Mutual Exclusion

1
Mutual Exclusion
2
Critical Section Problem

• acquire(c)
• CriticalSection c
• release(c)
• Critical Section correctness
• Mutual Exclusion safety (ME1)
• Deadlock-Freedom liveness (ME2)
• Starvation-Freedom fairness (ordering,ME3)

3
Analysis of Distributed ME Algs

• Bandwidth usage
• Client Delay
• Throughput

4
Algorithms

• Centralized algorithm
• Token Ring algorithm
• Distributed algorithm
• Voting algorithm

5
Distributed Algorithm

• No leader process
• Need total order on messages
• Broadcast all messages
• Lamports timestamps
• Two message types
• request(P,c,ts) P requests CS c at time ts
• acquire(P,Q,c) Q permits P to enter c
• P acquires c all proc permit P to enter c

6
Maekawas Voting Algorithm

• Each process pi has a voting set Vi of size K
• K ? ?n
• pi ? Vi
• All pairs of voting sets share at least one
  element
• Each process pi is contained in K voting sets

7
Questions (1/3)

• Which distributed ME algorithm provides
• Fairness?
• Liveness?
• Which distributed ME algorithm uses the least
  bandwidth?
• Which distributed ME algorithm uses the most
  bandwidth?

8
Questions (2/3)

• Which distributed ME alg can handle failure the
  easiest?
• Which type of failure can none of the distributed
  ME algorithms recover from if run in an
  asynchronous distributed system?
• If processes continually (or frequently) need to
  enter the CS, which algorithm would you recommend?

9
Questions (3/3)

• If CS entry requests are sporadic, which
  algorithm would you recommend?
• Why can voting algorithms handle failure of
  processes easier than other algorithms?

10
Leader Election

• Section 12.3

11
Leader Election

• Many distributed tasks need a coordinator
• Eg clock synchronization
• If all processes identical cannot elect leader
• Assume each process has unique identifier
• Elect process with largest identifier

12
Ring Algorithm (for process i)

• if (leader down)
• Leader ? send Election, i
• receive Election,j
• if (ij) send LeaderElected, i Leader i
• else if (j gt i) send Election,j Leader ?
• else if (Leader ? ? ) send Election, i
  Leader ?
• receive LeaderElected, j
• if (i ? j) send LeaderElected, j Leader j

13
Bully algorithm

• P gets no response from current leader
• P sends election message to processes with higher
  id
• If no responses
• P is new leader
• P sends leader message to all
• If higher-up Q receives election message
• Q sends bully message to P
• Q sends election message to higher-ups

14
Questions

• Why can we not elect a leader if all processes
  are identical?
• Which of the two leader election algorithms
  discussed can tolerate failure best?
• Which leader election algorithms uses the least
  bandwidth?