Mobile Video Delivery via Human Movement
0 likes490 views
This document proposes a mobile video delivery system called VideoFountain that leverages human movement between locations to distribute video content. The system would deploy WiFi access points with storage at popular venues. Analysis of location-based social network data shows that venue popularity and human traffic between venues follow Zipf-like distributions, and most venue pairs are within 2-3 hops of each other, indicating feasibility. The document outlines algorithms for initial content placement, routing content between venues, and evaluates performance based on trace-driven simulations.
Mobile Video Delivery via Human Movement
- 1. Mobile Video Delivery via Human Movement Gene Moo Lee, Swati Rallapalli, Wei Dong, Yi-Chao Chen, Lili Qiu, Yin Zhang University of Texas at Austin SECON 2013, New Orleans LA
- 2. Introduction ● Mobile devices have great local connectivities ○ WiFi, NFC, Bluetooth ● People bring mobile devices wherever they go ○ Home, workplace, restaurants, subways, bus stops Can we leverage human movements as a content distribution channel? 2/21 0.65 Mbps (in our LBSN data)
- 3. VideoFountain ● VideoFountain is a mobile content distribution system ○ To deploy kiosks (WiFi AP w/ storages) at popular venues ○ To let mobile users download/upload contents ○ To leverage mobile users to deliver contents between kiosks ● Challenges ○ To understand real user mobility ○ To bootstrap mobile contents ○ To route contents ○ To incentivize users ○ To protect copyrights ○ To preserve integrity of contents ○ so on... source: http://www.redbox.com 3/21
- 4. Roadmap 1. User mobility analysis a. Location-based social networks b. Various mobility properties 2. Content distribution via human movement a. Initial content placement b. Routing in VideoFountain 3. Feasibility study with trace-driven simulations a. Impacts of algorithms b. Impacts of various factors 4/21
- 5. Mobility analysis with LBSN ● LBSN: Location-Based Social Networks ○ Mobile users check-in to places where they visit ○ Foursquare, Gowalla, Facebook, Google+ ● Why LBSN? ○ Mobile devices are the targets of VideoFountain ○ Dwell time is long enough to do manual check-in ○ Massive data to understand human movement ● Limitations ○ No check-out time, missing check-ins 5/21
- 6. Venue popularity ● Popularity of venues (# of check-ins) follows Zipf-like distribution ○ Significant value of placing content at popular venues ● Venue popularity is stable over time ○ We can learn from historical data 6/21 Number of check-ins
- 7. Inter-venue human traffic ● Aggregated human traffic between venues ○ T(a,b,t) = # users moving from venue a to venue b in day t ■ user u check-ins at a then b ● Human traffic is stable over time ○ In two week data, T(a,b,t) doesn't change for 80-90% pairs ● Aggregated human traffic exhibits Zipf-like distribution 7/21 Number of human traffic
- 8. Degree of separation ● How well connected are the venues in a city? ○ 80-90% venue pairs are within 2-3 hops ○ Routing is relatively easy ● Methodology 8/21 Number of hops Fractionofvenuepairs
- 9. Individual mobility prediction ● Given a user is at venue X, can we predict the next check-in venue? ○ Somewhat. ○ Paris (55%), Manhattan (41%), Austin (26%) ● Use a very simple algorithm to predict user's next-checkin based on personal history plus movements from general population ○ See paper for algorithm details 9/21
- 10. Roadmap 1. User mobility analysis a. Location-based social networks b. Various mobility properties 2. Content distribution via human movement a. Initial content placement b. Routing in VideoFountain 3. Feasibility study with trace-driven simulations a. Impacts of algorithms b. Impacts of various factors 10/21
- 11. Initial content placement ● Given the set of destinations to deliver contents, where do we seed the contents? ● Placement algorithms ○ Popularity-based placement ■ Place the contents from the most popular venues ○ Utility-based placement ■ Place the contents from the venues that maximize our utility functions ● Ex. Minimize the distances to the destinations ■ Use greedy algorithm ○ Random placement 11/21
- 12. Routing algorithms How to route initial contents to destination venues? ● Utility-based replication ○ When a user visits a venue, calculate the marginal utilities of uploading (user->venue) or downloading (venue->user) contents ○ Execute download/upload with highest marginal utility ○ Marginal utility: an example with geographic distance ■ How much the content is getting closer to the final destination with this download/upload? 12/21
- 13. Utility functions Let Sc be the set of venues that currently have the content c Let dc be the destination venue of content c ● Expected delay ○ What is the expected delay from Sc to dc ? ● Geographic distance ○ How far is the Sc to dc ? ● Single-hop traffic ○ How many people move from Sc to dc in a day? ● Multi-hop traffic ○ Consider multi-hop traffic with decay factor 13/21
- 14. Benchmark algorithms ● Flooding ○ Whenever a user visits a venue, copy all the contents ● Epidemic routing ○ Flooding with hop count limit (2 in the experiments) ● Oracle routing (upperbound) ○ Assume that we know the users' next check-ins ○ Construct spatio-temporal graph ○ Run LP to optimize the throughput achievable in the graph 14/21
- 15. Roadmap 1. User mobility analysis a. Location-based social networks b. Various mobility properties 2. Content distribution via human movement a. Initial content placement b. Routing in VideoFountain 3. Feasibility study with trace-driven simulations a. Impacts of algorithms b. Impacts of various factors 15/21
- 16. Trace-driven simulation ● Foursquare traces ○ Train: January 2nd 2012 to January 16th 2012 (2 weeks) ○ Test: January 17th 2012 to January 30th 2012 (2 weeks) ● Default settings ○ Wireless capacity: 50 Mbps (802.11n) ○ Storage: 10 GB (mobile), 1 TB (venue) ○ Dwell time: exponential distribution with mean 60 mins ○ 50 flows generation (1GB each) ■ Sources: 1% or 5% venues (by placement algorithms) ■ Destinations: up to 10% random venues (non-source) ● Evaluation metrics ○ Flow delivery rate: how many flows completely delivered ○ Traffic delivery rate: consider partial delivery 16/21
- 17. Impact of initial placement ● Utility based placement outperforms others ○ Random 25%, popular 67%, utility 74% ● Utility based routing outperforms others ○ Different utility functions work for different cities ● Utility routing vs Oracle routing ○ Paris: 69% vs 81% ○ London: 69% vs 84% 17/21 Random placement Utility placement Flowdelivery(%) Flowdelivery(%)
- 18. Impact of contents ● Results from London ○ Consistent in other cities ● Number of flows ○ Increasing congestion ○ Utility routing degrades gracefully ○ Flooding/Epidemic suffer ● Content sizes ○ The system can support 500+ MB contents, which can cover most mobile contents 18/21
- 19. Impact of device specs ● Wireless capacity ○ The higher bandwidth, the better system performs ○ 50 Mbps is enough ● User storage ○ The larger storage, the better system performs ○ 10GB is enough ● Replacement strategy ○ Utility-based outperforms others 19/21
- 20. Impact of user behaviors ● User dwell time ○ Avg 30 mins is enough ● Comparing Oracle vs Prediction ○ Difference within 18%~38% 20/21
- 21. Conclusion ● Collected and analyzed LBSN data (Foursquare, Gowalla) ○ Venue popularity and human traffic exhibit Zipf-like distribution ○ Cities are well-connected by human movements ● Proposed VideoFountain ○ Mobile content distribution system leveraging human movements ○ Designed placement & routing algorithms ● Evaluated the system with trace-driven simulations ○ Utility-based placement & routing work well the best 21/21
- 22. Thank you!
- 23. Link capacity between venues ● Does inter-venue human movement link make considerable bandwidth? ○ Yes, it is. ● A simple back-of-the-envelope calculation: ○ Assume WiFi bandwidth L (50 Mbps) ○ If user u stays at venue a for ta then venue b for tb , ■ then the user can carry min(ta , tb ) X L / 2 ■ assuming she downloads and uploads equally ○ Latency T: inter-checkin time from a to b ● Avg capacity of all inter-venue movements = 0.65 Mbps ○ Avg global internet speed = 2.9 Mbps ○ Avg US internet speed = 7.4 Mbps ○ source : https://www.cabletechtalk.com/broadband-internet/u-s-moves-up-in-average-worldwide-internet-speed- rankings/ Back to main slides /21
- 24. Spatio-Temporal Graph ● Construct spatio-temporal graph ○ Two venues are linked if any user checked two venues in that day ○ optimistic vs conservative Back to main slides 7/21
- 25. Venue popularity ● Popularity of venues (# of checkins) follows Zipf-like distribution ○ Richer gets richer ○ Significant value of placing content at popular venues ● Venue popularity is stable over time ○ We can learn it from historical data ● Popular venues are spread across the city ○ pair-wise distances 6/21
- 26. Benchmark algorithms ● Flooding ○ Whenever a user visits a venue, copy all the contents ● Epidemic routing ○ Flooding with hop count limit (2 in the experiments) ● Oracle routing (upperbound) ○ Assume that we know the users' next checkins ○ Construct spatio-temporal graph ○ Run LP to optimize the throughput achievable in the graph 14/21
- 27. Impact of initial placement ● Seeds: 1% and 5% venues ● 1% seed results: ○ Random placement: 23% ○ Popular placement: 77% ○ Utility placement: 64% ● 5% seed results: ○ Random placement: 25% ○ Popular placement: 67% ○ Utility placement: 74% ● Utility routing vs Oracle routing ○ Paris: 69% vs 81% ○ London: 69% vs 84% 17/21
- 28. Introduction ● People bring mobile devices wherever they go ○ Home, workplace, restaurants, subways, bus stops ● Mobile devices have great local connectivities ○ WiFi, NFC, Bluetooth Can we leverage human movements as a content distribution channel? source: http://www.apple. com/ios/ios7/features/#airdrop source: http://www.androidcentral.com/hands-s-beam-samsung- galaxy-s-iii 2/21