Dynamic Pacing for Real-time Satellite Traffic

TL;DR

This work addresses real-time videoconferencing over LEO networks, where GCC-based QoS degrades due to handovers and dynamic delays. It introduces a handover-aware queue management policy trained via offline imitation learning, mapping network context to WebRTC pacing queue limits using a transformer encoder and clustering-derived expert actions. The approach yields significant gains, including up to 3x video bitrate and substantial reductions in freezes and packet loss in both emulated sparse LEO and live Starlink deployments, while remaining safe in terrestrial networks. The results demonstrate that adaptive, sender-side pacing informed by predicted handover dynamics can meaningfully enhance QoE for interactive RTC in satellite-enabled environments, with broad implications for deploying WebRTC in LEO constellations.

Abstract

Google's congestion control (GCC) has become a cornerstone for real-time video and audio communication, yet its performance remains fragile in emerging Low Earth Orbit (LEO) networks. In this paper, we study the behavior of videoconferencing systems in LEO constellations. We observe that video quality degrades due to inherent delays and network instability introduced by the high altitude and rapid movement of LEO satellites, with these effects exacerbated by WebRTC's conventional "one-size-fits-all" sender-side pacing queue management. To address these challenges, we introduce a data-driven queue management mechanism that tunes the maximum pacing queue capacity based on predicted handover activity, minimizing latency during no-handover periods and prioritizing stability when entering periods of increased handover activity. Our method yields up to 3x improvements in video bitrate and reduces freeze rate by 62% in emulation, while delivering up to a 41% reduction in freeze rate and 40% decrease in mean packet loss on real Starlink constellations compared to WebRTC's default pacing queue policy.

Figures (6)

• Figure 1: Two distinct failure modes of WebRTC's default queue policy in LEO networks: self-inflicted delays during stable periods (few handovers) and high instability during dynamic periods (frequent handovers).
• Figure 2: System Workflow: (1) An initial policy is deployed client-side. (2) Telemetry logs are aggregated and (3) clustered to build an expert target policy. (4) A final policy is trained to imitate this expert and is subsequently deployed.
• Figure 3: Handover characteristics in LEO deployments. (a) Distribution of time between handovers. (b) Distribution of handover counts per 2-minute window in Starlink.
• Figure 4: On emulated terrestrial links, we observe no impact on key metrics, which suggests that dynamic pacing (Ours) offers little benefit over the default policy (Def) in these low-latency environments.
• Figure 5: In emulated sparse LEO environments, our handover-aware pacing policy (Ours) outperforms WebRTC's default (Def), significantly reducing video freezes while improving video bitrate.