Study of Replication Approach for Video Streaming in P2P Networks

1
Study of Replication Approach for Video Streaming
in P2P Networks

• Wilson, W.F. Poon
• The Chinese University of Hong Kong

2
Content

• Introduction
• Related Works
• Full Replication Approach
• Replication Profile
• Placement Policy
• Selection Scheme
• Experimental Results
• Erasure Code Approach
• Conclusion

3
Introduction (1)

• Providing video streaming services have long been
  a research topic
• parallel server designs such as RAID
• multicast/broadcast transmission schemes
• distributed VoD systems (scalability and
  reliability)
• Tremendous growth in computer power in personal
  computers
• peer-to-peer (p2p) systems
• Peers contribute storage, content and bandwidth

4
Introduction (2)

• Two types of p2p systems
• Unstructured, e.g. KaZaA and Gnutella
• Peers are not organized into highly-structured
  overlays
• Content is randomly assigned to peers
• Structured, e.g. CAN, Chord
• Peers are organized into highly-structured
  overlays
• Distributed hash table (DHT) substrates are used
• Keys are deterministically assigned to peers
• However, most of these p2p systems are targets
  for file sharing/web caching services

5
Introduction (3)

• Previous work mainly focused
• Search mechanism
• Storage management
• The work on p2p video streaming has not been
  thoroughly studied
• Investigate whether such a p2p system is
  applicable to supporting video streaming
  applications

6
Introduction (4)

• One of major challenges of a p2p system
• Peer machines may be turned on and off in an
  unpredictable manner
• The system experiences very worse availability
• Two strategies are considered
• Full Replication (whole file)
• Erasure Code Replication (data blocks)

7
Related Works (1)

• Unlike the traditional distributed VoD systems
• The components of the p2p system experience very
  worse availability
• Works dealt mainly with the system/file
  availability
• Most Frequently Requested (MFR) KANG02 was
  proposed to maximize the file hit rate of the
  structured p2p system
• In LV02 and COHE02, it studied optimal
  replication in an unstructured peer-to-peer
  network
• Reduce random search times
• Lee et. al. LEE02 proposed a server-less
  architecture for video streaming
• Use erasure code replication to distribute the
  data among peers

8
Peers Behavior

• Majority of peers had availability rates of under
  20 percent
• (S. Saroiu, P. K. Gummadi, and S. D.
  Gribble, A measurement study of peer-to-peer
  file sharing systems, MMCN, 2002)

9
p2p Streaming System

• Similar to traditional VoD systems
• Replication strategy
• Full Replication VS Erasure Code
• Replication Profile
• Full Replication number of replicas of the
  videos
• Erasure Code ratio of original and redundant
  data (redundancy overhead)
• Data Placement
• Full Replication distribute the replicas of the
  videos
• Erasure code distribute the data block and
  redundant block
• Peer Selection Policy
• Select the peers (servers) to stream the
  requested video

10
Full Replication Approach (1)

• A network has G peers in which I peers (serving
  peers) stores a set of J different videos
• The other peers (free riders) just make requests
  but not contribute their resources
• Assume
• ? is the up probability of the peers
• Tup is the mean up time duration
• Tdown is the mean down time duration

• Assume
• Ni is the amount of shared storage in peer i
• bj is the size of video j
• qj is the request probability for video j
• Cj is the bit rate for video j

11
Full Replication Approach (2)

• sj is number of replicas for video j
• Requests to a serving peer for vj is given by

• System storage constraints

12
Full Replication Approach (3)

• How to determine the number of replicas, sj, for
  each videos
• Replication Profile
• Random
• Maximize the hit rate (MaxHit)
• Minimize the request rate per peer (MinReq)
• Random
• In a p2p system, the simplest way is to randomly
  determine the number replicas for each video
• e.g. a user may need to manually copy the videos
  into the shared storage

13
Full Replication Approach (4)

• MaxHit KANG02
• The optimal number of replicas can be determined
  by maximizing the system hit rate

Maximize
Subject to

• This optimization problem can be efficiently
  solved by resource allocation algorithms such as
  dynamic programming
• The optimal replication profile

14
Full Replication Approach (5)

• MinReq
• For video streaming, a video request that can be
  served requires
• The requested video is available in the system
• The serving peers have the available bandwidth
• The number of replicas should be determined by
  minimizing the load of the serving peers

Minimize
Subject to
15
Number of Replicas

• Assume
• qi follows zipf distribution, zipf factor 0.271
• All the peers have the same storage and streaming
  capacity
• The length (L) and bit rate (C) of all the videos
  are the same
• I1500, J100, ?0.1 (Tup3600s and Tdown32400s)

(a) Storage, S 2
(b) Storage, S 10
16
Full Replication Approach (6)

• How to store the replicas of videos among peers
• Placement Policy
• Random
• Smallest Load First (SLF) ZHOU02
• Random
• The replicas of the videos are randomly stored
  among the peers
• The load of the system is imbalance

17
Placement Policy

• Smallest Load First (SLF)
• Sort the replicas of the videos in a
  non-increasing order by request weight, wj
• For each iteration
• Select I replicas with the greatest request
  weight
• Distribute these I copies to the I peers
• (Rules greatest request weight is placed to the
  peers with the smallest load AND peer cannot
  store more than one copy of the same video)

18
Smallest Load First

• Example
• 4 peers, each can store 2 copies of the videos
• 3 videos

19
Requests per Peer (1)

• MinReq
• I1500, J100, ?0.1 (Tup3600s and Tdown32400s)

20
Requests per Peer (2)

• MaxHit
• I1500, J100, ?0.1 (Tup3600s and Tdown32400s)

21
Selection Scheme

• How to choose a peer to stream a video
• Inappropriate serving peers are selected, more
  requests will be rejected from the system

• When a peer requests for a video
• Get a list of serving peers storing the requested
  video
• Random
• Randomly select one peer in the list as a server
• Least Load First (LLF)
• Use the current available uplink bandwidth, Bup
• Choose a peer in the list with the maximum Bup
• Randomly choose a peer if more than one peer has
  the same Bup

• What if the peer leaves the system while serving
  the other peers
• If the serving peer disconnects, the peer who is
  being served should find another peer based on
  the above algorithm
• If no serving peer is available, the playback
  will be stopped

22
Simulation

• Run simulation experiments
• Request probabilities follow a Zipf distribution
  with parameters 0.271
• Tup and Tdown follow an exponential distribution
• Measure the successful playback of the system
• Peers can start the playback until the end of the
  video

23
Results Arrival Rate

• Number of peers1500
• Number of videos100
• Up time probability0.1
• Peer storage10 videos
• Video Length7200s

• Successful rate of MaxHit_SLF_LLF and Random
  is similar

24
Results Peer Availability

• Arrival rate0.04/s
• Number of peers1200
• Number of videos100
• Peer storage10 videos
• Video Length7200s

• The system provides reliable services
• More peers are willing to share the resources
• Peer availability is increased

25
Is Erasure Code Better?

• Erasure code such as Reed Solomon Erasure (RSE)
• A video, vj, is divided into I blocks
• (I-h) data blocks and h redundant blocks
• Blocks are evenly distributed to all the I peers
• To recover the video, the user should receive any
  (I-h) out of I blocks
• Using the erasure code replication
• System Gains
• Require less bandwidth and storage
• Maintain the load balance
• Overhead
• Larger number of messages than in a replicated
  system
• Require real-time scheduling to reassemble data
  blocks
• Maintain the system consistency (e.g. updating
  redundant data)

26
Erasure Code

• Assume the system maintains the same reliability
  level for all the videos
• Storage overhead for all the videos are the same
• If (I-h) out of I blocks can recover the video,
  the storage overhead is

27
Full Replication VS Erasure-Code
(a) Storage, S 2
(b) Storage, S 10

• A portion of peers uplink bandwidth is the
  overhead

28
Conclusion

• System availability and bandwidth capacity are
  important elements to design p2p streaming
  systems
• The system performance can be further improved by
  the erasure coding scheme
• The full replication system is better
• Overhead of erasure coding is too high
• Develop a model in order to closely examine
  different parameters
• system size
• overhead
• bandwidth capacity