AI service traffic patterns revealed for 6G networks with new testbed
Characterizing AI Service Traffic for 6G: An Open Testbed and Initial Results from 3GPP SA4
Networking and Internet Architecture
Summary
Building 6G mobile networks needs to understand how AI-powered services send and receive data. The authors created an open testbed combining network emulation and real AI models to measure AI traffic under controlled conditions. They found that the ratio of data sent uplink versus downlink changes dramatically across different AI services, and that signal delays and data loss impact AI performance in measurable ways. They also studied how network emulator quirks affect interpreting these measurements. This work helps future 6G designs better handle AI service demands.
What this means in practice
- •For mobile network engineers: Design and test 6G systems to efficiently support varying AI service traffic patterns using this open testbed's measurement insights.
- •For cloud service operators: Optimize deployment and scaling of hosted AI models by understanding uplink/downlink traffic balances and loss impacts.
Authors
Imed Bouazizi
Abstract
The design of sixth generation (6G) mobile systems requires measurements of artificial intelligence (AI) service traffic under controlled network conditions. The 3rd Generation Partnership Project (3GPP) Service and System Aspects Working Group 4 (SA4) documents such measurements in draft Technical Report 26.870. In this paper, we describe the open testbed used for the study and analyze the results recorded in version 0.6.2 and its companion workbook. The testbed combines network emulation, packet capture, application measurements, and scenario execution against hosted and self-hosted models. The reported uplink to downlink volume ratios span more than six orders of magnitude. A personal assistant agent has a ratio of 0.007 at its local control interface and 0.50 on its wide-area egress. Deterministic agent signaling follows the configured round-trip delay. In a tokenized video experiment with 125 questions per profile, the number of correct answers changes by at most three through 10% independent uplink loss. Larger losses reduce accuracy, but the available repetitions do not establish a tolerance threshold. We also analyze emulator behavior that affects the interpretation of the study. Per-packet jitter introduces reordering, correlated random-loss settings suppress rare loss events, and a Gilbert-Elliott parameter mapping changes the stationary loss rate. The results support characterization by scenario, measurement layer, and versioned network configuration.