Distributing Layered Encoded Video through Caches J Kangasharju
Distributing Layered Encoded Video through Caches J. Kangasharju, F. Hartanto, M. Ressiein and Keith W. Ross Presented by Felix Lam
Introduction n Web caching/Proxy ¡ n Efficient distribution of web objects Challenges of streaming media caching ¡ ¡ Require much larger storage Consume much higher bandwidth for a much longer time
Introduction n Problem ¡ Which layer(s) of which movie(s) should be cached to maximize the revenue when both the cache size and the network bandwidth is limited.
System Model n Assumptions ¡ ¡ ¡ Only complete layers of video objects are cached Cache access bandwidth (e. g. disk access speed) is not a bottleneck Clients would not choose more layers than its access bandwidth can afford
System Model n Notations M: # of movies L: # of layers per movie T(m): Movie length (in sec. ) of movie m rl(m): Effective bandwidth of layer l of movie m p(j, m): Popularity of the jquality (quality of receiving layers 1 to j) of movie m R(j, m): Revenue of serving j-quality of movie m cm : # of layers of movie m cached G: Total cache storage C: Total bottleneck link bandwidth
System Model n Procedures for handling a request ¡ ¡ A client requests a j-quality stream of video m. If cm >=j , all requested layers are streamed from the proxy. If cm <j , the proxy attempts to connect to the origin server for streaming the missing cm+1 to j layers. If NOT sufficient available bandwidth, the request is consider to be blocked.
Optimal Caching n n Let λbe the Poisson arrival rate of request The expected total rate of revenue is: Bc(j, m) is the blocking probability of the request for a j-quality stream of video m, given the cache configuration c = (c 1, …, c. M) Bc(j, m) is calculated by Stochastic Knapsack model
Optimal Caching n Optimization ¡ ¡ ¡ maxc. R(c) subject to S(c)<=G Analytically intractable Exhaustive search space is too large for large M and L
Utility Heuristics n n Assign a utility (ul, m) to each of the ML layer objects Cache the layer objects in decreasing order or ul, m until no more layer can be put into the cache
Performance Evaluation n Default System Settings ¡ ¡ ¡ M = 1000 L=2 Rate of each layer ~ U(0. 1 Mbps, 3 Mbps) T(m) : Exponentially distributed with mean equals 3600 seconds The requested video and quality us Zipf distributed with ζ= 1. 0 λ= 10. 8 requests / second
Performance Evaluation n Comparison of Heuristics to exhaustive search in small problem
Performance Evaluation n Comparison of different utility functions
Performance Evaluation n Impact of cache size and link capacity
Performance Evaluation n Effect of Zipf-parameter
Performance Evaluation n Revenue gain by renegotiation of stream quality
Performance Evaluation n Revenue gain from queuing requests
Conclusion n Formulated an analytical stochastic model for the layered video caching problem. Effective Utility Heuristics Shows little gain from renegotiation and queuing of requests
Comment n n n Lack of comparison with some traditional caching schemes used in web caching The cost difference of link capacity and storage is not accounted Using revenue per admitted streaming request in determining the cache allocation is very sensible and practical
- Slides: 18