Optimizing the Quality of Scalable Video Streams in P2P Networks

1. Bitrate Epoch 1 Epoch 5 Video bitrate Time Bandwidth Time Buffering Optimizing the Quality of Scalable Video Streams in P2P Networks Raj Kumar Rajendran Dan Rubenstein Dept of Electrical Engineering Columbia University Method The Problem • Discretized Model: • Video time-sliced fixed-bytesize epochs • How do we use available bandwidth chunks of current epoch? • Novel Prefetching Approach: • We identify an off-line algorithm when future bandwidth availability is known • We prove optimality: • Minimizes waste and variability • Maximizes smoothness • In practice: future bandwidth availability not known • We develop on-line algorithms that allocates current bandwidth to download current or future parts of the video • Challenge: what should available bandwidth be used to prefetch? • Downloading all layers of current portion can create unpleasant variation in video quality as bandwidth availability changes • Prefetching one layer at a time unnecessarily forces viewing of initial portion of video at lowest quality • Video streaming in P2P systems is a potential killer app • Bandwidth availability to a client fluctuates unpredictably and rapidly with time • Solution: use scalable coding (FGS) to break up video into M layers • Viewing more layers = higher fidelity viewing Solution Bandwidth from a single Epoch • Naïve Solutions • Same-Index (Greedy) • Allocate all currently available bandwidth to nearest future epoch • Problem: Large variation in the quality of video displayed • Smallest-Bin (Conservative) • Allocate all current bandwidths to future epoch with fewest layers • Problem: Wastes bandwidth Largest Hill: Allocate to earliest epoch maintaining slope • Our Solution: Constrained Allocators • Attempt to maximize utilization, given smoothness constraints • Bound the downhill slope of allocations • Three variants Approach Mean Hill: Most empty epoch smaller than mean, maintaining slope • Formulate measures of video quality • Waste: the amount of unused available bandwidth • Smoothness: the average change in video quality • Variability: the standard deviation • How do we maximize video quality time-varying bandwidth? Wide Hill: Earliest epoch smaller than mean maintaining slope Video Quality of Constrained Allocators Experimental Verification • Bandwidth Trace Experiments: • Experiment • Input: bandwidth traces obtained while downloading video from a P2P network • Tested on DSL and T1 • Video downloaded from multiple peers • Waste, smoothness, Variabilty measured with increasing epoch lengths • Results • Mean-hill and wide hill allocators perform close to the bound • Largest hill performs a little worse • Naïve Allocators perform poorly • Simulation Results • Experiment • Input bandwidth simulated • Increasing variance, constant mean • Waste,smoothness, variability measured • Results • Constrained Allocators vastly outperform naïve allocators and are close to the bound • The naïve allocators perform well on one of the measures but poorly overall Impact • Viewing scalably encoded videos in P2P systems without smart prefecthing strategies yields a poor viewing experience • We provide an off-line algorithm that provides the optimal performance given bandwidth constraints • We provide on-line algorithms that perform close to the optimum and vastly outperform naïve algorithms