Choir: Tackling RTBC Performance Impossible Triangle with 5G Collaboration

Abstract: Real-time broadband communication (RTBC) scenarios, such as cloud virtual reality and 8K live streaming, further raise the criteria of the performance triangle, requiring video bitrates exceeding 30 Mbps, tail delay below 50 ms, and fairness guarantees for multi-user concurrent access. Based on our testing and analysis, existing RTBC-oriented rate control solutions, including end-to-end algorithms and network-assisted algorithms, fail to simultaneously satisfy all performance metrics. The native dynamic delay and physical-layer resource allocation strategy inherent to the 5G radio access network (RAN) are the key reasons. These solutions lack adaptation to the 5G architecture, leading to reduced decision performance. This paper proposes Choir, an innovative collaborative solution mainly deployed on 5G base stations that deeply integrates 5G radio characteristics and video streaming traffic patterns to guide efficient sender-side rate control. Extensive simulation and testbed evaluations demonstrate Choir's significant performance in achieving high average bitrate, low tail delay, and inter-flow fairness across different 5G network scenarios.

• The paper introduces Choir, a base-station-assisted rate-control framework that uses TDD scheduling, retransmission data, and per-flow capacity estimates to guide sender bitrate and pacing decisions.
• Choir reduces 99.9th-percentile tail delay by up to 90.7% in simulation while keeping average bitrate within 0.4%–3.3% of the best baseline, and it remains effective across wired latencies and testbeds.
• The framework delegates fairness to 5G base-station scheduling, achieving nearly proportional allocation from 7 to 14 concurrent flows, while open questions remain around commercial-RAN deployment, mobility, and parameter tuning.

RTBC (Real-Time Broadband Communication) workloads such as cloud VR and 8K live streaming impose a "Performance Triangle" of requirements: video bitrates above 30 Mbps, tail delay below 50 ms, and fairness across concurrent users. This paper argues that existing rate control solutions—both end-to-end probing algorithms and network-assisted schemes—cannot satisfy all three simultaneously in 5G networks, and proposes Choir (2605.02510), a base-station-collaborated rate control framework that exploits 5G RAN characteristics rather than treating them as an opaque bottleneck.

Motivation: why 5G breaks existing rate control

The authors first demonstrate that state-of-the-art solutions (SCReAM, COPA, SQP, Pudica, ABC, L4S, SCONE) exhibit unstable transmission rates and persistent delay oscillations even in stable 5G environments. The root cause is what they term "native dynamic delay" from the 5G RAN architecture:

• TDD pattern dynamics: with the widely deployed DDDSU pattern (3.5:1.5 downlink-to-uplink ratio), packets arriving during uplink slots incur roughly 1.5 ms of additional queuing delay even without congestion.
• Retransmission-induced latency: with a measured BLER around 10%, HARQ retransmissions add ~6 ms for the first retransmission and ~16 ms after three; RLC Acknowledged Mode retransmissions can add tens of milliseconds. Because these mechanisms hide loss from senders, they amplify tail latency instead of triggering rate reduction.

Delay-based congestion signals are therefore unreliable in 5G, forcing end-side solutions into persistent oscillation. Network-assisted schemes such as L4S and ABC adapt quickly to bandwidth drops but converge slowly and continue fluctuating afterward.

The paper's key architectural observation is that this same RAN design also provides an opportunity: per-flow dedicated queues and polling-based scheduling at the base station enforce physical-layer resource fairness natively. This aligns with the CCAI principle, relieving sender-side CCAs from enforcing fairness so they can focus on bandwidth-delay optimization. The implication is that fairness can be delegated entirely to the base station, while rate control reduces to elastically tracking base-station-allocated capacity.

Choir design

Choir places two functions in the base station and a lightweight controller at the sender.

Radio-characteristics-oriented capacity estimation. At TTI granularity (0.5–1 ms), the base station computes each flow's PRB allocation based on active flow count and idle resources from the previous TTI, derives per-symbol rates from CQI-reported MCS values, and accounts for the TDD pattern by counting only downlink TTIs actually used for user data over a 100 ms window. It then corrects for retransmission overhead using a retransmission ratio $R_{ReTx} = HN/DN$ computed over a 10 ms window—a deliberately finer granularity than the standard 100 ms BLER statistic—and discounts transport block overhead via a measured $\gamma$ factor.

Flow-pattern-oriented available bandwidth prediction. The allocated bandwidth is converted into a guidance bandwidth targeting near-zero queuing: $$\widehat{BW_i} = (\eta \times \overline{BW_i} - DR_i)^{+}$$ with $\eta = 0.95$, where $DR_i$ is the draining rate needed to empty the queue before the next frame arrives. The central challenge is that the base station lacks knowledge of frame sizes and intervals, yet must predict queue occupancy over the decision-to-execution latency between itself and the sender. Choir infers frame intervals from inter-packet gaps (smoothed with EWMA, $\alpha = 0.8$), estimates the number of in-flight frames from the wired propagation delay, and reconstructs predicted queue length from historical feedback rates, an autoregressive error-correction term, and a sliding-window minimum of the RLC buffer sampled per TTI. The minimum-window statistic is chosen specifically because moving averages miss bursty load changes while instantaneous samples are noise-prone—an empirical design choice justified by video traffic burstiness rather than derived analytically.

User-experience-oriented sender rate control. The sender maps guidance bandwidth to encoder target bitrate and pacing rate ($\rho = 1.25$ times the max of receive rate and target bitrate). A smoothing coefficient $\epsilon$ trades off bitrate stability against delay responsiveness; the default $\epsilon = 1$ tracks guidance directly.

Feedback is delivered online: the base station marks the guidance rate in IP option fields on uplink packets, avoiding both QUIC payload encryption complications and receiver-side involvement. This shortens the feedback path relative to schemes like ABC and L4S that require receiver processing.

Implementation

Evaluation uses two platforms: trace-driven simulation with NS3's 5G LENA module (explicitly configuring TDD patterns, wired latencies of 1/10/20 ms), and two OAI-based software-defined testbeds—a USRP B210 system (~30 Mbps capacity) and a Witcomm OpenXG system (~300 Mbps)—with free5GC core, XQUIC datagram-mode transport, RTP-over-QUIC framing, and gstreamer/H.264 encoding at 60 fps. Baselines span end-side (Pudica, SQP, SCReAM, COPA) and network-assisted (L4S, ABC, SCONE) approaches; GCC and NADA are excluded on the grounds that prior comparisons show SCReAM dominates them in latency.

Single-flow results

In simulation, Choir reduces 99.9th percentile tail delay by 45.5%–90.7% versus all baselines except SCONE under 1 ms wired latency (where SCONE is 4.9% better), 33.8%–88.9% versus others under 10 ms (SCONE 8.1% better), and 51.4%–81.6% under 20 ms (SCONE 19.2% better). On average bitrate, Choir trails the best baseline by only 0.4%–3.3% while beating weaker ones by up to 89.9%. Testbed results confirm the pattern: on the B210 RAN, Choir achieves nearly the lowest frame delay with the highest average bitrate, reducing tail latency by 7.10%–68.64% against non-SCONE baselines.

The paper is candid that SCONE consistently achieves slightly lower tail latency—the cost of Choir's bandwidth-utilization emphasis—but positions Choir as occupying the upper-left region of the bitrate-delay Pareto frontier, balancing both objectives where SCONE drains queues aggressively at some throughput cost. Notably, Choir's tail delay increase as wired latency grows approximates the raw RTT growth, indicating limited sensitivity to feedback-path length; however, the authors acknowledge that when wired latency exceeds the encoder update interval, guidance lags behind bitrate adjustments, causing increased bitrate fluctuations and some residual queuing.

Multi-flow results and fairness

Under 7-flow concurrency on the openXG RAN, Choir achieves consistent performance across flows, whereas baselines exhibit uneven allocation. Against all baselines except SCONE, Choir reduces 99.9th percentile tail delay by 3.77%–71.67% depending on wired latency, with SCONE holding only a 2.15%–2.45% advantage; average bitrate improves by 2.51%–35.04%. Scaling to 14 flows at full load, Choir maintains fair allocation with no significant tail-delay rise—for example, at 1 ms wired latency, 7 flows achieve 34.85 Mbps each with 37.06 ms P99.9 delay, and 14 flows achieve 17.41 Mbps each with 37.72 ms P99.9 delay. This scaling behavior substantiates the paper's claim that base-station-enforced fairness makes sender-side fairness mechanisms unnecessary.

Sensitivity analysis

Three ablations probe robustness. First, ACK frequency reduction (one ACK per 1/2/3 frames, simulating selective acknowledgement) degrades Choir only mildly, with slight tail-delay increases. Second, increasing the smoothing factor $\epsilon$ improves bitrate stability but raises 99th percentile tail delay, quantifying the stability-latency trade-off inherent in the sender-side design. Third, higher wired latency increases bitrate fluctuations due to prediction error over longer decision-to-execution windows, though Choir remains the best performer among compared solutions.

Discussion points raised by the authors

On deployability, Choir requires only sender and base station cooperation—fewer parties than ABC or L4S—and works with IP options over RTP or QUIC/TCP without receiver modification. On adversarial behavior, the base station inherently bounds misbehaving senders: exceeding the guidance bandwidth yields only self-inflicted tail delay, and cannot degrade other flows. Computational overhead at the base station is claimed comparable to L4S, with lower communication overhead since feedback bypasses the UE.

Limitations and open questions

Several limitations deserve note. The evaluation relies on OAI-based testbeds and NS3 simulation rather than commercial RAN equipment; whether the fine-grained PRB, TBS, and retransmission statistics Choir depends on are accessible in production base stations remains unaddressed. The prediction model assumes periodic frame generation with stable frame intervals and uses ping-measured wired delay as ground truth—assumptions that may weaken under variable cross-traffic on the wired segment or adaptive frame-rate encoders. The smoothing coefficient $\gamma$0 and pacing factor $\gamma$1 are set empirically, and the paper does not provide a principled method for tuning them across deployment contexts. Finally, evaluation covers stationary single-cell scenarios; mobility, handover, and carrier aggregation—known sources of severe 5G capacity fluctuation—are outside the evaluated scope, as is interaction with background non-Choir traffic beyond brief fairness illustrations.

Conclusion

Choir reframes RTBC rate control in 5G as a collaboration problem: the base station, uniquely positioned to observe TDD patterns, HARQ/RLC retransmissions, and per-flow scheduling, predicts each flow's available bandwidth at TTI granularity and guides sender-side bitrate and pacing decisions through in-band feedback. Across trace-driven simulation and two open-source 5G testbeds, it reduces tail latency by up to 90.7% relative to existing solutions while matching their best bitrates within a few percent, and achieves near-perfect inter-flow fairness at full cell load by delegating fairness to physical-layer scheduling. The main open questions concern portability to commercial RANs, robustness under mobility, and principled tuning of its empirical parameters.