1 Contention Management and Obstruction-free Algorithms Niloufar Shafiei.

1

Contention Management and Obstruction-free Algorithms

Niloufar Shafiei

2

Agenda

Introduction Non-blocking approaches Contention management Two obstruction-free shared deque

algorithms

3

Distributed shared memory systems

Mutual exclusion (using locks) Delay of processes Not fault-tolerant (failure of processes) Deadlock Priority inversion

Non-blocking Delay and failure of processes do not cause

performance problems

4

Non-blocking

Wait-free All processes complete in a finite number of steps.

Lock-free Some processes complete in a finite number of

steps.

Significant overhead Complex and subtle

5

Non-blocking

Obstruction-free Each process completes if it runs without

interference for sufficiently long.

Not guarantee the progress Simpler Easier to design

6

Contention management

Decide which process can take steps at a time

A simple method is back-off when they encounter interference.

Contention management policy: Guarantee the progress of all process

7

+

=Contention

ManagementObstruction-free

Algorithms

Wait-freeAlgorithms

8

Contention Management Policies Aggressive

Always abort

Polite Exponential back-off

KillBlocked Abort if an operation is blocked or waits for maximum waiting

time

Karma Set a priority (priority = amount of work + number of retries)

9

Contention Management Policies Eruption

Set a priority and increase priority of running operation

Randomized Flip a coin

Kindergarten Keep a list

Timestamp Abort earlier timestamp

QueueONBlock Set a flag of waiting operation

10

Obstruction-free Deque on Linear Array

0 1 … MAX+1

LN … LN RN … RN

value

counter

Invariant: array always consists of at least one LN, followed by zero or more data values, followed by at least one RN.

11

Rightpush(v) Change the leftmost RN to v

1. Find index of leftmost RN

LN … LN RN … RN

k

12

2. Check if k=MAX+1 then deque is full

3. Increase the counter of A[k-1]

LN … … … … RN

K=MAX+1

LN … … RN … RN

k

A[k-1]

13

4. Change A[k] to v and increase its counter value

LN … … v … RN

k

14

Rightpop Change the rightmost data value to RN.

1. Find index of leftmost RN

LN LN RN RN

k

15

2. Check if deque is empty

3. Increase the counter of A[k]

LN LN LN LN RN RN

k-1 k

LN LN RN RN

k

A[k]

16

4. Change A[k-1] to RN and increase its

counter value and return the popped value

LN LN RN RN RN

k

17

Problem: push on one end and pop from the other end

LN LN LN LN LN RN

18

Obstruction-free Deque on Circular Array

LN

RN

DN

RN

RN

LN

LN

LN

Invariants:1. All null values are in a contiguous sequence of locations.2. The sequence consists of zero or more RN, followed by zero

or one DN, followed by zero or more LN.3. At least two different null values are in the sequence.

19

Rightpush(v) Change the leftmost RN to v

1. Find index of leftmost RN

v1

RN

DN

RN

v2

LN

LN

LN

k

20

2. Read A[k+1] If A[k+1] = RN

v1

RN

DN

RN

v2

LN

LN

LN

k

k+1

1. Increase the counter value of A[k-1]2. Change A[k] to v

k-1

21

If A[k+1] = LN

v1

LN

LN

RN

v2

LN

LN

LN

k

k+1

1. Increase the counter value of A[k]2. Change A[k+1] to DN

k-1

22

If A[k+1] = DN and A[k+2] Null

v1

DN

v6

RN

v2

v3

v4

v5

k

k+1

Return FULL

k-1

k+2

23

If A[k+1] = DN and A[k+2] = LN

v1

DN

LN

RN

v2

v3

v4

v5

k

k+1

k-1

k+2

1. Increase the counter value of A[k+2]2. Change A[k+1] to RN

24

Conclusion Obstruction-free weakens the progress property

but it is simpler.

Contention management policies improve the progress of obstruction-free algorithms.

Obstruction-free algorithms with appropriate contention management policy can perform like wait-free algorithms in practice.