PaperlayerView exact revision on arXiv

Source-linked AI summary

The World's First Real-Time Testbed for Massive MIMO: Design, Implementation, and Validation

Steffen Malkowsky, Joao Vieira, Liang Liu, Paul Harris, Karl Nieman, Nikhil Kundargi, Ian Wong, Fredrik Tufvesson, Viktor Öwall, Ove Edfors

arXiv:1701.01161v2cs.IT

TL;DR

Massive MIMO required practical testbeds with systematic design guidance and real-world, real-time evaluation beyond its theoretical promise. The paper develops a scalable hardware and processing-partitioning framework and validates it through the LuMaMi platform, which serves up to 12 users on the same time/frequency resource in static indoor and outdoor tests.

  • Problem

    Systematic methodology for designing massive MIMO testbeds and real-time evaluation in real-world scenarios had not been reported.

  • Method

    The paper evaluates hardware and system requirements, proposes scalable processing and hardware partitioning, and implements the reconfigurable LuMaMi testbed.

  • Results

    LuMaMi supports up to 100 antennas and 12 simultaneously served users with real-time OFDM-TDD transmission, validated through static indoor and outdoor measurements.

  • Takeaways & Limitations

    The testbed provides a platform for prototyping massive MIMO, accelerating algorithm design, and validating digital base-band solutions with real measurement data.

Abstract

from arXiv · show

This paper sets up a framework for designing a massive multiple-input multiple-output (MIMO) testbed by investigating hardware (HW) and system-level requirements such as processing complexity, duplexing mode and frame structure. Taking these into account, a generic system and processing partitioning is proposed which allows flexible scaling and processing distribution onto a multitude of physically separated devices. Based on the given HW constraints such as maximum number of links and maximum throughput for peer-to-peer interconnections combined with processing capabilities, the framework allows to evaluate modular HW components. To verify our design approach, we present the LuMaMi (Lund University Massive MIMO) testbed which constitutes the first reconfigurable real-time HW platform for prototyping massive MIMO. Utilizing up to 100 base station antennas and more than 50 Field Programmable Gate Arrays, up to 12 user equipments are served on the same time/frequency resource using an LTE-like Orthogonal Frequency Division Multiplexing time-division duplex-based transmission scheme. Proof-of-concept tests with this system show that massive MIMO can simultaneously serve a multitude of users in a static indoor and static outdoor environment utilizing the same time/frequency resource.

I. INTRODUCTION

The introduction identifies a gap in real-time, real-world massive MIMO testbed methodology and evaluation, then presents LuMaMi as a scalable, reconfigurable platform addressing it. The platform supports real-time processing and simultaneous multi-user transmission on the same time/frequency resource.

  • Motivation: Massive MIMO uses many base-station antennas to spatially separate multiple users sharing the same time/frequency resource.The excess antennas enable linear processing to focus energy spatially and separate users.
  • Motivation: Prototype systems are needed to evaluate massive MIMO under real-world conditions and identify practical challenges beyond theory.Existing efforts included several testbeds, but systematic design methodology and real-time real-scenario evaluation had not been reported.
  • Approach: The paper analyzes hardware and system requirements, then proposes a signal-processing and hardware partitioning strategy for scalable testbed implementation.The framework distributes data shuffling and processing across physically separated components subject to interconnection and processing constraints.
  • Validation: The testbed supports real-time OFDM-TDD transmission with up to 12 simultaneously served user equipments on a shared time/frequency resource.The paper reports uncoded uplink and downlink measurements in static indoor and outdoor scenarios.
  • Validation: LuMaMi provides real measurement data for accelerating algorithm design and validating digital base-band solutions in real-world settings.The reported results suggest increased spectral efficiency compared with traditional point-to-point MIMO systems.
  • LuMaMi testbed: LuMaMi is presented as the world’s first real-time 100-antenna massive MIMO testbed built with software-defined radio technology.The platform is fully reconfigurable and designed to support prototyping and validation of massive MIMO systems.

B. Down-link

Downlink transmission uses uplink channel estimates for precoding, but hardware-induced non-reciprocity between uplink and downlink channels must be calibrated. The received downlink symbols are modeled through the downlink channel and transmit vector.

  • Downlink model: The downlink received-symbol vector is modeled using the K × M downlink channel H, transmit vector x, and complex Gaussian receive noise.H includes the propagation channel and downlink hardware transfer functions.
  • Reciprocity-based precoding: Because explicit downlink channel estimation is resource intensive, the estimated uplink channel G is reused for downlink transmission under propagation reciprocity.The transmit vector is constructed by precoding the intended user symbols and applying reciprocity calibration.
  • Reciprocity-based precoding: Uplink and downlink radio channels generally differ because their transceiver hardware responses are non-reciprocal.The hardware responses are represented by diagonal transmitter and receiver matrices at the BS and UEs.
  • Reciprocity calibration: A calibration matrix estimated internally at the BS compensates non-reciprocal components and can cancel inter-user interference caused by non-reciprocity.This compensation enables construction of a downlink precoder from uplink channel estimates.

D. Linear Detection & Precoding Schemes

The design focuses on LTE-like OFDM and TDD, while mobility, pilot timing, calibration, and frame organization constrain practical operation. TDD is selected because reciprocity avoids the extensive downlink pilots and feedback required by FDD.

  • A. Modulation Scheme: OFDM creates frequency-flat narrowband subcarriers, enabling the single-channel equalization strategy used in the testbed.The paper uses OFDM because it is common in modern wireless communication systems.
  • A. Modulation Scheme: LTE-like OFDM parameters are used to simplify comparison and assess how massive MIMO could be added to current cellular systems.The parameters are intended to remain familiar to existing LTE-based designs.
  • Duplexing mode: TDD is chosen over FDD because reciprocity requires only orthogonal uplink pilots, whereas FDD requires downlink pilots and CSI feedback.The paper therefore focuses entirely on TDD operation.
  • 1) Mobility:: The pilot rate determines how well channel variations are tracked and indirectly limits the highest supported UE speed.The frame structure must accommodate calibration, pilots, data, control, and guard symbols.
  • 1) Mobility:: A mobility constraint is derived by requiring channel correlation of 0.9 under a 2D wide-sense stationary channel model with uncorrelated isotropic scattering.The model links pilot spacing to the maximum supportable UE speed through a Jakes fading assumption.

2) Processing latency:

High mobility requires frequent uplink pilots, leaving limited time for channel estimation and downlink precoding. The resulting TDD turnaround constraint must include all hardware, processing, and routing delays.

  • 2) Processing latency:: High mobility demands a high pilot rate, forcing uplink data, downlink data, and guard symbols into the interval between consecutive uplink pilots.This reduces the time between uplink pilot reception and downlink transmission.
  • 2) Processing latency:: The TDD precoder turnaround time includes analog front-end, OFDM, CSI-estimation, precoding, reciprocity-compensation, and routing delays.These contributions are represented by Δrf,TX, Δrf,RX, ΔOFDM, ΔCSI, Δprecode, and Δrout.
  • 2) Processing latency:: CSI estimation and precoding must be optimized so their processing latency does not violate the turnaround constraint Δ.The constraint depends on OFDM-symbol arrangement and pilot repetition in the frame structure.
  • 2) Processing latency:: The BS acquires CSI from orthogonal uplink pilots transmitted by the K UEs.Possible designs include orthogonal subcarriers, orthogonal pilot sequences, and semi-blind or blind techniques.
  • 2) Processing latency:: The generic frame hierarchy places BS reciprocity calibration at the start of each reciprocity cycle and encapsulates downlink pilot cycles within it.The reciprocity-cycle length follows BS transceiver stability, while downlink pilot frequency follows UE-side stability.

1) Processing Capabilites:

The paper quantifies the substantial processing, data-transfer, and connectivity demands of massive MIMO, then frames testbed design around scalable processing and hardware constraints.

  • Processing Requirements: 126 Gops/s is required for the M antenna FFTs or IFFTs performed close to the antennas.The estimate assumes four real multiplications per complex multiplication.
  • Processing Requirements: Up to 80 Gops/s is required for data precoding, detection, and reciprocity compensation involving large matrix and vector multiplications.One example is multiplying an M × K matrix by a K × 1 vector.
  • Processing Requirements: Approximately 1 Tops/s is needed for zero-forcing processing when pseudo-inverse computation must finish within two OFDM symbols.The estimate includes Gram-matrix formation, matrix inversion, and multiplication with the Hermitian channel matrix.
  • Data Transfer Requirements: 2M peer-to-peer links are needed between antenna units and centralized MIMO processing when uplink and downlink paths use separate links.The subcarrier data rate is M · w · 16.8 MB/s, while information symbols require K · 16.8 MB/s under the stated assumptions.
  • Design Framework: Scalable testbed design must jointly consider processing throughput, transfer latency, and the number of peer-to-peer links.The proposed partitioning also supports modular hardware, reconfigurability, and integration of in-house designs.

2) Switches:

The proposed architecture distributes massive MIMO processing across SDRs, switches, and centralized co-processors while grouping data streams to satisfy hardware link and throughput limits.

  • System Components: SDRs interface between the digital and RF domains and perform local processing, while switches aggregate or disaggregate data between system components.Co-processing modules provide centralized MIMO processing, and higher-layer processing controls configuration and runtime status.
  • SDR Partitioning: ⌈M/nant⌉ SDRs are required for an M-antenna system when each SDR supports nant antennas.This sizing follows the per-antenna processing partitioning.
  • SDR Partitioning: Per-antenna RF, OFDM, and reciprocity processing on SDRs reduces the data rate before samples cross the bus.Grouping is introduced because directly connecting 2M antenna links to co-processors would likely exceed practical peer-to-peer link limits.
  • Subsystem Constraints: Outer SDRs implement routing for grouped antenna streams and therefore face the highest peer-to-peer link counts and throughput demands.Their subsystem sizes must satisfy maximum-link and maximum-bidirectional-throughput inequalities, including RSDRmax > nant · nsub · w · Fsub.
  • Co-processor Partitioning: Co-processors centralize detection, precoding, CSI acquisition, and symbol mapping while splitting the bandwidth across sub-bands.Reconfigurable routers aggregate and disaggregate data from all antennas, and co-processor constraints are expressed through maximum links and throughput.
  • Scalability: The modular partitioning enables hardware platforms to be evaluated under different system parameters by redefining quantities such as Fs and Fsub.The framework is designed to preserve scalability while accounting for throughput, latency, and peer-to-peer link constraints.

V. LUMAMI TESTBED IMPLEMENTATION

The LuMaMi implementation instantiates the modular architecture as a reconfigurable 100-antenna testbed serving up to 12 UEs, with hardware selected and dimensioned against the framework’s constraints.

  • System Configuration: 100 BS antennas and up to 12 simultaneous UEs define the LuMaMi testbed design target.The implementation is described as an initial version with planned add-ons and improvements.
  • Hardware Platform: The selected SDRs provide up to 15 peer-to-peer links, 830 MB/s bidirectional throughput, 1.2–6 GHz center-frequency tuning, and 15 dBm transmit power.Each SDR contains two RF chains and a Kintex-7 FPGA.
  • Hardware Platform: The selected co-processors provide 2.4 GB/s bidirectional peer-to-peer throughput, up to 32 links, and up to 2.845 GMAC/s FPGA performance.These capabilities support centralized massive MIMO processing within the modular architecture.
  • Control and Reconfigurability: LabVIEW provides system configuration and control while abstracting peer-to-peer setup, device communication, FPGA programming, and host processing.The stack also supports integration of intellectual property and in-house hardware designs.
  • Subsystem Dimensioning: 50 SDRs are required for M = 100 antennas, with nsub = 8 selected and a final two-SDR subsystem because 8 does not divide 50.The subsystem size minimizes utilization of available peer-to-peer links at the co-processors.
  • Validation: The design remains within the stated constraints when scaling from 100 to 128 BS antennas.The 128-antenna scaling was implemented in subsequent designs using the same hardware.

C. Frame Structure

LuMaMi uses a 10 ms LTE-like frame with dedicated control, pilot, and data timing, and its measured latency is below the default frame-structure requirement.

  • Frame Structure: 10 ms frames contain ten 1 ms subframes, each divided into two 0.5 ms slots.The first subframe carries control signals, while the remaining 18 slots contain seven OFDM symbols each.
  • Control and Calibration: The first subframe supports over-the-air synchronization, uplink power control, and other control signaling.Reciprocity calibration is defined over the full base-station runtime, while pilot and control cycles follow the frame structure.
  • Mobility Support: A pilot spacing of approximately 430 µs, or six OFDM symbols, supports a maximum speed of 70 km/h at a 3.7 GHz carrier frequency.The stated correlation target is 0.9, corresponding to νmax ≈ 240 Hz.
  • Latency Budget: The precoding turnaround budget is 214 µs after receiving uplink pilots, with approximately 140 µs remaining for channel estimation, MIMO processing, and routing.The TX/RX 2048-point FFT/IFFT operations consume about 70 µs of the budget.
  • Validation: Measured implementation latency is far below the default frame-structure requirement, supporting higher-mobility use from a latency perspective.This conclusion concerns the latency requirement associated with the default frame structure.

1) Base-band Processing:

The LuMaMi base-band processing design distributes demanding detection, estimation, monitoring, and storage tasks across reconfigurable co-processors and host resources. Its architecture supports configurable antenna scaling, frame arrangements, and measurements under interconnection and memory constraints.

  • CSI estimation: Orthogonal pilot allocation gives each UE every K-th subcarrier, reducing estimated CSI-matrix storage by a factor of K.A least-square estimator uses zeroth-order hold over K = 12 subcarriers.
  • Detection: Two detector implementations use QR decomposition with a systolic-array partial parallelization or a Neumann-series approach.The QR implementation processes channel-matrix columns with modified Gram-Schmidt steps.
  • Data movement: The co-processors provide 1 GB/s and 14 P2P links for system visualization, performance metrics, host data, and real-time BER transmission.Additional P2P links transmit and store real-time BERs for all 12 UEs.
  • Data movement: Each co-processor uses a 2 GB DRAM buffer to capture CSI dynamics at millisecond intervals without exceeding P2P bandwidth limits.Snapshots cover 60 s at 5 ms intervals or 12 s at 1 ms intervals.
  • Scalability/Reconfigurability: The testbed can configure 4–100 active BS antennas and arbitrarily rearrange all 140 OFDM symbols before startup.This supports experiments such as static UL-only operation with one pilot symbol and remaining data symbols.
  • Reciprocity calibration: Calibration coefficients are estimated on the host for floating-point precision and applied distributively on the SDRs after the hardware reaches operating temperature.Hardware drift is described as insignificant once operating temperature is reached.
  • Hardware integration: The complete system is housed in two 0.8 × 1.2 × 1 m racks mounted on a four-wheel trolley for indoor and outdoor testing.The movable rack mount is an explicit requirement for testing different scenarios.

H. Antenna Array

The testbed combines a configurable dual-polarized antenna array with synchronized SDR-based BS and UE hardware for indoor and outdoor validation. Experiments use 12 densely grouped indoor UEs and evaluate real-time BER under multiple gains and modulations.

  • Antenna Array: The in-house planar T-shaped array has 160 dual-polarized half-wavelength patch elements and 320 possible antenna ports.The default arrangement is 4 × 25, while alternative configurations such as 10 × 10 can be explored.
  • User Equipment: Six SDRs implement 12 single-antenna UEs, with each SDR serving two independent users.UE processing includes OFDM modulation/demodulation, symbol mapping/demapping, least-square CSI acquisition, and ZF equalization.
  • Synchronization: A 10 MHz reference clock and trigger-distribution network provide RF-chain time synchronization and phase coherence at the BS.The reference clock drives radio local oscillators, while the trigger supplies a common time reference.
  • Synchronization: LTE Zadoff-Chu PSS synchronization occupies the center 1.2 MHz and supports over-the-air timing synchronization and frequency-offset compensation.The procedure first scans the radio frame coarsely, then tracks the PSS in a narrowed window.
  • Measurement scenarios: The validation uses a 4 × 25 BS antenna configuration and linear-polarized ultra-wideband UE antennas in indoor and outdoor scenarios.The indoor test stresses dense-user spatial multiplexing, whereas the outdoor test targets range and multiplexing.
  • Measurement procedure: Real-time uncoded BER curves are measured with MRC/MRT and ZF while synchronously sweeping UE or BS transmit-chain PA gains.Each gain step lasts about 4 s and transmits approximately 36 × 10^6 QPSK bits or 108 × 10^6 64-QAM bits.
  • Indoor scenario: Twelve UEs are arranged in groups of four on three lecture-hall tables, creating high user density per unit area.The BS is positioned at the front of the hall.
  • Indoor BER results: For indoor ZF detection, UE0–UE3 have the highest BER, while UE9–UE12 nearest the BS perform best.UE0 and UE1 reach BER 0.5 because of PA saturation; all other UEs remain below 10% at 15 dB QPSK and 25 dB 64-QAM gain.

3) DL BERs:

Downlink measurements show strong dependence on user position, modulation, reciprocity calibration, and precoding choice. Across the tested configurations, BER curves resemble AWGN behavior and support simultaneous service of 12 densely deployed UEs, while several measurement limitations constrain interpretation.

  • DL BERs: UE9–UE12, closest to the BS, achieve considerably better DL QPSK performance than the other user groups.With 64-QAM, all UEs exhibit an error floor at higher transmit gains.
  • DL BERs: The 64-QAM DL error floor is attributed to imperfect reciprocity calibration and leakage caused by imperfect channel knowledge.QPSK performs better with the higher transmit power available from 100 active RF chains.
  • Limitations: The experiments primarily prove functionality rather than maximize reciprocity-calibration accuracy.Individual system components were still being tested for improvement.
  • Limitations: Fig. 10 BER curves do not directly indicate performance versus SNR because the implementation lacks real-time SNR estimates.UE4 was selected for an indirect SNR estimate, but interference or weak signal levels limit that procedure for other users.
  • Performance evaluation: 12 UEs can share the same time/frequency resource in high-density conditions, with BER curves closely resembling an AWGN channel.The paper presents this as evidence that the MaMi concept works and improves spectral efficiency relative to current cellular standards.
  • MRC/MRT versus ZF: For UE7, ZF achieves more than an order of magnitude lower BER than MRC on the UL, while ZF also substantially outperforms MRT on the DL.MRC levels off with increasing PA gain, and MRT shows a significant high-gain error floor.
  • MRC/MRT versus ZF: Direct UL–DL performance comparison is difficult because UL depends on UL transmit power, whereas DL also depends on channel-estimate quality, DL power, and reciprocity accuracy.The DL precoder uses UL channel estimates, making overall performance depend on multiple coupled factors.

B. Outdoor Test

The outdoor campaign served up to eight UEs across two building wings, evaluating MRC and ZF in uplink and downlink. ZF separated the users more clearly, while downlink remained the more challenging duplex case.

  • Outdoor setup: Up to eight UEs were served simultaneously 18–22 meters from the rooftop testbed across two floors of the opposite building wing.Six UEs were on the second floor and two on the first floor.
  • Uplink: Without error-correcting coding, MRC produced significant interference for the six outdoor UEs.The comparison used received uplink QPSK constellations.
  • Uplink: ZF separated up to eight uplink UEs and produced very clear constellations through interference suppression.The result is shown for the outdoor test setup in Fig. 15.
  • Downlink: On the downlink, ZF outperformed MRC for all eight UEs, and downlink was consistently the more challenging duplex case.The received downlink constellations are shown in Fig. 16.
  • Scope: The LuMaMi testbed also supported outdoor mobility measurements, but those results are outside this paper’s scope.The mobility results are reported separately.
Loading 1701.01161v2…