Distributed Computing:

1
Fault-Tolerant
Distributed Computing
Atomic Broadcast
2
Outline

• Why distributed computing?
• Atomic Broadcast
• The atom system
• Relevance for e-textiles
• Whats next?
• QA

3
Why Distributed Computing?

• Spread and balance the computational weight of
  applications
• Solve bigger problems
• Deal with problems locally instead of
  centralizing all the data

4
Example

• Space filtering vs. raw consensus
• Acoustic Beam Forming master collects
  information from slaves and decides according to
  the relevance of data
• Consensus no master, all processes decide upon
  one common value

5
Atomic Broadcast Definition (1)

• Atomic Broadcast the same set of messages is
  delivered by all the processes in the same order
• Consensus all processes decide upon one common
  value among those proposed

6
Atomic Broadcast Definition (2)

• Validity If a correct process broadcasts a
  message m it will eventually receive it
• Uniform agreement If a process delivers a
  message m then every correct process will deliver
  it
• Uniform integrity Every message m is delivered
  at most once and only if it was reliably
  broadcasted by sender(m)
• Total order If 2 correct processes p and q
  deliver 2 messages m and m then p delivers m
  before m iff q delivers m before m

7
Atomic Broadcast Bad News

• Impossibly to achieve in a totally asynchronous
  system Fisher, Lynch, Patterson 85

8
Atomic Broadcast Good News

• Can be done using unreliable failure detectors
• Based on a Consensus algorithm described in
  Chandra, Toueg 96

9
Atom

• Open source Atomic Broadcast system

10
Atom
Producer
A-broadcast
AB task1
FD suspect
transmission
do_Consensus
R-broadcast
start
AB task 3
One_run
start
do_decide
cancel
AB task 2
RB
A-deliver
Consumer
FD trust
11
Relevance to E-textiles

• Synchronization of data
• Coordination of decisions and actions
• Light-weight process
• Buffer sizes can be predicted

12
Whats Next?

• Scalability is a problem for classic
  fault-tolerant distributed algorithms
• Bimodal Multicast Ken Birman, Mark Hayden, Oznur
  Ozkasap, Zhen Xiao, Mihai Budiu, Yaron Minsky
  1998
• Gossip protocol
• Relaxes the strong reliability guarantees
  replacing them with probabilistic guarantees
• Converges to strong reliability in the absence
  of failures
• Scalable with steady throughput

13
Questions