We have considered the effects of a linear phase drifton an OFDM encoded packet in Equation 6. Denoteby ωn = 2πn

NTsthe subcarrier frequency and with ωc the

carrier frequency. Let the timing error of the samplingclock be ∆ts and the timing error of the carrier clockbe ∆tc and assume that they are on the same order ofmagnitude. The phase error due to the sampling clockwill be ωn∆ts while the phase error due to the carrier

clock will be φi = ωc∆tc. The term ej(2πNTs

τin+φi) in (6)can be rewritten as ej(ωn∆ts+ωc∆tc). Since ωc is muchgreater than ωn, the dominant phase rotation is due tothe carrier clock and does not depend on the subcarrierfrequency. Moreover, since time errors are additive, ifthe time error is approximately linear in time (linearclock drift) then the phase error will also be linear intime and almost equal for all subcarriers.

The assumptions behind the above statement are ver-ified by the results presented in Figure 2. We have con-structed an experiment in which a transmitter sendsseveral tone signals, i.e., simple unmodulated sine waves,corresponding to several different subcarrier frequen-cies. In the absence of phase drift these tone signalswould exhibit a constant phase when measured over sev-eral OFDM frames. In reality, the phase is not constantand the frame to frame phase drift of the tone signalscan be measured and recorded. In the figure the phasedrift has been plotted over the duration of a few tensof frames, a time length comparable to that of a packettransmission in a WLAN standard. As evidenced bythese plots, our experiment confirms what was antic-ipated above: the drift is indeed linear and does notdepend on the subcarrier frequency. This allows us todesign a scheme for which the drift can be tracked andpredicted.

The fact that the common phase drift of all subcarri-

ers can be predicted by observing only a few pilots tonesprompts the following approach to achieving phase syn-chronization between access points: a main access point(master) is chosen to transmit a reference signal con-sisting of several pilot tones placed outside the datatransmission band, in a reserved portion of the systembandwidth. An initial channel probing header, trans-mitted by the master access point, is used by the othertransmitters in order to get an initial phase estimate foreach carrier. After this initial estimate is obtained, thephase estimates will be updated using the phase driftmeasured by tracking the pilot signals. After the initialchannel estimation header, all access points start trans-mitting simultaneously in the data band, making use ofthe continuously updated phase estimates in order tocreate phase synchronous signals.

The achievable precision of this synchronization methoddepends on two main parameters: the SNR qualityof the channel linking the secondary access points tothe master access point and the jitter characteristics ofthe oscillator clocks. The impact of jitter can be esti-mated using the following back-of-the envelope calcula-tion. Assume the use of an oscillator having a typicalprecision of 0.1 ppm (parts per million) over short timedurations. The phase error of the synchronization cir-cuit due to the oscillator can be estimated by multiply-ing the precision value with the time length of the syn-chronization loop. In our system, this loop has a timelength corresponding to five OFDM symbols, or 80 mi-croseconds. When assuming a carrier frequency of 2.4GHz the resulting predicted phase offset is 3.5 degrees,which is more than adequate for our purposes as is ev-ident from the experimental results that we present inSection 5. Capacity region calculations show that withthis precision of synchronization, ZFBF can create, foran uniform user power allocation, parallel channels withup to a 27 dB SINR value.

4. IMPLEMENTING AIRSYNCSoftware Radio Implementation. We have im-

plemented AirSync as a digital circuit in the FPGA ofthe WARP radio platform [26]. The WARP radio is amodular software radio platform composed of a centralmotherboard hosting an FPGA and several daughter-boards containing radio frequency (RF) front-ends. Theentire timing of the platform is derived from only tworeference oscillators, hosted on a separate clock board:a 20 MHz oscillator serving as a source for all samplingsignals and a 40 MHz oscillator which feeds the carrierclock inputs of the transceivers present on the RF front-ends. The shared clocks assure that all signals sent andreceived using the different front-ends are phase syn-chronous. Phase synchronicity for all sent signals orfor all received signals is a common characteristic ofMIMO systems. However, the fact that the design of the

6

WARP ensures phase synchronicity among the sent andreceived signals, as opposed to using separate oscillatorsfor modulation and demodulation, greatly simplifies thesynchronization task. The system’s data bandwidth is5 MHz. We place the synchronization tones outside thedata bandwidth, at about 7.5 MHz above and belowthe carrier frequency. The placement of the carriers al-lows us to exploit the adjustable baseband sender filterpresent in the transmit signal path in order to avoid, inthe case of the pilot tones, any self-interference at thesecondary transmitters.

We have implemented a complete system-on-chip de-sign in the FPGA, taking advantage of the presenceof hard-coded ASIC cores such as a PowerPC proces-sor, a memory controller capable of supporting trans-fers through direct memory access over wide data busesand a gigabit Ethernet controller. Atop this system-on-chip architecture we have ported the NetBSD operatingsystem and created drivers for all the hardware com-ponents hosted on the platform, capable of setting allsystem and radio board configuration parameters. Theoperating system runs locally but mounts a remote rootfilesystem through NFS. In the same system-on-chip ar-chitecture we integrated a signal processing componentcreated in Simulink which provides interfaces for fastdirect memory access. This later component is respon-sible for all the waveform processing and for the synthe-sis of a phase synchronous signal and interfaces directlywith the digital ports of the radio front-ends. We inter-faced the Ethernet controller and the signal processingcomponent using an operating system kernel extensionresponsible for performing zero-copy, direct memory ac-cess data transfers between the two, with the purposeof passing back and forth waveform data at high ratesbetween a host machine and the WARP platform. Thelarge data rates needed (320 Mbps for a 10MHz wirelesssignal sampled with 16 bit precision) required optimiz-ing the packet transfers into and out of the WARP. Forexample, consider the direct memory access ring asso-ciated with the receive end of the Ethernet controlleron the board, which is shared between packets destinedto the signal processing component and packets des-tined to the upper layers of the operating system stack.We do not release and reallocate the memory buffersoccupied by packets destined to the signal processingcomponent. Instead, we use a lazy garbage collectionalgorithm in order to reclaim these buffers when theyare consumed in a timely manner or reallocate them at alater point if they are not consumed before the memoryring runs low on available memory buffers. The ratio-nale for this particular optimization is that the overheadof managing the virtual-memory based reallocation ofmemory buffers of tens of thousands of packets everysecond would bring the processor of the software radioplatform to a halt.

All transmitting WARP radios are connected to acentral processing server through individual Ethernetconnections operating at gigabit speeds. Most of thesignal synthesis for the packet transmission is done of-fline, using Matlab code. We produce precoded packetsin the form of frequency domain soft symbols. How-ever, the synchronization step and the subsequent sig-nal generation is left to the FPGA. The server, a fastmachine with 32 processor cores and 64GB of RAM, en-codes the transmitted packets and streams the resultingwaveforms to the radios.

The Synchronization Circuit in Detail. AirSyncoperates similarly to other OFDM-based, distributedtransmission systems such as SourceSync [25] or FineGrained Channel Access [30], but extends them by achiev-ing phase synchronization among transmitters. An im-portant component of those systems, essential in or-der to avoid leakage from one carrier to another dur-ing the decoding process, is the realization of framealignment that arranges frame starting points at thereceivers within an interval shorter than a cyclic prefixlength. In other words, the overlap of the frames sentby different senders must be greater than the length of aframe without cyclic prefix in order to allow the receiverto perform a full-length discrete Fourier transform onthe received signal. Note that the use of zero-forcingdoes not relax this requirement. Zero-forcing is achievedby arranging the phases of several transmitted signals tosum up to almost zero at one of the receivers. The nat-ural way in which these signals may add up to zero afterapplying the discrete Fourier transform is for the phasealignment between the different signals to be consis-tent for the whole duration of the transform. Thus theframes must overlap over the entire time interval asso-ciated with the transform. AirSync achieves frame syn-chronization through a technique used in block bound-ary detection, namely the insertion of pseudo-noise (PN)sequences in the master access point’s packet header inorder to allow the secondary transmitters and the re-ceivers to obtain a time reference. For reasons that willbecome clear, achieving frame synchronization withinthe length of the cyclic prefix is a sufficient startingpoint for also achieving phase synchronization.

In the following we will say that two signals are phasesynchronous when the pure tones (that is tones thathave not been multiplied with a constellation symbolfor data transmission) transmitted by their senders overeach subcarrier, have a constant phase difference overthe duration of successive frames and this difference canbe known a priori by the senders. Naturally, due to thephase offsets induced by propagation delays the valueof the phase difference depends on the location wherethe two signals are received. Thus the phase differencecan be considered constant only when the two signalsare compared at the same receive location.

7

Figure 3: AirSync Schematic. The baseband sig-nals are processed through an FFT which feeds phaseestimates into a Kalman Filter. The IFFT produces aphase-adjusted data signal, with the same phase driftas the main transmitter. The modulation and demod-ulation use the same carrier clock.

AirSync implements the idea of observing phase driftusing pilot tone signals. In order to reduce self-interferenceat the secondary transmitters, the tone signals are placedoutside the data band, from which they are separatedby a large guard interval. The secondary transmittersplace an analog baseband filter around their data bandfurther limiting their interference with the pilots. Self-interference could have been avoided using a number ofother techniques such as antenna placement [7], digitalcompensation [10], or simply relying on the OFDMA-like property of a frame aligned system [30] and pre-venting the secondary transmitters from using the pilotsubcarriers.

Figure 3 illustrates the process of creating a phasesynchronous signal at the secondary transmitter whileFigure 6 in Section 5.1 presents the initial synchroniza-tion sequence. The secondary transmitter overhears apacket sent by the primary transmitter and uses theinitial PN sequence in order to determine the blockboundary timing of this packet. Using a discrete Fouriertransform the secondary transmitter decodes the suc-cessive frames of the incoming packet. It then em-ploys the CORDIC algorithm on the complex-valuedreceived soft symbols in order to obtain their phasesin radians. The phases of the out-of-band pilot signalsare tracked throughout the entire packet transmissionin order to estimate the phase drift from the primarysender. The measurements from the four different pilotsare averaged and passed through a simplified Kalmanfilter which maintains an accurate estimate and pre-dicts, based on the current estimate, the phase drift af-ter the passage of a few further frames. In addition, theheader sent by the primary sender contains a numberof channel estimation symbols, used to obtain an initialphase offset estimate for each subcarrier. As mentionedpreviously, the phase drift is almost identical for all car-riers, therefore these two measurements suffice in orderto predict the phase rotation induced by the main trans-mitter on any subcarrier tone for the entire period of apacket.

The phase estimates are used in synthesizing a syn-chronized signal. The secondary transmitter uses an in-verse discrete Fourier transform, whose output framesare timed such that they align with the frames of themain sender’s signal. For every subcarrier the secondarytransmitter rotates the soft symbol to be sent by an an-gle corresponding to the subcarrier’s estimated phaseoffset. The result is a tone that, while not having thesame phase as the corresponding tone from the maintransmitter, follows that tone at a fixed, pre-knownphase difference.

The synchronization circuit could have been constructedin different ways. For example consider SourceSync [25],a recent work which has implemented frame alignment.AirSync differs in the implementation approach in threeimportant points. SourceSync performs fine frequencyoffset correction. AirSync avoids this correction. Fre-quency offset correction prevents power leakage fromneighboring carriers during decoding. Our synchroniza-tion circuit does not decode the subcarriers in the databand but only the pilot tones, for which power leak-age from neighboring carriers is not a concern, andsubsumes frequency correction on the transmit side byphase synchronization. Another design decision differ-ent from SourceSync is the use of a PN sequence forblock boundary detection [33] instead of measuring theslope of the phase rotation induced by the timing mis-alignment between the sources on the decoded frames.The final difference from SourceSync is that since thesenders are phase synchronous, the receivers do notneed to monitor the evolution of the sender’s pilots sep-arately through joint channel estimation.

Centralized joint encoding. By transmitting phasesynchronous signals from multiple access points we havecreated the equivalent of a distributed MIMO transmit-ter, capable of employing multiuser MIMO precodingstrategies in order to transmit to multiple users at thesame time. However, the use of multiple access pointscomplicates the design of the transmitter system. Formost of the precoding schemes available, the encodingof the waveforms to be transmitted over the antennasmust be done jointly, since reaching a single user usuallyinvolves transmitting over multiple antennas. While intheory the joint encoding process could be duplicated ateach access point given the binary information destinedto each user, we chose to do the encoding only once,at a central server and send the resulting waveforms toeach access point for transmission2.

Our central server has an individual gigabit Ether-net connection to each of the WARP radios serving asaccess points. We divide the downlink time into slotsand in each slot schedule for transmission a number of

2This approach is practical in enterprise networks where anumber of access points are already connected to a commonserver.

8

Figure 4: Testbed diagram. The central server isconnected to the two transmitters, the main transmitteron the left and the secondary transmitter on the right.

packets destined to various users, according to an algo-rithm that will be presented in Section 6. The mediumaccess encoding of the packets is presented in the samesection. For each of the access points, the server com-putes the waveform of the signal to be transmitted inthe next downlink slot. However, it does not performany phase correction at this point. The only informa-tion used in the precoding is the data to be transmittedand the channel state information between each accesspoint antenna and each user antenna. The server as-sumes that all access points are phase synchronous, likein a normal MIMO system. The server transmits theircorresponding waveforms to all secondary transmittersand finishes by sending the last waveform to the primarytransmitter. The primary transmitter starts transmit-ting right away and the secondary transmitters follow.

The design of AirSync ensures its scalability. There isno added overhead for synchronizing a larger number ofsecondary transmitters, while the overhead for channelestimation is the same as in regular MIMO systems.

In comparison to simple point-to-point transmission,AirSync uses about 10 more frames per packet in or-der to achieve synchronization. This number should betaken with a grain of salt in computing the overhead,since multiuser transmissions involve multiple packetsbroadcasted at the same time. When compared to asingle packet duration, the overhead is about 4%.

5. PERFORMANCE EVALUATIONOur system setup is presented in Figure 4. It con-

sists of a primary transmitter, a secondary transmitterand two receivers. The main sender uses a single RFfront-end configured in transmit mode, placing an 18MHz shaping filter around the transmitted signal. The

−10 −5 0 5 100

0.2

0.4

0.6

0.8

1

Phase Error (Degrees)

Pe

rce

nta

ge

of

Exp

erim

en

ts

Figure 5: The Precision of the Phase Synchro-nization. AirSync achieves phase synchronizationwithin a few degrees of the source signal.

secondary sender uses an RF front-end in receive modeand a second RF front-end in transmit mode, with a 12MHz shaping filter. As mentioned previously, the pi-lots used in phase tracking are outside the secondary’stransmission band, therefore the secondary transmitterwill not interfere with the pilot signals from the maintransmitter. The series of experiments is intended totest the accuracy of the synchronization and the effi-ciency of channel separation.

5.1 Synchronization AccuracyIn this particular experiment we have placed the two

transmitters and the two receivers at random locations.We placed a third RF front-end on the secondary senderand configured it in receive mode. The secondary trans-mitter samples its own synthesized signal over a wiredfeedback loop and compares it with the main transmit-ter’s signal. The synchronization circuit measures andrecords the phase differences between these two signals.Since we use the primary transmission as a reference, inthis experiment we do not broadcast the signal synthe-sized by the secondary transmitter in order to protectthe primary transmission from unintended interference.We note that the use of a third RF front-end is notneeded in the general case.

We have modified the synchronization circuit to pro-duce a signal that is not only phase synchronous withthat of the primary transmitter but has the exact samephase when observed from the secondary transmitter.To achieve this, the circuit estimates the phase rota-tion that is induced between the DAC of the secondarytransmitter and the ADC through which the synthe-sized signal is resampled. It then compensates for thisrotation by subtracting this value from the initial phaseestimate. It is worth noting that this rotation corre-sponds to the propagation delay through the feedbackcircuit and is constant for different packet transmis-sions, as determined through measurements. The resultwas a synthesized signal that closely follows the phase

9

−0.5

0

0.5A

mp

litu

de

I Component

−0.5

0

0.5

Am

plit

ud

e

Q Component

−π

0

π

Ph

ase

(Ra

d)

Initial Phase Estimate

−π

0

π

Ph

ase

(Ra

d)

Current Phase Drift

Predicted Phase Drift

1 5 10 15 20 25

−π

0

π

OFDM Symbols

Ph

ase

(Ra

d)

Master Phase

Secondary Phase

Figure 6: Phase Synchronization Acquisition. The secondary transmitter obtains an initial phase estimate.It then tracks the phase drift of the subcarriers and uses a Kalman filter to predict its value a few samples later.

of the signal broadcast by the master transmitter, asillustrated in Figure 6. The figure illustrates the initialphase acquisition process, the initial phase estimation,the tracking and estimation of the phase drift, as wellas the synthesis of the new signal. The phase discon-tinuities appearing in the main transmitter’s signal aredue to the presence of PN sequence along with a tempo-rary disturbance needed in order to tune the feedbackcircuit.

Figure 5 illustrates the CDF of the synchronizationerror between the secondary transmitter and the pri-mary transmitter. The error is measured on a frame-to-frame basis using the feedback circuit. In decimaldegree values, the standard deviation is 2.37 degrees.The 95th percentile of the synchronization error is atmost 4.5 degrees.

We have measured the SNR value of the synchroniza-tion pilots in the signal received by the secondary trans-mitter to be around 28.5 dB. This is easily achievablebetween typically placed access points.

5.2 Beamforming gainOur second experiment was done using the complete

four radio setup with the secondary transmitter broad-casting a secondary signal over the air. We measuredthe channel coefficients between the two transmittersand the receivers using standard downlink channel es-timation techniques and arranged the amplitudes andthe phases of the transmitted signals such that at one

of the receivers the amplitudes of the two transmittedsignals would be equal while the phases would align.The maximal theoretic power gain over transmitting thetwo signals independently is 3.01dB. We compared theaverage power of the individual transmissions from thetwo senders to the average power of a beamformed jointtransmission. Our measurements show an average gainof 2.98 dB, which is consistent with the precision of thesynchronization determined in the previous experiment.

This result shows that for all practical purposes weare able to achieve the full beamforming gain in ourtestbed.

5.3 Zero-Forcing AccuracyThe following experiment measures the amount of

power which is inadvertently leaked when using Zero-Forcing to non-targeted receivers due to synchroniza-tion errors. Again we have placed our radios at randomlocations in our testbed. We have estimated the channelcoefficients and arranged for two equal amplitude tonesfrom the two transmitters to sum as closely as possibleto zero. The residual power is the leaked power dueto angle mismatching. Figure 7 illustrates the CDF ofthis residual power for different measurements. The av-erage power leaked is -24.46 dB of the total transmittedpower.

This establishes that Zero-Forcing is capable of al-most completely eliminating interference at non-targetedreceiver locations.

10

−32 −30 −28 −26 −24 −22 −200

0.2

0.4

0.6

0.8

1

Zero Forcing Leaked Power (dB)

Pe

rce

nta

ge

of

Exp

erim

en

ts

Figure 7: The Power Leakage of Zero-Forcing.The leaked power is significantly smaller than the totaltransmitted power, transforming each receiver’s channelinto a high SINR channel.

Receiver 1 Receiver 2

Figure 8: Scattering Diagram. The scattering di-agram for two independent data streams transmittedconcurrently demonstrates that AirSync achieves com-plete separation of the user channels.

5.4 Zero-Forcing Beamforming Data Transmis-sion

The final experiment transmits data to the two re-ceivers. We have used symbols chosen independentlyfrom a QAM-16 constellation at similar power levels.The scattering plots in Figure 8 illustrates the receivedsignal at the two receivers.

The SINR values at the two receivers are 29 dB and 26dB respectively. It is evident that the testbed achievesthe full MIMO multiplexing gain.

6. MEDIUM ACCESS CONTROLGiven that we have achieved the necessary synchro-

nization accuracy between access points, we turn tothe large body of work on optimal scheduling for cen-tralized multiuser MIMO systems (see for example [9,19]). Inspired by this work, we propose a MAC layerthat significantly departs from the classic networkinglayered architectural model and adopts a cross-layer“PHY/MAC” design strategy.

6.1 High level description

Time Division Duplexing. First, we consider theissue of allocating air time and frequency spectrum be-tween the uplink and the downlink. We can choosebetween two natural strategies for separating the up-link from the downlink: time division duplex (TDD)and frequency division duplex (FDD). We choose touse TDD for two reasons. First, with TDD we canexploit channel reciprocity at the access point and mea-sure the uplink channel, using pilots from the users, toinfer the downlink channel as previously described. Onthe other hand, in FDD the uplink and downlink car-rier frequencies are separated by much more than thechannel coherence bandwidth, and therefore the chan-nel matrix coefficients of the uplink and downlink chan-nel are essentially statistically independent. Thus, inFDD no useful information about the downlink chan-nel matrix can be learned from the uplink pilots. Inthis case, an explicit closed-loop channel estimation andfeedback needs to be implemented, with a protocol over-head that increases linearly with the number of jointlyprecoded access point antennas [17]. Second, TDD isideally suited for the transport of asymmetric traffic,as is typical in an enterprise WiFi scenario, whereas anFDD system provides less flexibility for managing dif-ferent traffic patterns. We shall consider the schedulingof users in the uplink and downlink periods separately.In the uplink, clients compete for bandwidth using reg-ular CSMA/CA. Thus, in the rest of this section wefocus on the downlink.

Downlink scheduling. The central server keepstrack of packet queue sizes and other readily availableQoS information, e.g. the time since these queues havebeen served last. It then selects a subset of users totransmit to at each downlink time slot. At the startof the downlink period all access points send a jam-ming signal, causing the clients to backoff. The ensuingdownlink transmission will silence the clients until theend of the downlink period. In the following we discussin detail how the central server selects these users ateach time slot.

The selection and power allocation problem for lin-ear Zero-Forcing precoding has a rich literature (forexample [9, 38]) which follows theoretical models sim-ilar to the one introduced in Section 3. Conceptually,this optimization problem can be solved by exhaustivelysearching over all feasible subsets of users, optimizing aweighted rate function under some general power con-straints. In practice, greedy algorithms have proven toprovide excellent results at moderate complexity [9,19].

To simplify the design of the scheduling algorithm, wecan use a greedy algorithm like the one described in [19].However, real world considerations prompt a number ofchanges. Following the example of the de facto MIMOstandard 802.11n, we may allocate the same power to allsubcarriers instead of solving a complicated dual mul-

11

tiuser allocation and waterfilling problem. Note thatthis simplifies the optimization problem significantly.Second, to achieve fairness among flows the construc-tion of the utility function must take into account thequeue delays experienced by packets. This matters as itis well known that scheduling decisions which are solelybased on queue sizes, while guaranteeing stability, leadto starvation and may lead to timeouts at higher layers,for example in TCP.

It is beyond the scope of this paper to further opti-mize the scheduling algorithm. In general, though, anon-flat spectrum will allow the system to get closer tothe theoretically optimal performance. Instead, we turnour attention to the practical issues associated with ourMAC protocol, and in particular, to the design of thepacket headers.

6.2 Protocol DesignOur protocol design focuses on the downlink chan-

nel. Figure 9 presents a simplified schematic of thedownlink data packets and corresponding uplink ac-knowledgments. The MAC layer packet design and theprotocol’s sequence of actions are tuned for enablingmultiuser MIMO broadcasts. The crucial design con-straint is to provide the central server with timely esti-mates of the channel state information for all clients towhich it is about to transmit or which are consideredfor the next round of transmissions. For this purpose,we schedule downlink transmissions to closely follow up-link acknowledgments and require the clients to providethe server with channel estimates during the uplink pe-riod. The mechanism through which this is achievedwill be described in the following paragraphs. The cen-tral server uses the uplink estimates to select a set ofclients for the following transmission slots, according tothe scheduling algorithms introduced earlier.

The downlink packet starts with a transmission fromthe main sender containing a pseudo-noise sequence usedto achieve frame alignment by the transmitters and forblock boundary detection by the receivers. The masteraccess point then transmits the first set of channel esti-mation pilots which are used by the other access pointsto determine the initial phases of the subcarrier tones,as described in Section 4. After this point, all accesspoints take part in the downlink transmission. Thepacket header that follows is broadcast to all clients,including the non-targeted ones, using the Alamoutiencoding [1]. Due to phase alignment between trans-mitters, the clients do not need to track the secondarysenders in order to decode this header. The MAC ad-dresses of the hosts targeted in the current transmissionand the MAC addresses of the clients that are requiredto provide the server with channel estimates during thenext acknowledgment are the most important pieces ofinformation contained in the header fields. The posi-

tions of the addresses in the header fields create an im-plicit ordering of the clients, which will be used in theuplink period. The following part of the header is an al-location map, similar to the one found in the LTE stan-dard, which assigns carriers to small groups of differentclients and specifies the constellations used in broad-casting to them. The header is followed by a secondset of channel estimation pilots, transmitted this timearound by all access points using ZFBF, which are usedby all clients in order to obtain the channel estimatesfor their individual downlink channels. The clients usethe downlink estimates together with the synchroniza-tion pilot tones in order to gain a lock on the subcarri-ers. The downlink transmission continues with payloadtransmission.

In current 802.11 MIMO implementations, the chan-nel estimates are obtained using downlink pilots whichare in turn quantized by the receivers and communi-cated back in numerical form to the transmitter. Thequantization and communication steps incur a large over-head. Using the reciprocity property of wireless chan-nels, we can reduce the complexity of the channel esti-mation process significantly. First, we prefer to performuplink channel estimation since uplink estimates can bereceived simultaneously by all access points, reducingthe number of pilot transmissions needed by a factorequal to the total number of access point antennas. Sec-ond, uplink estimates are sent using analog pilot signalsin an unquantized form, leaving the quantization stepto the access points. This reduces the overhead of thetransmission significantly. Third, while the usual esti-mation pilots are full OFDM frames, we choose to sendpulse-like signals, measure the channel response, andfill the non-significant taps with zeros before taking aFourier transform in order to determine the frequencydomain response. This insures that our pilots need tobe spaced only by an interval that can accommodate along channel response, i.e. the length of a cyclic prefix.

After the downlink transmission has finished, the clientswho have been requested to send their channel estimatesstart sending these short estimation pilots in quick suc-cession. We note that there is a large degree of simi-larity between the functioning of the downlink channelestimation for receive decode purposes and the uplinkchannel estimation step. The timing of the system re-mains unchanged during the uplink slot and the rolesof the transmitters and the receivers are switched. Theuplink pilots are followed by smart acknowledgmentsfor the data packets sent using the technique detailedin [11].

We tested each component of the downlink and uplinkprotocol slots. However, since our radios do not switchfrom receive to transmit in a timely manner, we couldnot perform complete real-time MAC experiments.

Overhead. A note on the overhead of the above

12

Figure 9: Packet Design. Downlink data packet (left) and uplink acknowledgment (bottom right).

MAC is in order. As those familiar with the PHY/MACdetails of the 802.11 family of protocols would have rec-ognized already, the overhead of our MAC is not morethan that of 802.11n. The additional signaling overheadcomes from requiring a few frames to predict the initialphase, and a few frames to dictate the MAC addressesof the nodes from which we wish to request channelstate information for the next time slot. Even withvery conservative estimates this will be less than a 20%increase in header time duration over that of a tradi-tional 802.11 system. Note, however, that we get abandwidth increase that grows almost linearly in thenumber of clients. This means that our overhead, nor-malized such that we consider the total control bits overthe total data bits transmitted during a fixed airtimeslot, is much less than in a traditional 802.11 system.

7. DISCUSSION AND FUTURE WORKIn the future, we plan to extend our experiments to a

larger testbed. For demonstration purposes, we intendto showcase real time video streaming at high rates. Inthe rest of this section, we discuss approaches that gobeyond the results discussed in the previous sections.

In this paper we used linear ZF precoding because ofits conceptual and implementation simplicity, and near-optimal performance when allowing for flexible user se-lection [9,38]. In the future we plan to experiment withmore sophisticated precoding techniques, such as reg-ularized ZF [22], lattice reduction precoding [37] andmodified Tomlinson-Harashima precoding [5], which canbe regarded as a viable and very efficient approximationof the capacity-achieving Dirty Paper Coding (DPC)scheme [8]. The relative merit of these schemes is knownfrom a theoretical viewpoint, but a thorough compar-ison on the basis of an actual SDR implementation isnot available.

In wireless communication systems, different rates areusually supported using different codes. The currentstandard, 802.11n, offers many code combinations tofully utilize the capacity of the MIMO channel. Since

a multiuser MIMO system serves multiple users in thesame time slot, an even larger set of rates and codeswould have to be supported for efficiently using capac-ity. In this case, an attractive and innovative approachwould be the use of rateless codes (e.g., Raptor codes[12, 28] and the recently proposed Spinal codes [23]) atthe physical layer, in a so-called Incremental Redun-dancy (IR) configuration (see [6,21,27]), as already ex-emplified by Strider [16], to decrease the signaling andretransmission overhead. This is another interesting di-rection that we plan to pursue as future work.

8. ADDITIONAL AUTHORS

9. REFERENCES[1] S. Alamouti. A simple transmit diversity

technique for wireless communications. IEEE J.Sel. Areas Commun., 16(8):1451–1458, 1998.

[2] E. Aryafar, N. Anand, T. Salonidis, and E. W.Knightly. Design and experimental evaluation ofmulti-user beamforming in wireless LANs. InACM MobiCom, Chicago, IL, 2010.

[3] G. Caire, N. Jindal, M. Kobayashi, andN. Ravindran. Multiuser MIMO achievable rateswith downlink training and channel statefeedback. IEEE Trans. Inf. Theory,56(6):2845–2866, 2010.

[4] G. Caire and S. Shamai. On the achievablethroughput of a multiantenna gaussian broadcastchannel. IEEE Trans. Inf. Theory, 49(7):1691 –1706, Jul. 2003.

[5] G. Caire, S. S. Shamai, A. Shokrollahi, andS. Verdu. Fountain codes for lossless datacompression. In AMS DIMACS Workshop onAlgebraic Coding Theory and Information Theory,Jun. 2005.

[6] G. Caire and D. Tuninetti. The throughput ofhybrid-arq protocols for the gaussian collisionchannel. IEEE Trans. Inf. Theory, 47(5):1971–1988, Jul. 2001.

13

[7] J. I. Choi, M. Jain, K. Srinivasan, P. Levis, andS. Katti. Achieving single channel, full duplexwireless communication. In IEEE MobiCom,Chicago, IL, 2010.

[8] M. Costa. Writing on dirty paper (corresp.). IEEETrans. Inf. Theory, 29(3):439–441, May 1983.

[9] G. Dimic and N. Sidiropoulos. On downlinkbeamforming with greedy user selection:performance analysis and a simple new algorithm.IEEE Trans. Signal Process., 53(10):3857 – 3868,Oct. 2005.

[10] M. Duarte, C. Dick, and A. Sabharwal.Experiment-driven characterization of full-duplexwireless systems. CoRR, abs/1107.1276, 2011.

[11] A. Dutta, D. Saha, D. Grunwald, and D. Sicker.SMACK: a SMart ACKnowledgment scheme forbroadcast messages in wireless networks. In ACMSIGCOMM, Barcelona, Spain, 2009.

[12] O. Etesami and A. Shokrollahi. Raptor codes onbinary memoryless symmetric channels. IEEETrans. Inf. Theory, 52(5):2033 – 2051, May 2006.

[13] G. Foschini and M. Gans. On limits of wirelesscommunications in a fading environment whenusing multiple antennas. Wireless PersonalCommunications, 6:311–335, 1998.

[14] S. Gollakota and D. Katabi. Zigzag decoding:combating hidden terminals in wireless networks.In ACM SIGCOMM, Seattle, WA, 2008.

[15] S. Gollakota, S. D. Perli, and D. Katabi.Interference alignment and cancellation. In ACMSIGCOMM, Barcelona, Spain, 2009.

[16] A. Gudipati and S. Katti. Strider: automatic rateadaptation and collision handling. In ACMSIGCOMM, Toronto, Ontario, Canada, 2011.

[17] J. Jose, A. Ashikhmin, P. Whiting, andS. Vishwanath. Channel estimation and linearprecoding in multiuser multiple-antenna TDDsystems. IEEE Trans. Veh. Technol., 60(5):2102–2116, Jun. 2011.

[18] J. Jose, A. Ashikhmint, P. Whiting, andS. Vishwanath. Scheduling and pre-conditioningin multi-user MIMO TDD systems. In IEEE ICC,Beijing, China, 2008.

[19] M. Kobayashi and G. Caire. Joint beamformingand scheduling for a MIMO downlink withrandom arrivals. In IEEE ISIT, Seattle, WA, Jul.2006.

[20] A. Lapidoth, S. Shamai, and M. A. Wigger. Onthe capacity of fading MIMO broadcast channelswith imperfect transmitter side-information. In43rd Annual Allerton Conference, Monticello, IL,2005.

[21] C. Lott, O. Milenkovic, and E. Soljanin. HybridARQ: Theory, state of the art and futuredirections. In IEEE ITW on Information Theory

for Wireless Networks, Solstrand, Norway, Jul.2007.

[22] C. Peel, B. Hochwald, and A. Swindlehurst. Avector-perturbation technique for near-capacitymultiantenna multiuser communication-part i:channel inversion and regularization. IEEE Trans.Commun., 53(1):195 – 202, Jan. 2005.

[23] J. Perry, H. Balakrishnan, and D. Shah. Ratelessspinal codes. In ACM HotNets, Cambridge,Massachusetts, 2011.

[24] J. Proakis and M. Salehi. Digital communications.McGraw-Hill, New York, NY, 2007.

[25] H. Rahul, H. Hassanieh, and D. Katabi.SourceSync: a distributed wireless architecture forexploiting sender diversity. In ACM SIGCOMM,New Delhi, India, 2010.

[26] Rice University. Rice university warp project.[27] S. Sesia, G. Caire, and G. Vivier. Incremental

redundancy hybrid ARQ schemes based onlow-density parity-check codes. IEEE Trans.Commun., 52(8):1311 – 1321, Aug. 2004.

[28] A. Shokrollahi. Raptor codes. IEEE Trans. Inf.Theory, 52(6):2551 –2567, Jun. 2006.

[29] Q. Spencer, C. Peel, A. Swindlehurst, andM. Haardt. An introduction to the multi-userMIMO downlink. IEEE Commun. Mag.,42(10):60–67, 2004.

[30] K. Tan, J. Fang, Y. Zhang, S. Chen, L. Shi,J. Zhang, and Y. Zhang. Fine-grained channelaccess in wireless LAN. In ACM SIGCOMM, NewDelhi, India, 2010.

[31] E. Telatar. Capacity of multi-antenna gaussianchannels. European Transactions onTelecommunications, 10(6):585–595, 1999.

[32] D. Tse and P. Viswanath. Fundamentals ofWireless Communication. Cambridge UniversityPress, New York, NY, 2005.

[33] F. Tufvesson, O. Edfors, and M. Faulkner. Timeand frequency synchronization for OFDM usingPN-sequence preambles. In IEEE VTC, 1999.

[34] C. S. Vaze and M. K. Varanasi. The degrees offreedom regions of MIMO broadcast, interference,and cognitive radio channels with no CSIT.CoRR, abs/0909.5424, 2009.

[35] S. Verdu. Multiuser Detection. CambridgeUniversity Press, New York, NY, 1998.

[36] H. Weingarten, Y. Steinberg, and S. Shamai. Thecapacity region of the gaussian multiple-inputmultiple-output broadcast channel. IEEE Trans.Inf. Theory, 52(9):3936 –3964, Sept. 2006.

[37] H. Yao and G. Wornell. Lattice-reduction-aideddetectors for MIMO communication systems. InIEEE GLOBECOM, Hsinchu, Taiwan, Nov. 2002.

[38] T. Yoo and A. Goldsmith. On the optimality ofmultiantenna broadcast scheduling using

14

zero-forcing beamforming. IEEE J. Sel. AreasCommun., 24(3):528 – 541, Mar. 2006.

15