6G AI Traffic Characterization Testbed

A framework for measuring and analyzing AI/LLM service traffic patterns under various network conditions, designed to support 3GPP SA4 6G Media Study contributions.

Components

Component	Description
aitestbed/	Main testing framework for running AI traffic experiments
netemu/	Network emulation library wrapping Linux tc/netem

aitestbed

The core testing framework that orchestrates experiments across multiple AI providers and scenarios.

Features:

• 11 scenario types: Chat, agentic AI with MCP tools, image generation, multimodal, video understanding, realtime WebSocket/WebRTC
• LLM providers: OpenAI, Gemini, DeepSeek, vLLM, plus OpenAI Realtime (WebSocket/WebRTC)
• 60+ metrics: TTFT/TTLT, latency percentiles, UL/DL ratios, token rates, agent loop factors
• Multi-layer traffic capture: L3/L4 via tcpdump, L7 via mitmproxy
• SQLite logging with structured metrics schema

# From the repo root:
pip install -e netemu
pip install -r aitestbed/requirements.txt
cd aitestbed
python orchestrator.py --scenario chat_basic --profile 5g_urban --runs 10

For the full SA4 cross-check run (all scenarios × profiles, with PCAP capture and report generation), see the Cross-Checking for SA4 AI Traffic Characterization section in aitestbed/README.md.

netemu

A lightweight network emulation library providing a clean interface to Linux traffic control.

Features:

• Wraps tc and netem for delay, jitter, packet loss, and rate limiting
• Bidirectional shaping via IFB devices
• Predefined profiles including 3GPP 5QI mappings and SA4 S4-260848 reference conditions
• Context manager support for automatic cleanup

from netemu import NetworkEmulator

with NetworkEmulator(interface="eth0") as emu:
    emu.apply_profile("5g_urban")  # 20ms delay, 0.1% loss, 100 Mbps
    # Run your tests here
# Rules automatically cleared

Quick Start

# Clone and setup
cd testbed
python -m venv venv
source venv/bin/activate

# Install netemu first (separate package), then testbed dependencies
pip install -e netemu
pip install -r aitestbed/requirements.txt

# Set API keys
export OPENAI_API_KEY="your-key"

# Run a basic experiment
cd aitestbed
python orchestrator.py --scenario chat_basic --profile 6g_itu_hrllc --runs 5

Docker

docker build -t 6g-ai-testbed -f aitestbed/Dockerfile .
docker run --cap-add=NET_ADMIN -e OPENAI_API_KEY="..." \
  6g-ai-testbed python orchestrator.py --scenario all --runs 10

Network Profiles

Current test matrix (from aitestbed/configs/profiles.yaml, aligned with 3GPP SA4 S4-260848 Table C.Z-1):

Profile	Delay	Jitter	Loss	Loss Distribution	Rate	Use Case
no_emulation	0 ms	0 ms	0%	--	--	Reference (no tc/netem)
6g_itu_hrllc	1 ms	0.2 ms	0.001%	correlated (10%)	300 Mbps	6G HRLLC (ITU IMT-2030 / M.2160)
5g_urban	20 ms	5 ms	0.1%	correlated (25%)	100 Mbps	Mainstream urban cellular
wifi_good	30 ms	10 ms	0.1%	correlated (30%)	50 Mbps	Home/office WiFi
cell_edge	120 ms	30 ms	1%	Gilbert-Elliot	5 Mbps	Weak radio, heavy-tail jitter
satellite_leo	22 ms	7 ms DL / 8 ms UL	0.5% DL / 0.8% UL	correlated (40% DL / 45% UL)	100 DL / 15 UL Mbps	LEO satellite (asymmetric)
satellite_geo	340 ms	15 ms DL / 18 ms UL	0.1% DL / 0.2% UL	correlated (20% DL / 25% UL)	50 DL / 3 UL Mbps	GEO satellite (asymmetric)
congested	200 ms	50 ms	3%	Gilbert-Elliot	1 Mbps	Bufferbloat / heavy congestion
5qi_7	80 ms	10 ms	0.1%	correlated (20%)	--	5QI 7: voice / live streaming
5qi_80	8 ms	1 ms	1e-6	correlated (5%)	--	5QI 80: low-latency eMBB / AR

Asymmetric profiles (satellite_leo, satellite_geo) use an optional uplink: block that overrides egress-side fields. See aitestbed/README.md for the full table with jitter, loss models, and advanced netem parameters.

Requirements

• Python 3.10+
• Linux with iproute2 (for network emulation)
• Sudo access or Docker with NET_ADMIN capability