PaperlayerView exact revision on arXiv

Source-linked AI summary

AirSync: Enabling Distributed Multiuser MIMO with Full Spatial Multiplexing

Horia Vlad Balan, Ryan Rogalin, Antonios Michaloliakos, Konstantinos Psounis, Giuseppe Caire

arXiv:1205.6862v2cs.NI

TL;DR

Dense wireless deployments are limited by interference and synchronization challenges between distributed access points. AirSync provides phase synchronization for distributed MU-MIMO, and an eight-radio testbed shows that theoretically optimal spatial multiplexing gain is achievable in practice.

  • Problem

    Dense spectrum reuse is limited by multiuser interference and the lack of synchronization between jointly processed access points.

  • Method

    AirSync synchronizes distributed access points by achieving phase synchronization and correcting instantaneous phase offsets between geographically separated transmitters.

  • Results

    An eight-radio WARP testbed shows that AirSync makes the theoretically optimal spatial multiplexing gain achievable in practice.

  • Takeaways & Limitations

    AirSync demonstrates a practical path toward distributed MU-MIMO with full spatial multiplexing gain and motivates MAC-layer mechanisms for deployment.

Abstract

from arXiv · show

The enormous success of advanced wireless devices is pushing the demand for higher wireless data rates. Denser spectrum reuse through the deployment of more access points per square mile has the potential to successfully meet the increasing demand for more bandwidth. In theory, the best approach to density increase is via distributed multiuser MIMO, where several access points are connected to a central server and operate as a large distributed multi-antenna access point, ensuring that all transmitted signal power serves the purpose of data transmission, rather than creating "interference." In practice, while enterprise networks offer a natural setup in which distributed MIMO might be possible, there are serious implementation difficulties, the primary one being the need to eliminate phase and timing offsets between the jointly coordinated access points. In this paper we propose AirSync, a novel scheme which provides not only time but also phase synchronization, thus enabling distributed MIMO with full spatial multiplexing gains. AirSync locks the phase of all access points using a common reference broadcasted over the air in conjunction with a Kalman filter which closely tracks the phase drift. We have implemented AirSync as a digital circuit in the FPGA of the WARP radio platform. Our experimental testbed, comprised of two access points and two clients, shows that AirSync is able to achieve phase synchronization within a few degrees, and allows the system to nearly achieve the theoretical optimal multiplexing gain. We also discuss MAC and higher layer aspects of a practical deployment. To the best of our knowledge, AirSync offers the first ever realization of the full multiuser MIMO gain, namely the ability to increase the number of wireless clients linearly with the number of jointly coordinated access points, without reducing the per client rate.

I. INTRODUCTION

AirSync addresses the synchronization barrier that has kept distributed multiuser MIMO largely theoretical, enabling coordinated access points to realize optimal spatial multiplexing in practice. The paper presents the synchronization scheme, its WARP FPGA implementation, experimental validation, and deployment considerations.

  • I. INTRODUCTION: Distributed MU-MIMO can increase spectrum reuse by coordinating multiple access points as one distributed multi-antenna base station, but unsynchronized access points create the main practical bottleneck.Joint precoding makes transmitted power useful for data transmission rather than interference.
  • I. INTRODUCTION: AirSync tracks a master access point’s broadcast pilot and predicts phase correction during each precoded slot, allowing access points to maintain the phase coherence required for MU-MIMO precoding.Each access point receives the pilot on one antenna while transmitting data on the others.
  • I. INTRODUCTION: AirSync is introduced as a scheme for phase synchronization in distributed MU-MIMO and implemented as a digital circuit on the WARP FPGA platform.The contribution targets synchronization between geographically separated transmitters.
  • I. INTRODUCTION: The testbed demonstrates that AirSync makes the theoretically optimal spatial multiplexing gain achievable in practice.The implementation uses eight WARP radios.
  • I. INTRODUCTION: The paper also considers practical MAC issues including scheduling, client selection, adaptive modulation and coding, and rateless incremental redundancy.These mechanisms are discussed as part of deploying distributed MU-MIMO systems.

II. RELATED WORK

Multiuser MIMO extends multiplexing gains to multiple users, but practical distributed systems require accurate precoding, channel knowledge, and synchronization. Prior implementations addressed related alignment and coding problems without increasing degrees of freedom, while zero-forcing beamforming can create interference-free user channels when channel and timing references remain stable.

  • Theoretical Foundations: Multiuser MIMO can increase capacity by transmitting independent spatial streams, but requires precoding and accurate channel-state information to suppress inter-user interference.The achievable multiplexing factor is bounded by the relevant channel rank, such as min{N_t, KN_r} for a MIMO broadcast channel.
  • Practical Implementations: Distributed space-time coding and flexible orthogonal access provide diversity or sharing benefits but do not increase system degrees of freedom beyond one information symbol per time-frequency dimension.These approaches therefore differ from full spatial multiplexing, which requires simultaneous independent streams.
  • OFDM Signaling: OFDM converts suitable frequency-selective channels into parallel frequency-flat channels, and its cyclic prefix tolerates user timing offsets up to L − ℓ without destroying alignment after demodulation.Typical cyclic-prefix lengths are 16 to 64 samples, corresponding to roughly 0.7 to 3.2 µs for a 20 MHz signal.
  • Zero-Forcing Beamforming: Zero-forcing beamforming chooses a precoding matrix that diagonalizes the channel, so each receiver observes an interference-free Gaussian channel.When the channel has rank K, a normalized Moore-Penrose pseudo-inverse generally supplies the beamforming matrix, with transmit power controlled by column normalization.
  • Tomlinson-Harashima Precoding: Zero-forcing removes multiuser interference, but nonlinear Tomlinson-Harashima precoding instead introduces noise folding when the modulo operation folds Gaussian-noise tails.The modulo operation limits each precoded stream’s transmit power, while the resulting folded noise affects the received signal.

IV. SYNCHRONIZATION IN DISTRIBUTED MIMO SYSTEMS

Distributed MU-MIMO requires accurate timing and carrier-phase synchronization across jointly precoded access points because phase drift can destroy interference cancellation. The paper models these offsets and motivates tracking transmitter phase during downlink transmission.

  • Synchronization challenges: Timing alignment only needs to fall within the OFDM cyclic prefix, whereas carrier-phase synchronization remains the more challenging requirement.The paper assumes timing offsets and carrier-phase effects can be modeled separately under stated sampling-clock assumptions.
  • Synchronization challenges: Phase rotation across a downlink slot can destroy an initially ideal ZFBF or THP interference-nulling effect.The resulting signal mixing occurs over the actual transmission channel, so receivers cannot eliminate it.
  • Synchronization challenges: Synchronizing different transmitters is harder than conventional receiver synchronization because APs must track a time-varying RF reference while transmitting.The architecture must support simultaneous data transmission and carrier-reference reception.
  • Signal model: The effective channel includes transmitter and receiver phase terms that vary with subcarrier and OFDM-symbol indices.These terms make the channel seen during transmission differ from the channel used to compute precoding.
  • Phase tracking: Measured tone phases form approximately parallel, time-varying trajectories, allowing their spacing and common slope to estimate phase de-rotation coefficients.The coefficients must be predicted ahead of transmission to compensate for hardware delay.

V. AIRSYNC

AirSync synchronizes distributed APs by broadcasting pilot tones from a master and having secondary APs estimate, track, and predict their relative phase. The implementation uses dedicated pilot reception, FPGA processing, and centralized precoding over wired backhaul.

  • AirSync design: AirSync uses a master AP’s out-of-band pilot tones and an initial probing header to estimate each secondary AP’s relative phase.Multiple pilot tones provide frequency diversity and spread pilot power across frequency bins.
  • AirSync design: The phase estimate combines timing offset, initial phase offset, and master-to-secondary channel phase, while drift compensation keeps APs aligned to the master frequency.The implementation can also separate calibrated channel phase from recurring timing and initial-phase estimates.
  • AirSync design: Secondary APs compensate phase drift by estimating the relative slope with a four-sample smoothing filter and predicting phase several OFDM symbols ahead.Look-ahead aligns the correction with the actual transmission time rather than the measurement time.
  • Implementation: AirSync was implemented as a digital circuit in the WARP FPGA, with synchronization tones placed about 7.5 MHz above and below the 5 MHz data bandwidth.Each secondary AP dedicates one antenna to pilot reception while other antennas transmit phase-synchronous data.
  • Implementation constraints: In FDD operation, ADC and DAC dynamic range limits simultaneous reception of master pilots and transmission of a nearby secondary data signal.For relative antenna-distance factors α from 32 to 128, the data signal can exceed pilots by 30 dB to 42 dB; WARP’s ADCs provide 84 dB dynamic range.
  • Centralized joint encoding: Centralized precoding sends quantized I/Q baseband symbols to APs and becomes less demanding than decentralized processing for K > 5 clients in the stated example.The example uses 4 bit/s/Hz clients over W = 5 MHz and b = 16-bit quantization.

VI. PERFORMANCE EVALUATION

The performance testbed uses one primary transmitter, three secondary transmitters, and four receivers to evaluate synchronization accuracy, channel separation, and theoretical gain. Figure 5 depicts the acquisition, tracking, and prediction process.

  • Testbed: The testbed comprises one primary transmitter, three secondary transmitters, and four receivers, with experiments targeting synchronization accuracy, channel separation, and theoretical gain.The primary transmitter uses one RF front-end, while secondary senders use separate receive and transmit front-ends.

A. Synchronization Accuracy

AirSync synchronizes distributed transmitters by estimating and compensating phase differences against a master reference, achieving degree-level synchronization accuracy.

  • A. Synchronization Accuracy: The experiment uses two transmitters and two receivers, with a third RF front-end providing wired feedback for phase comparison.The secondary transmitter samples its synthesized signal and compares it with the primary transmitter’s signal.
  • A. Synchronization Accuracy: AirSync modifies the synchronization circuit to reproduce the master transmitter’s phase at the secondary transmitter.It compensates for the phase rotation introduced by the feedback path.
  • A. Synchronization Accuracy: Figure 6 presents the synchronization-error CDF and reports phase synchronization within a few degrees of the source signal.The error is evaluated between the secondary and primary transmitters.
  • A. Synchronization Accuracy: 2.37 degrees standard deviation and at most 4.5 degrees at the 95th percentile quantify the frame-to-frame synchronization error.The error is measured using a feedback circuit.
  • A. Synchronization Accuracy: Synchronization pilots achieve approximately 28.5 dB SNR above the noise floor in a typical office environment.The measured pilot SNR is described as achievable between typically placed access points.

B. Beamforming gain

AirSync preserves coherent transmission between distributed transmitters, enabling the testbed to achieve the full beamforming gain for practical purposes.

  • B. Beamforming gain: The testbed achieves the full beamforming gain for practical purposes.The measured average gain is 2.98 dB, consistent with the synchronization precision measured previously.

C. Zero-Forcing Accuracy

AirSync’s synchronization supports accurate Zero-Forcing, with substantially reduced leaked power and near-elimination of interference at non-targeted receivers.

  • C. Zero-Forcing Accuracy: Figure 7 shows that leaked power is significantly smaller than total transmitted power, producing high-SINR receiver channels.The figure summarizes the distribution of leaked power from Zero-Forcing.
  • C. Zero-Forcing Accuracy: The average power leaked by Zero-Forcing is -24.46 dB of the total transmitted power.The experiment measures residual power caused by angle mismatching from synchronization errors.
  • C. Zero-Forcing Accuracy: Zero-Forcing is capable of almost completely eliminating interference at non-targeted receiver locations.The result follows measurements of residual leaked power in the testbed.

D. Zero-Forcing Beamforming Data Transmission

AirSync enables concurrent Zero-Forcing Beamforming transmissions that separate user channels and approach the theoretical multiplexing gain in a two-user experiment.

  • D. Zero-Forcing Beamforming Data Transmission: A two-transmitter, two-receiver experiment creates separate channels and achieves a multiplexing factor of two over point-to-point transmissions.The experiment transmits independent QAM-16 symbols using ZFBF, while actual rates depend on channel quality.
  • D. Zero-Forcing Beamforming Data Transmission: The system is evaluated against TDMA, whose achievable rates are bounded by the best point-to-point rate.Current enterprise WiFi transmissions use single access points serving single clients separated in time.
  • D. Zero-Forcing Beamforming Data Transmission: The compound MIMO channel reaches 18.27 bps/Hz versus 9.96 bps/Hz for the point-to-point channel, yielding a 1.83 multiplexing rate gain.The gain is close to the theoretical value of 2 under ideal codes.
  • D. Zero-Forcing Beamforming Data Transmission: Using practical modulations, the experimental multiplexing gain equals the theoretical value of 2.Both Zero-Forcing channels support the highest commonly used 802.11g rate of 6 bps/Hz with 64-QAM.

E. Tomlinson-Harashima precoding

The four-transmitter, four-receiver experiment uses Tomlinson-Harashima precoding to create four wireless channels and achieves a multiplexing factor of 4. Under practical modulation, the system reaches a 16 bps/Hz sum rate, exceeding the cited 802.11g comparison.

  • E. Tomlinson-Harashima precoding: Figure 9 presents four distinct wireless channels created for the four users with Tomlinson-Harashima precoding.The figure uses QAM-16 constellations and reports 16 bits/second/Hz spectral efficiency.
  • E. Tomlinson-Harashima precoding: A multiplexing factor of 4 was achieved with four transmitters and four receivers using Tomlinson-Harashima precoding.The actual rate gains depend on channel quality.
  • E. Tomlinson-Harashima precoding: The four channels have SINRs of 16.8, 19.2, 21.4, and 20.8 dB, yielding a 26 bps/Hz sum rate from Shannon-rate estimates.Lower SINRs result from increased power leakage caused by additional transmissions to other receivers.
  • E. Tomlinson-Harashima precoding: 16 bps/Hz is achieved with four 16-QAM constellations, versus 6 bps/Hz for 802.11g at 32 dB SNR.This corresponds to a multiplexing gain of 2.66 under practical modulations.

VII. MEDIUM ACCESS CONTROL

The proposed MAC design builds on centralized multiuser MIMO scheduling and uses a cross-layer PHY/MAC strategy for distributed MU-MIMO broadcasts.

  • VII. MEDIUM ACCESS CONTROL: The MAC layer adopts a cross-layer PHY/MAC design inspired by centralized multiuser MIMO scheduling.The design follows the achievement of synchronization accuracy and full multiplexing gain.

A. High level description

The high-level MAC design addresses duplexing, channel estimation, client access, downlink scheduling, and realistic modulation and coding choices for distributed MU-MIMO.

  • A. High level description: TDD is favored because channel reciprocity supports downlink-channel inference from uplink pilots and handles asymmetric enterprise traffic flexibly.FDD requires closed-loop downlink estimation and feedback whose overhead grows linearly with jointly precoded antennas.
  • A. High level description: Clients contend for uplink bandwidth through regular CSMA/CA, while the section focuses its scheduling design on the downlink.Backward compatibility requires protection mechanisms and operating modes for legacy 802.11 devices.
  • A. High level description: The central server selects downlink users using queue and QoS information, with approaches applicable to both ZFBF and THP precoding.For ZFBF, the discussion focuses on selecting users in each slot; THP requires minimal changes.
  • A. High level description: Greedy algorithms provide excellent results at moderate complexity for the user-selection and power-allocation problem in linear zero-forcing precoding.The ideal optimization would exhaustively search feasible user subsets under power constraints.
  • A. High level description: Realistic rate selection maps SINR values to discrete modulation and coding schemes, using a mapping corresponding to the nine mandatory 802.11ac MCSs.The selected scheme is assumed to be the best one for the received SINR, although mappings vary by vendor or deployment.
  • A. High level description: The large gap between ZF-ACM and ideal ZF motivates more flexible rate allocation and broader rate-and-code support for multiuser MIMO.Because multiple users share a time slot, efficient capacity use requires a larger set of supported rates and codes.

B. Protocol Design

The protocol design supplies timely channel-state information for downlink scheduling, aligns transmitters through pilots, and adds synchronization and signaling overhead while remaining experimentally incomplete in real time.

  • B. Protocol Design: The MAC collects timely channel-state estimates so the central server can select clients for upcoming transmission slots.Estimates can come from uplink pilots using TDD reciprocity or standard downlink estimation.
  • B. Protocol Design: Downlink pilots require active phase alignment for every packet so the effective channel matrix remains constant between transmissions.This synchronization requirement differs from uplink estimation and directly constrains the synchronization system.
  • B. Protocol Design: Before downlink data, access points probe channels, request client feedback, and begin transmission after collecting the requested information.The procedure follows the broad structure of 802.11ac channel probing and feedback.
  • B. Protocol Design: The master access point broadcasts alignment and channel-estimation pilots before all access points join the downlink transmission.The initial pseudo-noise sequence supports frame alignment and receiver block-boundary detection.
  • B. Protocol Design: Complete real-time MAC experiments were not performed because the radios could not switch between reception and transmission quickly enough.Individual downlink and uplink protocol components were tested.
  • B. Protocol Design: The added signaling is estimated at under a 20% increase in header duration, while bandwidth growth is nearly linear in client count.Normalized over total data transmitted in a fixed airtime slot, the overhead is much lower than in traditional 802.11.
  • B. Protocol Design: Future work targets more robust synchronization estimators, lower synchronization overhead, scalable semi-decentralized precoding, and a complete MAC implementation.The planned MAC implementation would use uplink-pilot channel estimates based on channel reciprocity.
Loading 1205.6862v2…