Distributing Streaming Media Content Using Cooperative Networking

1
Distributing Streaming Media Content Using
Cooperative Networking

• VenkataN.Padmanabhan,
• Helen H.Wang,
• Philip A.Chou

2
Outline

• Introduction
• Related Work
• Cooperative networking
• Performance Evaluation
• Conclusions

3
Introduction

• Use CoopNet distributing streaming media content
  both live and on demand
• Cooperative Networking (CoopNet)
• An approach to content distribution that combine
  aspects of infrastructure-based and peer-to peer
• In infrastructure-based server may overwhelmed
  because of a large flash crowd,but CoopNet
  addresses it by have client cooperate with each
  other
• The distinction between CoopNet and pure P2P
  system is CoopNet still have a server

4
Related Work

• Infrastructure-based on CDN such as Akamai employ
  a dedicated network to serve content on behalf of
  server
• Peer-to-peer such as Napster and Guntella has no
  dedicated infrastructure,so the short life-time
  of individual nodes pose a challenge
• Use application-level multicast is directly
  relevant to live streaming aspect of CoopNet
• CoopNet uses a centralized protocol to repair
  quickly

5
Live streaming

• Live streaming
• the synchronized distribution of streaming media
  content to one or more clients
• Distribution tree
• Root is server , clients is its member
• Each member transmit the received stream to its
  children using unicast
• The out-degree is constrained by bandwidth
• Use MDC to reduce the disruption cause by node
  departure

6
Multiple Description Coding(MDC)

• Method of encoding the audio and/or video signal
  into separate streams or description
• Many different streams (description) for same
  GOF (group of frame) then packetized into packet
• Every subset of descriptions transmit by
  different tree
• Every subset of descriptions can be recovered a
  signal with distortion
• For extra flexibility,MDC incur performance
  penalty
• Signal quality relative to number of descriptions
  received

7
Multiple Description Coding(cont)
8
Quality During Multiple Failure

• The probability of host n receive mth description
  is
• An is the number of host ancestors,e is
  probability that a host fail
• So average probability is
  (N nodes)
• As N increase for fixed M,performance slowly
  degrades

9
Quality During Single Failure

• The probability of host n failed to receive mth
  description is An/(N-1)
• So average probability to receive is
• As N increase for fixed M,performance become
  nearly perfect,but difficult to repair the trees
  before a second failure occurs

10
Further Analyses

• When out-degree 1,the tree like a chain
• the average probability ?n(N1/2N)
• As N go to infinity ,the probability that any
  host receives any descriptions goes to zero

11
Tree management

• The goals for tree management algorithm
• Short and wide tree
• Minimize latency but high degree reduces rate for
  other
• Efficiency vs. tree diversity
• Close to underlying network topology but common
  points of make diverse tree
• Quick join and leave
• Minimizes delay for new nodes but balancing takes
  time
• Scalability
• scale to large number of users with high number
  of arrivals departures

12
Tree management(cont)

• When new node wants to join
• New node contacts server for content
• Server responds with list of designated parent
  nodes in each tree
• Starting at the server
• Go down tree until we get to a level with nodes
  containing spare capacity to serve as parent of
  new node.
• Chosen randomly from this list
• Node sends messages to parents to get added
• To leave, node sends server message. Server
  identifies children of node and moves each
  child to a new parent
• Node failure is identified through packet loss
  rate. Node is contact to parent, if responds,
  upstream of his parents. Else, affected node
  contact to server and a re-join takes place

13
Tree management(cont)

• Optimization
• First, Identify set of nodes physically close to
  new node using measurement to landmark hosts
• Choose a parent in this set of nodes
• Feasibility of the centralized protocol
• Tested the solution at MSNBC on 9/11
• number of distribution trees is 16,
  each tree average has
• 4 children
• Memory need 10MB,Bandwidth requirement 8 Mbps
• CPU 390 memory accesses per instruction
• the centralized approach can be scaled up by
  having a cluster of servers and partitioning the
  set of clients across the set of server nodes

14
On-demand streaming

• On-demand Streaming
• distribution of streaming media content on demand
• When server overload
• Server gives client list of IPs of clients
• Client redirect to clients to download the
  content
• A peer may only have a part of requested content
• Peer must know the information during initial
  handshake
• Peer may fail,depart
• Use distributed streaming to divided into sub
  stream then server by different peer

15
Performance evaluation

• Live streaming
• Trace characteristics
• Client may leave because the video quality is
  poor
• So quality were improved, life time may become
  longer

16
Performance evaluation(cont)

• Effectiveness of MDC
• In first experiment
• We make M random trees ,the out-degree of each
  node is random,then randomly remove d nodes from
  tree

17
Performance evaluation(cont)

• In second experiment
• We evolved the distribution trees
  by simulating tree management
• algorithm ,the root out-degree is
  to 100 ,the max out-degree of
• each node is 4 , then randomly
  remove d nodes from tree
• Impact of repair time
• The shorter the repair time, the probability of
  description received is better

18
Performance evaluation(cont)

• On-demand streaming
• Bandwidth load
• Load bandwidth usage
• The bandwidth contribution that only on some
  active nodes
• Reduce load at individual CoopNet clients
• Disjoint portion of the content can be retrieved
  in parallel from multiple in parallel from
  multiple peers use distributed streaming
• Use load-aware algorithm to redirect clients
  recently seen peers that are least loaded
• Client cooperation significantly reduce server
  load,free up bandwidth to support more clients

19
Performance evaluation(cont)
20
Conclusion

• CoopNet, a peer-to-peer content distribution
  scheme that help servers tide over crisis
  situation such as flash crowds
• CoopNet employs distributed streaming and
  multiple description coding improve the
  robustness of the distributed streaming content
• CoopNet is able to reduce serve load
  significantly without placing an unreasonable
  burden on clients
• Tested the solution using trace gathered at
  MSNBC on 9/11