How AutoRec Boosts Live Streaming Reliability: A Deep Dive into Timeliness‑Enhanced Loss Recovery

Overview

The paper “Toward Timeliness‑Enhanced Loss Recovery for Large‑Scale Live Streaming” (oral presentation at ACM Multimedia 2024) introduces AutoRec , a sender‑side loss‑recovery mechanism for large‑scale live video streams. AutoRec is implemented on the QUIC protocol (open‑source tquic library) and has been deployed in Tencent Cloud EdgeOne CDN.

Motivation

Live video traffic dominates the Internet, but packet loss causes stalls that degrade Quality of Experience (QoE). Existing Automatic Repeat reQuest (ARQ) solutions require coordinated changes on both server and client, which is infeasible in multi‑vendor CDN environments where only the server can be modified. Measurements on ~50 million streaming sessions show frequent on‑off transmission mode switches, making conventional ARQ insufficient.

Technical Challenges

Improve loss tolerance without requiring client‑side modifications.

Limit additional bandwidth overhead while accelerating recovery.

Adapt to spatial‑temporal variations in loss patterns and network conditions.

Core Contributions

Definition of a comprehensive loss‑recovery quality metric (Recovery Deterioration Rate, RDR).

Large‑scale measurement exposing the inadequacy of existing ARQ in live streaming.

Design of AutoRec, which injects a small number of redundant packet copies during off‑states, requiring only sender‑side changes.

Online‑learning based redundancy adaptation that dynamically selects the number of copies per loss event.

Prototype implementation on QUIC and evaluation in both test‑bed and production CDN deployments.

Design Details

4.1 Key Idea

AutoRec treats the off‑state of the on‑off transmission pattern as an opportunity to inject “few but enough” redundant copies of lost packets. When a loss is detected, the sender schedules these copies without any client modification.

4.2 Redundancy Adaptation

The system defines:

Redundancy level : number of redundant copies to inject for a specific lost packet.

Injection cost : total number of injected copies across the session.

An online‑learning adapter observes QoS signals (e.g., packet loss rate, RTT, stall frequency) and updates the redundancy level to balance recovery speed against overhead. The adapter uses a simple reinforcement‑learning rule: if recent stalls exceed a threshold, increase the redundancy level; otherwise, decay it.

4.3 Injection Control

After the redundancy level is determined, an injection controller schedules the copies:

Prefer transmission during off‑states to avoid competing with regular traffic.

opportunistically inject during on‑states when the estimated bandwidth headroom is sufficient.

Enforce a maximum injection rate (e.g., 5 % of total bandwidth) to keep competition safe.

This control mitigates uneven on‑off distribution and prevents bandwidth starvation for normal packets.

Evaluation

Metric: Recovery Deterioration Rate (RDR)

RDR = (loss volume that requires ≥2 time units to recover) / (total loss volume). A lower RDR indicates faster recovery.

Test‑bed Results

Average stall count per 100 s reduced by 11.4 %.

Average stall duration per 100 s reduced by 5.2 %.

Improvements are stable across bitrates, buffer sizes, and network conditions; higher loss rates and RTTs increase the difficulty but still yield gains.

Real‑world CDN Deployment

Utility (a composite QoE score) improved by 6.3 % on average, with 13.4 % improvement for the top 80 % of sessions.

Throughput degradation limited to ~5.1 %.

Retransmission rate increased by only 3.6 %.

For streams with large SRTT and loss rates, stall frequency reductions of 24.4 %–34.1 % (90th–95th percentiles) and stall duration reductions up to 16 % (≈80 ms) were observed.

Implementation

AutoRec is built on the user‑space QUIC library tquic (GitHub: https://github.com/Tencent/tquic). The prototype modifies only the sender side of the QUIC stack to add the redundancy adapter and injection controller. The code is open‑source and can be cloned via standard Git commands.

References

Paper DOI: https://dl.acm.org/doi/10.1145/3664647.3681423

Open‑source QUIC implementation: https://github.com/Tencent/tquic

Figures