PaperlayerView exact revision on arXiv

Source-linked AI summary

Colosseum: Large-Scale Wireless Experimentation Through Hardware-in-the-Loop Network Emulation

Leonardo Bonati, Pedram Johari, Michele Polese, Salvatore D'Oro, Subhramoy Mohanti, Miead Tehrani-Moayyed, Davide Villa, Shweta Shrivastava, Chinenye Tassie, Kurt Yoder, Ajeet Bagga, Paresh Patel, Ventz Petkov, Michael Seltser, Francesco Restuccia, Abhimanyu Gosain, Kaushik R. Chowdhury, Stefano Basagni, Tommaso Melodia

arXiv:2110.10617v2cs.NIcs.AIeess.SP

TL;DR

Large-scale wireless experimentation needs realistic, accessible environments, and spectrum-sharing research still lacks large-scale tests across heterogeneous network conditions. Colosseum provides an open testbed with programmable radios and emulated channels; its examples report repeatable cellular emulation and better URLLC resource allocation with DRL than with static scheduling.

  • Problem

    Spectrum-sharing research needs large-scale evaluation across heterogeneous network topologies, traffic, and channel conditions, while wireless spectrum demand is increasing.

    Source: C. Spectrum Sharing · p. 6

  • Method

    Colosseum combines programmable radios and protocol stacks with FPGA-based channel emulation to support large-scale experiments across wireless scenarios.

    Source: §I · p. 2, §II · p. 2, §VII · p. 8

  • Results

    The cellular example reports tight 95% confidence intervals for spectral efficiency, and DRL agents outperform static scheduling policies in URLLC resource allocation.

    Source: A. Cellular Networking · p. 5

  • Takeaways & Limitations

    Colosseum provides a publicly available platform for large-scale wireless experiments, including heterogeneous spectrum-sharing and UAV scenarios.

    Source: C. Spectrum Sharing · p. 6, §VII · p. 8

    Colosseum currently models terrains up to 1 × 1 km and sub-6 GHz omnidirectional transmissions, not directional transmissions or higher frequency bands.

    Source: A. Expansions to Support NRDZ and… · p. 7

Abstract

from arXiv · show

Colosseum is an open-access and publicly-available large-scale wireless testbed for experimental research via virtualized and softwarized waveforms and protocol stacks on a fully programmable, "white-box" platform. Through 256 state-of-the-art software-defined radios and a massive channel emulator core, Colosseum can model virtually any scenario, enabling the design, development and testing of solutions at scale in a variety of deployments and channel conditions. These Colosseum radio-frequency scenarios are reproduced through high-fidelity FPGA-based emulation with finite-impulse response filters. Filters model the taps of desired wireless channels and apply them to the signals generated by the radio nodes, faithfully mimicking the conditions of real-world wireless environments. In this paper, we introduce Colosseum as a testbed that is for the first time open to the research community. We describe the architecture of Colosseum and its experimentation and emulation capabilities. We then demonstrate the effectiveness of Colosseum for experimental research at scale through exemplary use cases including prevailing wireless technologies (e.g., cellular and Wi-Fi) in spectrum sharing and unmanned aerial vehicle scenarios. A roadmap for Colosseum future updates concludes the paper.

I. INTRODUCTION

Wireless research needs programmable, large-scale testbeds that combine realistic hardware, propagation, traffic, and topology conditions, but existing options are difficult to access, limited in scale, or costly. Colosseum is presented as a community-accessible platform for full-stack experimentation at scale in realistic emulated environments.

I. INTRODUCTION

  • High-fidelity wireless testbeds are hard to access, often limited in scale, and costly, while simulations and small labs each miss important real-world conditions.Simulations can lack hardware, interference, and channel-variation fidelity; small labs cannot capture deployment scale, real propagation, or interference.Source: §I · p. 1
  • Wireless testbed initiatives include city-scale PAWR platforms, while smaller indoor testbeds offer hardware and propagation capabilities but cannot capture large-deployment dynamics.The cited examples include POWDER-RENEW, COSMOS, AERPAW, and ARA, as well as the Drexel Grid, ORBIT, and Arena.Source: §I · p. 1
  • Colosseum combines 256 SDRs and 128 remotely accessible compute nodes for full-protocol-stack experiments with complex emulated channels.It is also described as an AI playground and wireless data factory for creating large-scale datasets and training or testing solutions.Source: §I · p. 2

II. COLOSSEUM ARCHITECTURE

Colosseum connects remotely controlled radio nodes to an FPGA channel emulator and supporting scenario, traffic, and management services. Its architecture supports softwarized protocol stacks and emulates channel effects by filtering and combining transmitted signals.

II. COLOSSEUM ARCHITECTURE

  • The 128 Standard Radio Nodes are synchronized in time and frequency, each using a 48-core server, GPU, and USRP X310 with two daughterboards operating from 10 MHz to 6 GHz.Containers on the bare-metal nodes can run varied protocol stacks and control their parameters and configurations.Source: §II · p. 2
  • MCHEM processes radio signals through FPGA-based FIR filters using precomputed channel impulse responses to model transmitter-to-receiver links.The emulator uses 128 USRP X310s connected one-to-one with the SRN radios; its four quadrants each contain four FPGA modules with 16 Virtex-7 FPGAs apiece.Source: §II · p. 2
  • The emulator convolves transmitted signals with their channel responses and combines the results at each receiving node, capturing interference and signal superposition.Signals are forwarded to all experiment SRNs, not only the intended receiver.Source: §II · pp. 2–3
  • Only 4 channel taps are nonzero because generating scenarios is computationally complex and storing them requires substantial space.The channel responses are represented with 512 precomputed complex-valued taps.Source: §II · p. 2
  • RF scenarios model path loss and fading over areas up to 1 km^2 and bandwidths up to 80 MHz, while TGEN supplies configurable IP traffic and management services support reservations, resource allocation, storage, and synchronization.MCHEM can emulate different scenarios concurrently for different experiments.Source: §II · p. 3

III. EXPERIMENTAL SCENARIOS IN COLOSSEUM

Colosseum scenarios let users configure and reproduce wireless channels at scale, including propagation effects and interference. The available examples span urban UAV communications, changing signal conditions, and cellular deployments based on real-world base-station locations.

III. EXPERIMENTAL SCENARIOS IN COLOSSEUM

  • RF scenarios emulate up to 65,535 wireless channels across areas up to 1 km^2, with user-controlled channel effects and repeatable conditions derived from models, field measurements, or ray tracing.FPGA FIR filters reproduce channel impulse responses and model interference and signal superposition.Source: §III · p. 3

A. Examples of Available Experimental Scenarios

  • The Alleys of Austin scenario places 50 nodes in five squads, each with nine pedestrians and a circling UAV, and changes traffic over three five-minute stages.It uses a 1 GHz center frequency and progresses from voice to voice plus image/video traffic, then to a higher transmission rate.Source: A. Examples of Available Experimental… · p. 3
  • The SCE Qualification scenario runs 10 nodes at 1 GHz for 10 minutes, lowering SNR by 5 dB every two minutes before a final frequency shift to 1.1 GHz.During the last two minutes, SNR resets to 20 dB; nodes exchange constant-bitrate UDP packets.Source: A. Examples of Available Experimental… · p. 3
  • Cellular scenarios model 8 to 10 base stations serving four mobile users each at real deployment locations in Rome, Boston, and Salt Lake City.Users vary by distance and mobility, and uplink and downlink traffic includes video packets.Source: A. Examples of Available Experimental… · p. 3

B. Colosseum as a Wireless Data Factory

Colosseum supports both repeatable bulk data collection and hands-on experimentation. Users can configure experiments interactively or submit batch jobs that automatically instantiate containers, run scenarios, and save results.

B. Colosseum as a Wireless Data Factory

  • Users can configure an experiment once and automatically rerun it across different environments, channel conditions, and traffic scenarios to collect large-scale datasets for machine-learning training and testing.The authors note that varied conditions are needed for ML algorithms to generalize and adapt to unforeseen situations.Source: B. Colosseum as a Wireless Data Factory · p. 3

IV. OPERATIONAL MODES OF COLOSSEUM

  • Interactive experiments begin with SRN reservations and image selection; resource use costs tokens, allocated through weekly team budgets.The system automatically instantiates containers, and users access their SRNs through the gateway.Source: §IV · p. 4
  • In interactive mode, users start RF and traffic scenarios, control USRPs through softwarized protocol stacks, and can snapshot containers and retrieve saved logs.The colosseumcli interface provides access to services including MCHEM and TGEN.Source: §IV · p. 4
  • Batch jobs are suited to benchmarking and bulk data collection, with JSON configuration for duration, scenarios, SRN count, images, and user-defined parameters.Jobs queue until resources are available, then run containers and startup scripts automatically and copy logs to the NAS.Source: §IV · p. 4

V. USE CASES OF COLOSSEUM

The section presents Colosseum use cases spanning cellular networking, software-defined Wi-Fi, spectrum sharing, and UAV networking, then details cellular experiments.

A. Cellular Networking

  • A 4-base-station, 24-user srsRAN network measures downlink throughput under 1, 2, and 4 Mbps per-user traffic loads.The base stations use a 10 MHz channel bandwidth.Source: A. Cellular Networking · p. 4, Figure 4 · p. 4
  • 95% confidence intervals remain tight as cellular spectral efficiency decreases with greater user distance or mobility, supporting emulation reliability and repeatability.The scenarios compare distances of 20, 50, and 100 m and speeds of 0, 3, and 5 m/s.Source: A. Cellular Networking · p. 5, Figure 5 · p. 5
  • Colosseum runs an O-RAN-managed cellular network with DRL xApps controlling slice allocation and scheduling through a near-real-time RIC.The experiment uses four softwarized base stations and 40 users in a dense urban scenario.Source: A. Cellular Networking · p. 5, Figure 6 · p. 5
  • DRL-controlled scheduling achieves superior URLLC resource allocation to static scheduling policies.The agents dynamically select policies based on channel conditions and traffic demand.Source: A. Cellular Networking · p. 5, Figure 7 · p. 5

B. Wi-Fi

The Wi-Fi use case evaluates a 50-node network using an open-source GNU Radio implementation of IEEE 802.11a/g/p. Mobility degrades measured SINR.

B. Wi-Fi

  • The use case addresses large-scale experimentation for emerging unlicensed 6 GHz Wi-Fi spectrum, including Wi-Fi 6.The passage identifies Wi-Fi 6 as the IEEE 802.11ax amendment.Source: B. Wi-Fi · p. 5
  • The experiment measures SINR in a 50-transceiver urban Wi-Fi scenario using an open-source GNU Radio implementation of IEEE 802.11a/g/p.It compares stationary nodes with nodes moving at an average speed of 3 m/s.Source: B. Wi-Fi · p. 5, Figure 8 · p. 6
  • With Wi-Fi node mobility, rapidly changing channel conditions cause SINR to degrade by 3.18 dB on average; Fig. 8 also shows a scenario without node mobility.Source: B. Wi-Fi · p. 6, Figure 8 · p. 6

C. Spectrum Sharing

Colosseum supports large-scale tests of heterogeneous wireless technologies sharing spectrum; an example shows Wi-Fi transmissions coinciding with degraded cellular link metrics.

C. Spectrum Sharing

  • Rising device counts and traffic demand intensify spectrum scarcity, while static allocation can leave unlicensed bands congested and licensed bands lightly loaded.The passage identifies limited large-scale experimentation across heterogeneous topologies, traffic, and channel conditions as a limitation of prior spectrum-sharing work.Source: C. Spectrum Sharing · p. 6
  • Colosseum lets researchers instantiate large-scale networks with heterogeneous protocol stacks, such as cellular and Wi-Fi, on the same spectrum bandwidth.This supports testing spectrum-sharing solutions across diverse network deployments and conditions.Source: C. Spectrum Sharing · p. 6
  • Wi-Fi activity coincides with rapidly dropping cellular-user CQI and increasing uplink errors in the shared-spectrum experiment.The experiment places a cellular base station serving five users alongside two Wi-Fi nodes on the same spectrum portion.Source: Figure 9 · p. 6, C. Spectrum Sharing · p. 6

D. Unmanned Aerial Vehicles

Colosseum models a 30-node UAV-and-relay network and compares relay-selection strategies under inter-user and inter-squad interference.

D. Unmanned Aerial Vehicles

  • The Alleys of Austin experiment has 30 nodes in three squads, each with one UAV and nine ground relays delivering data to squad members.The UAV trajectories are shown in the network-topology panel.Source: D. Unmanned Aerial Vehicles · p. 6, Figure 10 · p. 7
  • Relay nodes, aerial or ground, forward squad data to intended destinations while coping with inter-user and inter-squad interference.The compared strategies are SINR-based maximum weighted matching and random allocation.Source: D. Unmanned Aerial Vehicles · pp. 6–7
  • Maximum weighted matching yields higher average SINR than random relay allocation in the three-squad scenario.It selects relays with better channel conditions toward squad members and uses UAVs as aerial relays with line-of-sight to ground nodes.Source: Figure 10 · p. 7, D. Unmanned Aerial Vehicles · p. 7

VI. EVOLUTION OF COLOSSEUM

Colosseum is planned to evolve through added emulation capabilities and AI components.

VI. EVOLUTION OF COLOSSEUM

  • Planned Colosseum extensions will add new emulation capabilities and AI components.The paper introduces these as planned updates to meet future wireless-networking needs.Source: §VI · p. 7

A. Expansions to Support NRDZ and mmWave Research

Colosseum is positioned to support NRDZ research, with planned extensions addressing current limits in terrain scale, directional transmission, and frequency range.

A. Expansions to Support NRDZ and mmWave Research

  • NRDZ activities will expand Colosseum to new spectrum bands, scenarios, and power levels, with a role in spectrum-sharing protection, observation, validation, and automation.The planned extensions add two quadrants and improve MCHEM emulation.Source: A. Expansions to Support NRDZ and… · p. 7, Figure 11 · p. 7
  • Current emulation is limited to terrains up to 1 × 1 km and sub-6 GHz omnidirectional transmissions; directional links and higher bands are unsupported.These constraints limit Colosseum's use for NRDZ design and study.Source: A. Expansions to Support NRDZ and… · p. 7
  • A development environment will test MCHEM and radio-hardware extensions before scaling them to full Colosseum.Planned channel-emulation extensions target 100 × 100 km terrain and beamforming-based directional mmWave communications; a new quadrant is planned for baseband bandwidths up to 2 GHz.Source: A. Expansions to Support NRDZ and… · p. 7

B. AI Jumpstart Integration

Colosseum’s AI Jumpstart expansion adds AI-focused compute resources to the testbed, with proposed uses spanning wireless dataset training, 5G signal processing, and network control. The paper also presents Colosseum as a publicly available platform for large-scale experiments across multiple wireless scenarios.

B. AI Jumpstart Integration

  • The MassTech AI Jumpstart program is expanding Colosseum to support AI-based wireless research, including large-scale dataset training, real-time 5G signal processing, and model-free network adaptation and control.The program aims to connect Massachusetts firms with facilities and researchers.Source: B. AI Jumpstart Integration · p. 7
  • New equipment includes two NVIDIA DGX A100 nodes capable of up to 10petaFLOPS, plus a 3 TB memory node and an Infiniband switch supporting up to 16 Tbps aggregated traffic.The equipment is to be fully meshed with Colosseum’s compute and wireless resources.Source: B. AI Jumpstart Integration · pp. 7–8

VII. CONCLUSIONS

  • Colosseum is presented as a publicly available hardware-in-the-loop testbed for large-scale experiments in cellular, Wi-Fi, spectrum-sharing, and UAV scenarios, with future expansion planned.Source: §VII · p. 8

Related papers

Loading 2110.10617v2…