Baidu Live Streaming Startup Performance Optimization: Technical Practices and Solutions

This article provides a comprehensive analysis of Baidu's live streaming startup optimization efforts, covering technical challenges and solutions across multiple dimensions.

Background and Goals: Baidu's live streaming aims to replicate real-world experiences online and create beyond-reality experiences using 5G, VR, and AI technologies. The primary focus is on Quality of Experience (QOE), with startup latency being the first user perception metric targeted for optimization.

Current Status: Baidu's live streaming分为泛服务直播 (media, consulting, e-commerce) and 泛娱乐直播 (show, audio, voice room). The 泛服务直播 has complex characteristics: intricate processes, multiple states (live, playback, generating), involving multiple teams (streaming, player, kernel, network, CDN), requiring changes while the "car is running."

Data Analysis: The startup process is divided into three phases: business logic time, player time, and kernel time. Over 60 tracking points analyze each step's duration, with business logic consuming over 60% of total time.

Business Scenario Optimization: Two jump scenarios exist: external scheme jump and immersive in-app jump. Solution A (iPhone 8+): create player and start playback during slide. Solution B (iPhone 8 and below): create player during slide, prepare resources, destroy previous room and start new playback after slide stops. For external jumps, the scheme carries the streaming URL directly for parallel player creation and playback.

DNS Pre-resolution: Using HTTPDNS to prevent hijacking and reduce latency. Strategy triggers during cold start (10s), network switching, and foreground/background transitions. The model considers user viewing history, network status, and backend controls. IP validity defaults to 300s. Result: 90%+ of streaming DNS queries complete in under 3ms.

Kernel Optimizations: 1) Forced first-frame rendering bypassing audio-video sync; 2) Low-bitrate startup for weak networks; 3) Pre-loading next player kernel on high-end devices with frame-seeking (dropping 200ms audio every 2 seconds when buffer exceeds 3 seconds).

Media Information Parsing: For containers without video dimensions (like FLV), parsing H.264 NAL units (SPS/PPS) directly obtains width, height, and encoding format, avoiding software decoding.

HLS m3u8 Pre-fetching: For live playback (HLS format), pre-fetching m3u8 index files to local storage eliminates download time. A/B test shows 346ms improvement for cached m3u8 hits.