DataDriven Overlay Streaming System Implementation

1
Data-Driven Overlay Streaming System
Implementation

• Jerry, Wu

2
Concept

• A p2p media transfer platform
• Use gossip protocol to gossip data and control
  message
• Get streaming data from partners rather than
  specified servers or multicast nodes
• A good solution to carry high bit rate media
  through an overlay network

3
Architecture

• Borrowed from DONet

4
Implementation

• Whats the real streaming format ?
• Player behavior variability
• Time-synchronization among clients

5
Real Stream Format

• What we have tried are
• mp3, ogg, rm
• Some comparisons
• We choose MP3 finally

6
Player Behavior Variability

• Cache size
• Different cache size may influence write speed
• Starting play time variation
• Mplayer may not always start as we predicted

7
Time Synchronization

• Synchronize the Mplayer time with real time
• By experience, root node with 128 slot is not
  enough
• So we use 1282 256 bytes as root node data
  size
• So we use auto adjust to fine-tune the Buffer
  Map

8
Environment For Implementation (1/4)

• Platform Linux with kernel v2.6.11
• Live media encoder MPlayer
• MP3 encoder lame
• Using named pipe
• Simple socket

l ame mypipe mlayer cache 8192 -
9
Environment for Implementation (2/4)

• Network topology
• A. 5 nodes from NSLAB
• B. 7 nodes from ATT lab(???)
• C. 5 nodes from CSIE workstations
• D. 4 nodes from ADSL (2M/256k)
• A and D are under NAT

10
Environment for Implementation (3/4)

• Media
• 10 mins mp3 (192K) file
• 10 mins mp3 extract and get from web streaming
  broadcast

11
Environment for Implementation (4/4)

• Cross NAT
• Simple time synchronization
• Use simple STUN TERN server to get the next used
  port
• Send SYN packet to each other
• For ?? series user, sorry
• Result
• Good at most of the cases except
• NSLABs NAT server block the intra-domain SYN as
  bad connection

12
Experimental Result (1/2)

• Result
• Perceptional converge time (include skew
  correction)

13
Experimental Result (2/2)

• Packet miss deadline rate

14
What would be done

• A fancy GUI
• Better cross NAT solution
• Change the preference of the algo (from rare
  packet first to a balance mode)
• Find out if it can work at larger scale
• Use a higher level language
• No bit array in C
• Maybe delphi or java