An Efficient Implementation of Interactive Video-on-Demand Steven Carter and Darrell Long University of California, Santa Cruz Jehan-François Pâris University of Houston
Why Video-on-Demand? Increased customer convenience Few people enjoy returning video tapes Even fewer people enjoy paying late fees Improved selection of videos Current pay-per-view provides only a small selection of popular videos Savings in time and resources It takes time and fuel to drive to the video rental store
Why Now? Technology becoming available Processors are inexpensive Storage is nearly free ($200 for 40GB) Fast networking is seeing wide deployment Consider the success of Tivo Records live television using MPEG to disk Provides interactive access to recorded programs
Why Interactive? It s hard! It s more expensive! but it s what people expect They won t give up functionality they have come to expect They d like to pause to make microwave popcorn They d like to rewind to see the play again They d like to be able to fast forward past the boring parts
Related Research Conventional video-on-demand (VoD) Requires one stream per client Patching An independently developed version of stream tapping Batching Group the requests of several clients together Various near video-on-demand (NVoD) schemes
Key Observation For videos of non-trivial length, several clients will be viewing portions of that video One client watching a 120 minute video and a second client begins watching the same video 10 minutes later The server needs only send data for the non-overlapping portion The potential for savings is enormous
Assumptions A set-top-box with: A fast network connection A few gigabytes of local storage A modest processor Keep in mind that set top boxes with these features already exist
Our Solution Stream Tapping uses multicast to tap in existing video streams Server load is the primary difficult in making VOD a reality Stream Tapping reduces server load by allowing clients to tap into video streams created for other clients Cost per client is dramatically reduced Client waiting time is also reduced
Stream Types c c C Partial tap b b β Stream B Complete stream Full tap A 0 β 2β 3β 4β Time (since start of complete stream A)
Complete Streams Start at a particular position in a video and transmit the remainder of the video For non-interactive Stream Tapping, the starting position is the beginning of the video Used primarily by the first client in a group to view the video
Stream Types c c C Partial tap b b β Stream B Complete stream Full tap A 0 β 2β 3β 4β Time (since start of complete stream A)
Full Tap Streams Can be used if the delay ( ) is less than the buffer size (β) The full tap stream transmits the video from time 0 to The complete stream is tapped and written to the buffer while the full tap stream is played
Stream Types c c C Partial tap b b β Stream B Complete stream Full tap A 0 β 2β 3β 4β Time (since start of complete stream A)
Partial Tap Streams Can be used when a complete stream is available but >β Note that given current technology, β will be very large The client will tap the complete stream for β units while simultaneously viewing the first β from a partial tap stream Subsequently, partial tap streams of length β are used for the client to catch up to the complete stream
Stream Types c c C Partial tap b b β Stream B Complete stream Full tap A 0 β 2β 3β 4β Time (since start of complete stream A)
Tapping Options Extra Tapping Allows the client to tap data from any active video stream active, not just the complete stream of the video group Decreases server load by decreasing the length of full tap streams Stream Stacking If the server has streams available, the client can combine them to receive data at rate higher than the nominal rate Allows the server to service stream more quickly, which allows new streams to be scheduled
Interactive Stream Tapping When an interaction begins, Stream Tapping deallocates resources associated with a client If the client was the only one using a stream, then the stream is terminated Stream Tapping determines the resources needed for an interaction, and allocates an interaction stream Note: for rewind, the client s buffer can be used When the interaction is complete, the client is merged into a video group (tapping existing streams if available)
Contingency Streams These are streams that are held in reserve for interaction The pool of these streams can be managed using high and low watermarks for hysteresis Having such a reserve of streams is essential to avoid blocking
Simulation Model Stream Tapping is too complex to model analytically, so we used discrete event simulation The length of the videos was derived from empirical data and a gaussian with mean 102 minutes provided the best fit The popularity of videos was modeled using a Zipf-like distribution, which is the distribution used in most VoD studies
Comparison with Conventional Systems 700 600 Conventional System Staggered Broadcasting (2 min) 500 Staggered Broadcasting (5 min) Stream Tapping Bandwidth 400 300 200 100 0 0 50 100 150 200 250 300 350 400 Arrival Rate
Contingency Streams versus Start-up Latency 45 Average Start-up Latency (min) 40 35 30 25 20 15 10 λ=250 λ=300 λ=350 λ=400 5 0 0 10 20 30 40 50 60 70 80 90 100 Contingency Streams
Contingency Streams versus Resume Latency 1.8 Average Start-up Latency (min) 1.6 1.4 1.2 1 0.8 0.6 0.4 λ=250 λ=300 λ=350 λ=400 λ=450 0.2 0 0 10 20 30 40 50 60 70 80 90 100 Contingency Streams
Contingency Streams versus Blocking Probability 80 Interaction Blocking Probability (%) 70 60 50 40 30 20 λ=250 λ=300 λ=350 λ=400 10 0 0 10 20 30 40 50 60 70 80 Contingency Streams
Effect of Client Buffer Size 140 AcerageStart-up Latency (min) 120 100 80 60 40 λ = 250 λ = 300 λ = 350 λ = 400 λ = 500 20 0 0 10 20 30 40 50 60 STB Buffer Size (min)
Effect of Tapping Options 80 Average Start-up Latency (min) 70 60 50 40 30 20 Neither Stream Stacking Only Extra Tapping Both 10 0 0 10 20 30 40 50 60 STB Buffer Size (min)
Conclusions Stream Tapping has been shown to work well in the interactive environment We have shown that VCR-like controls are possible Previous work has ignored them or only provided course-grained control The use of storage in the STB is an enabling technology
Video Length Distribution