Reliable Multicast Using ReedSolomon and Gossip

1
Reliable Multicast Using Reed-Solomon and Gossip

• Continues Project to FEC using RS Project

Daniel Finchelstein and Noam Ivri
2
Project Goals

• Search and determine the best approach to add
  reliability and performances to our
  Frame-Distribution system and implement it.
• Analyze the new system results Vs. first
  projects results.

3
Recap on FEC using RS

• One lecturer (server) and many students (clients)

Server User Interface
FEC encoder
Server Application
Wireless Network
4
Using SR to send a File

• A file will be divided into a blocks
• Every Block is decoded separately
• Every block is K packets long before encoding
• Lets assume
• file size FS
• Packet Size PS
• Number of Blocks a FS/(PSK)
• Loosing one block ? loosing the whole file.

5
success rate as function of the distance from the
base station results from previous work
We would like to optimize this steep degradation
of the service without overloading the clients
CPU
6
What happens when we need more packets?

• Need more packets less then K were received
  after N were sent at least k are needed for
  correct decoding of the frame.
• Option1 (naïve) ask the server to send again
• Option2 change the (K,N) values for future
  packets
• Option3 get it from near-by clients!

7
Near-by clients?

• Reasonable assumptions
• different clients loose different packets
• Physical proximity allows good connection
• The clients have the resources to deal with low
  number of requests.

Using a background mechanism the clients can
exchange needed packets!
8
How does it work?

• Fixed redundant information set produced by the
  server (N-K rate)
• Clients join a Group Membership System (GMS)
• Gossip between sets of members
• Exchange needed packets from the redundant set.

9
Benefits of our approach

• Immediate correction of fault receives.
• Can work with low values of N,K.
• Does not bother the server.
• Extremely efficient for large number of clients.
  (epidemic rate of progress).
• Cost effective no need for complicated
  calculations.

10
The Gossip Stage

• Our work focuses on creating a highly effective
  gossip mechanism
• The selected solution is based on pbcast. (based
  on Cornell's project)

11
Gossip, Requirements

• Work independently- regardless of the RS system.
• Exchange needed information only.
• Exchange data MUCH faster than its production.
• Allow clients to join and leave at any stage.

12
Gossip, Design

• unsynchronized gossip rounds .
• In each round, 2 phases
• 1.detect message loss .
• 2 .correct such losses (runs only if needed) .

Members randomly choose other members ? send a
summary of their packets history ? ask for
packets they discover themselves to be lacking