Optimizing Playback Experience and Reducing Bandwidth Cost in Vivo Short Video Platform

The article summarizes the talks by Cai Chuangye and Ma Yunjie at the 2023 Vivo Developer Conference, focusing on how Vivo short video improves playback experience while reducing bandwidth costs.

Business Overview

Vivo short video’s workflow includes content production (shooting, import, editing, upload), video processing (enhancement, transcoding, compression), distribution, and consumption (pre‑load, playback). Supporting subsystems such as log collection, monitoring, and A/B testing are also described.

Experience Optimization

Key optimization directions identified are connection optimization, chunked download, pre‑rendering, and data pre‑load.

Connection optimization: reuse connections, keep‑alive, HTTP DNS, and server‑side IP fallback to reduce DNS/SSL/TCP latency.

Chunked download: a local proxy service between the player and CDN enables single‑threaded start‑up and multi‑threaded subsequent downloads, reducing first‑frame latency by 3.8% and failure rate by 9%.

Pre‑rendering: pre‑create H.264/H.265 codec instances and use a global player resource pool to cut start‑up time by ~50 ms and avoid OOM/ANR issues.

Pre‑load strategies: fixed‑size pre‑load (queue next 5 videos) and dynamic pre‑load (multi‑level cache, size based on video length and user scroll speed). Dynamic pre‑load lowered first‑frame latency by 2.3% and stutter rate by 19.5%.

Metrics are defined at multiple levels: P0 (first‑frame latency, failure rate, stutter rate) and P1 (cache size, hit rate, download speed). A hierarchical monitoring system (P0 → P1 → strategy metrics → video basics) is deployed.

Cost Optimization

Cost is dominated by CDN (≈80%). The cost breakdown includes price, traffic volume, bitrate, and traffic utilization rate. Three main cost‑reduction directions are price reduction, extreme compression, and utilization governance.

Price reduction: introduce PCDN (low‑cost edge devices) with intelligent buffering and 1‑byte probe packets to minimize 302 redirects; share idle bandwidth with other services to smooth peak usage.

Extreme compression: a self‑developed neural‑network‑based adaptive encoding pipeline (pre‑enhancement, per‑title scene segmentation, feature extraction, model‑driven bitrate selection) reduces average bitrate from 60% to 40% while preserving visual quality.

Utilization governance: build a utilization‑funnel to monitor traffic waste across versions, conduct A/B experiments to tune pre‑load thresholds, and raise traffic utilization from ~60% to ~70%.

Summary & Outlook

The optimizations combine big‑data analysis, A/B testing, and AI techniques to achieve a win‑win between user experience and cost. Future work includes adopting H.266 for further bitrate reduction and exploring edge‑cloud collaborative enhancement.