Predictable Audio‑Video Networks: An In‑Depth Review of Alibaba Cloud’s GSO‑Simulcast, LiveNet, and Zhuge Systems

In the mobile‑Internet era, audio‑video applications such as live shopping, remote meetings, short videos, and cloud gaming have reshaped daily life, and their success depends on robust audio‑video network transmission technology. At the SIGCOMM conference, Alibaba Cloud published three papers describing how they construct a “predictable audio‑video network” to support core services like Taobao, DingTalk, and cloud gaming.

The rapid evolution of video consumption—from 360p to 4K (a 40× bandwidth increase) and from several‑second latency to sub‑100 ms end‑to‑end latency—has exposed a large gap between advertised peak network parameters and real‑world performance, especially on mobile links where bandwidth can drop dramatically.

GSO‑Simulcast addresses the bottleneck effect in large‑scale video conferences. By encoding a single video source into multiple bitrate streams and using a global optimizer (the GSO Controller) that considers all users’ uplink/downlink bandwidth, subscription relationships, and codec capabilities, the system can automatically select the appropriate number of streams (up to 15 fine‑grained levels) and their bitrates, dramatically improving video‑stutter, audio‑stutter, and frame‑rate metrics while increasing user satisfaction by 7 % after deployment in DingTalk.

LiveNet is a low‑latency CDN overlay network for live streaming. Unlike the earlier hierarchical HIER architecture, LiveNet adopts a flat overlay where each node can act as a producer, consumer, or relay, and a centralized controller computes the optimal overlay path based on global network state. This design reduces path latency from ~393 ms to 188 ms, cuts hop count from four to two, and improves startup time, stall rate, and overall QoE for massive live‑stream audiences.

Zhuge targets low‑latency video over Wi‑Fi for gaming and other interactive scenarios. It decouples congestion feedback from the downstream queue by embedding early‑arrival ACK timing information into packets preceding the congested ones, enabling the sender to react faster to rising queue delay. Experiments show a reduction of long‑tail round‑trip time by 17 %–95 %.

Collectively, GSO‑Simulcast, LiveNet, and Zhuge demonstrate how predictable, fine‑grained network control can sustain high‑quality audio‑video experiences across conferencing, live streaming, and gaming, paving the way for future metaverse‑scale video services.