Pushing the limits of Full-duplex: Design and Real-time Implementation

Pushing the limits of Full-duple x: Design and Real-time Implementation Achaleshwar Sahai, Gaurav P atel and Ashutosh Sabharwal { as27,gpatel,ash u } @rice.edu Depar tment of Electrical and Computer Engineer ing Rice University T echnical Repor t TREE1104 ABSTRA CT Recent work has sho wn the feasibility of single-channel full- duplex wireless physical layer , allowing nodes to send and receiv e in the same frequency band at the same time. In this report, we first design and implement a real-time 64- subcarrier 10 MHz full-duplex OFDM physical layer , FD- PHY . The proposed FD-PHY not only allows synchronous full-duplex transmissions but also selective asynchronous full- duplex modes. Further , we show that in ov er-the-air exper - iments using optimal antenna placement on actual devices, the self-interference can be suppressed upto 80dB, which is 10dB more than prior reported results. Then we pro- pose a full-duplex MA C protocol, FD-MA C, which b uilds on IEEE 802.11 with three new mechanisms – shared ran- dom backof f, header snooping and virtual backof fs. The new mechanisms allow FD-MAC to disco ver and exploit full- duplex opportunities in a distributed manner . Our over -the- air tests show over 70% throughput gains from using full- duplex o ver half-duple x in realistically used cases. 1. INTR ODUCTION Half-duplex comm unication, where a no de can either transmit or receiv e in a single c hannel, is the com- monly imp osed constrain t in the design of all practi- cal wireless netw orks. In the last tw o decades, man y w orks [1–8] hav e rep orted exp erimen ts and/or mo dels for full-duplex comm unications. P erhaps the most en- couraging results were rep orted by tw o groups simul- taneously [7, 8] whic h used off-the-shelf hardw are to demonstrate that single-c hannel full-duplex wireless can in fact b e implemented and provides measurable gains o v er half-duplex systems. Ho w ev er, most w ork till date has limited its attention to tw o no des exchanging in- formation with each other, with the fo cus on physi- cal la y er feasibilit y – a crucial first step. Ho w ev er, there is no prior work on the design of medium access proto cols whic h leverage full-duplex communications in m ulti-no de netw orks. In this report, w e prop ose the first full-duplex random access proto col, FD-MA C. F urther, w e implemen t a real-time OFDM-based full-duplex physical la y er (FD- PHY for short) and the prop osed FD-MAC on a W ARP- based testb ed. Our ma jor contributions for FD-PHY and FD-MAC are as follows. FD-PHY : W e develop and implemen t a real-time full-duplex capable physical lay er, FD-PHY. The OFDM- based FD-PHY has 64 sub-carriers and o ccupies 10 MHz bandwidth. The key challenge in full-duplex communi- cations is the large self-interference caused by a node’s o wn transmissions, whic h can completely swamp the pac k ets from other nodes. Th us, analog cancellation (passiv e and/or active) is essen tial to reduce the p ow er of self-interference compared to the pack et of interest b efor e the analog-to-digital conv erter conv erts the sig- nal from the antenna. W e implement an active analog cancellation which injects an appropriately scaled can- celing signal at the receive antenna, to reduce the self- in terference. This activ e cancellation is implemen ted on a p er subcarrier basis, and can th us b e applied to an y OFDM PHY with arbitrary num b er of sub carriers. In [7, 8], the self-interference cancellation was p er- formed both in analog and in digital baseband, together pro viding nearly 70 dB of atten uation to self-interference signal. W e explore another av enue of attenuating the self-in terference – the role of ph ysical placement and orien tation of transmit and receiv e antennas on actual mobile devices, lik e laptops and tablets. W e conduct extensiv e exp erimen ts by moun ting the an tennas on an iP ad-sized device for different antenna configurations. The main finding is that device-induced attenuations com bined with analog cancellation can lead to 80 dB of self-interference suppression, ev en without baseband cancellation. This finding further strengthens the case for actual deploymen t of full-duplex comm unication in mobile devices. F urther, the proposed FD-PHY is also capable of enabling asynchr onous full-duplex communications in some cases, whic h further expands the design space for medium access proto cols. W e show that a full-duplex capable no de can b egin to receive a pack et from a no de while transmitting to another no de, alb eit with a 3 dB loss for the same bit error rate (BER). How ev er, the 1 other case where a full-duplex-capable no de wan ts to transmit a pack et while receiving a pac k et is not p os- sible to implement reliably . This imposes imp ortan t constrain ts on medium access proto cols, whic h the pro- p osed FD-MAC completely adheres to. FD-MA C : Lev eraging the capabilities of FD-PHY, w e develop and implement a random access proto col FD-MA C for infrastructure-based WiFi-like netw orks, where all flows are b et w een an access p oint and mobile units. W e use IEEE 802.11 pack et structure with an additional FD header. The key challenge in maximally using full-duplex capability is to disco ver the opportuni- ties to send and receiv e at the same time in a completely distributed manner. Since no des only hav e the kno wl- edge ab out the pack ets in their o wn queues, discov ering a full-duplex opp ortunity requires sharing queue infor- mation with neighboring no des. At the same time, an y MA C proto col has to allow opp ortunities for all no des to access the medium while trying to maximize net w ork throughput. The FD-MAC uses three mec hanisms to ac hiev e this balance. First mec hanism is the shar e d r andom b ackoff , whic h temp orarily couples the back off counter for tw o no des whic h hav e discov ered that they ha v e a pack et for eac h other. The discov ery that tw o nodes ha v e a pack et for eac h other is p erformed via the FD headers in every D A T A and ACK pack ets. If the tw o no des hav e man y pac k ets for each other, the FD header allows no des to k eep discov ering these full-duplex opp ortunities. One p ossibilit y is that the nodes can o ccup y the medium and con tin uously transmit to each other. Ho wev er, while this maximizes the use of full-duplex mo de, it can p o- ten tially starv e other no des. Thus, we prop ose that once tw o no des disco v er that they hav e more pack ets for eac h other, they use the 10-bit SRB (shared random bac k off ) field in FD header to share a back off counter with each other. The t w o no des back off for a common duration to sta y sync hronized and at the same time allo w other no des to contend and capture the c hannel. Th us the proto col balances access with the maximal use of full-duplex mo de. The second mechanism inv olves no des sno oping on headers of all ongoing transmissions within radio range, ev en when the no des hav e frozen their coun ters during net w ork allo cation vectors (NA V). The pac k et sno op- ing allows no des to estimate their lo cal top ology and in turn disc o ver if the ongoing transmissions b etw een the access p oint and other no des forms a clique or hid- den no de with themselv es. If a Mobile M 2 estimates that it will form a clique with the ongoing AP to Mo- bile M 1 flo w, then M 2 cannot exploit full-duplex since its new transmission will collide with ongoing flow, ei- ther at AP or at Mobile M 1 . Ho w ev er, if M 2 – AP – M 1 forms a hidden no de top ology , then the asynchronous full-duplex capabilities of FD-PHY enable injecting a new pack et to AP while AP is sending a pac k et to M 1 . Lastly , FD-MAC uses t w o virtual c ontention r esolu- tion mechanisms whic h further balance the ob jective to maximally exploit full-duplex mo de to allowing access to other competing flo ws. The salien t mechanism is the case where the AP looks at m ultiple pack ets in its buffer (not just head of line pack et) and statistically decides whic h pac k et it will serve first. By lo oking in to multiple pac k ets in the queue, AP can discov er more opportuni- ties to use the full-duplex mo de. This, of course, leads to the p ossibilit y of AP dela ying the transmission of its HOL pack et whic h could b e problematic at higher lay er proto cols. So we prop ose to send a non-HOL pack et with v anishing probabilit y . Of course, this mec hanism is optional and can be completely turned off at the cost of reduced use of full-duplex capabilities. Our exp eri- men tal results show that FD-MAC ac hiev es a through- put gain of up to 70% ov er comparable half-duplex sys- tems. The gain is a function of distance, pack et arriv al pattern, extent of con ten tion etc. The rest of the rep ort is organized as follows. In Section 2, we review the challenges and state-of-the-art in full-duplex wireless communications. In Section 3, w e describ e the OFDM-based full-duplex physical la y er (FD-PHY) study an tenna placement on actual devices and the p erformance of asynchronous full-duplex. In Section 4, we describ e the mec hanisms for FD-MAC and study its behavior in prototypical top ologies, finally presen ting its ev aluated performance. 2. REVIEW OF FULL-DUPLEX WIRELESS 2.1 Main Bottleneck in Enabling Full-duplex T o appreciate the key challenge in achieving full-duplex wireless, consider the tw o-wa y link shown in Figure 1, where the t w o no des are trying to send and receive a pack et simultaneously in the same frequency band. No de 1 has a pack et for No de 2 and vice v ersa. Since the situation is symmetric, we can fo cus our atten- tion on No de 1. Assuming that the transmit and re- ceiv e antenna are ph ysically different, the p o wer of sig- nal transmitted b y the node by Antenna T1 causes self-in terference at the receiving antenna of the no de, An tenna R1, which can b e anywhere from 15–100dB higher than the signal of in terest coming from transmit an tenna Antenna T2 of No de 2. Since most mo dern systems process the received signal digitally to deco de pac k ets, the analog received signal is conv erted to the digital form using an analog-to-digital conv erter (ADC). With suc h large difference in the p o wers from the tw o signals, self-in terference and signal of interest, the fi- nite resolution of the ADC is the main b ottlenec k in enabling full-duplex comm unications. When tw o radio signals impinge on the an tenna, the v oltage generated at the an tenna is the sum of the t w o signals. That voltage is down-con v erted to the base- 2 RF DA C ADC x [ n ] y [ n ] h 0 SI RF Up Down x 0 [ n ] y 0 [ n ] RF Down ADC RF Up DA C h SI No de 1 No de 2 R1 T1 R2 T2 h S h 0 S Figure 1: A full-duplex transmission betw een t w o no des. band frequency and scaled suc h that the sum of the t w o signals occ upies a voltage range (nominally denoted as [-1,1]) such that full dynamic range of the ADC is used. This ensures the b est p ossible representation of the ana- log signal in the digital domain. If one of the signals is m uc h smaller than the other signal, then it effectively gets few er bits to represent its voltage levels compared to the case where the smaller signal arrived at the ADC b y itself. That is b ecause in the latter case, the au- tomatic gain control algorithm will scale the signal to o ccup y the whole ADC range and thus allow more bits of resolution for the smaller signal by itself. Th us, ev en if the SNR of the signals w as individually high, the wide discrepancy in their r elative amplitudes implies that the smaller signal will hav e lo w er effec- tiv e SNR in the digital domain, leading to the stronger signal swamping the weak er signal. Th us signal-to- interfer enc e-plus-noise r atio (SINR) is an imp ortan t met- ric to determine the performance of any metho d for full- duplex communications. 2.2 Reported Methods T o achiev e full-duplex communication ov er reason- able distances, it is thus important to suppress the self- in terference in the analog domain b efore it reaches the ADC. In 2010, t w o groups [7, 8] rep orted tw o different tec hniques to ac hiev e appro ximately 60-70dB of self- in terference suppression, thereby showing the feasibil- it y of full-duplex transmissions. In [7], the authors pro- p osed an an tenna cancellation metho d using tw o trans- mit an tennas to use b eamforming to create a null at the receiv e antenna. Three physical antennas are needed in the proposed method to achiev e SISO full-duplex. The prototype demonstration show ed full-duplex p er- formance for short inter-node distances like those en- coun tered in IEEE 802.15.4 (e.g ZigBee) equipp ed de- vices. Perhaps a key deplo ymen t challenge is the need for large an tenna separation to ac hiev e an tenna cancel- lation, esp ecially for IEEE 802.15.4 devices which are often targeted for small form-factor devices and thus ma y not ha v e the required physical space to accommo- date the an tennas. In [8], the authors repurp osed MIMO RF chains to generate a canceling signal and add it in analog at the an tenna using RF adder. While this technique does not ha v e the drawbac k of the additional an tennas like in [7], the prototype implemen tation in [8] had a very narrow bandwidth (0.625 MHz) and th us its applicability to wide-band systems lik e 802.11 was not established. 3. REAL-TIME FULL-DUPLEX PHY In this section, we first describ e our wideband, OFDM- based, full-duplex ph ysical la y er implemen ted on an off- the-shelf SDR platform and metho ds to optimize an- tenna placemen t on actual electronic devices to impro ve the capacity and range of full-duplex wireless ph ysi- cal lay er. Based on this implementation, we compare the performance of full-duplex wireless with half-duplex ph ysical lay ers. Finally , w e discuss the challenge in en- abling asynchronous full-duplex systems, and how our prop osed design achiev es partial asynchronous full-duplex transmissions. 3.1 Real-time OFDM T ransceiver The conceptual blo c k diagram of our full-duplex phys- ical la y er is shown in Figure 2. W e use the narro wband tec hnique proposed in [8] for reducing self-interference in the analog domain, and apply it to a wideband OFDM (orthogonal frequency division m ultiplexing) system b y pro cessing each subcarier indep enden tly . Consider No de 1 in Figure 1. Denote the channel b et w een transmit an tenna T1 and receive antenna R1 for sub-carrier k as h k , where k = 1 , . . . , K with K b eing the total num b er of sub-carriers in the OFDM system. F urther, let the signal sent in sub-carrier k b e denoted as x k . Then the self-interference seen at the receiv e antenna in the k th sub carrier, without any cancellation, is giv en by z SI ,k = h k ∗ x k . (1) The abov e represen tation assumes that cyclic prefix is longer than the time delay of the multipath. This as- sumption is easily satisfied for the self-in terference chan- nel since the distance b et w een the transmit and receive an tennas of the self-interference c hannel is very small, thereb y resulting in very limited multipath delay . In most systems, the cyclic prefix is designed for long dis- tances b etw een t w o no des, like N1 and N2 in Figure 1. F ollowing [8], we opt for active self-interference can- cellation by using the ph ysical lay er architecture shown in Figure 2, where w e compute the canceling signal and cancel it b efore the receiv ed signal from the receiv e an- tenna R1 reaches the analog to digital con v erter. This cancellation is not p erformed ov er the air [7] but using a wired assembly and th us does not need extra anten- nas. Let the wireline c hannel b et w een the cancellation transmit chain and receiv e antenna R1 b e represented as h c,k for sub-carrier k ; note that wires are also a c hannels and thus can attenuate and c hange phases like wireless c hannels. 3 Radio Cyclic Prefix IFFT Serial to Cyclic Prefix Radio QPSK to Parallel Serial Cyclic Prefix Remov e Radio Demod FFT Parallel Serial to h c,k h k x [ n ] y [ n ] Multi -ply Parallel IFFT Serial to Parallel to Serial Parallel Modulation QPSK ˆ h k ˆ h c,k x k by Figure 2: A blo ck diagram of the PHY design with self in terference cancellation. The cancellation signal x c,k for the k th sub carrier is computed as x c,k = − ˆ h k ˆ h c,k x k , (2) where ˆ h k and ˆ h c,k represen t the estimates of channels h k and h c,k . In general, the estimates hav e errors and thus not equal to the quantit y they are estimating. So, the self-in terference signal received at the receive antenna after active analo g c anc el lation is z SI , cancel = z SI , k + x c . (3) F rom (3), it is clear that if the c hannel estimates were p erfect, the self-interference can b e completely suppressed in this technique. This is equiv alent to p erfect nulling in the ideal case for the antenna cancellation tec hnique prop osed in [7]. Since w e need to estimate t wo sets of channels h k and h c,k , w e can view the system as a tw o-transmit chain system (like in IEEE 802.11n MIMO mo des) and can exploit the already av ailable physical lay er headers in MIMO pack ets. Thus, no sp ecial PHY headers need to b e added to estimate the required channels to compute the canceling signal. W e leveraged the op en-source MIMO physical lay er designs a v ailable at the W ARP w ebsite [9] as the start- ing p oint for our implementation. The op en-source de- sign o ccupies 10 MHz bandwidth using 64 sub-carriers and also supp orts 2 × 2 MIMO transmissions. One of the mo des in the op en-source design is spatial multiplex- ing, where the transmitter sends t w o different streams of the data to tw o transmit antennas. W e repurp osed the spatial m ultiplexing mo de to implemen t the abov e sc heme, where the second stream in the MIMO design is replaced by the canceling signal x c , which requires m ultiplying the first signal by appropriate canceling co- efficien ts ˆ h k ˆ h c,k . The other ma jor component in our de- sign is the design of estimation pro cedure to obtain the required c hannel estimates ˆ h k and ˆ h c,k . Here again, w e used the MIMO channel estimation blo cks in the op en- source design [9] and hence the details are not pro vided in this rep ort due to lack of space. the 3.2 Antenna Placement on a Mobile Devices W e next inv estigate how full-duplex will p erform on actual mobile devices. The form factor of the mobile device limits its an- tenna placement, distance betw een transmit and receiv e an tennas, and orientation of the antennas. A t present none of the small form factor mobile devices, lik e smart- phones, use 802.11n MIMO mo des since they cannot accommo date tw o RF chains on one device. Th us, we limit our attention to larger form factor devices, lik e tablets and laptops. The driving questions are how should we place the transmit and receiv e antennas on a mobile device to op- timize the p erformance of full-duplex no des. W e con- sider three configurations as shown in Figure 3, with eac h configuration including tw o an tennas – one for transmit and one for receive. Laptop Configuration-C Configuration-B Configuration-A Tx antenna antenna Rx 13 in 9 in Figure 3: Different antenna configurations. The same an tenna configuration w as tested in the presence and absence of the device Configuration A : While most omni-directional an- tennas used in commercial devices (laptops and tablets) are reasonably omni-directional in the far field, they are almost never truly omni-directional in the near field. Most omnidirectional an tennas ha ve small energy trans- mission along the z-axis (i.e, ab ov e and b elow the an- tenna) [10]. The an tenna pattern immediately sug- gests a p otential deploymen t scenario, where the trans- mit and receiv e can b e mounted on top of eac h other; this is lab eled as Configuration A in Figure 3. Configuration B : In man y 802.11n equipped de- vices which hav e tw o antennas to supp ort MIMO mo des, the an tennas are often installed on the opposite end of the device (like the opp osite edges of the screen) to cre- ate sufficient separation b et w een the antennas. This is lab eled Configuration B in Figure 3. The maximal sepa- ration betw een the antennas creates statistically nearly- indep enden t channels to ac hiev e MIMO spatial multi- plexing gains. While Configuration B was not designed for full-duplex op eration, the presence of the actual de- vice (e.g laptop) b et w een the an tennas has the p oten tial 4 to create additional path loss b et w een the t w o anten- nas and thereb y increase the attenuation of the self- in terference. Configuration C : Finally , we will also test the case when one of the an tennas is installed p erp endicular to the other an tenna, lab eled Configuration C in Figure 3. This configuration aims to exploit the potential differ- ence in radiation pattern along different axes. The exp eriments are performed by strapping the tw o 2.4 GHz 7 dBi Desktop Omni An tenna (typical Wifi An tenna) to a iP ad-sized device in different configura- tions. The dimensions are shown in Figure 3. W e fix the transmit p o w er at 6 dBm. F or each configuration, w e test the impact of antenna configuration and the device. The results are summarized in T able 1. The full-duplex PHY was implemented on W ARP b oards, eac h with three radio cards. One radio was connected to the transmit an tenna, the second was connected to the receive antenna and the third pro vided the cancel- ing signal ( x c ) o v er a wire and added in analog after the receiv e antenna. T able 1: The transmit p o wer is 6 dBm. Config. Device In terference Interference T otal Presen t p o w er after analog sup- cancellation -pression A No -28dBm -52dBm 58dB A Y es -28dBm -52dBm 58dB B No -46dBm -71dBm 77dB B Y es -51dBm -75dBm 81dB C No -40dBm -63dBm 69dB C Y es -49dBm -73dBm 79dB F our main results stand out from the T able 1. R esult 1 (Devic e r e duc es self-interfer enc e) : Depend- ing on the configuration, the presence of a device (e.g laptop/iP ad) can make a significant impact on the p ow er of self-interference, by p assively attenuating the signal. The metallic comp onen ts in a laptop-lik e device can significan tly atten uate the signal and thus reduce self- in terference. In Configuration C, device results in an additional attenuation of 9dB attentuation compared to the case when the device is not presen t. The device related atten uation is 5 dB in Configuration B and 0 dB in Configuration A. R esult 2 (Best ful l-duplex c onfigur ation) : The b est configuration in terms of self-in terference p o w er, with and without analog cancellation is Configuration B, where the self-interference p o w er with and without the analog cancellation is low est compared to other configurations. This is, in fact, great news b ecause Configuration B is also the ideal configuration for MIMO systems. Thus, there is a p oten tial to use multiple antennas in either MIMO or full-duplex mo des in mobile devices. R esult 3 (Baseb and c anc elation) : In [7, 8], baseband cancellation was also prop osed to reduce the self-interference p o w er. In our design, we did not implement base-band cancellation due to lack of sufficient FPGA logic on our W ARP b oards, but w e can still achiev e a self-interference suppression which is mor e than the prior work due to added suppression b y the device. R esult 4 (RF r e quir ements for c anc eling signal p ath) : The self-interference p o w er b efore analog cancellation in all configurations is more than 30dB. F or example, in Configuration A, the received p ow er with device is -28dBm for transmit p ow er of 6dBm, which implies 34dB loss in signal pow er when the self-in terference reac hes receive antenna. This implies that the cancel- ing transmit RF chain do es not require a p o w er am- plifier, because the canceling signal trav els ov er a wire and thus suffers only minor attenuation. In fact, w e had to install 40 dB attenuators on our off-the-shelf radio cards, which essentially remo v ed all the pow er amplification b y the p o w er amplifiers. This is again an encouraging news, which sho ws that the full-duplex transceiv er needs one full transmit c hain (up-con v erter for transmit antenna), one radio chain (down-con verter for receive antenna) and a partial transmit chain with- out p ow er amplifier (for canceling signal). Thus, com- pared to SISO transceiver (one transmit and one receive RF chain), full-duplex only needs the additional partial transmit chain. 3.3 Asynchronous Full-duplex So far, the PHY analysis in prior works [7, 8] and in Section 3.2 has b een motiv ated by t w o no des exchanging pac k ets with eac h other as sho wn in Figure 1. How ev er, full-duplex can b e employ ed in more general cases. Con- sider the hidden no de top ology in Figure 4(b), where M 2 is out of radio range of M 1 . Assume AP has a pac k et for M 1 and M 2 has a pack et for AP . In this case, since the AP has to b e a full-duplex node, the k ey question is if the full-duplex mode can b e enabled in an asynchr onous manner. That is, can a new flow be added once a flo w starts transmission. In the hidden no de example, there are t w o p ossibilities for AP : (a) start receiving a pac k et from M 2 after having initiated a transmission to M 1 , (b) start a transmission to M 1 while receiving a pack et from M 2 . A new reception while transmitting : Assume that AP is actively transmitting to M 1 and is contin uously op erating its analog canceler to suppress its o wn self in terference. This ensures that when M 2 starts a pack et, it can b e deco ded by AP ’s receiver. The key c hallenge is that AP has to estimate the channel b etw een M 2 and AP in the presence of self-interference, whic h is required to b e able to deco de M 2 ’s pack et at AP . In almost all curren t systems, ev en with m ultiple users, this training is p erformed without any (in ten tional) interference. Ho w ev er, to enable async hronous full-duplex, w e are 5 required to estimate the channel b et w een M 2 and AP in the presence of self-interference caused by AP ’s ongoing transmission. W e lab el the physical la y er c hannel esti- mation in the presence of ongoing transmission as dirty estimation , and quan tify the loss compared to the con- v en tional systems, all of which ha v e cle an estimation . T able 2: Eac h pack et has a pa yload of 324 bytes and w as QPSK-enco ded. Signal transmit p o wer w as fixed at 6dBm. A total of 1 . 3 × 10 6 bits w ere transmitted. SINR BER BER dirt y clean (with canceler) estimation estimation 18 dB 2 × 10 − 6 0 14 dB 4 × 10 − 4 1 . 4 × 10 − 4 11 dB 9 × 10 − 3 1 . 8 × 10 − 3 8 dB 2 . 4 × 10 − 2 5 × 10 − 3 7 dB 2 . 5 × 10 − 2 9 × 10 − 3 In T able 2, w e rep ort the results for differen t v alues of SINR which were achiev ed by changing the distance b et w een the tw o no des M 2 and AP . F rom T able 2, it is clear that estimating the M 2 → AP channel in the presence of self-interference increases the bit error rate (BER) for all distances. The impact is worse as the SINR reduces; for high SINR, there is hardly an y mea- surable loss and for lo w SINR, the BER in dirt y estima- tion system can be 6 times compared to clean estima- tion, whic h turns out to b e up to 3 dB loss in effectiv e SINR for the asynchronous pack et. This implies that the c ap acity of the ful l-duplex tr ansmission is r e duc e d if ful l-duplex is use d in this asynchr onous mo de. A new transmission while receiving : Now we con- sider the conv erse case, where AP is already receiving a pac k et from M 2 and in tends to send a pac k et to M 1 to lev erage its full-duplex capabilities. Unfortunately, this mo de c annot b e enable d r eliably. The key challenge is calculation of the self-canceling signal in the presence of an ongoing reception. T o com- pute the canceling signal x c , we need to estimate the c hannel co efficien ts h k and h c,k . If the MIMO PHY header is transmitted (as described in Section 3.1) while PHY is receiving a pack et, then the large uncanceled self-in terference will completely swamp the ongoing re- ception. This is because self-interference b efor e cance- lation is almost alwa ys muc h bigger than signal of in- terest (as also discussed in Section 2). While receiving the pac k et, the automatic gain control (AGC) is set to ensure that the incoming signal o ccupies the full dy- namic range of the analog-to-digital conv erter (ADC). Th us the pro cess of estimating the channels to establish canceling signal causes a “self-collision” at the receiver. A p ossible approach is to bac k off on how muc h of the dynamic range is o ccupied by the receiving pack et, th us allowing a big uncanceled signal to not completely destro y the pac k et. The dra wbac k of lost bits of reso- lution is that the quantization noise of the receiver is increased, whic h decreases its effectiv e SINR, increasing BER and thereb y reducing ov erall throughput. Another approach will b e not estimate the self-interference c hannel and simply use older estimates for the desired c hannels. In our exp erimen ts, the self-in terference chan- nel with a device in the middle had sufficient v ariations o v er time, which implies that self-interference canceler can end up doing more harm than go o d if it has out- dated channel estimates. This again, leads to the same situation where full-duplex cannot b e enabled reliably . R esult 5 (Al lowable asynchr onous mo des) : The key result is that async hronous full-duplex can b e enabled to receiv e while transmitting (with some loss in the p er- formance of receiving pack et) but not transmit while receiving. 4. MA C PRO TOCOL DESIGN In this section, we will describe F ull-Duplex Medium Access Protocol (FD-MAC) whic h uses the full-duplex- capable physical lay er describ ed in Section 3. W e will limit our atten tion to infrastructure based systems and fo cus on the scenario inv olving one access p oint ( AP ). This will allow us to define the fundamental elements of a full-duplex MAC proto col. 4.1 Challenges in MA C Design The first challenge in designing full-duplex MAC is iden tification of the no des which can engage in a full- duplex mo de. In any net w ork of multiple nodes, m ulti- ple flo ws with random arriv als exist at the same time, leading to random instances when full-duplex can b e used. The second c hallenge is imp osed by the physical lay er. F rom Section 3.3, either full-duplex has to b e p erformed sync hronously betw een tw o no des (a pac k et exchange) or can b e done async hronously only if a full-duplex node receiv es a pack et while transmitting a pac k et to another no de. Any MAC design has to resp ect this constraint in its design. The third c hallenge is shared b y an y MA C protocol (full or half-duplex) and is to pro vide opportunity to all no des to access the medium. Thus, the access protocol should not unduely fav or full-duplex opportunities ov er half-duplex flows. 4.2 Overview of FD-MA C In the infrastructure-based netw ork, all flows hav e either AP as their source or destination. Th us, at an y giv en time, a maxim um of tw o flows can b e active among full-duplex capable no des. The tw o p ossible scenarios 6 whic h lev erage full-duplex capabilities are shown in Fig- ure 4(a) and 4(b), where (i) AP and mobile no de M 1 are exchanging pack ets or (ii) AP is sending and receiv- ing a pac k et simultaneously from tw o mobile no des M 1 and M 2 , which are hidden from eac h other. M 1 AP (a) The sim- plest net work with 2 no des. M 2 AP M 1 (b) Both the mobile no des are connected to the AP but are not in the radio range of one another AP M 1 M 2 (c) All three no des are in radio range of eac h other M 2 M 3 AP M 1 (d) M 2 and M 3 are hidden to M 1 Figure 4: A line connecting an y t w o nodes indi- cates that they are in radio range of one another FD-MA C is a random access proto col, which will use most of the dominant elements of the IEEE 802.11 DCF. Ho w ev er, while IEEE 802.11 is CSMA/CA, collision a v oidance in FD-MAC is done selectively to leverage full-duplex opp ortunities. FD-MA C introduces follow- ing three new proto col elemen ts. Shar e d r andom b ackoff : When t wo no des, sa y AP and M 1 in Figure 4(a), are in a situation where they hav e man y pack ets for eac h other and thus truly exploit full- duplex, they do not contin uously capture the medium in order to allow other no des to send or receiv e from AP . Instead they agree on a shared random back-off which allo ws other no des to con tend for the medium. If no one else wins, the tw o no des can contin ue with their full-duplex transmission. Sno oping to disc over ful l-duplex opp ortunities : In FD- MA C, no des deco de headers of all ongoing transmis- sions, even when net work allocation vector NA V is non- zero. This allo ws the no des to estimate the local top ol- ogy and initiate full-duplex opp ortunistically . Virtual c ontention r esolution : FD-MAC also has tw o virtual conten tion mechanisms to balance use of the full- duplex mo de with access for all nodes in the netw ork. While the FD-MAC can b e used with or without R TS/CTS, we will only describ e for the more p opular use case of infrastructure mo de of 802.11 which do es not use R TS/CTS. 4.3 FD-MA C Packet structure W e adopt IEEE 802.11 pack et structure and add a new FD header, for managing full-duplex transmissions as shown in Figure 5. Each pack et contains a PHY header, a MAC header, a full-duplex header, a payload and a cyclic redundancy chec k (CRC). Except for the full-duplex (FD) header, all other fields are iden tical to IEEE 802.11 pack ets. W e briefly explain the fields whic h are essen tial to describ e FD-MAC. The PHY header has a preamble and the training sym b ols necessary for the functioning of the physical la y er. The existing elemen ts of the MAC header that w e use in our FD-MAC proto col are Duration ID denoting the duration (DUR) of the pack et, source address (SA), destination address (D A) and FRAG (denoting if there are more fragments of the same pack et in line for the destination). The MAC header distinguishes b et w een data pack et and ackno wledgemen t. F or simplicity of description of the FD-MAC proto col, data pack ets will b e referred as DA T A and ac kno wledgemen t as A CK. The FD header has a one-bit field to distinguish pack et t yp e (DUPMODE) which can either assume v alues HD (indicating that it is a half-duplex pack et) or FD (in- dicating that it is a full-duplex pack et). Then there is a one-bit field, Head-of-line (HOL), indicating that the next pack et in the buffer is for the destination of the curren t pack et. The current 802.11 MAC header has a field lab eled ‘more data’ in F rame Control Field of MA C header, but to av oid any conflict with other uses this field, we ha v e defined HOL in the FD header. The o v erall ov erhead increase is minimal since the HOL is only 1-bit long. The next field rev eals the duration of head of line pac k et, DURNXT, and is useful when HOL = 1. It 2 b ytes long. The next field is mean t for revealing dura- tion of the full-duplex exchange, DURFD. I t to o is 2 b ytes long. The next one-bit is a Clear-T o-Send (CTS) indicating that destination of the current pac ket can send a pac k et to source of the curren t pack et. Finally the FD header has a field for a 10-bit n umber which is the Shared Ran- dom Back off (SRB). Fields DURNXT and DURFD are needed in order to coun ter the hidden no de problem in infrastructure mo de of 802.11. They are optional and in their absence, the FD header is only 13 bits. PHY 802.11 MAC FD header header 1bit CRC 1bit 1bit 10bits 2 bytes 2 bytes DUPMODE HOL DURNXT DURFD header header MAC Pa yload CTS SRB Figure 5: Structure of the pac k et b eing used for the FD-MA C proto col 7 4.4 Shared Random Backoff Consider the most basic tw o-no de example shown in Figure 4(a). It is p ossible that at any giv en time either b oth no des hav e a pac k et for each other or only one no de has a pac k et for the other. Note that in this case, asyn- c hronous full-duplex is not possible b ecause of the PHY constrain ts (Section 3.3), where a no de cannot start a new transmission while it is receiving a pac k et. Thus, no des hav e to find a wa y to sync hronizing their trans- missions, such that they can estimate the channel co- efficien ts for maximal self-interference cancellation (as discussed in Section 3.1). T o maximize the use of full-duplex mo de while re- sp ecting the constrain ts imp osed b y the ph ysical lay er, FD-MA C pro ceeds as follo ws. Assume that the no des con tend for the medium since they do not know if b oth no des ha v e a pac k et for each other or not. Without loss of generalit y , assume that AP wins the conten tion res- olution. Then if the AP has another pac k et lined up in the buffer for M 1 , it sets HOL=1 in the DA T A pack et . Here SRB AP = 0 and the DUPMODE = HD. Thus, the first pac k et in a tw o-w a y exchange is half-duplex. If M 1 receiv es the D A T A successfully and has a pack et for AP , it sends and ACK pac k et with HOL=1 and DURNXT set to the length of the head of the buffer pac k et. Also, SRB=0, CTS = 1. After receiving the A CK, b oth no des know that they can initiate a full- duplex. The PHY needs AP to train its self-interference c hannel, and th us AP sends an A CK pack et, with HOL = 1, and also reveals DURNXT, and set SRB=0, CTS=1. No w the tw o no des are set to b e in full-duplex. They w ait for max(SRB DA T A , SRB ACK ) (whic h is = 0 at this stage) and then send their resp ectiv e D A T A pack ets with the DUPMODE = FD, and DURFD= max(DURNXT AP , DURNXT M 1 ). Each no de sends an ACK only at the end of DURFD duration. Also, AP alwa ys sends the ACK after the M 1 in full- duplex mo de, whic h allows hidden no des to contend in the medium at the end of the ACK from AP ; see Sec- tion 4.5. After one full-duplex transmission, it is p ossible that b oth no des still hav e more pac kets for eac h other, which they will discov er b y setting the FD header fields as describ ed ab ov e. How ever, if the t w o no des contin ue to o ccupy the medium without any breaks, then other no des will get completely starv ed. On the other hand, if the tw o nodes kno w they ha v e a pack et for eac h other but giv e up the medium for other nodes, they will hav e to go through a conten tion resolution again follow ed by one half-duplex pac k et. Th us, it is imp ortan t for no des to retain the knowledge of queue state which they obtain b y ab ov e hand-shaking enabled b y FD header. So, w e introduce the idea of shar e d r andom b ack- off (SRB), where AP and M 1 handshak e on the ran- dom delay they will b oth wait b efore resuming full- duplex mo de. In the ACK sent after receiving first full-duplex pack et, AP pic ks a random back off from [0 , CW max , AP ] where CW max , AP is the current maximum con ten tion window width for AP and places that num- b er in SRB. The mobile no de M 1 also pic ks a ran- dom back off from its own maxim um conten tion windo w [0 , CW max , M 1 ], and places it the SRB field of its ACK pac k et FD header. After the t w o no des hav e finished sending ACKs, they w ait for max(SRB AP , SRB M 1 ). In the 802.11 DCF, back- off countdo wns are paused by carrier-sense even ts. In our work, we require distributed nodes to independently coun t do wn for the same duration and, as such, cannot emplo y this pausing mechanism since they each might see indep enden t channel busyness even ts. Hence, we prop ose a different kind of behavior for the shared ran- dom back off; no des do not pause their back off count- do wns in the presence of energy on the medium but in- stead p erform one final idle-for-DIFS chec k at the end of the interv al ensure that there is nothing curren tly using the medium when they are about to transmit. If no other no de in the netw ork wins the medium b efore this shared back off counter expires, the tw o no des en- ter the full-duplex mode again and contin ue the ab o v e pro cess till they hav e pack ets for eac h other. A time- line of the even ts is shown in Figure 6. Note that the proto col requires AP and M 1 to wait for at least max(SRB AP , SRB M 1 ) b efore transmitting, ho w ever it can tolerate more delay in start of DA T A pack ets of AP as the PHY lay er as already estimated the required chan- nels. Ho w ev er if another no de wins the medium b efore the expiry of the calculated bac k off, then b oth AP and M 1 purge their kno wledge ab out the other no des and start completely afresh. The reason to purge the states is b ecause up on dis- co v ering full-duplex opp ortunities with another no de, sa y M 2 , the AP will mo dify the ordering of pack ets in its buffer to place pack ets destined for M 2 in fron t of the buffer. In Section 4.6, we discuss the idea of reordering the buffer in more detail. Another reason for this purge is to account for the previously discussed mo dification to the back off process. In the presence of other traffic, the shared back off will effectively b e cancelled despite the remov al of the explicit pausing mechanism. Th us, the only difference b et w een traditional bac k offs and our shared back offs is the fact that our full-duplex no des will not pause their back offs in the presence of unde- co dable energy on the medium. at AP in more detail. F ailur e of DA T A or ACK: In a full-duplex exchange if a no de do es not decode D A T A correctly , it do es not send the corresp onding A CK. At this p oin t, synchro- nization of bac k offs is not possible and since both no des ha v e not receiv ed at least one of DA T A or ACK, b oth AP and M 1 purge the information ab out queue state of the no de and con tend for the medium once they do 8 DATA A CK A CK D A T A D A T A HOL = 1 HOL = 1 CTS = 1 CTS = 1 A CK HOL = 1 HOL = 1 DURMODE = FD CTS = 1 CTS = 1 SRB M 1 D A T A D A T A time DURMODE = FD A CK max(SRB M 1 , SRB AP ) A CK A CK SRB AP HOL = 1 D A T A DURMODE = HD max(SRB M 1 , SRB AP ) Source = AP Source = M 1 Figure 6: Timeline of pack ets sent from AP → M 1 and M 1 → AP . The relev ant fields for decision making are listed ab o v e and b elow the pack ets. not detect an y energy on the medium. On the other hand if only one of the ACK fails, the no de not re- ceiving the ACK purges its knowledge of the queue and contends for the medium at the end of tw o ACK p eriods after the DA T A pack et exc hange finishes. It is then a case of ph ysical medium conten tion by the no des with one of nodes ha ving a the bac k offs fixed to max(SRB AP , SRB M 1 ) and others having a random bac k- off. Both ACK failure simply calls for purging queue state information and thus results in another 802.11 - lik e conten tion. Therefore in the p oor channel condi- tions case too, the FD-MAC has a throughput at least as muc h as that of 802.11 (minus the throughput loss due to additional FD header). 4.5 Snooping to Leverage FD Mode Consider the case of three no des, one AP and tw o mobile no des M 1 and M 2 . With three no des such that b oth mobile units can communicate with the AP , there are tw o p ossible top ologies: (i) all nodes can hear each other and thus forming a clique and (ii) M 1 and M 2 are not in the radio range of each other and thus hidden from each other. W e discuss how sno oping headers of the ongoing transmissions can help no des identify op- p ortunities to leverage full-duplex mo dalit y . W e note that there is no explicit top ology disco v- ery mechanism in FD-MA C. Thus, no des estimate the top ology b y o v erhearing pack ets as follo ws. Assume AP sends a DA T A pac k et to M 1 . Since M 2 is asso ciated with AP , so it can deco de the headers and kno ws that the pack et is addressed to M 1 . If the ACK from M 1 is ov erheard by M 2 , then M 2 concludes that it forms a clique top ology with M 1 . Else it concludes that it is in hidden-no de topology with M 1 . Note that M 2 can mak e an error in its estimation due to random channel in- duced errors causing either the DA T A or A CK to drop, eac h leading to a wrong conclusion at M 2 . How ever, since MAC headers can b e enco ded at base-rate, the probabilit y of making errors is often negligibly small. If { M 1 , M 2 , AP } form a clique, then the only possible full-duplex combinations are AP  M 1 and AP  M 2 . The com binations { AP → M 1 , M 2 → AP } and { AP → M 2 , M 1 → AP } are not p ossible because they cause col- lisions (tw o simultaneous incoming pack ets) at one of the mobile no des due to the topology b eing a clique. No w consider the case of hidden no de top ology . In this case, all four full-duplex com binations are possible: (i) AP  M 1 , (ii) AP  M 2 , (iii) { AP → M 1 , M 2 → AP } and (iv) { AP → M 2 , M 1 → AP } . W e hav e dis- cussed how to establish the first tw o full-duplex op- p ortunities, (i) and (ii), in Section 4.4. The third and fourth cases are mirror reflections of eac h other, so we can fo cus on any one of the t w o. Without loss of gen- eralit y , consider case (iii). W e first recall that there is a PHY-imp osed constraint that a no de cannot initiate a new transmission if it is already receiving a pack et from another no de. Thus, since only AP will b e in full-duplex mo de in case (iii), this case is only p ossible if AP b egins to send its pack et to M 1 first. Assume that that is the case where AP wins the conten tion resolution and b egins sending its pack et to M 1 . By sno oping on the HOL field of the FD header, M 2 can learn if AP has another pack et for M 1 or not. If AP do es hav e a HOL line pack et for M 1 , then M 2 can tranmit a pac k et to AP while AP transmit its next pac k et to M 1 , if (a) M 1 is not the radio range of M 2 and (b) M 1 should not b e attempting to achiev e AP  M 1 . In order to ensure (a) M 2 w aits for one A CK duration after the finish of D A T A pac k et from AP . If M 2 do es not receiv e the ACK, it assumes that M 1 is not its radio range. In order to ensure (b), M 2 do es not contend for the medium and allows the AP to capture the media. It then decodes the FD header of DA T A pac k et being sent from AP . If its destination is M 1 and the DUPMODE is HD, then M 1 can transmit its own pac k et to AP . It de- co des the duration DUR of AP ’s pack et and fragments its pack et to ensure it ends no later than AP ’s trans- mission. It also sets the FRAG =1 in its pac k et. The fragmen tation is necessary to av oid collisions with the A CK from M 1 . The A CK from AP will arriv e one ACK p eriod after the finish of the D A T A pac k et. The same pro cedure contin ues as long as AP has a pac k et for M 1 , M 2 has a pac k et for AP , and M 1 do es not hav e a pac k et for AP . Ev en t timeline is shown in the Figure 7. A t an y p oin t of time if M 1 has a pack et for AP , it will co ordinate with AP via ACKs to enable AP  M 1 , and M 2 can disco v er this setup if the DUPMODE=FD for 9 DATA HOL = 1 A CK D A T A A CK A CK D A T A DA = M 1 , HOL = 1 Conten tion p eriod time D A T A D A T A A CK HOL=0 D A T A M 2 wins contention DA = AP , HOL = 1 DA = M 1 , HOL = 1 Source = AP Source = M 1 Source = M 2 DA = AP DUPMODE = HD DA = M 1 DA = M 1 Figure 7: AP → M 1 and M 2 is hidden from M 1 . A CKs from M 1 → AP are not receiv ed at M 2 . The dashed lines in D A T A pack et of AP signify the end of the header whic h M 2 can deco de. Corruption of D A T A implies no ACK from receiver the DA T A pack et from AP to M 1 . Moreov er, DURFD will let M 2 kno w that it should not contend for the medium at least for DURFD + 2ACK p erio ds. This gets rid of unnecessary collisions if pack ets of AP are m uc h smaller than that of M 1 . 4.6 V irtual Contention Resolution In the previous tw o sections, we in tro duced metho ds to allow mobile to AP flows to get a chance to con- tend (Section 4.4) and disco v er opp ortunities to exploit full-duplex capabilities at PHY (Section 4.5). In this section, we introduce tw o more mechanisms which al- lo w (i) AP to break aw a y from a full-duplex handshak e to send pac kets to other no des, since AP can ha ve do wn- link flows for any mobile no de asso ciated with it, and (ii) reduce the probability of collisions in sno oping based full-duplex access. First consider the case where AP in a full-duplex pac k et exchange with a mobile no de M 1 . In standard 802.11 MA C proto col, alw a ys the pack et at the front of the buffer is transmitted, i.e. the depth of the MAC buffer is one. In order to further increase possibility of op erating in full-duplex, the AP can hav e a larger MAC buffer suc h that it has more c hances to find a pac k et for M 1 as long as the no de M 1 has a pack et to send to the AP . It do es so by placing the next av ailable pac k et des- tined for M 1 in front of its buffer, as sho wn in Figure 8, making MAC no longer a FIFO lay er. Bufdepth is a parameter that can b e increased to ac hiev e full-duplex exc hange. W e note that if FIFO op eration is desired then Bufdepth can b e chosen to b e one and hence this mo de is optional . Increasing the depth of the buffer improv es the c hance of op eration in full-duplex. On the flip side, it can starv e transmission of pack ets to other mobile no des. In order to break aw ay from the full-duplex handshake and allo w AP to send pac k ets, virtual conten tion is ar- ranged b et w een the destination of the current head of the buffer and the destination with whom AP engaged in full-duplex exchange. Up on discov ering an oppor- tunit y of a full-duplex exchange with M 1 , AP searc hes for a pack et with M 1 as its destination in its buffer, and sends it if found. After the first full-duplex ex- c hange, AP searc hes through 2nd to Bufdepth pack ets in the buffer and with a probabilit y p pick pic ks the first pac k et with destination= M 1 as the new head of line pac k et. Since the probability of picking k consecutive out-of-order pack ets decays geometrically as p k pick , the AP c hooses to not send head-of-line pack ets with a fast deca ying probability . enables one more round of FD pack et for M 1 pack et for M 2 Head of Buffer Current Buffer Reordering maximizing throughput p pick switches to HD 1 − p pick Figure 8: Virtual con ten tion resolution b etw een pac k ets in the buffer of AP with Bufdepth = 3 . Virtual conten tion is a probabilistic reordering of the MA C buffer at the end of ev ery full-duplex exc hange Second, consider the case where multiple no des are sno oping on ongoing transmissions by AP as describ ed in Section 4.5, sa y M 3 in addition to M 2 as shown in Figure 4(d). If both M 2 and M 3 are hidden from M 1 , then they will b oth send a pack et to AP at the same time and end up colliding at AP since AP can only re- ceiv e one pack et at a time. Thus, it is imp ortan t that there is a mechanism to av oid such collisions. Since M 2 and M 3 do not know ho w many no des are there which ma y try to con tend, they only send a pack et to use the full-duplex mo de at AP probabilistically . That is, each no de whic h detects a full-duplex opp ortunit y , sends the pac k et with probability p i , where p i is computed based on the current maximum back off windo w as p i = β CW max , where β is a pre-c hosen constant whic h controls the ag- gressiv eness in the system. The motiv ation for using p i ∝ 1 CW max is that each no de can use their current maximum conten tion window as a pro xy for amount of exp ected comp etition in the system. Since each no de’s neighborho o d is different, all no des face a different amount of conten tion on the 10 a v erage. Of course, it is p ossible to fix p i = p where p is pre-chosen and allo ws equal c hance for eac h no des. R esult 6 (Imp act of lar ger buffer depth) : Increasing the buffer depth at AP increases the throughput. The increase in throughput comes at the cost of increased dela y due to pack et reordering. In order to understand the tradeoff b etw een dela y and throughput due to larger than one Bufdepth , and the probabilit y p pick w e simulate the buffer of the AP with pack ets for 5 mobile no des. All no des alwa ys hav e pack ets for the AP , and in ra- dio range of one another. Thus full-duplex exchange is alwa ys p ossible and can b e broken only via virtual con ten tion. Only type of conten tion allow ed was vir- tual conten tion in the buffer of AP . F or every Bufdepth , p pick w as ranged from 0 to 1. The AP had uniform traf- fic for all no des with pack ets lined up in an arbitrary order. Figure 4.6 shows a plot of throughput vs. av er- age delay (for the head of the buffer pack et) for different buffer depths. A key finding that the throughput and a v erage delay are linearly related. Larger Bufdepth can help in improving the throughput at the cost of delay . Also, it is often p ossible to obtain the same (through- put, a v erage dela y) pair for smaller Bufdepth by simply increasing the probabilt y of reordering, p pick . 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 Av e rage dela y Throughput Throughput vs. De lay for d iffe re n t depths of Buffe r a t the AP Bufdepth = 10 Bufdepth = 5 Bufdepth = 2 Bufdepth = 1 Figure 9: The throughput is normalized, with a maxim um and minimum b eing 2 (all FD pack- ets) and 1 (all HD pac k ets). Bufdepth = 1 implies 0 dela y The proto col description is now complete. In the next section, w e consider an example top ology to understand ho w all the prop osed mec hanisms in FD-MA C come in pla y . 4.7 State T ransitions in FD-MA C Consider the clique top ology shown in Figure 4(c). Assume that there are four flo ws in the netw ork, AP → M 1 , AP → M 2 , M 1 → AP , M 2 → AP . F our flows bring forth the p ossibilit y of tw o full-duplex scenarios AP  M 1 and AP  M 2 . Figure 10 sho ws the mechanisms which allo w the net- w ork to go from one mo de to another. Eac h transition is enabled b y the features in tro duced b y the FD-MA C proto col. The three no de net w ork with clique top ology can transition from one full-duplex mo de i.e. AP  M 1 to AP  M 2 only through half-duplex mo des. This is so b ecause the first pack et in tw o-wa y full-duplex ex- c hange, as discussed in Section 4.4, is alw a ys a half- duplex pack et. Supp ose that netw ork is in the mo de AP  M 1 . F rom this full-duplex mo de, the netw ork can transition to a half-duplex mo de due to different reasons: (a) if at least one of AP or M 1 has no more pac k ets in the buffer for the other, i.e. if(HOL AP = 0 or HOL M 1 = 0), b oth the AP and M 1 naturally give up full-duplex mo de (b) any of the DA T A or ACK pac k- ets is not deco ded righ t - failure in reception leads to purging of queue states of other no des to start 802.11 t yp e conten tion (c) the pac k ets with M 2 as destination win the virtual con tention resolution allo ws the net work to break aw a y from full-duplex entering a AP → M 2 mo de, and (d) M 2 wins the physical con ten tion during the silen t shared random bac k off p erio d, th us initiating M 2 → AP . All the half-duplex mo des can switch among each other with the 802.11 proto col. Consider the half-duplex mo de M 1 → AP . F rom this mo de the only p ossible tran- sition to a full-duplex mo de is AP  M 1 . The mec ha- nism of t w o-w a y setup is discussed in Section 4.4. The FD-MA C proto col therefore allows all mo des to o ccur b y switc hing b etw een v arious mo des through mecha- nisms introduced by FD-MAC, and some existing 802.11 capabilit y . M 2 M 2 M 1 AP M 2 M 1 AP M 2 M 1 AP AP M 1 Virtual conten tion Physical contention HOL AP = 0 HOL M 1 = 0 Two-w a y FD setup Figure 10: Switching b et w een differen t mo des of op eration in a clique top ology . The part of the state diagram illustrating all the key features of the FD-MA C is sho wn. State diagram has tw o more half-duplex mo des AP → M 1 and M 1 → AP The hidden node topology with t w o mobile nodes and an AP is sho wn in Figure 4(b). With M 1 and M 2 hid- 11 den with resp ect to eac h other, t w o full-duplex flo ws in addition to AP  M 1 , AP  M 2 are p ossible. They are { M 2 → AP , AP → M 1 } , and { AP → M 2 , M 1 → AP } . Eac h of four full-duplex mo des, whether tw o-wa y ex- c hange or otherwise start with a particular half-duplex mo de. F or instance { M 2 → AP , AP → M 1 } is possible only if there exists AP → M 1 as discussed in Section 4.5. In order to ensure that transition to all full-duplex mo des is p ossible, FD-MAC m ust ensure that the half- duplex mo de needed to kick start it is p ossible. Half- duplex mo des among themselves contend via 802.11 t yp e of physical con ten tion. The tw o-wa y full-duplex exc hanges hav e a p erio d of shared random bac k off for other half-duplex mo des to o ccur. Moreov er they also ha v e virtual con tention resolution at the AP to allow dif- feren t half-duplex mo des. On the other hand AP  M 2 , { M 2 → AP , AP → M 1 } t yp e of full-duplex, has AP al- w a ys con tending for the media after it finishes sending the ACK, th us allo wing all other no des to contend and establish a half-duplex communication with AP . Since all half-duplex mo des are p ossible from any starting state. Consequently the sno oping mechanism will al- lo w all full-duplex mo des to o. 4.8 FD-MA C evaluations on W ARP In this section we ev aluate the FD-MAC for a tw o no de full-duplex exchange by implemen ting it on a real time full-duplex system designed using W ARP . Figure 11 sho ws a full-duplex W ARP node, with one transmit and one receive antenna. Figure 11: A full-duplex W ARP no de The exp erimen tal set-up has tw o full-duplex no des exc hanging pac k ets with each other. FD-MAC ensures setting up of the full-duplex up on discov ering an op- p ortunit y to exc hange pack ets in full-duplex mo de. The buffer at b oth the no des alwa ys had a head of line pack et for the other. The ev aluation compares the throughput of full-duplex against half-duplex (again implemented on W ARP). The mo dulation used for transmission was QPSK. R esult 7 (Incr e ase in thr oughput due to ful l-duplex) : The encouraging result is that the throughput of full- duplex t w o-w a y exc hange using FD-MA C is 70% higher than that of half-duplex for identical transmit pow er. T able 3: Num ber of pack ets/sec SINR Throughput Throughput of FD of FD HD 9dB 285 158 8dB 276 165 7dB 253 169 5dB 269 151 5. DISCUSSION AND CONCLUSIONS W e note that FD-PHY and FD-MAC are first real- time design and implementation of full-duplex physical and medium access lay ers, and thus exp ect many av- en ues to further optimize the system performance. Per- haps the most promising is a joint design of transmitted signal, canceling mechanisms and baseband pro cessing on full-duplex no des. W e b eliev e that this could lead to further self-in terference suppression, and p erhaps push the p erformance to near-p erfect full-duplex systems. Considering that full-duplex is still viewed skeptically b y many , it is crucial to demonstrate real-time imple- men tations sho wing fully op erational net w ork stac ks. T ow ards that end, our w ork shows very promising re- sults, creating a strong case for practical use of full- duplex in deplo y ed netw orks. References [1] P .D.L. Beasley , A.G. Stove, B.J. Reits, and B.O. As. Solving the problems of a single antenna frequency modulated CW radar. In IEEE International R adar Confer enc e , 1990. [2] S.Chen, M.A. Beach, and J.P . McGeehan. Division-free du- plex for wireless applications. In IEEE Electr onics L etters , vol- ume 34, pages 147–148, 1998. [3] W. Schac herbauer, A. Springer, T. Ostertag, C.C.W Rupp el, and R. W eigel. A flexible m ultiband fron tend for softw are ra- dios using high if and active interference cancellation. In IEEE International Microwave Simp ossium , 2001. [4] A. Raghav an, E. Gebara, E. M. T entzeris, and J. Laskar. Anal- ysis and design of an interference canceller for collocated ra- dios. IEEE T ransactions on Micr owave The ory and T e ch- niques , 53(3498-3508), 2005. [5] D. W. Bliss, P .A. Park er, and A. R. Margetts. Simultaneous transmission and reception for impro ved wireless netw ork per- formance. In Statistic al Signal Pr o cessing, 2007. SSP ’07. IEEE/SP 14th Workshop on , pages 478–482, Aug. 2007. [6] B. Radunovic, D. Gunawardena, P . Key , A. Proutiereand N. Singh, V. Balan, and G. Dejean. Rethinking indo or wireless: Low power, low frequency , full duplex. T echnical rep ort, Mi- crosoft Researc h, 2009. [7] J. I. Choi, M. Jain, K. Srinivasan, P . Levis, and S. Katti. Achiev- ing single c hannel, full duplex wireless communications. In Pro- c ee dings of ACM Mobic om , 2010. [8] M. Duarte and A. Sabharwal. F ull-duplex wireless communica- tions using off-the-shelf radios: F easibility and first results. In Pr oc e e dings of Asilomar Confer enc e on Signals, Systems and Computers , 2010. [9] W ARP MIMO Reference Design. http://w arp.rice.edu/trac. [10] Constantine A. Balanis. A ntenna Theory: Analysis and Design . John Wiley , third edition, 2005. 12