Title: Synchronizing with the Scheduler: Dual-Loop Congestion Control for 5G Uplink on Commodity Devices
Authors: Qiang Wu, Yuxin Liu, Tianyang Zhang, Yaxiong Xie (University at Buffalo, SUNY); Haoran Wan, Kyle Jamieson (Princeton University)
Introduction
This paper studies congestion control for 5G and LTE uplinks, where real-time applications such as video conferencing, live streaming, and teleoperation need both high throughput and low latency. The main challenge is a timescale mismatch: cellular base-station schedulers reallocate radio resources within milliseconds, while conventional end-to-end congestion-control algorithms depend on RTT, packet loss, or delay feedback that may arrive hundreds of milliseconds later after queues have already built up. Consequently, existing algorithms such as CUBIC, BBR, and GCC either overshoot available capacity and cause bufferbloat or recover too slowly and waste bandwidth. Infrastructure-assisted approaches can expose more accurate radio information, but they require modifying proprietary base stations; SDR-based monitoring is also too expensive and impractical for widespread deployment.
Key idea and contribution
The paper identifies the Grant-to-Buffer Status Report Ratio (GBR), defined as the ratio between the uplink resources granted by the base station and the mobile device’s reported buffered demand. In the normal cellular BSR-to-grant procedure, a device reports how much data it wants to send and the scheduler allocates resources accordingly. When the cell is lightly loaded, grants meet or exceed demand, so GBR is at least one; when the cell is congested, grants are clipped and GBR falls below one. Because this exchange is mandatory in LTE and 5G and occurs about every 5 ms, GBR provides a fast scheduler-level congestion signal available from commodity-device diagnostic interfaces without changing the network infrastructure.
Based on this observation, the authors build GBR-CC, a dual-loop congestion controller. Its inner loop adjusts the sending rate on every fresh GBR sample, using bounded multiplicative updates to track radio-resource scarcity before queues grow. Its outer loop retains GCC-style end-to-end delay control to detect bottlenecks outside the radio access network, such as a congested backhaul. The authors also integrate the design into WebRTC. Rather than reacting to every small radio fluctuation, the WebRTC version uses QoE-aware tuning to stabilize the encoder bitrate and avoid inflating the receiver’s jitter buffer.
Evaluation
The authors evaluate GBR-CC on commercial AT&T, Verizon, and T-Mobile LTE and 5G FDD/TDD networks, using stationary and mobile scenarios. Their measurements show that GBR strongly tracks base-station load and remains available at millisecond granularity even when RTT is inflated by buffering. For bulk transport, GBR-CC reduces median RTT by up to 4× relative to loss-based schemes while providing up to 2.5× the throughput of delay-based alternatives; compared with BBR, it achieves 5% higher average throughput and halves median RTT. In WebRTC, it improves average throughput over GCC by 50%, reduces median end-to-end playout latency by 32–53%, and cuts video freeze rate by about 60%, while also improving visual quality. This result is significant because it shows that a deployable endpoint-only design can turn previously hidden scheduler decisions into immediate control signals, substantially improving both network efficiency and user-perceived video quality.
Personal thoughts
I like the paper’s central insight because it is simple, practical, and grounded in how cellular networks already work. Rather than asking operators to deploy a new protocol or requiring a separate SDR to observe the cell, the design repurposes the existing BSR-to-grant loop. The dual-loop architecture is also well motivated: GBR reacts to the wireless bottleneck at the scheduler’s timescale, while the end-to-end loop preserves protection against congestion elsewhere on the path. The WebRTC adaptation is especially thoughtful, since it recognizes that maximizing instantaneous transport performance does not necessarily maximize video-call QoE.
A key open question is deployability on ordinary user devices. The current implementation relies on modem diagnostic telemetry and, at present, root access; broad adoption would likely require operating systems or modem vendors to expose a safe, standardized GBR-like interface. Another question is how robust the fixed target GBR value is across different operators, scheduler policies, device chipsets, and future 5G/6G deployments. An adaptive target-selection mechanism could make the system more portable. It would also be valuable to examine interactions with more heterogeneous competing traffic and to extend the GBR signal to other congestion-control algorithms beyond GCC.