4596 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7

4596 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 11, NOVEMBER 2008

MAC/PHY Cross–Layer Design of MIMO Ad HocNetworks with Layered Multiuser Detection

Paolo Casari, Member, IEEE, Marco Levorato, Student Member, IEEE, and Michele Zorzi, Fellow, IEEE

Abstract—In this paper, we present a novel MAC approachfor ad hoc networks in which Multiple Input–Multiple Output(MIMO) techniques are used at the physical level (PHY). The useof MIMO in PHY point–to–point as well as multiuser commu-nications has been extensively studied in the recent literature.Here we go one step further, extending the approach to alsoinclude protocol design for decentralized and infrastructurelessad hoc networks. First, we study the impact of MIMO on packettransmission in an ad hoc network setting. Then, following across–layer design paradigm, we deploy a distributed accesscontrol protocol and characterize its performance, with the aimto improve spatial reuse and convey more information on a singlead hoc link. We also explore the interaction between the MediumAccess Control (MAC) and PHY layers, and use this knowledgeto implement proper policies for distributed traffic control androbustness against interference. Important tradeoffs that arisewhen managing radio resources are highlighted, and extensivesimulation results are presented and discussed.

Index Terms—MIMO ad hoc networks, cross–layer design,MAC protocols, V-BLAST, multiuser detection.

I. INTRODUCTION

AD HOC networks are made of autonomous nodes that canconnect to each other without the need for infrastruc-

tured administration or maintenance. Wireless technologiespotentially enable anytime–anywhere networking, allowingnodes to, e.g., share data and access distributed services ina seamless and easy way. Furthermore, such decentralizednetworks allow for fast deployment in emergency or militaryscenarios, besides being suited for commercial applicationsand for quick communications setup in any environment wherea cabled network is infeasible or not affordable.

Research interest on ad hoc networks has greatly increasedin the past decade, mainly due to the challenging task ofdesigning effective, distributed protocols that yield a goodthroughput, possibly addressing issues such as fairness, energysaving, and so on. Practical limitations, due to channel impair-ments and distributed access, may decrease the performanceof protocols for ad hoc networks. The main issue here ishow to perform decentralized channel access, while ensuringa reasonable probability of successful transmission.

With the shift towards higher frequency bands, the integra-tion of multiple antennas in a single terminal is progressively

Manuscript received June 11, 2007; revised March 25, 2008; acceptedMarch 25, 2008. The associate editor coordinating the review of this paperand approving it for publication was J. Hou. This work has been supportedin part by the US Army Research Office under the Multi-University ResearchInitiative (MURI) grant no. W911NF-04-1-0224. Part of this work has beenpresented at IEEE PIMRC 2005.

The authors are with the Department of Information Engineering, Univer-sity of Padova, Via G. Gradenigo 6/B, I–35131 Padova (PD), Italy (e-mail:{paolo.casari, marco.levorato, michele.zorzi}@dei.unipd.it). Michele Zorzi isalso with the University of California at San Diego, 9500 Gilman Drive, LaJolla, CA 92093-0407.

Digital Object Identifier 10.1109/T-WC.2008.070600

becoming feasible. The use of multiple antennas has showngreat promise in providing higher spectral efficiencies onwireless links than traditional communication systems, buttheir adaptation to ad hoc networks is non-trivial nonetheless.

Multiple antennas allow for more advanced communicationparadigms. Examples include: beamforming, whereby nodescan steer transmissions so as to cover a certain portion ofspace; diversity, which greatly mitigates the effects of multi-path propagation; SIR maximization through array processing,whereby the array reception pattern can be adapted to amplifyor suppress the power received from certain directions.

These techniques could have a great impact in ad hocnetworks. Directional transmissions would both decrease in-terference and amplify power gain toward wanted recipients,increasing spectral efficiency and spatial reuse. On the otherhand, they introduce further challenges, such as how to dealwith gain asymmetries (different array gains at different nodes)and deafness (a node is not aware of what other nodes do).Some works on this topic are summarized in Section II.

The transmission and reception of signals through multipleantennas can be holistically viewed as a MIMO system. Sincethe pioneering work by Foschini [1], MIMO has attractedsignificant attention as the key technology to achieve highspectral efficiency by exploiting rich scattering environments.MIMO enables the protection of communications in the“space” (i.e., antenna) domain, by processing and transmittingsignals through different antennas, according to predefinedschemes (e.g., Space–Time Codes, STC [2]). A subset ofSTCs, namely Layered STCs (LSTC), jointly use encodingand parallel transmissions, sending out multiple flows usingdifferent array elements. A special case of LSTC is V–BLAST [3], where the encoding component is absent, andall resources are used for parallelizing transmissions. Thisapproach is also called Spatial Multiplexing (SM). It has beenshown [4] that there exists a tradeoff between diversity andSM gain in MIMO networks: V–BLAST achieves the greatestSM depth, whereas codes such as [5] are optimal in a diversitysense.

MIMO techniques can be applied to ad hoc networks withsignificant benefits. If multiple bit sequences are sent bydifferent nodes, each using multiple antennas, all streams canbe taken as a separate contribution by the intended receiver. Ifsome channel information is available, the receive antennas’outputs can be recombined and processed such that the sentdata can be finally recovered. The primary consequence isthe coexistence of multiple data packets in the network (i.e.,without collisions), provided that some degree of coordinationis obtained among transmitters. Moreover, by splitting a singlepacket transmission among multiple antennas (e.g., with V–BLAST), a node is allowed a higher raw bit rate, which is

1536-1276/08$25.00 c© 2008 IEEE

Authorized licensed use limited to: ELETTRONICA E INFORMATICA PADOVA. Downloaded on January 23, 2009 at 10:54 from IEEE Xplore. Restrictions apply.

CASARI et al.: MAC/PHY CROSS–LAYER DESIGN OF MIMO AD HOC NETWORKS WITH LAYERED MULTIUSER DETECTION 4597

proportional to the number of antennas used [6], [7]. Wewish to highlight that if full channel state information at thetransmitter (CSIT) is available, beamforming techniques canexploit it and lead to better link-level performance. However,as explained more extensively in Section II, it might provehard to devise protocols that coordinate independent nodes sothat their use of beamforming does not give rise to deafness orsimilar problems. On the contrary, MIMO transmissions canprovide good performance even in the absence of CSIT, with-out resorting to explicit power gain shaping at the transmitter.In our work, we will focus on this specific scenario.

The advantages described above encourage the considera-tion of a MIMO physical layer in ad hoc networks, but leavemany issues open about the correct management of its po-tential. Using a more powerful physical layer in combinationwith existing MAC protocols for ad hoc networks (such as802.11 [8]) may not necessarily be the best choice. A betterdesign paradigm should jointly account for PHY and MACfeatures in a cross–layer fashion, and strive to take advantageof all available degrees of freedom, for example by allowingsome exchange of information between different layers.

In this paper we investigate the advantages, drawbacksand possible tradeoffs that arise at the MAC layer whendealing with ad hoc networks with V–BLAST at the PHYlayer. Exploiting this knowledge enables an effective design ofcompletely distributed channel access mechanisms that man-age transmission requests and grants1 to trade off higher bitrates (i.e., perceived throughput) for more resilient multiuserdetection (i.e., interference rejection) through a proper use ofV–BLAST’s successive interference cancellation capabilities.

This paper is organized as follows. In Section II wesummarize the literature on the use of multiple antennasin ad hoc networks, for both beamforming and MIMO. InSection III we give an overview of the decision feedbackmultiuser detector operating at the physical layer and assessits impact on networking. In Section IV we introduce a newMAC protocol, and describe in detail the proposed policiesfor channel access and traffic management in Section V. InSection VI we give extensive results on network performance,and finally conclude our paper in Section VII.

II. RELATED WORK

The integration of multiple antennas in ad hoc networksis a relatively recent topic. In [9], the authors focused onpurely directional transmissions and designed Multihop MAC(MMAC), a routing-aware protocol that bridges longer dis-tances by both coordinating farther nodes using RTS/CTSexchanges over multiple hops and exploiting the higher gainsand lower overall interference achieved by directional com-munications.

As introduced in Section I, such protocols suffer from“deafness.” For example, directional transmissions or half-duplex operations may leave nodes unaware of ongoingcommunications (deaf), which makes distributed coordinationdifficult. A solution based on busy tones is provided in [10],which requires more complex hardware. It should be noted

1By grant, we indicate a Clear–To–Send message that enables a transmis-sion. In the following, we will interchangeably use either term.

that receiver-side omnidirectional reception is possible even inthe presence of receive beamforming, e.g., by implementingparallel processors that weight the antenna outputs differently:each processor could beamform toward a different direction,ultimately obtaining omnidirectional coverage from the super-position of the processors’ outputs.

Ramanathan et al. [11] proposed UDAAN, a set of in-tegrated MAC, routing, neighbor discovery and signalingprotocols for ad hoc networks with directional antennas. Theyalso built a field demonstration using horn antennas that is, tothe best of our knowledge, the most comprehensive mobile adhoc network testbed deployed so far.

In [12], a MAC protocol is considered where nodes send andreceive data directionally thanks to some topology awareness.In [13] another MAC protocol is proposed with directionalRTS/CTS exchange. Thus, RTSs possess a longer reach,but are sent directionally in one beam at a time, so thatmany transmissions are needed to cover the whole horizon.This approach notifies farther nodes, thus mitigating deafnessand establishing longer links, but incurs longer handshakelatencies.

These papers present very interesting contributions, showingbenefits of directional communications in ad hoc networks.However, very simplified propagation and antenna models aretypically taken into account. This may not be sufficientlyaccurate, especially when achieving directionality througharrays of simple (e.g., dipole) antennas.

A different approach is to regard the whole set of multipletransmit and receive antennas as a MIMO system. In this case,all transmissions are omnidirectional, and multiple superim-posed signals can be separated and detected through some spe-cific signal processing at the receiver. The problem of studyingand optimizing MIMO links under different objectives andconstraints has been widely addressed in the literature. Tworecent books on the topic are [2], [14]. The specific applicationof MIMO systems to ad hoc networks, however, has receivedless attention.

MIMO systems are indeed a means of performing Multi-Packet Reception (MPR). In [15] and [16] (see also thereferences therein), the impact of MPR on random accessMAC protocols is considered. A Multi-Queue Service Roomprotocol is envisioned as the optimal solution for reachingmaximum throughput with random access, and is comparedto less complex but suboptimal protocols. One of the mainconclusions is that cross–layer sharing of simple parameters iscrucial to successfully design protocols for MPR channels. Re-cently, an access scheme to exploit MPR with CDMA, whilemeeting QoS requirements, was also proposed in [17], whilethe effect of MPR capabilities on the throughput capacity ofad hoc networks was studied in [18].

In many papers an information-theoretic point of viewis taken, by defining throughput as the maximum mutualinformation between a received and a transmitted signal.Throughput is then optimized, e.g., under maximum powerconstraints [19]. These works are interesting, but typicallyrequire the ideal assumption that the channel capacity isexactly reached, and often neglect specific networking issues,for instance a particular MAC implementation.

A networking-based approach is carried out in [20] with



MIMA-MAC, an access protocol specifically designed for adhoc networks with up to two antennas per node. The devisedMAC includes a contention-based and a contention-free pe-riod, used to set up links among receivers using two antennasto decode data coming from up to two transmitters using oneantenna each. The small number of nodes considered andthe constraint to use at most one antenna for transmissionrepresent significant limitations.

In [21], the authors propose that nodes transmit busy tonesover one of the sub-bands provided by a MIMO–OFDM sys-tem for signaling their intention to transmit. Prior to sendingthe busy tone, each transmitter chooses a random channelsense time. If it hears more than a certain number of tonesit defers transmission, otherwise it activates its own tone andsets up the link.

Another very interesting work on MIMO ad hoc networksis [22]. A MAC protocol is designed based on IEEE 802.11DCF [8], and is modified to exploit spatial diversity. RTSsand CTSs are used along with PHY preambles that allowto estimate the channel and consequently decide the correcttransmission data rate. STCs are used to achieve full diversity.An analysis of the impact of MAC on routing is also carriedout, evaluating the relation between the delay incurred beforesensing a free channel and the advancement obtained with aone-hop transmission. Directional antennas are explored as aspecial case of multiple antenna communications.

In [23], Time-Reversal (TR) STCs are considered. Aftershowing that an optimal maximum likelihood decoder achievesthe maximum diversity order in an intersymbol-interferencemultiple-access channel, the authors prove that using lowercomplexity linear MMSE detectors based on the TR-STCstructure achieves only slightly deteriorated performance.MIMO links and STCs have also been used in [24] foraddressing the problem of efficient broadcasting in ad hocnetworks with multiple antennas.

A different method for managing radio links with multipleantennas is given in [25]. There, a centralized controller is ableto estimate concurrent resource usage and to schedule links toexploit the benefits of MIMO such as SM and interferencesuppression, along with increased transmit rate. The finalobjective is a proportional fair scheduling of transmissions,that accounts for bottleneck links, and is achieved by graphcoloring. An online algorithm is also designed. This lastcontribution, although interesting, makes some very strongassumptions on the PHY layer, e.g., that any transmission usesthe full channel capacity and that signaling at the MAC levelis perfect.

When assessing the performance of an ad hoc networkwith multiple antennas, an accurate characterization of theunderlying PHY level is of paramount importance. In MIMOcommunications, a detailed model of the receiver capabilitiesis needed, in particular for nonlinear multiuser detectors likeBLAST [1], [26]. A low complexity quasi-analytical approachto model the performance of BLAST and its suitability to theevaluation of ad hoc networks was considered in [27].

From the MAC point of view, some of the works above relyon the exchange of signaling messages among communicationparties. MIMO links are inherently omnidirectional and notprone to deafness in the sense used for directional antennas.

Moreover, with the use of a proper receiver and with asufficient degree of spatial diversity available, SM would allowsignificant bit rate improvements.

A number of issues arise when designing MAC protocolsfor MIMO ad hoc networks [28]. In our approach, we startfrom an accurate PHY model and construct MAC and linkmanagement protocols in a completely cross–layer fashion,in order to exploit all available degrees of freedom. Ourprotocol heavily relies on the exchange of information betweenthe PHY and MAC layers, with a twofold objective. First,unlike 802.11, we want the MAC to coordinate transmissionsin order to favor parallel communications, while avoidingchannel overload. Secondly, we want to drive the reception ofSM signals so that wanted ones are sufficiently protected frominterference, using a mechanism to prevent some nodes fromtransmitting if needed. In order to do this, we let the MAC usethe knowledge of ongoing neighboring handshakes to decidewhether or not to grant some requested transmissions, so thatthe interference cancellation capabilities of the MIMO receiverare properly exploited without being overloaded. We will alsoshow how our cross–layer approach significantly outperformsa traditional layered solution.

III. PHYSICAL LAYER MODEL

For the reader’s convenience and for a better understand-ing of protocol design issues, this Section provides a shortsummary of the PHY model we assume for communications.More specific details can be found in [26].

Transmitting nodes — Any node splits transmit data intosub-packets, called PDUs. We suppose that uj PDUs aresent through spatial multiplexing, i.e., using uj antennas,one per PDU, where j is the node index. If NTx nodesare transmitting, the total number of simultaneous PDUs isU =

∑NTx

t=1 ut. Let us assign a progressive index to all transmitantennas, from user 1’s first antenna to antenna uNTx of nodeNTx . Call s′ = [s′1, . . . , s

′U ]T the U -length column vector

containing the symbols transmitted from each antenna, whereT denotes transposition, and let σ2

s′i

= E[|s′i|2] = Ptot/uj

be the power of the ith antenna, given that it belongs touser j, and that the maximum transmit power of any node isconstrained to Ptot . Note that Ptot is equally divided amongactive antennas, as this is the best choice in the absence ofchannel state information at the transmitter [2].

Receiving nodes — Any receiver, say node j, uses allits available antennas NA. Thus, the received signal can bedenoted using the NA-length column vector r(j) = H̃(j)s′ +ν ′(j), where ν ′(j) represents channel noise, and H̃(j) isthe NA × U channel gain matrix. Under a Rayleigh fadingassumption, H̃

(j)�,m is a circularly Gaussian complex random

variable, including fading gain and path loss between the mthtransmit and the �th receive antenna.

We assume that the nodes’ channel knowledge is limited,i.e., at most Nmax

S channels related to as many transmitantennas can be estimated at the beginning of each reception.The set N (j) = {n1, . . . , nNmax

S} contains the indices of such

known antennas (KAs), for which we assume perfect channelestimation. Without loss of generality, we also assume thatN (j) = {1, . . . , Nmax

S }. Note that, due to multiple access,



during each transmission N (j) includes the indices of thetransmissions meant for node j, as well as those of some otherinterfering transmissions. We call H(j) the matrix composedof the columns of H̃(j) with indices in N (j), and similarlydefine the Nmax

S -length vector of data symbols coming fromKAs as s(j) = [s′1, . . . , s

′Nmax

S]T .

The signals whose channel gains cannot be estimated aresaid to come from unknown antennas (UAs). From node j’sperspective, their indices are placed in N̄ (j) = {Nmax

S +1, . . . , U}. Let H̄(j) contain the columns of H̃(j) with indicesin N̄ (j). The corresponding transmitted symbols are groupedin the vector s̄(j) = [s′Nmax

S +1, . . . , s′U ]T . We then rewrite the

received signal as

r(j) = H(j)s(j) + H̄(j)s̄(j) + ν ′(j) . (1)

Modeling the pure interference term H̄(j)s̄(j) is important inan ad hoc scenario, because it accounts for limited terminalcapabilities. Furthermore, nodes may decide to neglect low-power interference and reduce processing, e.g., for energysavings purposes, thereby increasing H̄(j)s̄(j).

In order to simplify the notation, we will omit the index ofthe receive node (j) in the following.

A. The BLAST receiver

Without delving too much into the details, BLAST performsmultiuser detection through space-matched2 signal filteringand successive interference cancellation. The signals are de-tected in order of decreasing received power. All operationsare performed on an Nmax

S × 1 vector of the type z =H†r = Rs + ν + iUA, where † is the transpose-conjugateoperator, R = H†H is the signal correlation matrix, whereasν = H†ν ′ and iUA = HHH̄s̄ are the space-matched filterednoise and interference, respectively.

The BLAST receiver performs a decision feedback detec-tion of the PDUs. At each stage, one PDU is detected, andits contribution is removed from vector z [3]. Let ki be theindex of the PDU detected at stage i, and let z(i) be thevector obtained from z after removing the PDUs with indicesin K(i) = {k1, k2, . . . , ki−1}. The PDU ki can be detectedby combining z(i) with a proper weighing vector w(i). Oneway to do so is to set w(i) equal to the ith column of theMoore–Penrose pseudoinverse of the correlation matrix [29],R+(i), in order to perform zero forcing (ZF) [1]. The decisionon the received symbol is finally performed on the samples̃ki = w(i)T z(i) whose quantized value yields the symbolestimate b̂ki . The estimated symbol is then removed from theprocessed signal, yielding the vector

z(i+1) = z(i)−R·,ki σs(ki)b̂ki , i = 1, . . . , NmaxS − 1 , (2)

where z(1) = z, R·,ki denotes the ki-th column of R, and

σs(ki) =√

E[s2ki

]. The detection (and thus, cancellation)order is based on the PDUs’ signal to noise ratio (SNR) atthe detector input [1].

Note that a receiver with improved performance has beenderived in [26] for use with real constellations, which accounts

2The space-matched filtering corresponds here to multiplying the receivedsignal by the conjugate-transpose of the channel matrix, H†, as explained inthe following.

only for the real part Re [R] instead of R. Accordingly, weshall account for this improvement in the following.

In order to provide at least some insight on the performanceof the system, we report in Fig. 1 the Bit Error Rate (BER)in a distributed context, where a single receiver with NA = 8antennas tries to decode U incoming PDUs of fixed length,each sent from one antenna at full power by a different user.All users are placed at the same distance from the receiver.The simulation has been conducted with 1000-bit PDUs, overmany different channel realizations. As suggested by Fig. 1,there is a fairly low probability of bit error even in the presenceof substantial receiver overload. With 14 incoming PDUs, forinstance, the BER falls below 10−5 for sufficiently high SNR.Note that in a more realistic ad hoc network scenario, wherethe transmitters are randomly placed in the network area, thedifferent average received powers that result would lead toeven better performance of the decision-feedback detector.Furthermore, resorting to a pseudoinverse–based decorrelatingreceiver translates into a soft limit on the number of incomingflows, which is not limited to NA. The next Section is devotedto assessing the impact of this PHY layer on MAC design.

IV. CROSS–LAYER DESIGN OF MEDIUM ACCESS

CONTROL FOR MIMO AD HOC NETWORKS

A. Impact of PHY on MAC

The well known collision avoidance approach describedin the 802.11 standard [8] makes use of control messages(RTS/CTS) in order to mitigate the hidden terminal problem,thus preventing collisions that would result in loss of dataand waste of resources. In a MIMO ad hoc network, however,this is not always the best solution. Specifically, the receiverstructure we presented in Section III is able, given somechannel knowledge, to separate incoming PDUs which wouldthen not result in a collision, but could instead be detectedseparately. This crucial channel knowledge at the receiver isobtained through training preambles preceding packet trans-mission.3 The networking protocols may then choose howmany and which channels to estimate, taking into accountthat the limited receiver capabilities allow locking onto atmost Nmax

S sequences simultaneously. While doing this, theprotocols must be aware of the tradeoff existing between theamount of wanted data to detect and the interference protectiongranted to those data. In other words, trying to detect toomany wanted data packets could leave limited resources forinterference cancellation, leading to data loss. Note that, evenwith channel estimation and spatial demultiplexing, the MIMOreceiver itself is still vulnerable to “hidden terminals” in somesense: if the receiver is not aware of interfering nodes nearby,it cannot estimate their channel and cancel them.

A properly designed MAC protocol can offer much helphere. In particular, the concurrent channel access typicallyfound in ad hoc networks can be exploited, instead of beingsuppressed. Collision avoidance schemes, such as 802.11, tryto avoid concurrency by blocking the nodes that receive an

3Estimating the channel usually requires to calculate correlation over knownsignals preambles, that are typically provided by PN sequences. For a morein-depth discussion on channel estimation, the interested reader is referredto [28], [30], [31].



−5 0 5 10 1510

−5

10−4

10−3

10−2

10−1

100

SNR / receive antenna

Bit

Err

or R

ate

(BE

R)

(U, NA) = (4,8)

(U, NA) = (10,8)

(U, NA) = (14,8)

(U, NA) = (16,8)

(U, NA) = (22,8)

Fig. 1. BER performance of decision-feedback multiuser detection withNA = 8 receive antennas for different numbers of transmit antennas, U , allat the same distance from the receiving node.

RTS or a CTS. Instead of blocking, we want to encouragesimultaneous transmissions. We also want to make the re-ceivers aware of potential interferers, and to exploit the spatialdemultiplexing capabilities of MIMO processing. To this aim,we start with an assessment of the receiver performancewhen receiving data PDUs and signaling packets. Even if notexhaustive, this study is indeed important for two reasons.First, it yields some insight on data transmission capabilitiesand, more specifically, on how many PDUs can be spatiallymultiplexed. Second, it allows to understand the probabilitythat superimposed signaling messages are correctly received.This latter parameter is quite crucial, since if this probabilityis sufficiently high, signaling packets can be relied upon as asource of information on neighboring traffic and handshakes.

For this study, we place a node in the center of a circulararea of given radius R to act as a receiver. The intendedtransmitter is moved from 40 to 140 m away from this receiver.This node sends data in blocks of 1000 bits per antenna(e.g., 2000 bits spatially multiplexed through 2 antennas,4000 bits through 4 antennas, and so on). The maximumpower is constrained, and equally divided among the usedantennas, since this is the best choice in the absence ofchannel state information at the transmitter. Moreover, werandomly place inside the area some further (interfering) datasenders, which always transmit 1000 bits of data at full powerthrough one antenna. Those nodes falling below a thresholddistance d∗ < R are considered KAs, thus their contributioncan be detected and canceled (provided that the limit on themaximum number of channel estimations is not exceeded,after which these nodes become UAs). Conversely, the nodesfalling beyond d∗ are always treated as unknown interferers(UAs). The reasoning here is that a node could either have alimited knowledge of its neighborhood, or not wish to detectall incoming signals, but only those with a sufficiently highreceived power, in order to guarantee detection performance.We have set R and d∗ so that the probability that a signal isdetected is 50%.

The results of this test are given in Fig. 2, using 8 interferersin addition to the intended transmitter. The curves show thatthere exists a tight relationship between the number of used

40 50 60 70 80 90 100 110 120 130 14010

−1

100

Distance of transmit node

Pac

ket S

ucce

ss R

ate

(1−

PE

R)

2 transmit antennas, 8 further nodes4 transmit antennas, 8 further nodes8 transmit antennas, 8 further nodes

Fig. 2. Probability of capturing a data packet in the presence of interferingtraffic versus the distance of the transmitter, for varying number of antennasused by the transmitter.

antennas (thus, bit rate) and the average received power, thusthe maximum coverage distance affordable. For example, witha 90% minimum success ratio objective, a transmitter couldreach 70 m, 90 m and 110 m, using 8, 4 and 2 antennasrespectively. This maximum number of antennas as related tothe distance of a node is called the “class” of that node, andis especially useful in a multiple-receiver context, where thetransmitter could send data to many neighbors at once. In thiscase, the class of a neighbor represents the maximum numberof antennas allowed when transmitting to a set of receiversincluding that neighbor.4

To encourage parallelism, RTS/CTS messages do not blocktransmissions in our scheme, but rather are used for traffic loadestimation. Since signaling packets are shorter and transmittedwith a single antenna at full power, we expect them to bedetectable in large quantities without significant errors. Toverify this intuition, we have considered a similar scenarioas before, with 1 to 20 nodes transmitting simultaneously200 bits long RTSs to a receiver placed at the center ofa circular area. Again, all nodes beyond d∗ are consideredunknown interferers. Besides the previously explained reason,here d∗ also functions as a measure of a node’s knowledgeof its neighboring network activity. We simulate such highertracking capabilities through a greater d∗, and vary it suchthat the average number of interfering nodes (IN) over allnodes is 30%, 10% and, as a limit case, 2%. Recall that thesenodes are all UAs, and thus increase the interference level forthe reception of wanted packets. Conversely, the nodes withind∗ are UAs only if their transmissions exceed the estimationcapabilities at the receiver. Fig. 3 summarizes signaling packetcapture performance. With the same settings as in Fig. 2 (IN= 30%), there is still a fairly high probability of detecting agood percentage of the signaling packets, translating into 13to 15 correct detections with NA = 8 antennas, even if morepackets are sent. This value improves if the node can afford toincrease its neighborhood knowledge (IN = 10%, 2%). The

4This choice of setting classes based on distance is made for the sake of aclearer explanation, but is by no means necessary for correct protocol oper-ation. For example, one could choose the class based on average (estimated)link error probability, link utilization, and so forth.



2 4 6 8 10 12 14 16 18 200.5

0.6

0.7

0.8

0.9

1

Number of nodes within range of the receiver

Pac

ket S

ucce

ss R

ate

(1−

PE

R)

R

IN = 30%

IN = 10%

IN = 2%

Fig. 3. Probability of capturing a signaling packet versus the number of nodeswithin range R of the receiver, for varying number of interfering nodes overthe total number of transmit nodes.

insight gained here is that relying on the exchange of signalingpackets prior to data transmission is in fact possible, becauseit is highly likely that a substantial fraction of these packetsis received correctly.

B. Cross–layer MAC Design

To gather most of these advantages, we resort to a framedcommunication structure, with four phases. For this schemeto work correctly, all nodes have to share the same framesynchronization. These phases are designed according to thestandard sequence of messages in a collision avoidance mech-anism, and are summarized as follows.

RTS phase—In this phase, all senders look into their back-log queue, and if it is not empty they compose transmissionrequests and pack them into a single RTS message. Eachpacket in the queue is split into multiple PDUs of fixed length,such that each PDU can be transmitted through one antenna.For this reason, any request has to specify the number ofPDUs to be sent simultaneously, in addition to the intendeddestination node. How to associate a destination node with asuitable number of transmit antennas is an RTS policy, anddepends on the degree of spatial multiplexing sought, as wellas the local traffic intensity, thus the queue level of the sender.Any RTS may contain several such requests. Moreover, anRTS is always sent with one antenna and at full power.

CTS phase—During the RTS phase, all nodes that were nottransmitters themselves receive multiple simultaneous RTSs,and apply the reception algorithm of Section III to separateand decode them. In the CTS phase, when responding tothe correctly received RTSs, nodes have to account for theneed to both receive intended traffic (thus increasing through-put) and protect it from interfering PDUs (thus improvingreliability). The constraint in this tradeoff is the maximumnumber of trackable channels, i.e., the maximum number oftraining sequences a node can lock onto. We name a CTSpolicy the way the former is traded off for the latter, e.g.,controlling the number of allowed senders and/or the numberof allowed antennas. Since this is a design decision, we deferthe description of the compared CTS policies to the following

Section. CTSs are also sent out using one antenna and at fullpower.

DATA phase—All transmitters receive superimposed CTSsand, after BLAST detection, they follow CTS indications andsend their PDUs. Each PDU has a fixed predefined lengthand is transmitted through one antenna, but a node can sendmultiple PDUs simultaneously, possibly to different receivers.

ACK phase—After detection, all receivers evaluate whichPDUs have been correctly received, compose a cumulativePDU–wise ACK, and send it back to the transmitters. Afterthis last phase, the data handshake exchange is complete,the current frame ends and the next is started. Note thatthis corresponds to the implementation of a Selective RepeatAutomatic Repeat reQuest (SR–ARQ) protocol, where PDUsare individually acknowledged and, if necessary, retransmitted.

Before going more deeply into CTS policy definition, weremark that a random backoff is needed for nodes that donot receive a CTS, as otherwise persistent attempts may leadthe system into deadlock. Here, we make use of a standardexponential backoff. Accordingly, before transmitting, nodeswait for a random number of frames, uniformly distributed inthe interval [1, BW (i)], where i tracks the current attempt, andBW (i) = 2i−1W , with W a fixed backoff window parameter.An accurate study of the effects of different backoff strategiescan be found in [32].

Before proceeding, we highlight that we only require thatnodes be frame–synchronous, even if for simplifying thesystem description, we have referred to [26] in Section III,where synchronization is assumed at the symbol level. In fact,instead of operating on a per–symbol basis, the receiver canfirst detect one whole PDU and then cancel it, detect andcancel the second PDU and so forth, until the last one isdetected. Frame synchronization is not a strong requirement,and can be easily implemented with current technology.

V. RTS AND CTS POLICIES

The last details we need to specify about our MAC protocolare RTS and CTS policies, which are especially important inthis context, since efficient data exchange requires that thereceivers’ detection capabilities are not exceeded, and thatsufficient knowledge of the neighborhood is available. Beforedealing with MIMO-specific policies in Sections V-A and V-B,we introduce here a simpler baseline protocol that we willuse later for comparison. The definition of this protocol isnecessary, since the approaches described in Section II cannotbe directly compared to our solution, because of either theabsence of a specific MAC scheme [19], the optimizationof MAC around some fixed PHY parameters such as thenumber of antennas [20], the diverse issues related to differentmodulation and signaling schemes [21], the attention devotedto achieving full diversity instead of full parallelism [22], orthe idealized assumptions about a MIMO PHY level and MACsignaling [25].

Our baseline, instead, is meant as an example of how alayered networking solution would behave when set up ontop of a SM-capable MIMO PHY level. Furthermore, it isdirectly comparable with our policies, as it takes into accountthe PHY used (unlike [25], that focuses on link capacity) and



is sufficiently general not to depend on the number of antennasper node (unlike [20]). Our baseline works as follows. Whena node has a packet to transmit, it senses the channel, gainingaccess if it finds it free. In order to obtain an optimistic upperbound on the performance of this protocol, we assume thatthe transmitter selection is “ideal”, in the sense that one nodeamong the RTS senders is chosen to transmit, whereas theothers back off. Also, note that the random choice of a nodedoes not yield significant drawbacks, because the nodes of thenetwork simulated in Section VI are within coverage of oneanother (see the same Section for details about this choice) andtherefore any transmission would silence all other handshakes.When a node is granted access, it sends an RTS and waitsfor a CTS from the recipient. To be consistent with the fol-lowing MIMO transmission policies, data packets are dividedin PDUs, each 1000 bits long. Now, the best transmissionenhancement obtainable within this protocol is to increase theraw bit rate as much as possible. To this aim, PDUs are splitin chunks, one per each available antenna, and transmitted inparallel through all antennas. If more than one PDU belongsto the same packet, all PDUs are transmitted sequentially inthe same way. For example, if a packet is formed of a numberof 1000-bits PDUs and NA = 8, each antenna will send one125-bits chunk per PDU. Before returning to the idle stateor performing another transmission attempt, the node waitsfor an ACK from the receiver, reporting which PDUs weredetected correctly. In case of errors, only the erroneous PDUsare retransmitted.

This baseline is a reasonably simple protocol, yet it makesuse of MIMO capabilities and maintains other features similarto 802.11, such as carrier sense and contention-based channelaccess, with no cancellation of interference coming fromother nodes. Basically, the baseline protocol is a carrier-sensemultiple access scheme with collision avoidance, just usinga more powerful MIMO PHY layer. Results based on thisscheme will show that a straightforward use of a layeredsolution on top of the more powerful MIMO PHY is asignificantly suboptimal choice.

A. RTS policy

Let the set of neighbors of a given node s be denoted asV = {v1, v2, . . .}. Also let asvj be the class of vj , j = 1, 2, . . .,which indicates the maximum number of antennas that s canuse when transmitting to any set of nodes that includes vj .Since we wish to encourage spatial multiplexing, we restrictasvj to be either α1 = 2, α2 = 4, or α3 = 8. For clarity, werefer to Fig. 2, and set the class of the neighbors accordingto three threshold distances δ1, δ2, δ3 corresponding to themaximum reach achievable with α1, α2, α3 transmit antennas,respectively. Then, s sets asvj = αm if and only if δm−1 <d(s, vj) ≤ δm, where d(x, y) is the distance between nodex and node y, and δ0 = 0. Note that distances are directlyrelated to the average received SNR value, so that an objectivefunction can be chosen for setting threshold distances basedon either metric.5

5Later, we will use a 90% target success rate, and accordingly set δ1 =70m, δ2 = 90m and δ3 = 110m as per the results of Fig. 2.

Let us focus on node n. The algorithm begins with stepi = 1. If the node queue is non-empty, a request is createdas follows. The node sets k1 = 1 and reads the k1thpacket’s destination, dk1 , and the number of packet PDUs stillunsent, pk1 . Then, it compares pk1 with k1’s class, andk1

. Ifpk1 ≥ andk1

, the unsent PDUs saturate the node class, henceforbidding any further spatial multiplexing. In this case, therequest pair (dk1 , andk1

) is inserted in the RTS, and the RTSis sent right away.

Conversely, if pk1 < andk1, the pair (dk1 , pk1) is put in

the RTS. Node n keeps memory of the queue indices of allpackets selected for transmission, maintaining them in set Si,where i is the step index. It also accumulates in the variableA(i) the total number of antennas allotted until step i. Atstep 1, S1 = {k1}, and by calling M(1) = min{andk1

, pk1}the number of antennas allocated to packet k1, we haveA(1) = M(1). Since pk1 < andk1

, node dk1 could potentiallysustain a transmission with andk1

− pk1 further antennas (inthe absence of interference). In this case there is still room forsending one or more further PDUs taken from other packets.Therefore, the node proceeds to step i = 2 and scans its queue,until it finds a packet k2 whose destination’s class matchesthe condition andk2

> A(1). This means that the destinationdk2 can stand the transmission of the A(1) PDUs alreadyallocated, plus one or more of its own. The sender then setsS2 = S1 ∪ {k2}, calculates the number of PDUs allotted topacket k2 as M(2) = min

{min{andk1

, andk2}−A(1), pk2

},

so as not to violate the antenna constraints andk1and andk2

and taking into account that A(1) antennas have already beenallotted. Then, it inserts in the RTS the pair (dk2 , M(2)),and finally updates A(2) = A(1) + M(2). If there is stillroom for transmission without violating antenna constraints,i.e., if minj∈S2{andj} > A(2), the node proceeds to stepi = 3, searching again for a packet k3 in its queue whoseclass andk3

> A(2), and so on. In general, at step i,the node explores the queue for a packet ki with a fea-sible class andki

> A(i − 1). Then Si = Si−1 ∪ {ki},M(i) = min

{minj∈Si{andj} − A(i − 1), pki

}, and A(i) =

A(i − 1) + M(i). The request(dki , M(i)

)is put in the RTS.

The algorithm then proceeds to step i + 1 if and only ifminj∈Si{andj} > A(i) and a packet such that andki+1

> A(i)is found in the queue. As an example, consider Figure 4.The RTS is formed by first allotting the 2 PDUs required bynode 9, whose class is 8. Node 15 cannot be accommodated,because its class is 2, and two antennas have been alreadyallotted to send two PDUs to node 9. However, node 9can sustain six more antenna transmissions, which allows toaccommodate one PDU for node 7. Notice that node 7’s class(4) now represents the most restrictive constraint, and thatthree PDUs have already been requested so far. This allows athird request for one PDU to node 18, which completes theRTS construction. Note that while this excludes a packet forthe low class node 15 from being served, it will be the firstone to be considered for transmission in the next frame.

B. CTS policies

At first, the node sorts all requests contained in every cor-rectly decoded RTS in order of decreasing received power, and



PDUs#RX id destination

415 2

29 8

7 1 4

418 8

443

RX id # PDUs RX id # PDUs RX id # PDUs

9 2 7 1 118

Request 1 Request 2 Request 3

RTS packet

Class of the

Fig. 4. Example of application of the RTS policy. No request for nodes15 and 3 is included in the RTS, because the maximum number of antennasallowed toward these nodes is too small. In addition, allowing transmissionto node 3 or allowing a request of more than one PDU to node 18 wouldoverload the reception capability of node 7.

divides them in two subsets, namely W and U , respectivelystanding for wanted and unwanted. The first set containsall requests directed to the node, the second set all otherrequests. Recall that if a request by node sk implies thetransmission of, say, rk PDUs, the receiver has to account forchannel estimation resources that will be needed for all PDUtransmissions. Since the maximum number of simultaneousPDUs per receive antenna is limited to Nmax

S , each timea transmission is granted the number of available trackingresources is decreased by rk. Therefore, for each requestconsidered, the receiver inserts in the CTS the pair (sk, r̄k),where r̄k = min{rk, Nmax

S − ∑k−1j=1 rj}, until there are no

more available tracking resources. Grants are given accordingto one of the following policies, with the understanding thatno more than Nmax

S PDUs can be granted.do Not Follow Traffic (NFT)—In this case, the node grants

the requests in W until either they are all granted or allavailable channel estimation resources are used, and does notconsider U at all.

Follow Traffic (FT)—In FT, the node always grants thefirst (highest-power) request in W and then considers all otherrequests in W ∪ U , re-ordered by decreasing received power.At step k, if the processed request belongs to U , no grantsare given in the CTS, but the number of estimation resourcesavailable is decreased according to r̄k . An example is given inFigure 5, for 3 wanted and 3 unwanted traffic requests, eachwith a different number of associated PDUs. By assigningthe channel estimation resources to the requesting nodes inorder of decreasing SNR, the receiver can accommodate allwanted requests. Moreover, it can detect (and cancel) theinterference from U1 and U2, whereas only one PDU from U3

can be canceled, due to lack of further resources. This policystrives to guarantee some throughput through the allowanceof one transmission in W but prioritizes protection fromstrong interference by merging W and U when choosingwhich channels to track. In order to show that these are bothnecessary, we also consider the two following modificationsof FT.

Partially Follow Traffic (PFT)—With PFT, a node givespriority to wanted transmissions, processing first all requestsin W . If there is any tracking resource left, it then beginsto consider requests in U until all resources are exhausted,enabling the cancellation of some neighboring interference.This variant privileges wanted traffic over protection againstinterference.

FT Without Interference Cancellation (FT–WIC)—This

W1

W2

W3

U1

U2

U3

WantedTransmissions

ExpectedInterference

W1

U2

W3

U3

W2

U1 SNR

Cannot be canceled

Fig. 5. Example of application of the FT policy. Darker shades of grayrepresent higher receive SNRs. Some of the unwanted PDUs by U3 cannotbe canceled due to limited channel estimation capabilities, and are left asunknown interference.

policy operates as FT, but does not perform cancellation ofinterfering requests in U . This implies that the only means ofprotection given to data is refraining from transmission if thereare too many powerful interferers. This scheme is thereforeexpected to have poor performance and is considered here tostress the importance of interference cancellation.

For a better understanding of transmitter- and receiver-sideoperations, we report in Figure 6 a pseudo-code description ofour MAC algorithm, where FT has been chosen as the specificCTS policy. Observe that all described policies are cross–layer.On the one hand, they manage network access by selectingwhich incoming PDUs to decode. They perform this controlby accounting for the physical layer, which processes PDUsin order of average received power. On the other hand, theyforce the multiuser detector to decode subsets of PDUs thatcorrespond to different operating points on the throughput–reliability tradeoff. All these decisions are taken based oninformation about per–stream powers, a parameter providedby PHY which is simple, as suggested in [16], but crucial.Note also that i) CTS policies are the only way to reducedata traffic in this kind of networks, since RTSs/CTSs arenot used for channel reservation, but rather as an indicationof intention/clearance to transmit, and ii) both RTS and CTSpolicies favor the creation of multiple point–to–point links,all potentially making use of SM. This is made possible byinserting multiple requests (grants) in the RTS (CTS), eachcomposed of multiple PDUs. An accurate comparison of theCTS policies is carried out in the following Section.

Notice that the policies proposed here are not bound to aMIMO PHY but, on the contrary, they are suitable for use withany decision-feedback multiuser detection-capable PHY level.In other words, our policies can operate on top of any PHYthat successively detects multiple signals, and cancels theircontribution from the received signal prior to the followingdetections. We chose V–BLAST as one such PHY, since itis a good representative and has recently received a lot ofattention [7].

Before proceeding to the description of simulation results,we remark that the cross–layer design of a network includinga MIMO PHY level brings many details into the picture. Thepresented results have been obtained by varying those that



Transmitter-side MAC Operations// Initialize the step index i, the number of allotted antennas A, the set of// receivers S , and the number of failures Nfail

i = 1; A(0) = 0; S0 = ∅; Nfail = 0// RTS phase: add users until class constraints are violatedwhile minj∈Si

andj> A(i − 1) do

// Is there a packet in the queue that complies with the current// constraints?if ∃ a packet ki s.t. andki

> A(i − 1) then// Add user as receiverSi = Si−1 ∪ {ki}// Determine number of PDUs to send that does not violate any// current class constraintM(i) = min{minj∈Si

{andkj} − A(i − 1), pki

}A(i) = A(i − 1) + M(i)Insert request (dki

, M(i)) in RTSend if

end whileSend RTS// Data phase: check CTSif One or more CTS received then

Send data PDUs according to CTSsNfail = 1

elseBackoff for b frames, b uniformly distributed in [1, W · 2Nfail−1]Nfail = Nfail + 1

end ifif ACK received then

Mark all ACK’ed PDUsRemove from the queue all packets whose PDUs have been all ACK’ed

end if

Fig. 6. Pseudo-code description of transmitter- and receiver-side MAC operations. The chosen CTS policy is FT.

most affect the protocol performance. Other minor parametershave been fixed for this purpose, based on a wide range ofpreliminary simulations.

VI. RESULTS

A. Simulation Setup

In order to evaluate our MAC schemes specifically designedfor use with a decision feedback multiuser detector receiver, aswell as the related RTS/CTS policies, we deploy 25 nodes with8 antennas each in a square grid topology with 5 × 5 nodesand nearest neighbors 25 m apart. All nodes are static, andwe assume that the frame synchronization assumption holdsthroughout the simulation. Traffic is generated according to aPoisson process of rate λ packets per second per node. Eachgenerated packet is made of k 1000 bits-long PDUs, withk randomly chosen in the set {1, 2, 3, 4}. Unsent packets arebuffered. We test this specific configuration because nodes areall within coverage range of each other: this is a demandingscenario in terms of interference, required resources, and effi-cient protocol design. Transmissions follow the MAC protocoldescribed in Section IV-B and the policies of Section V. Allother relevant simulation parameters are given in Table I. Wehave built a fully detailed MATLAB simulator that accuratelyreproduces the multiuser detection algorithm at the symbollevel, on top of which we stack the framed MAC describedin Section IV-B and either the baseline or one of the MIMO-specific RTS/CTS policies.

B. Comparison among CTS Policies

In Fig. 7 we compare all CTS policies in terms of aggregatenetwork throughput as a function of traffic. Throughput is

Receiver-side MAC Operations// Initialize number of trackable training sequences, Ns

Ns = Nmaxs

// CTS phase: apply CTS policyif One or more RTSs received then

Create ordered sets W and ULet IW be the ordered set with the indices of the packets in WLet IU be the ordered set with the indices of the packets in U// Grant at least one wanted requesti = IW (1)Read source si and number of PDUs pi for the packet with index iInsert grant (di, pi) in CTSNs = Ns − pi

IW = IW � {i}// Manage other requests in order of decreasing received powerwhile Ns > 0 ∧ (IW �= ∅∨ IU �= ∅) do

Let i be the request with greatest power between IW(1) and IU (1)N = min{pi, Ns}Ns = Ns − Nif i ∈ IW then

Insert grant (di, pi) in the CTSend if

end whileend ifSend CTS// Data phase: receive data PDUsif Data PDUs received then

De-multiplex PDUs and extract wanted onesSend ACK for correctly received PDUs belonging to requests in W

end if

measured here in Mbit/s in the whole network. NFT showsvery poor performance for all traffic values, for two reasons:i) it allows the transmission of all requested PDUs, regardlessof whether the receivers can separate them, and ii) it does notcancel any interferer. PFT performs slightly better since, whilestill granting every requested PDU, it incorporates a mecha-nism that exploits unused estimation resources for cancelingthe strongest interfering PDUs (recall that every policy alwaysconsiders decreasing received powers when selecting what togrant or to cancel). Yet, PFT cannot cope with excessive trafficload. In fact, beginning from λ between 700 and 800, theamount of requested traffic leaves less room for cancellationof unwanted signals, and causes a throughput decrease. Thekey reason why FT performs better than PFT is that FT allotsestimation resources to both wanted and interfering PDUs,while still ensuring that at least one wanted PDU is granted.In the worst case, under exceedingly high traffic, 1 wantedPDU (the one with highest power) would be protected againstthe Nmax

S − 1 highest-power interferers, in an attempt to letsome wanted data get through. The net effect is to activatemore frequently short-distance links that can sustain more SM,as will be shown in Fig. 9. The importance of interferenceprotection is well highlighted by the FT-WIC policy results.From Fig. 7, we recognize the same trend experienced byNFT, just shifted up to some extent. This shift is explained bythe FT–like behavior, as FT-WIC limits traffic in the presenceof interfering PDUs. Nonetheless, FT-WIC still exhibits poorperformance, because it lacks the most important interferencecancellation feature. Note that the baseline transmission policyperforms poorly as well, because it does not make full useof the MIMO capabilities, resulting in very low throughput.We highlight that both NFT and FT-WIC experience a slight



TABLE IRELEVANT SIMULATION PARAMETERS

Description Value

Number of nodes 25

Antennas per node, NA 8

Operating band 5.8GHz ISMData rate per antenna 7.5Mbps

Digital modulation BPSKType of traffic Poisson, constant rate λ

Backoff window parameter, W 1

Maximum backoff window 32 framesSignaling packet length 200bits

PDU length 1000bits

PDUs per packet k, rand. chosen in {1, 2, 3, 4}Queue buffer capacity 120 PDUsPacket timeout 2500 frames (0.53s)No. of trackable sequences, Nmax

S 32

{δ1, δ2, δ3} {70, 90, 110}

linear increase at very low traffic, before reaching a saturationthroughput value that remains then constant at higher traffic.The initial increase is not shown here both because it is anexpected behavior, and because we wish to focus on morespecific policies such as PFT and FT. Moreover, it is expectedthat the throughput does not fall to zero, since some of thesignals are eventually transmitted to nearby nodes, where theSNR allows some non-zero probability of correct reception.

As a side remark, the simulations show that the decisionfeedback multiuser detector of Section III, with FT, is ableto support up to 12 successful PDUs per frame on average,which is larger than the maximum number of antennas pernode, i.e., 8, even in a fully connected network. This is avery interesting result: it substantiates the need for both awell-designed physical layer and a management protocol, andshows that the number of terminal antennas is a soft limitin MIMO ad hoc networks, if efficient RTS/CTS policiesfavoring the effective rejection of multiple access interferenceare provided.

Fig. 8 shows the average ratio of successfully received tosent PDUs, and basically confirms the previous statements.FT achieves the best results and still almost ensures a 90%probability of correct detection at the highest traffic. On thecontrary, NFT and FT-WIC incur a very low probability ofdetection success, and PFT stands in between, its chancesbeing smaller than 40% at high λ. Conversely, the success ratioof the baseline protocol is near 100% as expected, since veryfew transmissions take place due to the collision avoidancemechanism.

To corroborate the claim that the use of FT at high trafficturns into a more likely activation of short links, we depict inFig. 9 the number of grants given to each neighbor dependingon the maximum number of antennas allowed for use withthat neighbor (its “class”). We only show PFT and FT, whichachieve the most significant results, since they experiencemuch less congestion than NFT and FT-WIC. We observe that,after starting from nearly one grant per node per class, bothFT and PFT incur a progressively stronger decrease in thenumber of transmissions allowed toward neighbors in classes2 and 4. On the other hand, transmissions toward class 8

200 400 600 800 1000 1200 1400 16000

10

20

30

40

50

60

Packet arrival rate per node, λ [pkt/sec]

Ave

rage

thro

ughp

ut [M

Bits

/s]

NFTFTPFTFT−WICbaseline

Fig. 7. Throughput for all CTS policies versus traffic.

200 400 600 800 1000 1200 1400 16000

10

20

30

40

50

60

70

80

90

100


Ave

rage

str

eam

suc

cess

rat

io [%

]


Fig. 8. Transmission success ratio of a PDU for all CTS policies versustraffic. Notice the more effective interference protection capabilities of FT,that allow a good success ratio even at high traffic.

neighbors increase more steeply than the others decrease. Sucha behavior can be explained by observing that, in FT, it ishighly likely that the request with strongest power comes froma close class 8 node and that other resources are dedicatedto dealing with interference. The number of class 8 grantsincreases also for PFT, but this is only a consequence of thegreater flexibility given by class 8 nodes, that can afford higherSM and thus allow the receiver to give more grants (see theRTS policy in Section V-A).

The results described before are also confirmed by Figs. 10and 11, which show the average packet delay in secondsand the average queue length, respectively. Consistently withprevious results, we observe that only PFT and FT provide alimited delay, even for higher traffic values. More specifically,PFT reaches a saturation queueing plus transmission delayof approximately 0.08 s, corresponding to 370 frames beingnecessary for a packet to reach the head of the queue andbeing correctly transmitted. For the same traffic values, thehigher throughput achieved by FT is still capable of keepingthe network uncongested, explaining the smoother increasein delay. Similar considerations apply to the behavior ofthe queue length as a function of traffic. In this case, thelower PFT throughput does not allow sufficient packet delivery



200 400 600 800 1000 1200 1400 16000.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2


Ave

rage

num

ber

of g

rant

s pe

r fr

ame

per

clas

s ne

ighb

or

FT, class 2PFT, class 2FT, class 4PFT, class 4FT, class 8PFT, class 8

Fig. 9. Number of grants given to neighbors with different receptioncapabilities per frame versus traffic. Only the FT and PFT policies aredisplayed.

capabilities, hence the node buffers are filled at λ ≥ 800. InFT, instead, a higher amount of data gets through, resultingin a shorter queue length. All other policies, including thebaseline protocol, perform much worse, as their average delayis close to the upper bound imposed by the value of thetimeout.

Overall, the presented results show that the effective cross–layer design that led to FT achieves satisfactory performance,as it allows high throughput and success ratio, hence limiteddelay and backlog. The results also highlight the differencebetween FT-WIC and FT, thus the importance of interferencecancellation when protecting wanted data, especially in acontext where simultaneous channel access is encouraged inorder to exploit SM-capable receivers. Finally, we remark thatconsidering received requests in order of decreasing receivedpower tends to favor shorter links (with greater SINR) at hightraffic. This may have an impact on routing, as under heavytraffic it may be more convenient to set up longer paths withmultiple, more robust hops. A more detailed evaluation of thiseffect involves multihop topologies and routing issues, and isleft for future study.

VII. CONCLUSIONS

In this work, we have addressed the advantages, problems,and general issues arising when designing MAC protocols forMIMO ad hoc networks. We have first studied the performanceof the MIMO decision feedback multiuser detector at thephysical layer in networking scenarios, and used the obtainedresults to gain some understanding of the network behav-ior. We have considered cross–layer policies to drive trafficrequests and grants, with the aim of designing an efficientway to let multiple point–to–point links coexist while keepinginterference under control. During the entire process, we tookexplicitly into account the details and effects of the underlyingphysical layer. We have carried out extensive simulations ofMAC policies in a demanding network scenario with all nodeswithin coverage of each other, and have used these results tohighlight the key features that yield the best performance interms of throughput and other relevant network metrics.

200 400 600 800 1000 1200 1400 16000

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1


Ave

rage

del

ay p

er p

acke

t [s]


NFT, FT−WIC and the baseline protocolreach the maximum delay value (timeout)

Fig. 10. Delay before a correct packet transmission (including queueingdelay) versus traffic. Some curves are not displayed to focus on the moreinteresting comparison between FT and PFT.

200 400 600 800 1000 1200 1400 16000

20

40

60

80

100

120


Ave

rage

que

ue le

ngth

[# o

f 100

0−bi

t str

eam

s]


Fig. 11. Queue length for all CTS policies versus traffic.

Future work on this topic includes a study on the tradeoffsrelated to coding in MIMO ad hoc networks, the use of CDMA(with related changes in the receiver), and the extension tomultihop topologies and routing issues, not considered herewhere the focus was on MAC design and effects.

REFERENCES

[1] G. J. Foschini, “Layered space-time architecture for wireless communi-cation in a fading environment when using multiple antennas," Bell LabsTech. J., vol. 1, no. 2, pp. 41-59, 1996.

[2] H. Jafarkhani, Space-Time Coding: Theory and Practice. CambridgeUniversity Press, 2005.

[3] P. W. Wolniansky, G. J. Foschini, G. D. Golden, and R. A. Valenzuela,“V-BLAST: an architecture for realizing very high data rates over therich-scattering wireless channel," in Proc. IEEE ISSSE, Pisa, Italy, Sept.1998, pp. 295-300.

[4] L. Zheng and D. N. C. Tse, “Diversity and multiplexing: a fundamentaltradeoff in multiple-antenna channels," IEEE Trans. Inform. Theory,vol. 49, no. 5, pp. 1073-1096, May 2003.

[5] S. M. Alamouti, “A simple transmit diversity technique for wirelesscommunications,” IEEE Trans. Commun., vol. 16, no. 8, pp. 1451-1458,Oct. 1998.

[6] E. Telatar, “Capacity of multi-antenna Gaussian channels," Eur. Trans.Telecommun., vol. 10, no. 6, pp. 585-595, Nov. 1999.

[7] A. J. Paulraj, D. A. Gore, R. U. Nabar, and H. Bölcskei, “An overview ofMIMO communications: a key to gigabit wireless," Proc. IEEE, vol. 92,no. 2, pp. 198-218, Feb. 2004.



[8] IEEE Standards Department, ANSI / IEEE Standard 802.11. IEEE Press,1999.

[9] R. R. Choudhury, X. Yang, R. Ramanathan, and N. H. Vaidya, “On de-signing MAC protocols for wireless networks using directional antennas,"IEEE Trans. Mobile Comput., vol. 5, no. 5, pp. 477-491, May 2006.

[10] R. R. Choudhury and N. H. Vaidya, “Deafness: a MAC problem in adhoc networks when using directional antennas," in Proc. IEEE ICNP,Oct. 2004.

[11] R. Ramanathan, J. Redi, C. Santivanez, D. Viggins, and S. Polit, “Adhoc networking with directional antennas: a complete system solution,"IEEE J. Select. Areas Commun., vol. 23, no. 3, pp. 496-506, Mar. 2005.

[12] A. Nasipuri, S. Ye, J. You, and R. E. Hiromoto, “A MAC protocolfor mobile ad hoc networks using directional antennas," in Proc. IEEEWCNC, vol. 2, Chicago, IL, Sept. 2000, pp. 1214-1219.

[13] T. Korakis, G. Jakllari, and L. Tassiulas, “A MAC protocol for fullexploitation of directional antennas in ad hoc wireless networks," inProc. ACM MobiHoc, Annapolis, MD, June 2003.

[14] A. Paulraj, R. Nabar, and D. Gore, Introduction to Space-Time WirelessCommunications. Cambridge University Press, 2003.

[15] Q. Zhao and L. Tong, “A dynamic queue protocol for multiaccess wire-less networks with multipacket reception," IEEE/ACM Trans. Networking,vol. 11, no. 1, pp. 125-137, Feb. 2003.

[16] L. Tong, Q. Zhao, and G. Mergen, “Multipacket reception in randomaccess wireless networks: from signal processing to optimal mediumaccess control," IEEE Commun. Mag., vol. 39, no. 11, pp. 108-112, Nov.2001.

[17] H. Chen, F. Yu, H. Chan, and V. Leung, “A novel multiple access schemeover multi-packet reception channels for wireless multimedia networks,"IEEE Trans. Wireless Commun., vol. 6, no. 4, pp. 1501-1511, Apr. 2006.

[18] Z. Wang, H. R. Sadjadpour, and J. J. Garcia-Luna-Aceves, “Closing thecapacity gap in wireless ad hoc networks using multi-packet reception,"in Proc. ITA Workshop, San Diego, CA, Feb. 2008.

[19] B. Chen and M. Gans, “MIMO communications in ad hoc networks,"in Proc. IEEE VTC 2005-Spring, Stockholm, Sweden, May 2005, pp.2434-2438.

[20] M. Park, S.-H. Choi, and S. M. Nettles, “Cross-layer MAC designfor wireless networks using MIMO," in Proc. IEEE GlobeCom, vol. 2,St. Louis, MO, Nov. 2005, pp. 938-942.

[21] D. Vang and U. Tureli, “Cross-layer design for broadband ad hocnetworks with MIMO-OFDM," in Proc. Signal Processing Advances inWireless Commun., June 2005.

[22] M. Hu and J. Zhang, “MIMO ad hoc networks: medium access control,saturation throughput, and optimal hop distance," J. Commun. Networks,special issue on mobile ad hoc networks, pp. 317-330, Dec. 2004.

[23] S. N. Diggavi, N. Al-Dhahir, and A. R. Calderbank, “Algebraic proper-ties of space-time block codes in intersymbol interference multiple-accesschannels," IEEE Trans. Inform. Theory, vol. 49, no. 10, pp. 2403-2414,Oct. 2003.

[24] F. Rossetto and M. Zorzi, “On gain asymmetry and broadcast efficiencyin MIMO ad hoc networks," in Proc. IEEE ICC, Istanbul, Turkey, June2006.

[25] K. Sundaresan, R. Sivakumar, M. Ingram, and T.-Y. Chang, “Mediumaccess control in ad hoc networks with MIMO links: optimizationconsiderations and algorithms," IEEE Trans. Mobile Comput., vol. 3,no. 4, pp. 350-365, Oct. 2004.

[26] S. Sfar, R. D. Murch, and K. B. Letaief, “Layered space-time multiuserdetection over wireless uplink systems," IEEE Trans. Wireless Commun.,vol. 2, no. 4, pp. 653-668, July 2003.

[27] M. Levorato, P. Casari, S. Tomasin, and M. Zorzi, “Physical layerapproximations for cross-layer performance analysis in MIMO-BLASTad hoc networks," IEEE Trans. Wireless Commun., vol. 6, no. 12, pp.4390-4400, Dec. 2007.

[28] M. Zorzi, J. Zeidler, A. Anderson, B. Rao, J. Proakis, A. L. Swindle-hurst, M. Jensen, and S. Krishnamurthy, “Cross-layer issues in MACprotocol design for MIMO ad hoc networks," IEEE Wireless Commun.Mag., vol. 13, no. 4, pp. 62-76, Aug. 2006.

[29] G. H. Golub and C. F. van Loan, Matrix Comput.. Baltimore, MD: TheJohns Hopkins Univ. Press, 1983.

[30] M. Biguesh and A. B. Gershman, “Training-based MIMO channelestimation: a study of estimator tradeoffs and optimal training signals,"IEEE Trans. Signal Processing, vol. 54, no. 3, pp. 884-893, Mar. 2006.

[31] C. Peel and A. L. Swindlehurst, “Throughput-optimal training fora time-varying multi-antenna channel," IEEE Trans. Wireless Com-

mun., accepted. [Online]. Available: http://zeidler.ucsd.edu/~muri/pages/publications/lswindlehurst3_accepted.pdf

[32] M. Levorato, P. Casari, and M. Zorzi, “On the performance of accessstrategies for MIMO ad hoc networks," in Proc. IEEE GlobeCom, Nov.2006, pp. 1-5.

Paolo Casari (S’05, M’08) was born in Ferrara,Italy, on August 20th, 1980. He received the Laureadegree (BE) in Electronics and TelecommunicationsEngineering (2002) and the Laurea Specialisticadegree (ME) in Telecommunications Engineering(2004) summa cum laude, both from the Universityof Ferrara. From September to December 2004, hewas with the same University, studying geographicprotocols for wireless sensor networks. In March2008, he received the Ph.D. degree from the Uni-versity of Padova, Italy, where he was under the

supervision of Prof. Michele Zorzi. He held Teaching Assistantships atboth Universities, and is currently a postdoctoral research fellow at theUniversity of Padova. During the first half of 2007, he was on leave at theMassachussetts Institute of Technology, Cambridge, MA, working on energy-efficient protocol design for underwater acoustic networks. His main researchinterests are cross-layer protocol analysis and design for wireless networksthrough PHY/MAC/routing interactions, with a particular focus on MIMO adhoc networks, wireless sensor networks, and underwater acoustic networks.

Marco Levorato (S’06) was born in Venice onMarch 18th, 1980. He obtained both the BE (Elec-tronics and Telecommunications Engineering) andthe ME (Telecommunications Engineering) summacum laude from the University of Ferrara (Italy) in2002 and 2005, respectively. During 2005 he held afellowship at the University of Padova (Italy), andfrom January 2006 he has been a Ph.D. student inInformation Engineering at the University of Padovaunder the supervision of Prof. Michele Zorzi. Hisresearch interests include cooperative communica-

tions, design of ad hoc networks with multiuser detection and analysisof Hybrid ARQ techniques. He is currently on leave at the University ofSouthern California, Los Angeles, CA, working on cognitive algorithms forsimultaneous access in ad hoc networks.

Michele Zorzi (S’89, M’95, SM’98, F’07) wasborn in Venice, Italy, in 1966. He received theLaurea degree and the Ph.D. degree in ElectricalEngineering from the University of Padova, Italy, in1990 and 1994, respectively. During the AcademicYear 1992/93, he was on leave at the University ofCalifornia, San Diego (UCSD), attending graduatecourses and doing research on multiple access inmobile radio networks. In 1993, he joined the facultyof the Dipartimento di Elettronica e Informazione,Politecnico di Milano, Italy. After spending three

years with the Center for Wireless Communications at UCSD, in 1998 hejoined the School of Engineering of the University of Ferrara, Italy, and in2003 joined the Department of Information Engineering of the Universityof Padova, Italy, where he is currently a Professor. His present researchinterests include performance evaluation in mobile communications systems,random access in mobile radio networks, ad hoc and sensor networks, energyconstrained communications protocols, underwater networking and cognitivenetworks.

Dr. Zorzi was the Editor-In-Chief of the IEEE WIRELESS COMMUNI-CATIONS MAGAZINE from 2003 to 2005, and is currently the Editor-In-Chief of the IEEE TRANSACTIONS ON COMMUNICATIONS. He serves on theSteering Committee of the IEEE TRANSACTIONS ON MOBILE COMPUTING,and on the Editorial Boards of the IEEE TRANSACTIONS ON WIRELESS

COMMUNICATIONS, the WILEY JOURNAL OF WIRELESS COMMUNICA-TIONS AND MOBILE COMPUTING and the ACM/URSI/KLUWER JOURNAL

OF WIRELESS NETWORKS. He was also guest editor for special issues in theIEEE PERSONAL COMMUNICATIONS MAGAZINE (“Energy Management inPersonal Communications Systems”) and the IEEE JOURNAL ON SELECTED

AREAS IN COMMUNICATIONS (“Multi-media Network Radios” and “Under-water Wireless Communications and Networking”).


4596 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7

Documents

Transcript of 4596 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7