Transcript PowerPoint - New York University

Multimedia
Robert Grimm
New York University
Before We Get Started…
 Digest access authentication
 What is the basic idea?
 What is the encoding?
 What is the role of the nonce?
 Measurements (again)
 Groups 1 and 5 each compared their server with the
other group’s server
Comparing Results
Groups 1 and 5
GET 4 files (1K, 10K, 100K, 1M)
Server: ???
Client: ???
31 GETs for small files (< 24 KB)
Server: class20/25.scs.cs.nyu.edu
Client: Athlon XP 2200, 500 MB,
Windows XP, cable modem
Latency vs. Bandwidth
Group 5’s Results
Content: Multimedia
Content: Multimedia
Overview
 Multimedia = audio and video
 Saroiu et al.—An Analysis of Internet Content
Delivery Systems
 How is multimedia distributed over the Internet?
 How much is there?
 MacCanne et al.—Receiver-Driven Layered
Multicast
 How to best stream multimedia across the Internet?
Stefan Saroiu’s OSDI Talk
Streaming Multimedia
 Based on broadcast model
 One server, many clients
 Clients subscribe to streams
 Basic problem: Network heterogeneity
 One approach: Fixed rate, least common denominator
Better Approach:
Layered Transmission Scheme
 Basic idea: Encode signal in many layers
 Each layer provides better quality
 Sum of layers represents a session
 Cumulative layers
 Independent layers
 Simulcast
Underlying Network Model
 Three assumptions
 Best effort, multipoint packet delivery
 Efficiency of IP Multicast
 Group-oriented communication
 Important issue: router drop policy
The RLM Protocol
 Basic control loop
 On congestion, drop a layer
 On space capacity, add a layer
Capacity Inference
 One option: Monitor link utilization in network
 Problem: requires changes to network
 Another option: Actively probe network
 Join-experiments
 Issues with join-experiments
 Adaptability
 Scalability
Join-Experiment Adaptability
 Goal
 Perform infrequently when likely to fail
 Perform frequently when likely to succeed
 Algorithm
 Join-timer for each layer
 Exponential backoff for problematic layers
Join-Experiment Adaptability
(cont.)
 How to correlate join-experiment with outcome?
 Need to chose appropriate detection-time
 Unknown
 Variable
 Use estimator
 Initialize conservatively
 Adjust based on failed join-experiments
Join-Experiment Scalability
 Issue: interaction of independent join-experiments
 Add congestion
 Interfere with each other
 Approach: scale frequency with group size
 But, what about convergence?
Join-Experiment Scalability
Shared Learning
 Receiver notifies group of join-experiment
 On congestion, other receivers increase
corresponding join-timer
 Conservative
 Local
More on Shared Learning
 Join-experiments are not completely exclusionary
 Lower or equal level experiments may overlap
 What about router drop policy?
Evaluation
 Based on simulations (ns)
 Two metrics
 Worst-case short-term loss rate
 Convergence time to sustainable
throughput
 Four topologies




Latency scalability
Session scalability
Bandwidth heterogeneity
Superposition
Results
 RLM
 Is sensitive to transmission latency
 Scales with group size
 Though, convergence time increases!
 Supports bandwidth heterogeneity
 Though, with increased loss rate
 Supports simultaneous sessions
 Though, allocation was often unfair
What Did You Learn Today?
 Content distribution in the Internet
 Receiver-driven layered multicast