# Modeling Live Video Streaming: Real-Time Classification, QoE Inference, and Field Evaluation

**Authors:** Sharat Chandra Madanapalli, Alex Mathai, Hassan Habibi Gharakheili, Vijay Sivaraman  
**arXiv:** [2112.02637](https://arxiv.org/abs/2112.02637)  
**Date:** 2021-12-05  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2112.02637v1/

## Key points

- Live video streaming on platforms like Twitch and YouTube Live has grown rapidly, driven by social media, professional sports, and video games
- Network congestion can significantly impact the quality of experience (QoE) for users watching live streams since client playback buffers are often no more than a few seconds
- Identifying live streams and measuring their QoE is challenging since content providers use the same delivery infrastructure for both live and video-on-demand (VoD) streaming
- Researchers designed, built, and deployed ReCLive: a machine learning method for live video detection and QoE measurement based on network-level behavioral characteristics
- The team analyzed approximately 23,000 video streams from Twitch and YouTube to identify key features in their traffic profile that differentiate live and on-demand streaming
- They developed an LSTM-based binary classifier model that distinguishes live from on-demand streams in real-time with over 95% accuracy across providers
- They also developed a method that estimates QoE metrics of live streaming flows in terms of resolution and buffer stall events with overall accuracies of 93% and 90%, respectively
- The team prototyped their solution, trained it in the lab, and deployed it in a live ISP network serving more than 7,000 subscribers
- Their method provides ISPs with fine-grained visibility into live video streams enabling them to measure and improve user experience without introducing any risk to the operational network.
- The system architecture consists of upstream and downstream traffic received on separate optical tap links processed by a Linux server running Ubuntu 18.04 with DPDK support for high-speed packet processing.
- The tool FlowFetch interacts with DPDK to fetch raw packets executing telemetry functions to export request packet counters and chunk features.
- -The flow of events in the system involves detecting flows carrying video streams originating from Twitch and YouTube by performing pattern matches on the SNI field present in the TLS handshake.

## Summary

The rapid growth of live video streaming on platforms such as Twitch and YouTube Live, driven by social media, professional sports, and video games, has created challenges for network management. Short-time-scale network congestion can significantly impact the quality of experience (QoE) for users watching live streams since client playback buffers are often no more than a few seconds. However, identifying live streams and measuring their QoE is challenging since content providers use the same delivery infrastructure for both live and video-on-demand (VoD) streaming. Packet inspection techniques such as SNI/DNS query monitoring cannot always distinguish between the two. To address this challenge, a team of researchers designed, built, and deployed ReCLive: a machine learning method for live video detection and QoE measurement based on network-level behavioral characteristics. The team analyzed approximately 23,000 video streams from Twitch and YouTube to identify key features in their traffic profile that differentiate live and on-demand streaming. They released their traffic traces as open data to the public. The researchers developed an LSTM-based binary classifier model that distinguishes live from on-demand streams in real-time with over 95% accuracy across providers. They also developed a method that estimates QoE metrics of live streaming flows in terms of resolution and buffer stall events with overall accuracies of 93% and 90%, respectively. Finally, the team prototyped their solution, trained it in the lab, and deployed it in a live ISP network serving more than 7,000 subscribers. Their method provides ISPs with fine-grained visibility into live video streams enabling them to measure and improve user experience without introducing any risk to the operational network. The system architecture consists of upstream and downstream traffic received on separate optical tap links processed by a Linux server running Ubuntu 18.04 with DPDK support for high-speed packet processing. The tool FlowFetch interacts with DPDK to fetch raw packets executing telemetry functions to export request packet counters and chunk features. The flow of events in the system involves detecting flows carrying video streams originating from Twitch and YouTube by performing pattern matches on the SNI field present in the TLS handshake.