Spatio-Temporal Analysis of Bandwidth Maps for Geo-Predictive Video Streaming in Mobile Environments
DASH Background Server provides multiple qualities of the same video Client switch between video qualities based on available bandwidth B MPD Delivery HTTP client HTTP
Motivation Video streaming from mobile devices while commuting experience bandwidth fluctuations due to: Change of location with respect to cell tower (spatial aspect) Change in surrounding objects like buildings and tree (spatial aspect) Change in crowd levels (spatial and temporal aspect) Results in rapid change between video representations and frequent buffer stalls (affects overall QoE)
Research Question Can we predict future bandwidth when a user travels along a route ? Can we model bandwidth as function of location and time ? can guarantee the real-time performance for RT VMs, while allowing regular VMs to effectively utilizing the remaining CPU resources.
Related Work Explores relationship between bandwidth and location change but not time change (J. Hao et al) Proposed method explores difference in bandwidth at same location in morning, afternoon, evening Predicts bandwidth for users commuting along known path (Riiser et al.) Proposed method does not assume fixed route
Contributions Explore bandwidth changes with changing both location and time Predict future bandwidth values at unknown locations with better accuracy that previous work
Spatial Analysis Measure bandwidth at different locations and times across Singapore Draw variogram plot (relation between distance and semivariance) s1 s2 s3 10m 15m 15m s4 At distance 10m Semivariance = [B(s1) – B(s2)]^2 At distance 15m Semivariance = [B(s2) – B(s3)]2 + [B(s4) – B(s3)]2
Spatio-Temporal Analysis Dataset: measured along a major bus routes at three times of day (morning, after noon evening) 3D Variogram Plot: X-axis distance , y-axis time Dark blue indicates a strong correlation, yellow indicates weak correlation The further we deviate in time, measurements become less correlated The highest correlation is as expected on the bottom left (close proximity close time of day) s0 s1(t1) s2(t2) s3(t3) s4(t4)
Bandwidth Prediction Method Find bandwidth of s0 using Kriging interpolation: s0 s1(t1) s2(t2) s3(t3) s4(t4) Learned from pre-collected samples by minimizing the least square error between the predicted and measured bandwidth
Rate Stabilization Use ExoPlayer's (DASH-compliant Android media player) built-in stabilization Exoplayer takes predicted bandwidth as an input It outputs effective bandwidth which is a weighted average of current bandwidth and previous predictions Advantages: Smooth out abrupt changes Disadvantages: Slow to adapt in highly fluctuating networks
Evaluation Comparison between bandwidth prediction approaches: Kriging interpolation: Weights of neighboring samples are learned from the data KNN-IDW (Inverse distance weighting) Weights of K neighboring samples depends on the inverse of distance from the predicted sample Mean of neighboring samples s0 s1 s2 s3 10m 15m 15m s4
Kriging vs KNN-IDW Over estimated bandwidth Prediction with KNN-IDW Prediction with Kriging Interpolation (proposed) Prediction with KNN-IDW Abrupt and Frequent Quality changes
Comparison Between Three approaches Improvement of Kriging is more obvious in average buffer delay
Open Issues The size of the bandwidth map will grow dramatically if we keep all measurements Need Eviction policies The paper does not address which cell tower associated with moving device and how cell tower changes affect the bandwidth The paper does not take in account concurrent streaming users
Thoughts Pros Cons The paper shows interesting insight about the relationship between bandwidth and time (rush-hour could be more busy than midnight) The paper introduces Variogram plot to the community as a way to express correlation at time and space Kriging interpolation is more flexible method for predicting bandwidth and have better accuracy The paper does not compare between spatio-temporal vs spatial bandwidth prediction (what is the effect of adding the temporal aspect) The paper does not compare with methods that assume fixed route Results are only marginally better than previous work Lack of description of important parts of the paper ( How Kriging interpolation estimate weights)
Cons There is not even a brief description about the significant of the results in this table
Appendix – Kriging Interpolation Kriging Prediction formula: Goal: Minimize mean squared Gradient descent: [ - ]2