ABSTRACT. Wireless Full-Duplex: From Practice to Theory. Achaleshwar Sahai

Transcription

1

2 ABSTRACT Wireless Full-Duplex: From Practice to Theory by Achaleshwar Sahai Full-duplex is the ability of a node to transmit and receive simultaneously in the same band. Ideal wireless full-duplex communication can double the spectral efficiency compared to the traditional half-duplex communication. In this dissertation, we study the challenges in realizing full-duplex communication. We tackle the challenges from two different perspectives: node and network. Node perspective: Simultaneous transmission and reception results in a large selfinterference due to the proximity of transmit and receive antennas at the full-duplex node. To establish the feasibility of wireless full-duplex, we develop a wideband real-time physical layer and evaluate its performance on the WARP testbed. Selfinterference reduction in our proposed physical layer is achieved through passive suppression and active cancellation. Based on the constraints of the physical layer, we propose a MAC layer protocol which is designed specifically to discover and enhance opportunities to communicate in the full-duplex mode. Experimental evaluation of our physical layer design, as well as several other full-duplex designs proposed in literature, reveal that active cancellation does not

3 push self-interference all the way upto the thermal noise floor. In this dissertation, we explore the bottlenecks limiting active cancellation in full-duplex systems. We show that the amount of active cancellation is limited by transmitter side noise, particularly by the phase-noise in the local oscillator at the transmitter of the fullduplex node. Thus, unlike conventional half-duplex systems where receiver thermal noise is a limiting factor, full-duplex systems are limited by transmitter side noise. As a key by-product of our analysis, we propose a signal model for a wideband MIMO full-duplex system. We use our proposed signal model to study the performance limits of a system where the start of transmission and the start of reception at a full-duplex node are not synchronized. Interestingly, we discover that the bit-error-rate of the communication mode where the start of transmission precedes the start of reception is better than the mode where start of transmission follows reception of a packet at the full-duplex node. Network perspective: In order to extract gains in capacity from full-duplex operation in a multi-user network, we propose to use full-duplex capable nodes to simultaneously operate uplink and downlink in a network. Such operation results in a new type of interference in the network internode interference, i.e., the uplink transmission from each mobile user starts interfering with the downlink receptions at all the other mobile users. We show a physical layer coding strategy that aligns interference over time and extracts gains in degrees-of-freedom of the network. Finally, we recognize that larger gains from full-duplex operation are possible by leveraging the fact that the strength of the internode interference channel is often different from the uplink/downlink channel. By analysing the uplink/downlink capacity of a network composed of one base-station and two mobile users, we show that full-duplex not only out-performs half-duplex, but also recovers some of the degrees-

4 of-freedom lost due to lack/delay of channel state information of the network. iv

5 Acknowledgments This thesis would not been possible without the support of several wonderful people, and I would like to express my gratitude to each one of them. First and foremost, I would like to thank my advisor Professor Ashutosh Sabharwal for being an excellent guide. Ashu is curious and energetic. And yet he shows abundant patience when it comes to developing a student into a researcher. Through numerous examples, he has taught me how simple ideas are the most powerful ones, and that dissemination of knowledge is as important as its discovery. His ability to fearlessly foray into unfamiliar research territories and repeatedly come out successful is truly admirable. He is more than a researcher, he is an explorer who has always infused his sense of curiosity into me. Most of all, I consider myself fortunate to have an advisor who is not only a great mentor but also a very good friend. Next, I would like to thank members of my thesis committee: Dr. Behnaam Aazhang, Dr. Edward Knightly and Dr. Eugene Ng, all of who challenged me with questions of fundamental importance. Their questions, criticism and valuable suggestions played an important part in the development of this thesis. I would also like to thank all my collaborators: Dr. Suhas Diggavi, Dr. Melda Yüksel, Dr. Vaneet Aggarwal and Gaurav Patel. Throughout my graduate career, each one of them were instrumental in constantly fuelling my thought process with new ideas. A special note of thanks to Gaurav Patel, an amazing research engineer, whose contributions are an integral part of this thesis.

6 vi I spent more than 5 years in Houston and the credit for making my stay happy goes to the people around me. At Rice, I had a very lively research group. Melissa, Chris and Patrick always inspired me with their diligence. Dash made mundane math exciting. And Pedro, DK and Evan made sure that the conflicts during group meetings lasted long enough for us to settle them over beers at Valhalla. Outside my research group, I shared the wonderful company of my Duncan Hall buddies, Oscar, Bei, Sam, Cosmin, Corina, Kanes, Rajoshi, Mayank and many others who on multiple occasions let me unwind over a cup of coffee, making work feel like home. In Houston, I found a remarkable set of friends, Abhinav, Animesh, Chinmay, Deepti, Jatin, Manas, Naren, Ramdas, Richa and Sayantan. Each of you added an unique flavor to my life and without you Houston would not have been half as much fun. Special thanks goes to my roommate Naren, on whose Sambhar and Rasam I have survived for the last 5 years. I would like to thank my family, Bhaiya, Bhabhi, Ma and Papa, towards whom I feel tremendous gratitude. Ma, Papa, with your blessing, I have always felt confident about every decision that I made in my life. I owe all my success to your love and unconditional support. And finally, I would like to thank my sweetest girlfriend, Jingwen. Your presence has always let me coast through difficult times. You not only make my life happy, you make it meaningful.

7 Contents Abstract Acknowledgments List of Figures List of Tables ii v xiii xix 1 Introduction Motivation Background and Related Work Contributions of this thesis Node Perspective Network Perspective Outline of the thesis Design of a Real-Time Full-duplex System Introduction Real-time Full-duplex PHY Architecture of Full-duplex Node Real-time OFDM Transceiver Antenna Placement on Mobile Devices Partially Asynchronous PHY MAC protocol design

8 viii Challenges in MAC Design Overview of FD-MAC FD-MAC Packet structure Shared Random Backoff Snooping to Leverage FD Mode Virtual Contention Resolution State Transitions in FD-MAC FD-MAC evaluations on WARP Summary Phase-Noise: A Fundamental Bottleneck in Design of Full-duplex Node Introduction Unexplained Experimental Observations Classification of Known Architectures Methods of reducing self-interference First Attempt to Answer the Questions Narrowband Signal Model Amount of cancellation Identifying the Bottleneck in Active Cancellation Possible sources of bottleneck Experiment Mimicking active cancellation Experiment: Results and their explanation Answer 1. Impact of Phase Noise on Active Analog Cancellation... 62

9 ix Impact of phase noise on pre-mixer cancellers Performance of different active analog cancellers with imperfect channel estimates Answer 2. Benefit of Digital Cancellation after Active Analog Cancellation Digital cancellation when active analog cancellation uses perfect channel estimate Digital cancellation when active analog cancellation uses imperfect channel estimate Answer 3. Influence of Passive Suppression on Active Cancellation Signal Model for Full-Duplex Narrowband signal model Wideband signal model MIMO full-duplex signal model Asynchronous Full-duplex Introduction System Model Cancelling Self-Interference in Asynchronous Full-Duplex Training the Self-Interference Channel Impact of Asynchronous Full-Duplex on Training Channels Enabling Asynchronous Full-Duplex Receive-while-Sending Send-while-Receiving Summary

10 5 Degrees of Freedom of K-user Full-duplex Uplink/Downlink Channel Introduction New Challenge: Internode Interference System Model Network Encoding and Decoding Performance Metric Result Achievability Example scenario: K = 2L General Achievability Outer Bound Creating Opportunities for Interference Management in Full-duplex Uplink/Downlink Networks Challenges in the Existing Scheme Can longer coherence interval reduce delay? Different polarizations transmit and receive at BS Summary x 6 Generalized Degrees of Freedom of Two-User Uplink/Downlink Channel Introduction Full-duplex vs. Half-duplex Instantaneous CSI vs. Delayed CSI

11 xi Notations System Model Network Channel State Information Models Encoding and Decoding Performance Metric Main Results Instantaneous CSI Delayed CSI Outer Bounds Instantaneous CSI Delayed CSI Achievability Instantaneous CSI Delayed CSI Discussions and Conclusion No need for Fast-fading CSIT requirements for only Downlink communication Extending to multiple users Conclusion and Future Directions Innovations of the thesis Future directions A Appendix for Chapter 3 189

12 xii A.1 Lower bound for autocorrelation function A.2 Estimating the scaling for cancellation for a delay=d A.3 Calculating variance of phase noise A.4 Residual computations after active analog cancellations Bibliography 194

13 List of Figures 1.1 Depiction of self-interference at a full-duplex node Comparison of self-interference in cells of two different sizes Block diagram representation of all the self-interference reduction methods in concatenation A full-duplex transmission between two nodes A block diagram of the PHY design with self interference cancellation Different antenna configurations. The same antenna configuration was tested in the presence and absence of the device A line connecting any two nodes indicates that they are in radio range of one another Structure of the packet being used for the FD-MAC protocol Timeline of packets sent from AP M 1 and M 1 AP. The relevant fields for decision making are listed above and below the packets AP M 1 and M 2 is hidden from M 1. ACKs from M 1 AP are not received at M 2. The dashed lines in DATA packet of AP signify the end of the header which M 2 can decode. Corruption of DATA implies no ACK from receiver

14 xiv 2.8 Virtual contention resolution between packets in the buffer of AP with Bufdepth = 3. Virtual contention is a probabilistic reordering of the MAC buffer at the end of every full-duplex exchange. Prior to the virtual contention resolution, AP was engaged in a full-duplex packet exchange with mobile node M The throughput is normalized, with a maximum and minimum being 2 (all FD packets) and 1 (all HD packets). Bufdepth = 1 implies 0 delay Switching between different modes of operation in a clique topology. The part of the state diagram illustrating all the key features of the FD-MAC is shown. State diagram has two more half-duplex modes AP M 1 and M 1 AP A full-duplex WARP node Classification of methods of reducing self-interference Two architectures of analog cancellers differentiated based on whether the cancellation occurs at RF or analog baseband. The functions r up (.) and r down (.) represent the process of upconversion to RF and downconversion from RF respectively Schematic representation of the experiment in Section 3.5 to acquire copies of a signals using a vector signal analyzer. WARP and Vector Signal Generator were two different signal sources considered in the experiment

15 xv 3.4 Amount of cancellation as a function of the delay for different signal sources measured from the experiment in Section 3.5. Unit delay in the x-axis is the inverse of sampling frequency, which corresponds to 21.7ns Amount of active cancellation as a function of the training length for a delay d = 0 for WARP as the signal source measured from the experiment in Section Amount of active analog cancellation possible in different types of cancellers as function of phase noise. The solid curve is a plot of amount of cancellation possible in pre-mixer cancellers if LOs are not matched, as a function of the variance of phase noise The relationship between amount of active analog cancellation and the amount of digital cancellation in a pre-mixer canceller is shown. Also, we assume σsi 2 = σdown Total cancellation represents the sum of passive and active analog cancellation when operated in cascade in a pre-mixer canceller Timing diagrams of the two possible asynchronous full-duplex modes Although longer training length helps in cancelling the self-interference better, the cancellation saturates around 41 db indicating natural bottleneck in RwS The BER vs. pre-cancellation interference power curve for varying training length of RwS and synchronous full-duplex is shown. Increasing the training length lets reduces the gap between RwS and synchronous full-duplex

16 xvi 4.4 Very short training in presence intended signal offers very noisy estimates of the self-interference channel, thus leading to poor cancellation. Longer training (say, train len = 40) improves the cancellation of self-interference channel, but swamps as many as 40 symbols of intended signal. The modulation used is QPSK and SNR = 30 db RwS operates in two regimes, one where the noise and residual interference are approximately of the same power, and another where residual interference becomes the dominant. SwR has much worse BER than RwS when signal power and pre-cancelled interference are comparable, but SwR and RwS have comparable performance due to improved self-interference channel training. Modulation used is QPSK and SNR = 30 db Network with one 2 antenna base-station and 3 single-antenna mobile stations. The internode interference is between the mobile stations is highlighted The DoF-region in full-duplex network with L antenna base-station and K single antenna mobile-stations

17 xvii 5.3 Pictorial representation of interference alignment in (2, 4) full-duplex network. The internode interference channel at time t 1 and t 2 is same. Identical symbols are transmitted in uplink in the two time-slots such that, at each of the mobile users, subtracting the received signal at time t 2 from the received signal at time t 1 results in an interference free network. The bottom figure shows the resulting network formed by a combination of the network over two time-slots. Note that, the combined network has 4 antennas at the base-station and is also interference free The 4 node X transformation of an (L,K) full-duplex network Switching the polarizations of transmission/reception at the base-station in two consecutive time-slots changes the uplink/downlink channel it observes. Keeping the polarization of transmission/reception same in two consecutive time-slots at the mobile users lets the internode interference channel unchanged (as long as the coherence time is at least two symbol duration long) Two-user full-duplex uplink/downlink network GDoF of two-user full-duplex uplink/downlink channel with instantaneous CSI Modified channels to bound GDoF DL when only delayed CSI is available The relative power levels of downlink signal intended for 1 and uplink signal from 2, as observed at the receiver of 1. Noise floor at the receiver is also depicted

18 xviii 6.5 Achievable GDoF-region and its outer-bound when delayed CSI is available to all the nodes Comparison of the GDoF-region achieved by the two different sub-optimal strategies when delayed CSI is available at all the nodes. The parameter β = 5/ Depiction of codewords at the mobile user A and B in the first time-slot of different blocks to achieve the GDoF-pair (2,2β 2). A symmetric generation is performed in the second time-slot, where A generates one new codeword while B generates three codewords. The height of the bar indicates the power of the codewords, while the pre-log factor of the rate of the codebook is indicated inside the bar The rate and power at which the codewords are transmitted from both mobile users to achieve the non-trivial corner point ( min ( max(2β/3,2β 2),2 ),min ( 4β/3,2 )). The height of the bar indicates the power of the codewords, while the pre-log factor of the rate of the codebook is indicated inside the bar. Figures 6.8a and 6.8b depict the encoding strategy employed in the regime 1 β 3/2 achieving the desired corner point (2β/3,4β/3). Figure 6.8c and 6.8d depict the encoding strategy employed in the regime β > 3/2 to achieve the desired corner point (2β 2,2)

19 List of Tables 2.1 Amount of passive suppression and active cancellation in different configurations. The transmit power is 6 dbm BER for the signal received at the full-duplex node, in the Receive-while-Sending mode with clean and dirty channel estimation. Each packet has a payload of 324 bytes and was QPSK-encoded. Signal transmit power was fixed at 6 dbm. A total of bits were transmitted Number of packets/sec in real time full and half-duplex transmission modes Expected value of the strength of the residual self-interference after active analog cancellation with imperfect estimate of self-interference channel Parameters defining the signal models in (3.50) and (3.51) for SISO narrowband, (3.52) for SISO wideband and (3.53) for MIMO full-duplex for different types of cancellers. We assume that σ si = σ cancel 85

20 1 Chapter 1 Introduction 1.1 Motivation Improving the spectral efficiency of available wireless spectrum is the holy grail of wireless communication. One major stumbling block limiting the spectral efficiency that current wireless communication devices can achieve is due to their ability to communicate only in a half-duplex manner. Half-duplex communication implies that in one time-frequency block, a node can either transmit or receive, but not do both. The traditional paradigms of communication like time division duplex (TDD) and frequency division duplex (FDD) are different embodiments of half-duplex communication. The counterpart of half-duplex communication is full-duplex communication where a node can simultaneously transmit one signal and receive another signal in the same band. Compared to half-duplex communication, simultaneous transmission and reception enabled by full-duplex communication has significant potential of improving the spectral efficiency which makes it an attractive feature for future wireless communication devices. One might wonder why current wireless devices are not full-duplex capable. Due to the proximity of transmit and receive antennas on a device, simultaneous transmission and reception causes the transmitted signal to strongly couple with the signal being received(see Figure 1.1). This transmitted signal that couples with the received signal is called self-interference. Typically, the signal of interest being received is several

21 2 orders of magnitude ( db) weaker than the self-interference signal. Receiving a weak signal of interest in the presence of a very strong self-interference is the biggest challenge in enabling full-duplex communication on a device. Almost all current wireless communication devices are incapable of handling such large self-interference and hence do not operate as full-duplex nodes. Self-interference Full-duplex node Signal of interest Figure 1.1 : Depiction of self-interference at a full-duplex node In the past few years, there has been a growing interest in full-duplex communication and several research teams have successfully demonstrated the feasibility of full-duplex bidirectional communication [1, 2, 3, 4, 5, 6, 7, 8], at least for short ranges. The recent surge of interest in full-duplex communication can be partly attributed to the phenomenon of shrinking cell sizes due to growing prominence of Bluetooth, WiFi, Femto-cell, etc. To understand the connection between the two, consider the following. To achieve a target received signal-to-noise ratio (SNR), transmit power should scale with pathloss. Pathloss increases with distance. Thus, to achieve the samesnrindownlinkattheedgeofthecell, abase-stationmusttransmitatahigher power in a large cell compared to a small cell. Similarly, to achieve a target SNR in uplink, a mobile node at the cell edge in large cell needs to transmit at a higher power compared to a smaller cell. As shown in Figure 1.2, if a base-station and a

22 3 mobile node communicate with each other in uplink/downlink using full-duplex, they will observe larger receive power differential between the self-interference and signal of interest in a large cell compared to a small cell. As cell sizes become smaller, the incoming strength of self-interference has moved into manageable range which makes the study of full-duplex communication important and worthwhile. A A (a) Bigger cell implies bigger power differential (b) Small cell implies smaller power differential Figure 1.2 : Comparison of self-interference in cells of two different sizes 1.2 Background and Related Work The main challenge in enabling a transceiver to operate as a full-duplex node is managing the self-interference. Self-interference is usually reduced by a combination of passive and active methods. Passive methods, which use antenna designs, aim to increase the pathloss for the self-interference signal. Active methods employ the knowledge of self-interference to cancel it from the received signal. Active methods are usually classified into active analog cancellation and digital cancellation. Almost all experimental demonstrations of full-duplex [1, 2, 3, 4, 5, 6, 7, 8, 9] contain some combination of passive suppression, active analog cancellation and digital cancellation. A pictorial abstraction of self-interference reduction methods is shown in Figure 1.3.

23 4 For detailed classification of methods of self-interference reduction, see Chapter 3. Signal of interest thermal noise Self-interference Passive Active Analog ADC Active Digital Figure 1.3 : Block diagram representation of all the self-interference reduction methods in concatenation. Historically, full-duplex has been used in wireless repeaters to receive, amplify and re-transmit a wireless signal. Most designs of full-duplex wireless repeaters [10, 11, 12, 13] relied on passive suppression techniques like physically isolating the transmit and receive antenna from one another. To improve the amount of selfinterference reduction, active cancellation (both analog and digital) designs and algorithmsforwirelessrepeaters/relayshavebeenproposedin[1,14,15,16,17,18,19,20, 21, 22, 23, 24, 25, 26, 27, 28]. One of the earlier works which showed the feasibility of full-duplex through experimental demonstration is [29]. More recently, the feasibility of bi-directional full-duplex, where two full-duplex enabled nodes communicate with one another, has been experimentally demonstrated in[3, 4, 6, 5, 30, 7]. Improvements to initial bi-directional full-duplex architecture were shown by demonstrating the feasibility of full-duplex for multiple input multiple output (MIMO) systems in [31, 32]. Most of the early architectures were confined to short ranges of 8-10 meters (indoor). Range extension for full-duplex via sophisticated passive suppression techniques were proposed and evaluated in [33, 34, 35]. The requirement of separate transmit and receive antenna for enabling full-duplex was eliminated in recent full-duplex archi- Chapter 2 of this thesis is based on [6]

24 5 tectures [36, 9]. Simultaneous work on extensive experimental characterization of full-duplex channel was carried out in [37, 38]. Experimental evaluations indicate that most of the current full-duplex architectures do not suppress self-interference all the way upto thermal noise floor. Thus, residual self-interference is a common phenomenon observed in most full-duplex architectures. Some recent work has also progressed in characterizing the bottlenecks of self-interference reduction [35, 39] to understand the source of residual self-interference. Information-theoretic analysis of the capacity of bi-directional fullduplex in presence of residual self-interference due to dynamic range limitations can be found in [40]. A large body of literature [41, 42, 43, 44, 24, 45, 46] studies the impact on the capacity of a single-source single-destination relay network equipped with a full-duplex relay (with residual self-interference). The impact on capacity of a network with multiple user pairs, all of which are perfect full-duplex (no residual self-interference), has been captured from an information theoretic perspective in [47, 48, 49]. In-band cooperation induced via full-duplex among transmitters/receivers in an interference channel is studied in [50, 51, 52]. Beyond obtaining physical layer gains from full-duplex, MAC layer protocols proposed in [53, 54, 55] identify opportunities of full-duplex in a network and leverage its gains. Further, cross layer optimizations for the purpose of network scheduling in presence of full-duplex enabled transceivers is studied in [56, 57, 58, 59]. A detailed survey which comprehensively describes the literature of full-duplex can be found in a recent article [60].

25 6 1.3 Contributions of this thesis In this thesis, we study full-duplex from two different perspectives (a) node perspective (b) network perspective. In the following, we highlight our contribution through both the perspective Node Perspective Design of a real-time full-duplex node: As we have noted, the main challenge in enabling a node to communicate in a full-duplex manner is to handle the large incoming self-interference. Many researchers still remain skeptical of the possibility of handling such large power differential between the signal of interest and self-interference. In this thesis, we present an experimental demonstration of the feasibility of a real-time full-duplex communication node. The OFDM based real-time physical layer designed with WARP hardware [61] operates as a full-duplex node and supports communication over a bandwidth of 10 MHz. Self-interference management in the physical layer is achieved through a combination of passive suppression and active analog cancellation. Passive suppression in our design is improved by exploring the placement of transmit and receive antenna, on an actual device, relative to one another such that maximum isolation is achieved. Active analog cancellation builds on the design proposed in [3], and uses a separate radio chain to generate the cancelling signal which is injected into the received signal at the radio frequency (RF). Overall reduction in self-interference achieved in our design is 73 db. Based on the constraints posed by our physical layer, we also propose a full-duplex MAC layer, which enhances the opportunities for a node to operate in a full-duplex mode. Based on our work in [6]

26 7 Bottlenecks in active cancellation: In several experimental demonstrations of full-duplex, including ours and [3, 4, 7, 38], active cancellation falls short of reducing the self-interference to thermal noise floor by a large margin. In [37], residual selfinterference is reported to be even as high as 15 db stronger than the thermal noise floor. In this thesis, we investigate the fundamental source of the bottleneck in active cancellation. Through experiments and analysis, we determine that the transmit side noise due to phase noise in the local oscillator successfully explain the amount of active cancellation possible in a full-duplex system. We show that by incorporating the effect of phase noise at the transmitter, we can not only explain the bottleneck in active analog cancellation but also predict the connection between active analog cancellation and digital cancellation. Through our analytical prediction, we show that the sum of active analog cancellation and digital cancellation are bounded by a quantity that depends on the phase noise properties, which matches the trend observed in [37]. As a by product of our analysis, we determine full-duplex physical layer architectures that are better suited to combat phase noise. Further, we propose a parsimonious signal model that allows us to abstract away different forms of active cancellation, captures the dominant noise terms and can be used for signal design and analysis of full-duplex systems. Feasibility of asynchronous full-duplex: Almost all the architectures of fullduplex nodes assume control over synchronization in the time of arrival (reception) and departure(transmission) of packets at the full-duplex node. Such synchronization simplifies the self-interference channel training required for self-interference management. However, packets in a network typically arrive in an asynchronous manner. Based on our work in [62, 63]

27 8 In this thesis we study the feasibility of enabling asynchronous full-duplex to accommodate asynchronous reception and transmission of packets at the full-duplex node. The signal model developed from uncovering the bottlenecks in full-duplex is used to analyse the impact of self-interference on asynchronous full-duplex. We consider two natural modes of asynchronous full-duplex (a) Receive-while-sending (RwS) (b) Sendwhile-Receiving (SwR). In RwS mode, the start of transmission of a packet at the full-duplex node begins prior to the start of reception of another packet. In contrast in the SwR mode, the start of transmission of a packet at the full-duplex node begins after the start of reception. Through analysis, we show that both asynchronous modes of full-duplex can be enabled. The presence of residual self-interference captured by the signal model also reveals that both RwS and SwR suffer some hit in bit-error-rate (BER) compared to half-duplex transmission. While RwS mode has similar BER performance as synchronous full-duplex, SwR suffers significant hit in BER. The difference in performance is due to initial lack of knowledge of self-interference channel, prior to the training, performed when packet is already being received in the SwR mode Network Perspective In order to extract the gains in capacity from full-duplex, the impact of using fullduplex enabled nodes on the capacity of a network needs to be analyzed. In bidirectional full-duplex, rate gains over half-duplex are evident as long as self-interference is sufficiently suppressed. In fact, bi-directional full-duplex can have as much as twice the temporal degrees of freedom compared to half-duplex. However, in a multi-user network, using full-duplex enabled nodes need not necessarily increase the degrees of Based on our work [64]

28 9 freedom of communication. As an example, consider a network where a full-duplex base-station communicates with several full-duplex users in uplink and downlink simultaneously. Due to simultaneous uplink and downlink, the uplink transmission from each user will interfere with the downlink reception at all other users, thus causing a new kind of interference internode interference. In this thesis, we present strategies to tackle internode interference such that rate gains over half-duplex can be achieved. To capture the impact of internode interference, we assume that all nodes are perfect full-duplex nodes, i.e. self-interference is sufficiently suppressed. Internode interference management via ergodic alignment: As a first step, we analyze the degrees of freedom of full-duplex uplink/downlink network, where one multi-antenna full-duplex base-station communicates with several single-antenna full-duplex mobile nodes. We derive an information theoretic outer-bound for the degree-of-freedom region of the full-duplex uplink/downlink network. To achieve the outer-bound we propose an ergodic interference alignment [65] based strategy and show that there exist scenarios where even in the presence of multiple users, fullduplex can achieve twice the degrees of freedom compared to its half-duplex counterpart. Ergodic interference alignment has very strict network channel requirements as well as long delay. We innovate upon the ergodic alignment strategy by drawing connections to methods of passive suppression used for self-interference management, thereby proposing simpler strategies with lesser channel knowledge requirement and significantly lower delay. Opportunistic internode interference management: The presence of internode interference substantially degrades the degree of freedom of a full-duplex up- Partly based on our work in [66]

29 10 link/downlink network, and sometimes even results in half-duplex operation being degree of freedom optimal. However, a degree of freedom analysis fails to capture the difference in the strength of internode interference channel compared to the uplink/downlink channel. In this thesis, we show that even for a network with one multiantenna full-duplex base-station and two single-antenna full-duplex mobile users, the difference in the strength of internode interference channel and uplink/downlink channel impact the capacity in favorable way. We show that as long as the strength of the uplink/downlink channel is not identical to the internode interference channel, full-duplex operation out-performs half-duplex. For the two-user full-duplex uplink/downlink channel, we relax assumptions on channel knowledge and analyze the generalized degrees of freedom [67] possible when the transmitters are equipped with delayed channel state information [68]. Surprisingly, we discover that under fullduplex operation, even with delayed channel state information, the same generalized degrees of freedom can be achieved as with instantaneous channel state information, when the strength of internode interference channel is stronger than uplink/downlink channel. The positive result of full-duplex almost always out-performing half-duplex make a strong case in favor of deploying full-duplex even in multi-user setting Outline of the thesis In Chapter 2 of this thesis, we present the design of a real-time full-duplex system and evaluate its performance. Based on the constraints of the physical layer, we also propose a full-duplex MAC layer protocol. In Chapter 3, we unearth phase noise in the local oscillator as the source of residual self-interference observed in experimental demonstrations of full-duplex. Further, based on the analysis, we explain the connection between active analog cancellation and digital cancellation, and finally

30 11 propose a signal model for MIMO wideband full-duplex. In Chapter 4, we study the feasibility of asynchronous full-duplex operation, i.e., full-duplex communication where transmission and reception are not synchronized. The signal model proposed in Chapter 3 is used to capture the effect of residual self-interference on feasibility of different modes of asynchronous full-duplex operation. In Chapter 5, we switch gears to understand the impact of using full-duplex enabled nodes on the capacity of a network. In particular, we study an uplink/downlink network with one multi-antenna full-duplex base-station and several single-antenna full-duplex users and derive optimal achievable schemes to leverage the degree of freedom gains from full-duplex operation. In Chapter 6, we extend the analysis of Chapter 5 to capture the effect of difference in the strength of internode interference channel and uplink/downlink channel and show further gains in capacity. Finally, we conclude in Chapter 7 with some remarks on future directions.

31 12 Chapter 2 Design of a Real-Time Full-duplex System 2.1 Introduction In this chapter, we present the design of a real-time full-duplex capable physical layer, FD-PHY. Moreover, we present the architecture of a full-duplex MAC layer, FD-MAC, compatible with FD-PHY. The key features of the FD-PHY and FD-MAC are as follows. FD-PHY: The orthogonal frequency division multiplexing (OFDM) based FD- PHY proposed in this chapter has 64 sub-carriers and occupies 10 MHz bandwidth. Passive suppression and active analog cancellation are two key elements of FD-PHY which enable it to reduce the strength of self-interference prior to quantization of the received signal by an analog to digital converter (ADC). In FD-PHY, the active analog cancellation is implemented on a per subcarrier basis, and therefore can extend easily to an OFDM PHY with arbitrary number of subcarriers. Attenuation of self-interference by passive methods like increasing the distance between transmit and receive antenna has been explored in [3]. Here, we explore another avenue of attenuating the self-interference the role of physical placement and orientation of transmit and receive antennas on actual mobile devices, like laptops and tablets and discover that suitable placement/orientation of transmit and receive antenna can provide significant boost to the feasibility full-duplex.

32 13 FD-MAC: Leveraging the capabilities of FD-PHY, we develop a random access protocol, FD-MAC, for infrastructure-based WiFi-like networks, where all flows are between an access point and mobile units. We use IEEE packet structure with an additional full-duplex header. The key challenge in maximally using full-duplex capability is to discover the opportunities to send and receive at the same time in a completely distributed manner. Since nodes have knowledge about the packets only in their own queues, discovering a full-duplex opportunity requires sharing queue information with neighboring nodes. At the same time, any MAC protocol has to allow opportunities for all nodes to access the medium while trying to maximize network throughput. To achieve this balance in FD-MAC, we introduce three new mechanisms (a) shared random backoff, (b) snooping and (c) virtual contention resolution. The rest of the chapter is organized as follows. In Section 2.2, we describe the OFDM-based full-duplex physical layer, and the study impact of intelligent antenna placement on actual devices. Further, we discuss the feasibility of operating FD-PHY in a partially asynchronous manner. In Section 2.3, we describe FD-MAC and study its behavior in prototypical topologies, finally presenting its evaluated performance. We conclude with some remarks in Section Real-time Full-duplex PHY In this section, we first review general design principles behind the architecture of a full-duplex node. Then, we describe our wideband OFDM-based full-duplex physical layer implemented on an off-the-shelf software defined radio platform and methods to optimize antenna placement on actual electronic devices to improve the capacity and range of full-duplex wireless physical layer. Based on this implementation, we compare the performance of full-duplex wireless with half-duplex physical layers. Finally, we

33 14 describe the challenge in enabling asynchronous full-duplex systems in our FD-PHY, and highlight how our proposed design achieves partially asynchronous full-duplex transmissions Architecture of Full-duplex Node The architecture of a full-duplex transceiver is abstracted out in Figure 1.3, where a serial concatenation of passive suppression, active analog cancellation and digital cancellation are shown as methods of self-interference reduction. Passive Suppression Passive suppression acts as the first line of defence against self-interference (see Chapter 3 for a detailed classification). Passive suppression is the reduction of electromagnetic coupling between the transmit and the receive path at the full-duplex node. The amount of passive suppression is calculated as difference between the received selfinterference power before and after the method of passive suppression is introduced. One simple method of achieving electromagnetic isolation is using separate antennas to transmit and receive. In this chapter, we improve that technique by investigating the placement on transmit and receive antennas relative to one another on actual devices (e.g. ipad). A similar study of the improvement achieved through intelligent antenna placement is described in [38]. Active Analog Cancellation Since self-interference is several orders of magnitude larger than the signal of interest, even sophisticated passive suppression techniques [35] do not achieve more than db of reduction. Thus, active cancellation of self-interference is critical for reducing

34 15 it further. Active analog cancellation is the injection of a cancelling signal into the received signal in the analog domain at the full-duplex node in order to negate the self-interference. In order to appreciate the need for cancelling self-interference in the analog domain, we need to understand its connection to the limited dynamic range of analog components in a receive radio chain. The received signal at the full-duplex node is combination of a strong self-interference signal and a weak signal of interest. In small form factor devices, large passive suppression is not possible. Thus, it is likely that while operating as a full-duplex node, the power of the received signal(dominated by self-interference) can be outside the dynamic range of analog components like low noise amplifier, or even saturate the radio front-end. If the received signal is outside the dynamic range of the analog components, then non linearities will be introduced which are different to eliminate in the digital domain. This necessitates active analog cancellation, i.e., cancellation of self-interference prior to digitization of the received signal. Digital Cancellation Digital cancellation is the injection of a cancelling signal into the received signal at the full-duplex node after the received signal has been digitized. Digital signal processing is easier to implement compared to analog signal processing. Thus, digital cancellation is a useful method for further reducing the residual self-interference. It also acts as a safety net whenever active analog cancellation fails Real-time OFDM Transceiver The conceptual block diagram of our full-duplex physical layer is shown in Figure 2.2. We use the narrowband technique proposed in [3] for reducing self-interference

35 16 T1 T2 x[n] DAC RF Up h si h signal h si RF Up DAC x [n] y[n] ADC RF Down R1 h signal R2 RF Down ADC Node 1 Node 2 Figure 2.1 : A full-duplex transmission between two nodes. y [n] in the analog domain, and apply it to a wideband OFDM system by processing each subcarrier independently. Consider Node 1 in Figure 2.1. Denote the channel between transmit antenna T1 and receive antenna R1 for subcarrier k as h k, where k = 1,...,K with K being the total number of subcarriers in the OFDM system. Further, let the signal sent in subcarrier k be denoted as x k. Then, the self-interference seen at the receive antenna in the k th subcarrier, without any cancellation, is given by z si,k = h k x k. (2.1) The above representation assumes that cyclic prefix is longer than the time delay of the multipath. This assumption is easily satisfied for the self-interference channel since the distance between the transmit and receive antennas of the self-interference channel is very small, thereby resulting in very limited multipath delay. In fact in most systems, the cyclic prefix is designed for long distances between two nodes, like N1 and N2 in Figure 2.1 where multipath delay can be significantly longer than self-interference channels. Following [3], we opt for active self-interference cancellation by using the physical

36 17 x[n] Serial to Parallel Modulation QPSK Multi -ply x k by ĥ k ĥ c,k IFFT IFFT Parallel to Serial Parallel to Serial Cyclic Cyclic Prefix Prefix Radio Radio h k h c,k y[n] Parallel to Serial Demod QPSK FFT Serial to Parallel Remove Cyclic Prefix Radio Figure 2.2 : A block diagram of the PHY design with self interference cancellation. layer architecture shown in Figure 2.2, where we inject the cancelling signal into the received signal from the receive antenna R1 even before it is processed by the receiver radio chain. Unlike [4], the active cancellation in our design is not performed over the air, rather uses a wired assembly and thus does not need extra antennas. Let the wireline channel between the cancellation transmit chain and receive antenna R1 be represented as h c,k for subcarrier k; note that wires are also a channels and thus can attenuate and change phases like wireless channels. The cancellation signal x c,k, for the k th subcarrier is computed as x c,k = ĥk ĥ c,k x k, (2.2) where ĥk and ĥc,k represent the estimates of channels h k and h c,k. In general, the estimates have errors and thus not equal to the quantity they are estimating. So, the

37 18 residual self-interference signal after active analog cancellation is z si,cancel = z si,k +x c,k. (2.3) From (2.3), it is apparent that if the channel estimates were perfect, self-interference can be completely suppressed using this technique. This is equivalent to perfect nulling in the ideal case for the antenna cancellation technique proposed in [4]. Since we need to estimate two sets of channels h k and h c,k, we can view the system as a two-transmit chain system (like in IEEE n MIMO modes) and can exploit the already available physical layer headers in MIMO packets. Thus, no special PHY headers need to be added to estimate the required channels to compute the cancelling signal. We leveraged the open-source MIMO physical layer designs available on the WARP website [61] as the starting point for our implementation. The open-source design occupies 10 MHz bandwidth using 64 subcarriers and also supports 2 2 MIMO transmissions. One of the modes in the open-source design is spatial multiplexing, where the transmitter sends two different streams of data to two transmit antennas. We repurposed the spatial multiplexing mode to implement the above scheme, where the second stream in the MIMO design is replaced by the cancelling signal x c,k, which requires multiplying the first signal by appropriate cancelling coefficients ĥk ĥ c,k. The other major component in our design is the design of estimation procedure to obtain the required channel estimates ĥk and ĥc,k. Here again, we used the MIMO channel estimation blocks in the open-source design [61].

38 Antenna Placement on Mobile Devices We next investigate how full-duplex will perform on actual mobile devices. The form factor of a mobile device limits its antenna placement, distance between transmit and receive antennas, and orientation of the antennas. At present none of the small form factor mobile devices, like smartphones, use n MIMO modes since they cannot accommodate two RF chains on one device. Thus, we limit our attention to larger form factor devices, like tablets and laptops. The driving question is How should we place the transmit and receive antennas on a mobile device to optimize the performance of full-duplex nodes?. We consider three configurations as shown in Figure 2.3, where each configuration uses two antennas one for transmission and the other for reception. Configuration-A Configuration-B Configuration-C Rx antenna Laptop 13 in Tx antenna 9 in Figure 2.3 : Different antenna configurations. The same antenna configuration was tested in the presence and absence of the device Configuration A: While most omni-directional antennas 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 omni-directional antennas have small energy transmission along the z-axis (i.e, above and below the antenna) [69]. The antenna pattern immediately suggests a potential deployment sce-

39 20 nario, where the transmit and receive can be mounted on top of each other; this is labeled as Configuration A in Figure 2.3. Configuration B: In many n equipped devices which have two antennas to support MIMO modes, the antennas are often installed on the opposite end of the device (like the opposite edges of the screen) to create sufficient separation between the antennas. This is labeled Configuration B in Figure 2.3. The maximal separation between the antennas creates statistically nearly-independent channels to achieve MIMO spatial multiplexing gains. While Configuration B was not designed for fullduplex operation, the presence of the actual device(e.g. laptop) between the antennas has the potential to create additional path loss between the two antennas and thereby increase the attenuation of the self-interference. Configuration C: Finally, we also test the configuration when one of the antennas is installed perpendicular to the other antenna, labeled Configuration C in Figure 2.3. This configuration aims to exploit the potential difference in radiation pattern along different axes. Theexperimentsareperformedbystrappingthetwo2.4GHz7dBiDesktopOmni Antenna (typical Wifi Antenna) to a ipad-sized device in different configurations. The dimensions are shown in Figure 2.3. We fix the transmit power at 6 dbm. For each configuration, we test the impact of antenna configuration and the device. The results are summarized in Table 2.1. The full-duplex PHY was implemented on WARP boards, each with three radio cards. One radio was connected to the transmit antenna, the second was connected to the receive antenna and the third provided the cancelling signal (x c ) over a wire and added in analog after the receive antenna. Four main results stand out from the Table 2.1.

40 21 Configuration Device Interference Interference power Total Present power after analog cancellation suppression A No -28 dbm -52 dbm 58 db A Yes -28 dbm -52 dbm 58 db B No -46 dbm -71 dbm 77 db B Yes -51 dbm -75 dbm 81 db C No -40 dbm -63 dbm 69 db C Yes -49 dbm -73 dbm 79 db Table 2.1 : Amount of passive suppression and active cancellation in different configurations. The transmit power is 6 dbm. Result 2.1 (Device reduces self-interference) Depending on the configuration, the presence of a device (e.g. laptop/ipad) can make a significant impact on the power of self-interference, by passively attenuating the signal. The metallic components in a laptop-like device can significantly attenuate the signal and thus reduce selfinterference. In Configuration C, device results in an additional attenuation of 9 db compared to the case when the device is not present. The device related attenuation is 5 db in Configuration B and 0 db in Configuration A. Result 2.2 (Best full-duplex configuration) The best configuration in terms of self-interference power, with and without analog cancellation is Configuration B, where the self-interference power with and without the analog cancellation is lowest compared to other configurations. This is, in fact, great news because Configuration B is also the ideal configuration for MIMO systems. Thus, there is a potential to use multiple antennas in either MIMO or full-duplex modes in mobile devices. Result 2.3 (Need for baseband cancelation) In[3, 4], baseband cancellation was also proposed to reduce the self-interference power. In our design, we did not implement base-band cancellation due to lack of sufficient FPGA logic on our WARP

41 22 boards, but we can still achieve a self-interference suppression which is more than the prior work due to added suppression by the device. From Table 2.1, we observe that even prior to analog cancellation, the selfinterference is suppression in all configurations is more than 30 db. For example, in Configuration A, corresponding to a transmit power of 6 dbm, the received selfinterference power when a device is present between transmit and receive antenna is -28 dbm, implying a 34 db passive suppression. Result 2.4 (RF requirements for cancelling signal path) The large passive suppression of self-interference implies that the cancelling transmit RF chain does not require a power amplifier, because the cancelling signal travels over a wire and thus suffers only minor attenuation. In fact, we had to install 40 db attenuators on our off-the-shelf radio cards, which essentially removed all the power amplification by the power amplifiers. Further, it implies that full-duplex transceiver needs one full transmit chain (up-converter for transmit antenna), one radio chain (down-converter for receive antenna) and a partial transmit chain without power amplifier (for cancelling signal). Thus, compared to SISO transceiver (one transmit and one receive RF chain), full-duplex only needs an additional partial transmit chain Partially Asynchronous PHY So far, the PHY analysis in prior works [3, 4] and in Section has been motivated by two nodes exchanging packets with each other as shown in Figure 2.1. However, full-duplex can be employed in more general cases. Consider the hidden node topology in Figure 2.4b, where M 2 is out of radio range of M 1. Assume AP has a packet for M 1 and M 2 has a packet for AP. In this case, since the AP has to be a full-duplex

42 23 node, the key question is if the full-duplex mode can be enabled in an asynchronous manner. That is, Can a new flow be added once a flow starts transmission?. In the hidden node example, there are two possibilities for AP: (a) start receiving a packet from M 2 after having initiated a transmission to M 1, (b) start a transmission to M 1 while receiving a packet from M 2. Receive-while-Sending: Assume that AP is actively transmitting to M 1 and is continuously operating its analog canceller to suppress its own self interference. This ensures that when M 2 starts a packet, it can be decoded by AP s receiver. The key challenge is that AP has to estimate the channel between M 2 and AP in the presence of self-interference, so that it can reliably decode M 2 s packet. In almost all current systems, even with multiple users, this training is performed without any(intentional) interference. However, to enable asynchronous full-duplex, we are required to estimate the channel between M 2 and AP in the presence of self-interference caused by AP s ongoing transmission. We label the physical layer channel estimation in the presence of ongoing transmission as dirty estimation, and quantify the loss compared to the conventional systems, all of which have clean estimation. In Table 2.2, we report the results for different values of SINR which were achieved bychangingthedistancebetweenthetwonodesm 2 andap.fromtable2.2, itisclear 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 any measurable loss and for low SINR, the BER in dirty estimation system can be 6 times compared to clean estimation, which turns out to be up to 3 db loss in effective SINR for the asynchronous packet. This implies

43 24 SINR BER BER (with canceller) dirty clean estimation estimation 18 db db db db db Table 2.2 : BER for the signal received at the full-duplex node, in the Receive-while- Sending mode with clean and dirty channel estimation. Each packet has a payload of 324 bytes and was QPSK-encoded. Signal transmit power was fixed at 6 dbm. A total of bits were transmitted. that the capacity of the full-duplex transmission is reduced if full-duplex is used in this asynchronous mode. Send-while-Receiving: Now we consider the converse case, where AP is already receiving a packet from M 2 and intends to send a packet to M 1 to leverage its fullduplex capabilities. Unfortunately, FD-PHY cannot enable this mode reliably. The key challenge is calculation of the cancelling signal in the presence of an ongoing reception. To compute the cancelling signal x c, we need to estimate the channel coefficients h k and h c,k. If the MIMO PHY header is transmitted (as described in Section 2.2.2) while PHY is receiving a packet, then the large uncancelled self-interference will completely swamp the ongoing reception. This is because selfinterference before cancellation is almost always much bigger than signal of interest. While receiving the packet, the automatic gain control (AGC) is set to ensure that the incoming signal occupies the full dynamic range of the ADC. Thus, the process of estimating the channels to establish cancelling signal causes a self-collision at the

44 25 receiver. In Chapter 4, we study different mechanisms for reliably enabling both modes of asynchronous full-duplex in detail. However, for the purpose of developing and evaluating a MAC layer based on the constraints of FD-PHY, we present the following result. Result 2.5 (Allowable asynchronous modes in FD-PHY) The key result is that asynchronous full-duplex can be enabled to receive while transmitting (with some loss in the performance of receiving packet) but not transmit while receiving. 2.3 MAC protocol design In this section, we will describe Full-Duplex Medium Access Protocol (FD-MAC) which uses the full-duplex-capable physical layer described in Section 2.2. We will limit our attention to infrastructure based systems and focus on the scenario involving one access point (AP). This will allow us to define the fundamental elements of a fullduplex MAC protocol Challenges in MAC Design The first challenge in designing full-duplex MAC is identification of the nodes which can engage in a full-duplex mode. In any network of multiple nodes, multiple flows with random arrivals exist at the same time, leading to random instances when fullduplex can be used. The second challenge is imposed by the physical layer. From Section 2.2.4, either full-duplex has to be performed synchronously between two nodes(a packet exchange) or can be done asynchronously only if the full-duplex node starts receiving a packet

45 26 while it has started transmitting a packet to another node. Any MAC design has to respect this constraint in its design. The third challenge is shared by any MAC protocol (full or half-duplex) which is to provide opportunity to all nodes to access the medium. That is, the access protocol should not unduely favor full-duplex opportunities over half-duplex flows Overview of FD-MAC In the infrastructure-based network, all flows have either AP as their source or destination. Thus, at any given time, a maximum of two flows can be active among full-duplex capable nodes. The two possible scenarios which leverage full-duplex capabilities are shown in (i) Figure 2.4a, where AP and mobile node M 1 are exchanging packets and (ii) Figure 2.4b, where AP is sending a packet to M 1 and simultaneously receiving a packet from M 2 (or vice versa); M 1 and M 2 are two mobile nodes which are outside each other s radio range. FD-MAC is a random access protocol, which will use most of the dominant elements of the IEEE DCF. However, while IEEE is CSMA/CA, collision avoidance in FD-MAC is done selectively to leverage full-duplex opportunities. FD- MAC introduces following three new protocol elements. Shared random backoff: When two nodes, say AP and M 1 in Figure 2.4a simultaneously have several packets for each other, they can truly exploit full-duplex two-way exchange. However, in order to allow other nodes to send or receive from AP, they do not continuously capture the medium. Instead they agree on a shared random backoff which allows other nodes to contend for the medium. In the event that no other node wins the contention, AP and M 1 continue with their full-duplex transmission. Snooping to discover full-duplex opportunities: In FD-MAC, nodes decode headers

46 27 AP M 1 M 2 AP M 1 (a) The simplest network with 2 nodes. (b) Both the mobile nodes are connected to the AP but are not in the radio range of one another AP M 1 M 2 AP M 1 M 2 (c) All three nodes are in radio range of each other M 3 (d) M 2 and M 3 are hidden to M 1 Figure 2.4 : A line connecting any two nodes indicates that they are in radio range of one another of all ongoing transmissions, even when network allocation vector NAV is non-zero. This allows the nodes to estimate the local topology and initiate full-duplex opportunistically. Virtual contention resolution: FD-MAC also has two virtual contention mechanisms to balance use of the full-duplex mode with access for all nodes in the network. While FD-MAC can be used with or without RTS/CTS, we will restrict our discussion to the more popular use case of infrastructure mode of which does not use RTS/CTS FD-MAC Packet structure We adopt IEEE packet structure and add a new FD header, for managing full-duplex transmissions as shown in Figure 2.5. Each packet contains a PHY header, a MAC header, a full-duplex header, a pay-

47 28 load and a cyclic redundancy check (CRC). Except for the full-duplex (FD) header, all other fields are identical to IEEE packets. We briefly explain the fields which are essential to describe FD-MAC. The PHY header has a preamble and the training symbols necessary for the functioning of the physical layer. The existing elements of the MAC header that we use in our FD-MAC protocol are Duration ID denoting the duration (DUR) of the packet, source address (SA), destination address (DA) and FRAG (denoting if there are more fragments of the same packet in line for the destination). The MAC header distinguishes between data packet and acknowledgement. For simplicity of description of the FD-MAC protocol, data packets will be referred as DATA and acknowledgement as ACK. The FD header has a one-bit field to distinguish packet type (DUPMODE) which can either assume values HD (indicating that it is a half-duplex packet) or FD (indicating that it is a full-duplex packet). Then there is a one-bit field, Head-of-line (HOL), indicating that the next packet in the buffer is for the destination of the current packet. The current MAC header has a field labeled more data in Frame Control Field of MAC header, but to avoid any conflict with other uses of this field, we have included HOL in the FD header. Since HOL is only 1-bit long, the overall overhead due to its inclusion is minimal. The next field in FD header, DURNXT, reveals the duration of head of line packet, and is useful when HOL = 1. It 2 bytes long. The next field, DURFD, is meant for revealing duration of the full-duplex exchange. It too is 2 bytes long. The next one-bit is a Clear-To-Send (CTS) indicating that destination of the current packet can send a packet to source of the current packet. Finally the FD header has a field for a 10-bit number which is the Shared Random Backoff (SRB).

48 29 Fields DURNXT and DURFD are needed in order to counter the hidden node problem in infrastructure mode of They are optional and in their absence, the FD header is only 13 bits. PHYheader MAC header Payload CRC MAC header FD header DUPMODE HOL DURNXT DURFD CTS SRB 1bit 1bit 2 bytes 2 bytes 1bit 10bits Figure 2.5 : Structure of the packet being used for the FD-MAC protocol Shared Random Backoff Consider the most basic two-node example shown in Figure 2.4a. It is possible that at any given time either both nodes have a packet for each other or only one node has a packet for the other. Note that in this case, asynchronous full-duplex is not possible because of the PHY constraints (Section 2.2.4), where a node cannot start a new transmission while it is receiving a packet. Thus, nodes have to find a way to synchronize their transmissions, such that they can estimate the channel coefficients for maximal self-interference cancellation (as discussed in Section 2.2.2). To maximize the use of full-duplex mode while respecting the constraints imposed by the physical layer, FD-MAC proceeds as follows. Assume that the nodes contend for the medium since they do not know if both nodes have a packet for each other or not. Without loss of generality, assume that AP wins the contention resolution.

49 30 Then, if the AP has another packet lined up in the buffer for M 1, it sets HOL=1 in the DATA packet. Here SRB AP = 0 and the DUPMODE = HD. Thus, the first packet in a two-way exchange is half-duplex. Source = AP Source = M 1 HOL = 1 DATA time ACK HOL = 1 CTS = 1 CTS = 1 ACK DURMODE = FD HOL = 1 DATA DATA HOL = 1 DURMODE = FD ACK SRB M1 CTS = 1 ACK DATA DATA ACK ACK max(srb M1, SRB AP ) DATA CTS = 1 HOL = 1 DURMODE = HD SRB AP max(srb M1, SRB AP ) Figure 2.6 : Timeline of packets sent from AP M 1 and M 1 AP. The relevant fields for decision making are listed above and below the packets. If M 1 receives the DATA successfully and has a packet for AP, it sends an ACK packet with HOL=1 and DURNXT set to the length of the head of the buffer packet. Also, SRB=0, CTS = 1. After receiving the ACK, both nodes know that they can initiate a full-duplex exchange. The PHY needs AP to train its self-interference channel, and thus AP sends an ACK packet, with HOL = 1, and also reveals DURNXT, and sets SRB=0, CTS=1. Now the two nodes are set to be in full-duplex. They wait for max(srb DATA,SRB ACK ) (which is = 0 at this stage) and then send their respective DATA packets with the DUPMODE = FD, and DURFD= max(durnxt AP,DURNXT M1 ). Each node sends an ACK only at the end of DURFD duration. Also, AP always sends the ACK after M 1 is in full-duplex mode, which allows hidden nodes to contend in the medium at the end of the ACK from AP; see Section After one full-duplex transmission, it is possible that both nodes still have more packets for each other, which they will discover by setting the FD header fields as described above. However, if the two nodes continue to occupy the medium without

50 31 any breaks, then other nodes will be completely starved. On the other hand, if the two nodes know they have a packet for each other but give up the medium for other nodes, they will have to go through a contention resolution again followed by one half-duplex packet. Thus, it is important for nodes to retain the knowledge of the queue state which they obtain by above hand-shaking enabled by FD header. So, weintroducetheideaofshared random backoff (SRB),whereAPandM 1 handshake on the random delay that they will both wait before resuming to full-duplex mode. In the ACK sent after receiving first full-duplex packet, AP picks a random backoff from [0,CW max,ap ] where CW max,ap is the current maximum contention window width for AP and places that number in SRB. The mobile node M 1 also picks a random backoff from its own maximum contention window [0,CW max,m1 ], and places it the SRB field of its ACK packet FD header. AfterthetwonodeshavefinishedsendingACKs,theywaitformax(SRB AP,SRB M1 ). In the DCF, backoff countdowns are paused by carrier-sense events. In our work, we require distributed nodes to independently count down for the same duration and, as such, cannot employ this pausing mechanism since they each might see independent channel busyness events. Hence, we propose a different kind of behavior for the shared random backoff; nodes do not pause their backoff countdowns in the presence of energy on the medium but instead perform one final idle-for-difs check at the end of the interval to ensure that there is nothing currently using the medium when they are about to transmit. If no other node in the network wins the medium before this shared backoff counter expires, the two nodes enter the full-duplex mode again and continue the above process till they have packets for each other. A timeline of the events is shown in Figure 6. Note that the protocol requires AP and M 1 to wait for at least max(srb AP,SRB M1 ) before transmitting, however it can tolerate more

51 32 delay in start of DATA packets of AP as the PHY layer has already estimated the required channels. However, if another node wins the medium before the expiry of the calculated backoff, then both AP and M 1 purge their knowledge about the other nodes and start completely afresh. The reason to purge the states is because upon discovering full-duplex opportunities with another node, say M 2, the AP will modify the ordering of packets in its buffer to place packets destined for M 2 in front of the buffer. In Section 2.3.6, we discuss the idea of reordering the buffer in more detail. Another reason for this purge is to account for the previously discussed modification to the backoff process. In presence of other traffic, the shared backoff will effectively be cancelled despite the removal of the explicit pausing mechanism. Thus, the only difference between traditional backoffs and our shared backoffs is the fact that our full-duplex nodes will not pause their backoffs in the presence of undecodable energy on the medium. Failure of DATA or ACK: In a full-duplex exchange, if a node does not decode DATA correctly, it does not send the corresponding ACK. At this point, synchronization of backoffs is not possible and since both nodes have not received at least one of DATA or ACK, both AP and M 1 purge the information about queue state of the other node and contend for the medium if they do not detect any energy on the medium. On the other hand, if only one of the ACKs fail, the node not receiving the ACK purges its knowledge of the queue and contends for the medium at the end of two ACK periods after the DATA packet exchange finishes. It is then a case of physical medium contention by the nodes with one of nodes having a backoff fixed to max(srb AP,SRB M1 ) and others having a random backoff. Failure of both ACKs simply calls for purging queue state information and thus results in another like contention. Therefore, in the poor channel conditions case too, the FD-MAC

52 33 Source = AP Source = M 1 Source = M 2 DA = M 1, HOL = 1 DATA time ACK DUPMODE = HD DA = M 1 Contention period HOL = 1 DA = M 1 DA = M 1, HOL = 1 HOL=0 DATA ACK DATA DATA DA = AP ACK DATA DA = AP, HOL = 1 ACK DATA M 2 wins contention Figure 2.7 : AP M 1 and M 2 is hidden from M 1. ACKs from M 1 AP are not received at M 2. The dashed lines in DATA packet of AP signify the end of the header which M 2 can decode. Corruption of DATA implies no ACK from receiver. has a throughput at least as much as that of (minus the throughput loss due to additional FD header) Snooping to Leverage FD Mode Consider the case of three nodes, one AP and two mobile nodes M 1 and M 2. With three nodes such that both mobile units can communicate with the AP, there are two possible topologies: (i) all nodes can hear each other and thus form a clique and (ii) M 1 and M 2 are not in the radio range of each other and thus hidden from each other. We discuss how snooping headers of the ongoing transmissions can help nodes identify opportunities to leverage full-duplex modality. We note that there is no explicit topology discovery mechanism in FD-MAC. Thus, nodes estimate the topology by overhearing packets as follows. Assume AP sends a DATA packet to M 1. Since M 2 is associated with AP, it can decode the headers and knows that the packet is addressed to M 1. If the ACK from M 1 is overheard by M 2, then M 2 concludes that it forms a clique topology with M 1. Else it concludes that it is in hidden-node topology with M 1. Note that M 2 can make an error in its estimation due to random channel induced errors causing either the DATA or ACK to drop, each leading to a wrong conclusion at M 2. However, since MAC headers are encoded at

53 34 base-rate, the probability 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 combinations {AP M 1,M 2 AP} and {AP M 2,M 1 AP} are not possible because they cause collisions (two simultaneous incoming packets) at one of the mobile nodes due to the topology being a clique. Now consider the case of hidden node topology. In this case, all four full-duplex combinations 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}. We have discussed how to establish the first two full-duplex opportunities, (i) and (ii), in Section The third and fourth cases are mirror reflections of each other, thus without loss of generality, we discuss case (iii). We first recall that there is a PHY-imposed constraint that a node cannot initiate a new transmission if it is already receiving a packet from another node. In case (iii), only AP is operating in the full-duplex mode, and therefore it has to begin the transmission of a packet to M 1 no later than the start of reception of a packet from M 2. Assume that that is the case where AP wins the contention resolution and begins sending its packet to M 1. By snooping on the HOL field of the FD header, M 2 can learn if AP has another packet for M 1 or not. If AP does have a HOL line packet for M 1, then M 2 can transmit a packet to AP while AP transmits its next packet to M 1, if (a) M 1 is not the radio range of M 2 and (b) M 1 should not be attempting to achieve AP M 1. In order to ensure (a) M 2 waits for one ACK duration after the finish of DATA packet from AP. If M 2 does not receive the ACK, it assumes that M 1 is not its radio range. To ensure (b), M 2 does not contend for the medium and allows the AP to capture the media. Then, M 2 decodes the FD header of DATA packet being sent from AP. If the destination of the DATA packet is M 1 and the DUPMODE is HD,

54 35 then M 2 can transmit its own packet to AP. Additionally, M 2 decodes the duration DUR of AP s packet and fragments its packet to ensure it ends no later than AP s transmission. It also sets the FRAG =1 in its packet. The fragmentation is necessary to avoid collisions with the ACK from M 1. The ACK from AP will arrive one ACK period after the finish of the DATA packet. The same procedure continues as long as AP has a packet for M 1, M 2 has a packet for AP, and M 1 does not have a packet for AP. Event timeline is shown in the Figure 2.7. At any point of time if M 1 has a packet for AP, it will coordinate with AP via ACKs to enable AP M 1, and M 2 can discover this setup if the DUPMODE=FD for the DATA packet from AP to M 1. Moreover, DURFD will let M 2 know that it should not contend for the medium at least for DURFD + 2ACK periods. This gets rid of unnecessary collisions if packets of AP are much smaller than that of M Virtual Contention Resolution In the previous two sections, we introduced methods to allow flows from mobile nodes to AP to get a chance to contend (Section 2.3.4) and discover opportunities to exploit full-duplex capabilities at PHY (Section 2.3.5). In this section, we introduce two more mechanisms which allow (i) AP to break away from a full-duplex handshake to send packets to other nodes, since AP can have downlink flows for any mobile node associated with it, and (ii) reduce the probability of collisions in snooping based full-duplex access. First, consider the case where AP is in a full-duplex packet exchange with a mobile node M 1. In standard MAC protocol, always the packet at the front of the buffer is transmitted, i.e., the depth of the MAC buffer is one. In order to further increase the possibility of operating in the full-duplex mode, the AP can have a larger

55 36 MAC buffer such that it has more chances to find a packet for M 1 as long as the node M 1 has a packet to send to the AP. The AP can do so by placing the next available packet destined for M 1 in front of its buffer, as shown in Figure 2.8, making MAC no longer a FIFO layer. Bufdepth is a parameter that can be increased to achieve full-duplex exchange. We note that if FIFO operation is desired then Bufdepth can be chosen to be one and hence this mode is optional. Increasing the depth of the buffer improves the chance of operation in full-duplex. On the flip side, it can starve transmission of packets to other mobile nodes. In order to break away from the full-duplex handshake and allow AP to send packets, virtual contention is arranged between the destination of the current head of the buffer and the destination with whom AP engaged in full-duplex exchange. Upon discovering an opportunity of a full-duplex exchange with M 1, AP searches for a packet with M 1 as its destination in its buffer, and sends it if found. After the first full-duplex exchange, if the packet at the head of the buffer is not destined for M 1, then AP searches through the 2nd to Bufdepth packets in the buffer and with a probability p pick picks the first packet with destination=m 1 as the new head of line packet. Since the probability of picking k consecutive out-of-order packets decays geometrically as p k pick, the AP chooses to not send head-of-line packets with a fast decaying probability. Second, consider the case where multiple nodes are snooping on ongoing transmissions by AP as described in Section 2.3.5, say M 3 in addition to M 2 as shown in Figure 2.4d. If both M 2 and M 3 are hidden from M 1, then they will both send a packet to AP at the same time and end up colliding at AP since AP can only receive one packet at a time. Thus, it is important that there is a mechanism to avoid such collisions. Since M 2 and M 3 do not know how many nodes are there which may try to contend, they only send a packet to use the full-duplex mode at AP probabilistically.

56 37 Current Buffer Head of buffer p pick Reordering maximizing throughput 1 p pick packet packet for M 1 for M 2 New head of buffer enables one more round of FD switches to HD Figure 2.8 : Virtual contention resolution between packets in the buffer of AP with Bufdepth = 3. Virtual contention is a probabilistic reordering of the MAC buffer at the end of every full-duplex exchange. Prior to the virtual contention resolution, AP was engaged in a full-duplex packet exchange with mobile node M 1 That is, each node which detects a full-duplex opportunity, sends the packet with probability p i, where p i is computed based on the current maximum backoff window as p i = β CW max, where β is a pre-chosen constant which controls the aggressiveness in the system. The motivation for using p i 1 CW max is that each node can use their current maximum contention window as a proxy for amount of expected competition in the system. Since each node s neighbourhood is different, all nodes face a different amount of contention on the average. Of course, it is possible to fix p i = p where p is prechosen and allows equal chance for each nodes. Result 2.6 (Impact of larger buffer depth) Increasing the buffer depth at AP increases the throughput. The increase in throughput comes at the cost of increased delay due to packet reordering. In order to understand the tradeoff between delay and throughput due to larger

57 38 than one Bufdepth, and the probability p pick we simulate the buffer of the AP with packets for 5 mobile nodes. All nodes always have packets for the AP, and are in radio range of one another. Thus, full-duplex exchange is always possible and can be broken only via virtual contention. The only type of contention that was allowed was virtual contention in the buffer of AP. For every Bufdepth, p pick was ranged from 0 to 1. The AP had uniform traffic for all nodes with packets lined up in an arbitrary order. Figure 2.9 shows a plot of throughput vs. average delay (for the head of the buffer packet) for different buffer depths. A key finding from this experiment is that the throughput and average delay are linearly related. Larger Bufdepth can help in improving the throughput at the cost of delay. Also, it is often possible to obtain the same (throughput, average delay) pair for smaller Bufdepth by simply increasing the probabilty of reordering, p pick. The protocol description is now complete. In the next section, we consider an example topology to understand how all the proposed mechanisms in FD-MAC come in play State Transitions in FD-MAC Consider the clique topology shown in Figure 2.4c. Assume that there are four flows in the network, AP M 1, AP M 2, M 1 AP, M 2 AP. Four flows bring forth the possibility of two full-duplex scenarios AP M 1 and AP M 2. Figure 2.10 shows the mechanisms which allow the network to go from one mode to another. Each transition is enabled by the features introduced by the FD-MAC protocol. The three node network with clique topology can transition from one fullduplex mode, i.e., AP M 1 to AP M 2 only through half-duplex modes. This is so because the first packet in two-way full-duplex exchange, as discussed in Section 2.3.4,

58 Throughput vs. Delay for different depths of Buffer at the AP Throughput Bufdepth = 10 Bufdepth = 5 Bufdepth = 2 Bufdepth = Average delay Figure 2.9 : The throughput is normalized, with a maximum and minimum being 2 (all FD packets) and 1 (all HD packets). Bufdepth = 1 implies 0 delay is always a half-duplex packet. Suppose that network is in the mode AP M 1. From this full-duplex mode, the network can transition to a half-duplex mode due to different reasons: (a) if at least one of AP or M 1 has no more packets in the buffer for the other, i.e., if(hol AP = 0 or HOL M1 = 0), both the AP and M 1 naturally give up full-duplex mode (b) any of the DATA or ACK packets is not decoded right - failure in reception leads to purging of queue states of other nodes to start type contention(c)thepacketswithm 2 asdestinationwinthevirtualcontentionresolution allowing the network to break away from full-duplex entering a AP M 2 mode, and (d) M 2 wins the physical contention during the silent shared random backoff period,

59 40 thus initiating M 2 AP. All the half-duplex modes can switch among each other with the protocol. Consider the half-duplex mode M 1 AP. From this mode the only possible transition to a full-duplex mode is AP M 1. The mechanism of two-way setup is discussed in Section The FD-MAC protocol therefore allows all modes to occur by switching between various modes through mechanisms introduced by FD-MAC, and some existing capability. AP M 1 Virtual contention M 2 Physical contention AP M 1 AP M 1 M 2 Two-way FD setup M 2 HOL M1 = 0 AP M 1 HOL AP = 0 M 2 Figure 2.10 : Switching between different modes of operation in a clique topology. The part of the state diagram illustrating all the key features of the FD-MAC is shown. State diagram has two more half-duplex modes AP M 1 and M 1 AP The hidden node topology with two mobile nodes and an AP is shown in Figure 2.4b. With M 1 and M 2 hidden with respect to each other, two full-duplex flows in

60 41 addition to AP M 1, AP M 2 are possible. They are {M 2 AP,AP M 1 }, and {AP M 2,M 1 AP}. Each of four full-duplex modes, whether two-way exchange or otherwise start with a particular half-duplex mode. For instance {M 2 AP,AP M 1 } is possible only if there exists AP M 1 as discussed in Section In order to ensure that transition to all full-duplex modes is possible, FD-MAC must ensure that the half-duplex mode needed to kick start it is possible. Halfduplex modes among themselves contend via type of physical contention. The two-way full-duplex exchanges have a period of shared random backoff for other halfduplex modes to occur. Moreover, they also have virtual contention resolution at the AP to allow different half-duplex modes. On the other hand AP M 2, {M 2 AP,AP M 1 } type of full-duplex, has AP always contending for the media after it finishes sending the ACK, thus allowing all other nodes to contend and establish a half-duplex communication with the AP. Since all half-duplex modes are possible from any starting state, therefore the snooping mechanism will allow all full-duplex modes too FD-MAC evaluations on WARP In this section, we evaluate the FD-MAC for a two node full-duplex exchange by implementing it on a real time full-duplex system designed using WARP. Figure 2.11 shows a full-duplex WARP node, with one transmit and one receive antenna. The experimental set-up has two full-duplex nodes exchanging packets with each other. FD-MAC ensures setting up of the full-duplex upon discovering an opportunity to exchange packets in full-duplex mode. The buffer at both the nodes always had a head of line packet for the other. The evaluation compares the throughput of fullduplex against half-duplex (again implemented on WARP). The modulation used for

61 42 Figure 2.11 : A full-duplex WARP node transmission was QPSK. Result 2.7 (Increase in throughput due to full-duplex) The throughput of fullduplex two-way exchange using FD-MAC is 70% higher than that of half-duplex for identical transmit power. 2.4 Summary In this chapter, we described the design of FD-PHY, the first real-time design and implementation of full-duplex physical and proposed FD-MAC, a medium access layer for full-duplex enabled nodes. Considering that full-duplex is still nascent and many researchers are skeptical about its feasibility, it is crucial to demonstrate real-time implementations showing fully operational network stacks. Towards that end, FD- PHY shows very promising results, creating a strong case for practical use of full-