Abstract
With the rapid development of wireless networking and mobile devices, crowdsourced content sharing has shifted toward shorter multimedia formats. Twitter's Vine service pioneered this trend, enabling ultra-short video clips for instant sharing. This paper presents an initial study on modeling and optimizing these instant video clip services. We investigate unique mobile data access patterns and propose a scheduling policy designed to maximize user experience while minimizing monetary and energy costs. Our solution, involving pre-fetching and watch-time download scheduling, is validated through extensive trace-driven simulations.
The Challenge of "Instant" Video
Traditional video platforms like YouTube involve long-form content where a few seconds of startup delay is annoying but tolerable. However, the rise of "instant" video clips (6-second loops) changed the game. On these platforms, users engage in Batch Viewing and Passive Viewing, rapidly scrolling through a feed where the video itself is often shorter than a typical buffering period.
The challenge is three-fold:
- Zero Tolerance for Delay: If a 6-second video buffers for 3 seconds, the user experience is ruined.
- Resource Constraints: Mobile users have limited data plans and battery life.
- Unpredictability: Traditional "next-one" caching fails when a user "flings" the screen, skipping ten videos in a second.
Methodology: Physics-Based Scheduling
The general scheduling problem is NP-complete, leading us to propose a two-part practical solution:
1. Popularity-Aware Pre-fetching
When a device is on WiFi, the system proactively downloads a fraction of the most popular videos in the user's feed. This reduces the "start-up" burden for the most likely-to-be-watched clips. Our analysis of Vine data revealed that popularity is extremely skewed; the top 5% of videos account for over 99% of total reposts, making popularity a highly reliable signal for pre-fetching.
2. Gesture-Based Watch-time Prediction
This is the paper's most novel contribution. By modeling the physics of screen scrolling (friction and deceleration) within the mobile OS, the system can predict user behavior. By capturing the initial speed of a user's "fling" or "drag" gesture, the system calculates exactly how many videos will pass through the viewport and for how long each will remain visible.
3. Dynamic Scheduling
The system uses these physics-based predictions to prioritize downloads. If the model predicts a video will only be on screen for 0.5 seconds before being scrolled past, it avoids wasting bandwidth on a full download.
Key Results
The proposed WT+PF (Watch-Time + Pre-Fetching) approach was compared against "Sequential" and "Next-one" strategies using a custom Vine crawler and user gesture traces from volunteers.
- Stability: WT+PF maintained low playback discontinuity even under low bandwidth (< 0.5 MB/s), where naive methods failed.
- Efficiency: The system saved at least 40% in monetary costs and 30% in energy consumption. Under high-bandwidth conditions, savings reached over 90% by avoiding "over-downloading" skipped content.
- Optimal Aggressiveness: We identified that pre-fetching roughly 20% of a video's length provides the best balance between battery life and smooth playback.
Conclusion and Future Work
This study demonstrates that UI physics can serve as a deterministic predictor of network needs. While the current model is robust, future work could integrate mobile sensors (accelerometers, GPS) to predict "batch viewing" sessions or utilize machine learning to personalize content predictions beyond global popularity metrics. As platforms evolve toward longer clips, hybrid approaches combining these heuristics with traditional streaming methods will be essential.
Summary updated