How AutoRec Cuts Live‑Streaming Stalls by Over 11% with Smarter Packet Recovery

Research Overview

Tencent and the Ministry of Education Key Laboratory of Data Engineering and Knowledge Engineering at Renmin University co‑authored a paper titled “Toward Timeliness‑Enhanced Loss Recovery for Large‑Scale Live Streaming,” which was accepted as an oral presentation at ACM Multimedia 2024 (acceptance rate 3.97%). The proposed AutoRec mechanism has been deployed in Tencent Cloud EdgeOne, achieving an 11.4% reduction in average stall occurrences and a 5.2% reduction in stall duration.

Motivation

Live video streaming services rely heavily on packet‑loss tolerance to ensure quality of experience (QoE). Existing CDN providers can only modify server‑side strategies, while client‑side controls remain untouched, making large‑scale deployment of traditional ARQ‑based solutions difficult. Large‑scale measurements (up to 50 million streams) revealed frequent on‑off transmission mode switches that degrade recovery latency.

Key Challenges

Improving loss‑recovery tolerance without modifying both server and client.

Limiting additional overhead while enhancing tolerance.

Adapting to dynamic network conditions across regions and time.

Core Contributions

The study defines evaluation metrics for packet‑loss recovery quality and conducts extensive measurements on 50 million live‑stream packets, confirming that current recovery mechanisms cannot meet application demands. AutoRec’s core ideas are:

Turning on‑off mode disadvantages into advantages by injecting a small, sufficient number of redundant copies only when the sender detects loss.

Redundancy adaptation using an online‑learning adapter that dynamically selects the optimal number of redundant copies based on observed loss patterns.

Injection control that schedules the transmission of redundant copies to avoid bandwidth competition with non‑lost traffic, even during on‑state periods.

AutoRec is implemented on a user‑space QUIC stack and evaluated on both a test platform and real CDN edge servers.

Experimental Results

In controlled experiments, AutoRec consistently reduces recovery latency across various network conditions, with minimal impact from video bitrate or bottleneck buffer variations. In real‑world deployments, AutoRec lowers average stall frequency and duration by 11.4% and 5.2% respectively, with higher gains (up to 24.4% and 34.1% at the 90th/95th percentiles) for streams with large RTT and loss rates.

Utility optimization shows an average gain of 6.3% (80% of cases achieving 13.4% gain), while incurring only about 5.1% effective throughput degradation and a 3.6% increase in retransmission rate.

Architecture Diagram

Conclusion

AutoRec demonstrates that intelligent, sender‑side redundancy can transform the on‑off transmission pattern from a weakness into a strength, accelerating packet recovery without harming ongoing traffic. Deployed on Tencent’s CDN and EdgeOne, it now serves millions of live‑stream users worldwide, though further research is needed for more complex network environments.