Notes on Theory of Distributed Systems

CS 465/565: Spring 2016

James Aspnes

2016-01-30 23:20
Contents

Table of contents i

List of figures xii

List of tables xiii

List of algorithms xiv

Preface xviii

Syllabus xix

Lecture schedule xxii

1 Introduction 1
1.1 Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

I Message passing 7

2 Model 8
2.1 Basic message-passing model . . . . . . . . . . . . . . . . . . 8
2.1.1 Formal details . . . . . . . . . . . . . . . . . . . . . . 9
2.1.2 Network structure . . . . . . . . . . . . . . . . . . . . 10
2.2 Asynchronous systems . . . . . . . . . . . . . . . . . . . . . . 10
2.2.1 Example: client-server computing . . . . . . . . . . . . 11
2.3 Synchronous systems . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Drawing message-passing executions . . . . . . . . . . . . . . 12
2.5 Complexity measures . . . . . . . . . . . . . . . . . . . . . . . 14

i
CONTENTS ii

3 Coordinated attack 16
3.1 Formal description . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Impossibility proof . . . . . . . . . . . . . . . . . . . . . . . . 17
3.3 Randomized coordinated attack . . . . . . . . . . . . . . . . . 18
3.3.1 An algorithm . . . . . . . . . . . . . . . . . . . . . . . 19
3.3.2 Why it works . . . . . . . . . . . . . . . . . . . . . . . 20
3.3.3 Almost-matching lower bound . . . . . . . . . . . . . . 21

4 Broadcast and convergecast 22

4.1 Flooding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.1.1 Basic algorithm . . . . . . . . . . . . . . . . . . . . . . 22
4.1.2 Adding parent pointers . . . . . . . . . . . . . . . . . 24
4.1.3 Termination . . . . . . . . . . . . . . . . . . . . . . . . 25
4.2 Convergecast . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.3 Flooding and convergecast together . . . . . . . . . . . . . . . 27

5 Distributed breadth-first search 29

5.1 Using explicit distances . . . . . . . . . . . . . . . . . . . . . 29
5.2 Using layering . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.3 Using local synchronization . . . . . . . . . . . . . . . . . . . 31

6 Leader election 35
6.1 Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6.2 Leader election in rings . . . . . . . . . . . . . . . . . . . . . 37
6.2.1 The Le-Lann-Chang-Roberts algorithm . . . . . . . . 37
6.2.1.1 Proof of correctness for synchronous executions 38
6.2.1.2 Performance . . . . . . . . . . . . . . . . . . 38
6.2.2 The Hirschberg-Sinclair algorithm . . . . . . . . . . . 39
6.2.3 Peterson’s algorithm for the unidirectional ring . . . . 39
6.2.4 A simple randomized O(n log n)-message algorithm . . 42
6.3 Leader election in general networks . . . . . . . . . . . . . . . 42
6.4 Lower bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.4.1 Lower bound on asynchronous message complexity . . 43
6.4.2 Lower bound for comparison-based algorithms . . . . 44

7 Logical clocks 47
7.1 Causal ordering . . . . . . . . . . . . . . . . . . . . . . . . . . 47
7.2 Implementations . . . . . . . . . . . . . . . . . . . . . . . . . 49
7.2.1 Lamport clock . . . . . . . . . . . . . . . . . . . . . . 49
7.2.2 Neiger-Toueg-Welch clock . . . . . . . . . . . . . . . . 50
CONTENTS iii

7.2.3 Vector clocks . . . . . . . . . . . . . . . . . . . . . . . 51

7.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
7.3.1 Consistent snapshots . . . . . . . . . . . . . . . . . . . 52
7.3.1.1 Property testing . . . . . . . . . . . . . . . . 53
7.3.2 Replicated state machines . . . . . . . . . . . . . . . . 54

8 Synchronizers 55
8.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
8.2 Implementations . . . . . . . . . . . . . . . . . . . . . . . . . 56
8.2.1 The alpha synchronizer . . . . . . . . . . . . . . . . . 57
8.2.2 The beta synchronizer . . . . . . . . . . . . . . . . . . 57
8.2.3 The gamma synchronizer . . . . . . . . . . . . . . . . 58
8.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
8.4 Limitations of synchronizers . . . . . . . . . . . . . . . . . . . 59
8.4.1 Impossibility with crash failures . . . . . . . . . . . . 59
8.4.2 Unavoidable slowdown with global synchronization . . 60

9 Synchronous agreement 62
9.1 Problem definition . . . . . . . . . . . . . . . . . . . . . . . . 62
9.2 Lower bound on rounds . . . . . . . . . . . . . . . . . . . . . 63
9.3 Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
9.3.1 Flooding . . . . . . . . . . . . . . . . . . . . . . . . . . 65
9.4 Exponential information gathering . . . . . . . . . . . . . . . 66
9.4.1 Basic invariants . . . . . . . . . . . . . . . . . . . . . . 67
9.4.2 Stronger facts . . . . . . . . . . . . . . . . . . . . . . . 68
9.4.3 The payoff . . . . . . . . . . . . . . . . . . . . . . . . 68
9.4.4 The real payoff . . . . . . . . . . . . . . . . . . . . . . 68
9.5 Variants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

10 Byzantine agreement 69
10.1 Lower bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
10.1.1 Minimum number of rounds . . . . . . . . . . . . . . . 69
10.1.2 Minimum number of processes . . . . . . . . . . . . . 69
10.1.3 Minimum connectivity . . . . . . . . . . . . . . . . . . 71
10.1.4 Weak Byzantine agreement . . . . . . . . . . . . . . . 72
10.2 Upper bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
10.2.1 Exponential information gathering gets n = 3f + 1 . . 74
10.2.1.1 Proof of correctness . . . . . . . . . . . . . . 74
10.2.2 Phase king gets constant-size messages . . . . . . . . . 76
10.2.2.1 The algorithm . . . . . . . . . . . . . . . . . 76
CONTENTS iv

10.2.2.2 Proof of correctness . . . . . . . . . . . . . . 78

10.2.2.3 Performance of phase king . . . . . . . . . . 78

11 Impossibility of asynchronous agreement 80

11.1 Agreement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
11.2 Failures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
11.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
11.4 Bivalence and univalence . . . . . . . . . . . . . . . . . . . . . 82
11.5 Existence of an initial bivalent configuration . . . . . . . . . . 82
11.6 Staying in a bivalent configuration . . . . . . . . . . . . . . . 83
11.7 Generalization to other models . . . . . . . . . . . . . . . . . 84

12 Paxos 85
12.1 Motivation: replicated state machines . . . . . . . . . . . . . 85
12.2 The Paxos algorithm . . . . . . . . . . . . . . . . . . . . . . . 86
12.3 Informal analysis: how information flows between rounds . . 88
12.4 Safety properties . . . . . . . . . . . . . . . . . . . . . . . . . 88
12.5 Learning the results . . . . . . . . . . . . . . . . . . . . . . . 90
12.6 Liveness properties . . . . . . . . . . . . . . . . . . . . . . . . 90

13 Failure detectors 92
13.1 How to build a failure detector . . . . . . . . . . . . . . . . . 93
13.2 Classification of failure detectors . . . . . . . . . . . . . . . . 93
13.2.1 Degrees of completeness . . . . . . . . . . . . . . . . . 93
13.2.2 Degrees of accuracy . . . . . . . . . . . . . . . . . . . 93
13.2.3 Boosting completeness . . . . . . . . . . . . . . . . . . 94
13.2.4 Failure detector classes . . . . . . . . . . . . . . . . . . 95
13.3 Consensus with S . . . . . . . . . . . . . . . . . . . . . . . . . 96
13.3.1 Proof of correctness . . . . . . . . . . . . . . . . . . . 97
13.4 Consensus with ♦S and f < n/2 . . . . . . . . . . . . . . . . 98
13.4.1 Proof of correctness . . . . . . . . . . . . . . . . . . . 100
13.5 f < n/2 is still required even with ♦P . . . . . . . . . . . . . 101
13.6 Relationships among the classes . . . . . . . . . . . . . . . . . 102

14 Quorum systems 104

14.1 Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
14.2 Simple quorum systems . . . . . . . . . . . . . . . . . . . . . 104
14.3 Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
14.4 Paths system . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
14.5 Byzantine quorum systems . . . . . . . . . . . . . . . . . . . 107
CONTENTS v

14.6 Probabilistic quorum systems . . . . . . . . . . . . . . . . . . 108

14.6.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 109
14.6.2 Performance . . . . . . . . . . . . . . . . . . . . . . . 109
14.7 Signed quorum systems . . . . . . . . . . . . . . . . . . . . . 110

II Shared memory 111

15 Model 112
15.1 Atomic registers . . . . . . . . . . . . . . . . . . . . . . . . . 112
15.2 Single-writer versus multi-writer registers . . . . . . . . . . . 113
15.3 Fairness and crashes . . . . . . . . . . . . . . . . . . . . . . . 114
15.4 Concurrent executions . . . . . . . . . . . . . . . . . . . . . . 114
15.5 Consistency properties . . . . . . . . . . . . . . . . . . . . . . 115
15.6 Complexity measures . . . . . . . . . . . . . . . . . . . . . . . 116
15.7 Fancier registers . . . . . . . . . . . . . . . . . . . . . . . . . 117

16 Distributed shared memory 119

16.1 Message passing from shared memory . . . . . . . . . . . . . 120
16.2 The Attiya-Bar-Noy-Dolev algorithm . . . . . . . . . . . . . . 120
16.3 Proof of linearizability . . . . . . . . . . . . . . . . . . . . . . 122
16.4 Proof that f < n/2 is necessary . . . . . . . . . . . . . . . . . 123
16.5 Multiple writers . . . . . . . . . . . . . . . . . . . . . . . . . . 123
16.6 Other operations . . . . . . . . . . . . . . . . . . . . . . . . . 124

17 Mutual exclusion 125

17.1 The problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
17.2 Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
17.3 Mutual exclusion using strong primitives . . . . . . . . . . . . 126
17.3.1 Test and set . . . . . . . . . . . . . . . . . . . . . . . . 126
17.3.2 A lockout-free algorithm using an atomic queue . . . . 127
17.3.2.1 Reducing space complexity . . . . . . . . . . 128
17.4 Mutual exclusion using only atomic registers . . . . . . . . . 129
17.4.1 Peterson’s tournament algorithm . . . . . . . . . . . . 129
17.4.1.1 Correctness of Peterson’s protocol . . . . . . 130
17.4.1.2 Generalization to n processes . . . . . . . . . 133
17.4.2 Fast mutual exclusion . . . . . . . . . . . . . . . . . . 133
17.4.3 Lamport’s Bakery algorithm . . . . . . . . . . . . . . 136
17.4.4 Lower bound on the number of registers . . . . . . . . 137
17.5 RMR complexity . . . . . . . . . . . . . . . . . . . . . . . . . 139
CONTENTS vi

17.5.1 Cache-coherence vs. distributed shared memory . . . . 139

17.5.2 RMR complexity of Peterson’s algorithm . . . . . . . 140
17.5.3 Mutual exclusion in the DSM model . . . . . . . . . . 141
17.5.4 Lower bounds . . . . . . . . . . . . . . . . . . . . . . . 143

18 The wait-free hierarchy 144

18.1 Classification by consensus number . . . . . . . . . . . . . . . 145
18.1.1 Level 1: registers etc. . . . . . . . . . . . . . . . . . . 146
18.1.2 Level 2: interfering RMW objects etc. . . . . . . . . . 147
18.1.3 Level ∞: objects where first write wins . . . . . . . . 149
18.1.4 Level 2m − 2: simultaneous m-register write . . . . . . 150
18.1.4.1 Matching impossibility result . . . . . . . . . 152
18.1.5 Level m: m-process consensus objects . . . . . . . . . 153
18.2 Universality of consensus . . . . . . . . . . . . . . . . . . . . . 154

19 Atomic snapshots 157

19.1 The basic trick: two identical collects equals a snapshot . . . 157
19.2 The Gang of Six algorithm . . . . . . . . . . . . . . . . . . . 158
19.2.1 Linearizability . . . . . . . . . . . . . . . . . . . . . . 159
19.2.2 Using bounded registers . . . . . . . . . . . . . . . . . 160
19.3 Faster snapshots using lattice agreement . . . . . . . . . . . . 163
19.3.1 Lattice agreement . . . . . . . . . . . . . . . . . . . . 163
19.3.2 Connection to vector clocks . . . . . . . . . . . . . . . 164
19.3.3 The full reduction . . . . . . . . . . . . . . . . . . . . 164
19.3.4 Why this works . . . . . . . . . . . . . . . . . . . . . . 166
19.3.5 Implementing lattice agreement . . . . . . . . . . . . . 167
19.4 Practical snapshots using LL/SC . . . . . . . . . . . . . . . . 171
19.4.1 Details of the single-scanner snapshot . . . . . . . . . 172
19.4.2 Extension to multiple scanners . . . . . . . . . . . . . 174
19.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
19.5.1 Multi-writer registers from single-writer registers . . . 174
19.5.2 Counters and accumulators . . . . . . . . . . . . . . . 175
19.5.3 Resilient snapshot objects . . . . . . . . . . . . . . . . 175

20 Lower bounds on perturbable objects 177

21 Restricted-use objects 180

21.1 Implementing bounded max registers . . . . . . . . . . . . . . 180
21.2 Encoding the set of values . . . . . . . . . . . . . . . . . . . . 182
21.3 Unbounded max registers . . . . . . . . . . . . . . . . . . . . 183
CONTENTS vii

21.4 Lower bound . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

21.5 Max-register snapshots . . . . . . . . . . . . . . . . . . . . . . 184
21.5.1 Linearizability . . . . . . . . . . . . . . . . . . . . . . 187
21.5.2 Application to standard snapshots . . . . . . . . . . . 187

22 Common2 190
22.1 Test-and-set and swap for two processes . . . . . . . . . . . . 191
22.2 Building n-process TAS from 2-process TAS . . . . . . . . . . 192
22.3 Single-use swap objects . . . . . . . . . . . . . . . . . . . . . 192

23 Randomized consensus and test-and-set 196

23.1 Role of the adversary in randomized algorithms . . . . . . . . 196
23.2 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
23.3 Reduction to simpler primitives . . . . . . . . . . . . . . . . . 198
23.3.1 Adopt-commit objects . . . . . . . . . . . . . . . . . . 199
23.3.2 Conciliators . . . . . . . . . . . . . . . . . . . . . . . . 200
23.4 Implementing an adopt-commit object . . . . . . . . . . . . . 200
23.5 A one-register conciliator for an oblivious adversary . . . . . 201
23.6 Sifters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
23.6.1 Test-and-set using sifters . . . . . . . . . . . . . . . . 205
23.6.2 Consensus using sifters . . . . . . . . . . . . . . . . . . 205
23.7 O(log∗ n) Randomized test-and-set . . . . . . . . . . . . . . . 207
23.8 Space bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

24 Renaming 211
24.1 Renaming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
24.2 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
24.3 Order-preserving renaming . . . . . . . . . . . . . . . . . . . 213
24.4 Deterministic renaming . . . . . . . . . . . . . . . . . . . . . 213
24.4.1 Wait-free renaming with 2n − 1 names . . . . . . . . . 214
24.4.2 Long-lived renaming . . . . . . . . . . . . . . . . . . . 215
24.4.3 Renaming without snapshots . . . . . . . . . . . . . . 216
24.4.3.1 Splitters . . . . . . . . . . . . . . . . . . . . . 216
24.4.3.2 Splitters in a grid . . . . . . . . . . . . . . . 217
24.4.4 Getting to 2n − 1 names in polynomial space . . . . . 219
24.4.5 Renaming with test-and-set . . . . . . . . . . . . . . . 220
24.5 Randomized renaming . . . . . . . . . . . . . . . . . . . . . . 220
24.5.1 Randomized splitters . . . . . . . . . . . . . . . . . . . 221
24.5.2 Randomized test-and-set plus sampling . . . . . . . . 221
24.5.3 Renaming with sorting networks . . . . . . . . . . . . 222
CONTENTS viii

24.5.3.1 Sorting networks . . . . . . . . . . . . . . . . 222

24.5.3.2 Renaming networks . . . . . . . . . . . . . . 223
24.5.4 Randomized loose renaming . . . . . . . . . . . . . . . 224

25 Software transactional memory 226

25.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
25.2 Basic approaches . . . . . . . . . . . . . . . . . . . . . . . . . 227
25.3 Implementing multi-word RMW . . . . . . . . . . . . . . . . 228
25.3.1 Overlapping LL/SC . . . . . . . . . . . . . . . . . . . 229
25.3.2 Representing a transaction . . . . . . . . . . . . . . . 229
25.3.3 Executing a transaction . . . . . . . . . . . . . . . . . 230
25.3.4 Proof of linearizability . . . . . . . . . . . . . . . . . . 230
25.3.5 Proof of non-blockingness . . . . . . . . . . . . . . . . 231
25.4 Improvements . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
25.5 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232

26 Obstruction-freedom 233
26.1 Why build obstruction-free algorithms? . . . . . . . . . . . . 234
26.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
26.2.1 Lock-free implementations . . . . . . . . . . . . . . . . 234
26.2.2 Double-collect snapshots . . . . . . . . . . . . . . . . . 234
26.2.3 Software transactional memory . . . . . . . . . . . . . 235
26.2.4 Obstruction-free test-and-set . . . . . . . . . . . . . . 235
26.2.5 An obstruction-free deque . . . . . . . . . . . . . . . . 237
26.3 Boosting obstruction-freedom to wait-freedom . . . . . . . . . 239
26.3.1 Cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
26.4 Lower bounds for lock-free protocols . . . . . . . . . . . . . . 244
26.4.1 Contention . . . . . . . . . . . . . . . . . . . . . . . . 244
26.4.2 The class G . . . . . . . . . . . . . . . . . . . . . . . . 245
26.4.3 The lower bound proof . . . . . . . . . . . . . . . . . . 247
26.4.4 Consequences . . . . . . . . . . . . . . . . . . . . . . . 251
26.4.5 More lower bounds . . . . . . . . . . . . . . . . . . . . 251
26.5 Practical considerations . . . . . . . . . . . . . . . . . . . . . 252

27 BG simulation 253
27.1 Safe agreement . . . . . . . . . . . . . . . . . . . . . . . . . . 253
27.2 The basic simulation algorithm . . . . . . . . . . . . . . . . . 255
27.3 Effect of failures . . . . . . . . . . . . . . . . . . . . . . . . . 256
27.4 Inputs and outputs . . . . . . . . . . . . . . . . . . . . . . . . 256
27.5 Correctness of the simulation . . . . . . . . . . . . . . . . . . 257
CONTENTS ix

27.6 BG simulation and consensus . . . . . . . . . . . . . . . . . . 258

28 Topological methods 259

28.1 Basic idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
28.2 k-set agreement . . . . . . . . . . . . . . . . . . . . . . . . . . 260
28.3 Representing distributed computations using topology . . . . 261
28.3.1 Simplicial complexes and process states . . . . . . . . 261
28.3.2 Subdivisions . . . . . . . . . . . . . . . . . . . . . . . 265
28.4 Impossibility of k-set agreement . . . . . . . . . . . . . . . . . 269
28.5 Simplicial maps and specifications . . . . . . . . . . . . . . . 270
28.5.1 Mapping inputs to outputs . . . . . . . . . . . . . . . 271
28.6 The asynchronous computability theorem . . . . . . . . . . . 272
28.6.1 The participating set protocol . . . . . . . . . . . . . . 273
28.7 Proving impossibility results . . . . . . . . . . . . . . . . . . . 275
28.7.1 k-connectivity . . . . . . . . . . . . . . . . . . . . . . . 275
28.7.2 Impossibility proofs for specific problems . . . . . . . 276

29 Approximate agreement 278

29.1 Algorithms for approximate agreement . . . . . . . . . . . . . 278
29.2 Lower bound on step complexity . . . . . . . . . . . . . . . . 281

III Direct interaction 283

30 Self-stabilization 284

31 Population protocols 285

32 Chemical reaction networks 286

33 Mobile robots 287

34 Self-assembly 288

Appendix 290

A Assignments 290
A.1 Assignment 1: due Wednesday, 2016-02-17, at 5:00pm . . . . 290
A.2 Assignment 2: due Wednesday, 2016-03-09, at 5:00pm . . . . 290
A.3 Assignment 3: due Wednesday, 2016-04-20, at 5:00pm . . . . 291
CONTENTS x

B Sample assignments from Spring 2014 292

B.1 Assignment 1: due Wednesday, 2014-01-29, at 5:00pm . . . . 292
B.1.1 Counting evil processes . . . . . . . . . . . . . . . . . 292
B.1.2 Avoiding expensive processes . . . . . . . . . . . . . . 293
B.2 Assignment 2: due Wednesday, 2014-02-12, at 5:00pm . . . . 295
B.2.1 Synchronous agreement with weak failures . . . . . . . 295
B.2.2 Byzantine agreement with contiguous faults . . . . . . 296
B.3 Assignment 3: due Wednesday, 2014-02-26, at 5:00pm . . . . 297
B.3.1 Among the elect . . . . . . . . . . . . . . . . . . . . . 297
B.3.2 Failure detectors on the cheap . . . . . . . . . . . . . . 298
B.4 Assignment 4: due Wednesday, 2014-03-26, at 5:00pm . . . . 299
B.4.1 A global synchronizer with a global clock . . . . . . . 299
B.4.2 A message-passing counter . . . . . . . . . . . . . . . 299
B.5 Assignment 5: due Wednesday, 2014-04-09, at 5:00pm . . . . 300
B.5.1 A concurrency detector . . . . . . . . . . . . . . . . . 300
B.5.2 Two-writer sticky bits . . . . . . . . . . . . . . . . . . 302
B.6 Assignment 6: due Wednesday, 2014-04-23, at 5:00pm . . . . 303
B.6.1 A rotate register . . . . . . . . . . . . . . . . . . . . . 303
B.6.2 A randomized two-process test-and-set . . . . . . . . . 304
B.7 CS465/CS565 Final Exam, May 2nd, 2014 . . . . . . . . . . . 307
B.7.1 Maxima (20 points) . . . . . . . . . . . . . . . . . . . 307
B.7.2 Historyless objects (20 points) . . . . . . . . . . . . . 308
B.7.3 Hams (20 points) . . . . . . . . . . . . . . . . . . . . . 309
B.7.4 Mutexes (20 points) . . . . . . . . . . . . . . . . . . . 310

C Sample assignments from Fall 2011 312

C.1 Assignment 1: due Wednesday, 2011-09-28, at 17:00 . . . . . 312
C.1.1 Anonymous algorithms on a torus . . . . . . . . . . . 312
C.1.2 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . 313
C.1.3 Negotiation . . . . . . . . . . . . . . . . . . . . . . . . 314
C.2 Assignment 2: due Wednesday, 2011-11-02, at 17:00 . . . . . 315
C.2.1 Consensus with delivery notifications . . . . . . . . . . 315
C.2.2 A circular failure detector . . . . . . . . . . . . . . . . 316
C.2.3 An odd problem . . . . . . . . . . . . . . . . . . . . . 318
C.3 Assignment 3: due Friday, 2011-12-02, at 17:00 . . . . . . . . 319
C.3.1 A restricted queue . . . . . . . . . . . . . . . . . . . . 319
C.3.2 Writable fetch-and-increment . . . . . . . . . . . . . . 320
C.3.3 A box object . . . . . . . . . . . . . . . . . . . . . . . 321
C.4 CS465/CS565 Final Exam, December 12th, 2011 . . . . . . . 322
C.4.1 Lockable registers (20 points) . . . . . . . . . . . . . . 322
CONTENTS xi

C.4.2 Byzantine timestamps (20 points) . . . . . . . . . . . 323

C.4.3 Failure detectors and k-set agreement (20 points) . . . 324
C.4.4 A set data structure (20 points) . . . . . . . . . . . . . 325

D Additional sample final exams 326

D.1 CS425/CS525 Final Exam, December 15th, 2005 . . . . . . . 326
D.1.1 Consensus by attrition (20 points) . . . . . . . . . . . 326
D.1.2 Long-distance agreement (20 points) . . . . . . . . . . 327
D.1.3 Mutex appendages (20 points) . . . . . . . . . . . . . 329
D.2 CS425/CS525 Final Exam, May 8th, 2008 . . . . . . . . . . . 330
D.2.1 Message passing without failures (20 points) . . . . . . 330
D.2.2 A ring buffer (20 points) . . . . . . . . . . . . . . . . . 330
D.2.3 Leader election on a torus (20 points) . . . . . . . . . 331
D.2.4 An overlay network (20 points) . . . . . . . . . . . . . 332
D.3 CS425/CS525 Final Exam, May 10th, 2010 . . . . . . . . . . 333
D.3.1 Anti-consensus (20 points) . . . . . . . . . . . . . . . . 333
D.3.2 Odd or even (20 points) . . . . . . . . . . . . . . . . . 334
D.3.3 Atomic snapshot arrays using message-passing (20 points)334
D.3.4 Priority queues (20 points) . . . . . . . . . . . . . . . 336

E I/O automata 338

E.1 Low-level view: I/O automata . . . . . . . . . . . . . . . . . . 338
E.1.1 Enabled actions . . . . . . . . . . . . . . . . . . . . . . 338
E.1.2 Executions, fairness, and traces . . . . . . . . . . . . . 339
E.1.3 Composition of automata . . . . . . . . . . . . . . . . 339
E.1.4 Hiding actions . . . . . . . . . . . . . . . . . . . . . . 340
E.1.5 Fairness . . . . . . . . . . . . . . . . . . . . . . . . . . 340
E.1.6 Specifying an automaton . . . . . . . . . . . . . . . . 341
E.2 High-level view: traces . . . . . . . . . . . . . . . . . . . . . . 341
E.2.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 342
E.2.2 Types of trace properties . . . . . . . . . . . . . . . . 342
E.2.2.1 Safety properties . . . . . . . . . . . . . . . . 342
E.2.2.2 Liveness properties . . . . . . . . . . . . . . . 343
E.2.2.3 Other properties . . . . . . . . . . . . . . . . 344
E.2.3 Compositional arguments . . . . . . . . . . . . . . . . 344
E.2.3.1 Example . . . . . . . . . . . . . . . . . . . . 345
E.2.4 Simulation arguments . . . . . . . . . . . . . . . . . . 345
E.2.4.1 Example . . . . . . . . . . . . . . . . . . . . 346
CONTENTS xii

Bibliography 347

Index 364
List of Figures

2.1 Asynchronous message-passing execution . . . . . . . . . . . . 13

2.2 Asynchronous message-passing execution with FIFO channels 13
2.3 Synchronous message-passing execution . . . . . . . . . . . . 14
2.4 Asynchronous time . . . . . . . . . . . . . . . . . . . . . . . . 15

6.1 Labels in the bit-reversal ring with n = 32 . . . . . . . . . . . 46

10.1 Synthetic execution for Byzantine agreement lower bound . . 70

10.2 Synthetic execution for Byzantine agreement connectivity . . 71

13.1 Failure detector classes . . . . . . . . . . . . . . . . . . . . . . 96

14.1 Figure 2 from [NW98] . . . . . . . . . . . . . . . . . . . . . . 106

21.1 Snapshot from max arrays [AACHE12] . . . . . . . . . . . . . 189

24.1 A 6 × 6 Moir-Anderson grid . . . . . . . . . . . . . . . . . . . 218

24.2 Path through a Moir-Anderson grid . . . . . . . . . . . . . . 219
24.3 A sorting network . . . . . . . . . . . . . . . . . . . . . . . . 223

28.1 Subdivision corresponding to one round of immediate snapshot267

28.2 Subdivision corresponding to two rounds of immediate snapshot268
28.3 An attempt at 2-set agreement . . . . . . . . . . . . . . . . . 269
28.4 Output complex for renaming with n = 3, m = 4 . . . . . . . 277

B.1 Connected Byzantine nodes take over half a cut . . . . . . . . 297

xiii
List of Tables

18.1 Position of various types in the wait-free hierarchy . . . . . . 146

xiv
List of Algorithms

2.1 Client-server computation: client code . . . . . . . . . . . . . . 11

2.2 Client-server computation: server code . . . . . . . . . . . . . 11

4.1 Basic flooding algorithm . . . . . . . . . . . . . . . . . . . . . 23

4.2 Flooding with parent pointers . . . . . . . . . . . . . . . . . . 24
4.3 Convergecast . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.4 Flooding and convergecast combined . . . . . . . . . . . . . . . 28

5.1 AsynchBFS algorithm (from [Lyn96]) . . . . . . . . . . . . . . 30

6.1 LCR leader election . . . . . . . . . . . . . . . . . . . . . . . . 38

6.2 Peterson’s leader-election algorithm . . . . . . . . . . . . . . . 41

10.1 Byzantine agreement: phase king . . . . . . . . . . . . . . . . . 77

13.1 Boosting completeness . . . . . . . . . . . . . . . . . . . . . . . 94

13.2 Consensus with a strong failure detector . . . . . . . . . . . . . 97
13.3 Reliable broadcast . . . . . . . . . . . . . . . . . . . . . . . . . 99

17.1 Mutual exclusion using test-and-set . . . . . . . . . . . . . . . 127

17.2 Mutual exclusion using a queue . . . . . . . . . . . . . . . . . 128
17.3 Mutual exclusion using read-modify-write . . . . . . . . . . . . 129
17.4 Peterson’s mutual exclusion algorithm for two processes . . . . 130
17.5 Implementation of a splitter . . . . . . . . . . . . . . . . . . . 134
17.6 Lamport’s Bakery algorithm . . . . . . . . . . . . . . . . . . . 136
17.7 Yang-Anderson mutex for two processes . . . . . . . . . . . . . 141

18.1 Determining the winner of a race between 2-register writes . . 151

18.2 A universal construction based on consensus . . . . . . . . . . 155

19.1 Snapshot of [AAD+ 93] using unbounded registers . . . . . . . 159

xv
LIST OF ALGORITHMS xvi

19.2 Lattice agreement snapshot . . . . . . . . . . . . . . . . . . . . 165

19.3 Update for lattice agreement snapshot . . . . . . . . . . . . . . 166
19.4 Increasing set data structure . . . . . . . . . . . . . . . . . . . 169
19.5 Single-scanner snapshot: scan . . . . . . . . . . . . . . . . . . 172
19.6 Single-scanner snapshot: update . . . . . . . . . . . . . . . . . 172

21.1 Max register read operation . . . . . . . . . . . . . . . . . . . . 181

21.2 Max register write operations . . . . . . . . . . . . . . . . . . . 181
21.3 Recursive construction of a 2-component max array . . . . . . 186

22.1 Building 2-process TAS from 2-process consensus . . . . . . . . 191

22.2 Two-process one-shot swap from TAS . . . . . . . . . . . . . . 191
22.3 Tournament algorithm with gate . . . . . . . . . . . . . . . . . 193
22.4 Trap implementation from [AWW93] . . . . . . . . . . . . . . 194
22.5 Single-use swap from [AWW93] . . . . . . . . . . . . . . . . . . 195

23.1 Consensus using adopt-commit . . . . . . . . . . . . . . . . . . 199

23.2 A 2-valued adopt-commit object . . . . . . . . . . . . . . . . . 201
23.3 Impatient first-mover conciliator from [Asp12b] . . . . . . . . . 201
23.4 A sifter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
23.5 Test-and-set in O(log log n) expected time . . . . . . . . . . . . 206
23.6 Sifting conciliator (from [Asp12a]) . . . . . . . . . . . . . . . . 207
23.7 Giakkoupis-Woelfel sifter [GW12a] . . . . . . . . . . . . . . . . 208

24.1 Wait-free deterministic renaming . . . . . . . . . . . . . . . . . 214

24.2 Releasing a name . . . . . . . . . . . . . . . . . . . . . . . . . 216
24.3 Implementation of a splitter . . . . . . . . . . . . . . . . . . . 217

25.1 Overlapping LL/SC . . . . . . . . . . . . . . . . . . . . . . . . 229

26.1 Obstruction-free 2-process test-and-set . . . . . . . . . . . . . 236

26.2 Obstruction-free deque . . . . . . . . . . . . . . . . . . . . . . 238
26.3 Obstruction-freedom booster from [FLMS05] . . . . . . . . . . 241

27.1 Safe agreement (adapted from [BGLR01]) . . . . . . . . . . . . 254

28.1 Participating set . . . . . . . . . . . . . . . . . . . . . . . . . . 273

29.1 Approximate agreement . . . . . . . . . . . . . . . . . . . . . . 279

B.1 Counter algorithm for Problem B.4.2. . . . . . . . . . . . . . . 300

B.2 Two-process consensus using the object from Problem B.5.1 . . 301
LIST OF ALGORITHMS xvii

B.3 Implementation of a rotate register . . . . . . . . . . . . . . . 305

B.4 Randomized two-process test-and-set for B.6.2 . . . . . . . . . 305
B.5 Mutex using a swap object and register . . . . . . . . . . . . . 310

C.1 Resettable fetch-and-increment . . . . . . . . . . . . . . . . . . 321

C.2 Consensus using a lockable register . . . . . . . . . . . . . . . 322
C.3 Timestamps with n ≥ 3 and one Byzantine process . . . . . . . 324
C.4 Counter from set object . . . . . . . . . . . . . . . . . . . . . . 325

E.1 Spambot as an I/O automaton . . . . . . . . . . . . . . . . . . 341

Preface

These are notes for the Spring 2016 semester version of the Yale course CPSC
465/565 Theory of Distributed Systems. This document also incorporates
the lecture schedule and assignments, as well as some sample assignments
from previous semesters. Because this is a work in progress, it will be
updated frequently over the course of the semester.
Notes from Fall 2011 can be found at http://www.cs.yale.edu/homes/
aspnes/classes/469/notes-2011.pdf.
Notes from Spring 2014 can be found at http://www.cs.yale.edu/
homes/aspnes/classes/469/notes-2014.pdf.
Notes from earlier semesters can be found at http://pine.cs.yale.
edu/pinewiki/465/.
Much of the structure of the course follows the textbook, Attiya and
Welch’s Distributed Computing [AW04], with some topics based on Lynch’s
Distributed Algorithms [Lyn96] and additional readings from the research
literature. In most cases you’ll find these materials contain much more
detail than what is presented here, so it is better to consider this document
a supplement to them than to treat it as your primary source of information.

Acknowledgments
Many parts of these notes were improved by feedback from students taking
various versions of this course. I’d like to thank Mike Marmar and Hao Pan
in particular for suggesting improvements to some of the posted solutions.
I’d also like to apologize to the many other students who should be thanked
here but whose names I didn’t keep track of in the past.

xviii
Syllabus

Description
Models of asynchronous distributed computing systems. Fundamental con-
cepts of concurrency and synchronization, communication, reliability, topo-
logical and geometric constraints, time and space complexity, and distributed
algorithms.

Meeting times
Lectures are MW 2:30–3:45 in AKW 200.

On-line course information

The lecture schedule, course notes, and all assignments can be found in a sin-
gle gigantic PDF file at http://www.cs.yale.edu/homes/aspnes/classes/
465/notes.pdf. You should probably bookmark this file, as it will be up-
dated frequently.

Staff
The instructor for the course is James Aspnes. Office: AKW 401. Email:
james.aspnes@gmail.com. URL: http://www.cs.yale.edu/homes/aspnes/.
Office hours can be found in the course calendar at Google Calendar,
which can also be reached through James Aspnes’s web page.

xix
SYLLABUS xx

Textbook
Hagit Attiya and Jennifer Welch, Distributed Computing: Fundamentals,
Simulations, and Advanced Topics, second edition. Wiley, 2004. QA76.9.D5
A75X 2004 (LC). ISBN 0471453242.
On-line version: http://dx.doi.org/10.1002/0471478210. (This may
not work outside Yale.)
Errata: http://www.cs.technion.ac.il/~hagit/DC/2nd-errata.html.

Reserved books at Bass Library

Nancy A. Lynch, Distributed Algorithms. Morgan Kaufmann, 1996. ISBN
1558603484. QA76.9 D5 L963X 1996 (LC). Definitive textbook on formal
analysis of distributed systems.
Ajay D. Kshemkalyani and Mukesh Singhal. Distributed Computing:
Principles, Algorithms, and Systems. Cambridge University Press, 2008.
QA76.9.D5 K74 2008 (LC). ISBN 9780521876346. A practical manual of
algorithms with an emphasis on message-passing models.

Course requirements
If you are taking this as CPSC 465: Three homework assignments (60% of
the semester grade) plus a final exam (40%).
If you are taking this as CPSC 565: Three homework assignments (45%
of the semester grade), a presentation (15%) and a final exam (40%).
Each presentation will be a short description of the main results in a
relevant paper chosen in consultation with the instructor, and will be done
in front of the class during one of the last few lecture slots. If numbers and
time permit, it may be possible to negotiate doing a presentation even if
you are taking this as CPSC 465.

Use of outside help

Students are free to discuss homework problems and course material with
each other, and to consult with the instructor or a TF. Solutions handed in,
however, should be the student’s own work. If a student benefits substan-
tially from hints or solutions received from fellow students or from outside
sources, then the student should hand in their solution but acknowledge
the outside sources, and we will apportion credit accordingly. Using outside
resources in solving a problem is acceptable but plagiarism is not.
SYLLABUS xxi

Questions and comments

Please feel free to send questions or comments on the class or anything
connected to it to the instructor at james.aspnes@gmail.com.
For questions about assignments, you may be able to get a faster re-
sponse using Piazza, at http://piazza.com/yale/spring2016/cpsc465.
Note that questions you ask there are visible to other students if not specifi-
cally marked private, so be careful about broadcasting your draft solutions.

Late assignments
Late assignments will not be accepted without a Dean’s Excuse.

Academic integrity statement

The graduate school asks that the following statement be included in all
graduate course syllabi:

Academic integrity is a core institutional value at Yale. It

means, among other things, truth in presentation, diligence and
precision in citing works and ideas we have used, and acknowl-
edging our collaborations with others. In view of our commit-
ment to maintaining the highest standards of academic integrity,
the Graduate School Code of Conduct specifically prohibits the
following forms of behavior: cheating on examinations, problem
sets and all other forms of assessment; falsification and/or fab-
rication of data; plagiarism, that is, the failure in a dissertation,
essay or other written exercise to acknowledge ideas, research, or
language taken from others; and multiple submission of the same
work without obtaining explicit written permission from both in-
structors before the material is submitted. Students found guilty
of violations of academic integrity are subject to one or more of
the following penalties: written reprimand, probation, suspen-
sion (noted on a student’s transcript) or dismissal (noted on a
student’s transcript).
Lecture schedule

As always, the future is uncertain, so you should take parts of the schedule
that haven’t happened yet with a grain of salt. Readings refer to chapters
or sections in the course notes, except for those specified as in AW, which
refer to the course textbook Attiya and Welch [AW04].

2016-01-20 What is distributed computing and why do we have a theory of

it? Basic models: message passing, shared memory, local interactions.
Configurations, events, executions, and schedules. The adversary. Ba-
sic message-passing model. Fairness. Performance measures. Safety
and liveness. Examples of some simple protocols. Indistinguishability
and the Two Generals impossibility result. Readings: Chapters 1, 2,
and 3 up through §3.2; AW Chapter 1.

2016-01-22 Drawing message-passing executions. Synchronous vs. asyn-

chronous systems. Broadcast, convergecast, and distributed breadth-
first search. Proving correctness results and complexity bounds. Read-
ings: §§2.2–2.4, Chapter 4, Chapter 5 (but mostly just §5.1); AW
Chapter 2.

2016-01-25 Leader election. Impossibility without symmetry breaking.

Algorithms for rings and general networks. Lower bounds under vari-
ous conditions. Readings: Chapter 6; AW Chapter 3.

2016-01-27 Causal ordering, logical clocks, Chandy-Lamport snapshots.

Readings: Chapter 7; AW §§6.1–6.2, Chapter 11.

2016-02-01 Synchronizers and the session problem. Readings: Chapter 8;

AW Chapter 11.

2016-02-03 [[[ To be announced. ]]][[[ synchronous agreement with

failures ]]]

xxii
LECTURE SCHEDULE xxiii

2016-02-08 [[[ To be announced. ]]]

2016-02-10 [[[ To be announced. ]]]

2016-02-15 [[[ To be announced. ]]]

2016-02-17 [[[ To be announced. ]]]

2016-02-22 [[[ To be announced. ]]]

2016-02-24 [[[ To be announced. ]]]

2016-02-29 [[[ To be announced. ]]]

2016-03-02 [[[ To be announced. ]]]

2016-03-07 [[[ To be announced. ]]]

2016-03-09 [[[ To be announced. ]]]

2016-03-28 [[[ To be announced. ]]]

2016-03-30 [[[ To be announced. ]]]

2016-04-04 [[[ To be announced. ]]]

2016-04-06 [[[ To be announced. ]]]

2016-04-11 [[[ To be announced. ]]]

2016-04-13 [[[ To be announced. ]]]

2016-04-18 [[[ To be announced. ]]]

2016-04-20 [[[ To be announced. ]]]

2016-04-25 [[[ To be announced. ]]]

2016-04-27 [[[ To be announced. ]]]

2016-05-10 The final exam will be given Tuesday, 2016-05-10, starting

at 9:00, in a location to be determined by the registrar. It will be a
closed-book exam covering all material discussed during the semester.
Past final exams and their solutions can be found in Appendix D.
LECTURE SCHEDULE xxiv

Topics not yet scheduled

The order of this list is based on the previous incarnation of the course, and
may change.

• Synchronous agreement: lower bounds and algorithms for the crash-

failure model. Impossibility of Byzantine agreement with n ≤ 3f .
Readings: Chapter 9, §10.1.2; AW §§5.1, 5.2.1–5.2.3.

• More Byzantine agreement: additional impossibility results, the expo-

nential information gathering algorithm. Readings: §§10.1.3–10.2.1;
AW §5.2.4.

• Phase king algorithm for Byzantine agreement. Bivalence arguments

and the Fischer-Lynch-Paterson impossibility proof for asynchronous
agreement with one crash failure. Doing asynchronous agreement
anyway using Paxos. Readings: §10.2.2, Chapter 11, Chapter 12;
AW §5.2.5–5.3, [Lam01].

• Failure detectors: classification of failure detectors, consensus using S

and ♦S. Readings: Chapter 13 up through 13.4; [CT96].

• Impossibility results for failure detectors. Readings: §§13.5–13.6, Chap-

ter 7 through §7.2.2; AW §§6.1.1–6.1.2.

• Shared memory and distributed shared memory. Readings: Chap-

ter 15, Chapter 16; AW §§9.1 and 9.3.

• Quorum systems. Readings: Chapter 14; [NW98].

• Start of mutual exclusion: problem definition, algorithms for strong

primitives, Peterson’s tournament algorithm. A bit about splitters
and fast mutex, although I ran over before getting to the punchline
(see §17.4.2). Readings: Chapter 17 through §17.4.2; AW §§4.1–4.3.2,
4.4.2–4.4.3, 4.4.5.

• More mutex: return of the splitters, Lamport’s bakery algorithm,

Burns-Lynch lower bound on space, RMR complexity. Readings: §§17.4.2–
17.5.2. AW 4.4.4, [YA95].

• End of mutex: The Yang-Anderson algorithm for low RMR mutex in

the distributed shared memory model, a few more comments on lower
LECTURE SCHEDULE xxv

bounds. Wait-free computation and universality of consensus. Read-

ings: §§17.5.3, 17.5.4, and 18.2, also a bit of the start of Chapter 18;
[Her91b].

• The wait-free hierarchy: consensus number of various objects. Read-

ings: rest of Chapter 18 (except §18.1.4).

• Consensus number of simultaneous m-register write. Atomic snapshots

of shared memory: definition, the Afek et al. algorithm, applications,
reduction to lattice agreement. Readings: §18.1.4, Chapter 19 through
§19.2.1, §19.5, §§19.3.1–19.3.3; AW §10.3, [AHR95].

• Implementing lattice agreement. Perturbable objects and the Jayanti-

Tan-Toueg lower bound. How bounded max registers escape the bound.
Readings: §19.3.5, Chapter 20, §21.1. [IMCT94, JTT00, AAC09].

• More on restricted-use objects: max register variants, lower bounds for

max registers, max arrays and restricted-use snapshots with polyloga-
rithmic cost. How 2-process test-and-set (and by extension any object
with consensus number 2) implements all historyless objects. Read-
ings: rest of Chapter 21, Chapter 22; [AAC09, AACHE12, AWW93].

• Randomized consensus: adversaries, adopt-commits, and conciliators.

Reaching agreement in the Bracha-Rachman (for adaptive adversary)
and Chor-Israeli-Li (for weak adversary) algorithms. Readings: Chap-
ter 23 through §23.5.

• Faster algorithms for randomized test-and-set. Splitters as weak test-

and-set algorithms. Adaptive algorithms and RatRace. Readings:
§23.6 through §23.6.1, §24.4.3.1, §24.5.2; [AAG+ 10, AA11, GW12a].

• More randomized consensus: O(log log n)-time consensus for an obliv-

ious adversary. Renaming: definition, renaming to 2n − 1 names using
snapshots. Readings: §§23.6.2, 24.1, 24.2, and 24.4.1; [Asp12a], AW
§16.3.1.

• More renaming: renaming using splitters, randomized renaming. Read-

ings: §§24.4.3–24.5; [MA95, AAGG11, AAGW13].

• Solvability of asynchronous decision tasks: BG simulation of n-process

executions with f failures with f +1-process wait-free executions, start
of topological methods. Readings: Chapter 27, §28.3.1; [BG97].
LECTURE SCHEDULE xxvi

• Rest of topological methods: iterated immediate snapshots, subdivi-

sions, impossibility of k-set agreement with k failures, how renaming
is like trying to turn a sphere into a torus. Readings: rest of Chap-
ter 28; AW §16.1 if you want to see a non-topological proof of the k-set
agreement result, [HS99] for more of the topological approach.
Chapter 1

Introduction

Distributed systems are characterized by their structure: a typical dis-

tributed system will consist of some large number of interacting devices that
each run their own programs but that are affected by receiving messages,
or observing shared-memory updates or the states of other devices. Exam-
ples of distributed computing systems range from simple systems in which
a single client talks to a single server to huge amorphous networks like the
Internet as a whole.
As distributed systems get larger, it becomes harder and harder to pre-
dict or even understand their behavior. Part of the reason for this is that
we as programmers have not yet developed a standardized set of tools for
managing complexity (like subroutines or objects with narrow interfaces,
or even simple structured programming mechanisms like loops or if/then
statements) as are found in sequential programming. Part of the reason is
that large distributed systems bring with them large amounts of inherent
nondeterminism—unpredictable events like delays in message arrivals, the
sudden failure of components, or in extreme cases the nefarious actions of
faulty or malicious machines opposed to the goals of the system as a whole.
Because of the unpredictability and scale of large distributed systems, it can
often be difficult to test or simulate them adequately. Thus there is a need
for theoretical tools that allow us to prove properties of these systems that
will let us use them with confidence.
The first task of any theory of distributed systems is modeling: defining a
mathematical structure that abstracts out all relevant properties of a large
distributed system. There are many foundational models for distributed
systems, but for this class we will follow [AW04] and use simple automaton-
based models.

1
CHAPTER 1. INTRODUCTION 2

What this means is that we model each process in the system as an

automaton that has some sort of local state, and model local computation
as a transition rule that tells us how to update this state in response to
various events. Depending on what kinds of system we are modeling, these
events might correspond to local computation, to delivery of a message by
a network, carrying out some operation on a shared memory, or even some-
thing like a chemical reaction between two molecules. The transition rule
for a system specifies how the states of all processes involved in the event
are updated, based on their previous states. We can think of the transition
rule as an arbitrary mathematical function (or relation if the processes are
nondeterministic); this corresponds in programming terms to implementing
local computation by processes as a gigantic table lookup.
Obviously this is not how we program systems in practice. But what this
approach does is allow us to abstract away completely from how individual
processes work, and emphasize how all of the processes interact with each
other. This can lead to odd results: for example, it’s perfectly consistent
with this model for some process to be able to solve the halting problem,
or carry out arbitrarily complex calculations between receiving a message
and sending its response. A partial justification for this assumption is that
in practice, the multi-millisecond latencies in even reasonably fast networks
are eons in terms of local computation. And as with any assumption, we
can always modify it if it gets us into trouble.

1.1 Models
The global state consisting of all process states is called a configuration,
and we think of the system as a whole as passing from one global state
or configuration to another in response to each event. When this occurs
the processes participating in the event update their states, and the other
processes do nothing. This does not model concurrency directly; instead,
we interleave potentially concurrent events in some arbitrary way. The ad-
vantage of this interleaving approach is that it gives us essentially the same
behavior as we would get if we modeled simultaneous events explicitly, but
still allows us to consider only one event at a time and use induction to
prove various properties of the sequence of configurations we might reach.
We will often use lowercase Greek letters for individual events or se-
quences of events. Configurations are typically written as capital Latin
letters (often C). An execution of a schedule is an alternating sequence of
configurations and events C0 σ0 C1 σ1 C2 . . . , where Ci+1 is the configuration
CHAPTER 1. INTRODUCTION 3

that results from applying event σi to configuration C. A schedule is the

sequence of events σ0 σ1 . . . from some execution. We say that an event σ is
enabled in C if this event can be carried out in C; an example would be that
the event that we deliver a particular message in a message-passing system
is enabled only if that message has been sent and not yet delivered. When σ
is enabled in C, it is sometime convenient to write Cσ for the configuration
that results from applying σ to C.
What events are available, and what effects they have, will depend on
what kind of model we are considering. We may also have additional con-
straints on what kinds of schedules are admissible, which restricts the
schedules under considerations to those that have certain desirable proper-
ties (say, every message that is sent is eventually delivered). There are many
models in the distributed computing literature, which can be divided into a
handful of broad categories:
• Message passing models (which we will cover in Part I) correspond
to systems where processes communicate by sending messages through
a network. In synchronous message-passing, every process sends
out messages at time t that are delivered at time t + 1, at which
point more messages are sent out that are delivered at time t + 2,
and so on: the whole system runs in lockstep, marching forward in
perfect synchrony.1 Such systems are difficult to build when the com-
ponents become too numerous or too widely dispersed, but they are
often easier to analyze than asynchronous systems, where messages
are only delivered eventually after some unknown delay. Variants on
these models include semi-synchronous systems, where message de-
lays are unpredictable but bounded, and various sorts of timed sys-
tems. Further variations come from restricting which processes can
communicate with which others, by allowing various sorts of failures:
crash failures that stop a process dead, Byzantine failures that
turn a process evil, or omission failures that drop messages in tran-
sit. Or—on the helpful side—we may supply additional tools like fail-
ure detectors (Chapter 13) or randomization (Chapter 23).

• Shared-memory models (Part II) correspond to systems where pro-

cesses communicate by executing operations on shared objects that in
the simplest case are typically simple memory cells supporting read
1
In an interleaving model, these apparently simultaneous events are still recorded one
at a time. What makes the system synchronous is that we demand that, in any admissible
schedule, all the send events for time t occur as a sequential block, followed by all the
receive events for time t + 1, and so on.
CHAPTER 1. INTRODUCTION 4

and write operations (), but which could be more complex hardware
primitives like compare-and-swap (§18.1.3), load-linked/store-
conditional (§18.1.3), atomic queues, or more exotic objects from
the seldom-visited theoretical depths. Practical shared-memory sys-
tems may be implemented as distributed shared-memory (Chap-
ter 16) on top of a message-passing system in various ways.
Like message-passing systems, shared-memory systems must also deal
with issues of asynchrony and failures, both in the processes and in
the shared objects.
Realistic shared-memory systems have additional complications, in
that modern CPUs allow out-of-order execution in the absence of spe-
cial (and expensive) operations called fences or memory barriers.
We will effectively be assuming that our shared-memory code is liber-
ally sprinkled with these operations to ensure atomicity, but this is not
always true of real production code, and indeed there is work in the
theory of distributed computing literature on algorithms that don’t
require unlimited use of memory barriers.[[[ citation needed ]]]

• A third family of models involves direct interaction between processes,

where a process may be able to directly observe the states of other
processes. These models are used in analyzing self-stabilization,
for some biologically inspired systems, and for computation by
chemical reaction networks. We will discuss some of this work in
Part III.

• Other specialized models emphasize particular details of distributed

systems, such as the labeled-graph models used for analyzing routing
or the topological models used to represent some specialized agreement
problems (see Chapter 28.

We’ll see many of these at some point in this course, and examine which
of them can simulate each other under various conditions.

1.2 Properties
Properties we might want to prove about a system include:

• Safety properties, of the form “nothing bad ever happens” or more

precisely “there are no bad reachable states of the system.” These
include things like “at most one of the traffic lights at the intersection
CHAPTER 1. INTRODUCTION 5

of Busy and Main is ever green.” Such properties are typically proved
using invariants, properties of the state of the system that are true
initially and that are preserved by all transitions; this is essentially a
disguised induction proof.

• Liveness properties, of the form “something good eventually hap-

pens.” An example might be “my email is eventually either delivered or
returned to me.” These are not properties of particular states (I might
unhappily await the eventual delivery of my email for decades with-
out violating the liveness property just described), but of executions,
where the property must hold starting at some finite time. Liveness
properties are generally proved either from other liveness properties
(e.g., “all messages in this message-passing system are eventually de-
livered”) or from a combination of such properties and some sort of
timer argument where some progress metric improves with every tran-
sition and guarantees the desirable state when it reaches some bound
(also a disguised induction proof).

• Fairness properties are a strong kind of liveness property of the form

“something good eventually happens to everybody.” Such properties
exclude starvation, a situation where most of the kids are happily
chowing down at the orphanage (“some kid eventually eats something”
is a liveness property) but poor Oliver Twist is dying for lack of gruel
in the corner.

• Simulations show how to build one kind of system from another,

such as a reliable message-passing system built on top of an unreliable
system (TCP[[[ ref? ]]]), a shared-memory system built on top of
a message-passing system (distributed shared memory[[[ ref ]]]), or
a synchronous system build on top of an asynchronous system (syn-
chronizers—see Chapter 8).

• Impossibility results describe things we can’t do. For example, the

classic Two Generals impossibility result (Chapter 3) says that it’s
impossible to guarantee agreement between two processes across an
unreliable message-passing channel if even a single message can be
lost. Other results characterize what problems can be solved if various
fractions of the processes are unreliable, or if asynchrony makes timing
assumptions impossible. These results, and similar lower bounds that
describe things we can’t do quickly, include some of the most tech-
nically sophisticated results in distributed computing. They stand in
CHAPTER 1. INTRODUCTION 6

contrast to the situation with sequential computing, where the reli-

ability and predictability of the underlying hardware makes proving
lower bounds extremely difficult.

There are some basic proof techniques that we will see over and over
again in distributed computing.
For lower bound and impossibility proofs, the main tool is the in-
distinguishability argument. Here we construct two (or more) executions
in which some process has the same input and thus behaves the same way,
regardless of what algorithm it is running. This exploitation of process’s
ignorance is what makes impossibility results possible in distributed com-
puting despite being notoriously difficult in most areas of computer science.2
For safety properties, statements that some bad outcome never occurs,
the main proof technique is to construct an invariant. An invariant is es-
sentially an induction hypothesis on reachable configurations of the system;
an invariant proof shows that the invariant holds in all initial configurations,
and that if it holds in some configuration, it holds in any configuration that
is reachable in one step.
Induction is also useful for proving termination and liveness proper-
ties, statements that some good outcome occurs after a bounded amount of
time. Here we typically structure the induction hypothesis as a progress
measure, showing that some sort of partial progress holds by a particular
time, with the full guarantee implied after the time bound is reached.

2
An exception might be lower bounds for data structures, which also rely on a process’s
ignorance.
 Part I

Message passing

7
Chapter 2

Model

We’re going to use a simplified version of the model in [AW04, Chapter 2].
The main difference is that Attiya and Welch include an inbuf component in
the state of each process for holding incoming messages, and have delivery
events only move messages from one process’s outbuf to another’s inbuf. We
will have delivery events cause the receiving process to handle the incoming
message immediately, which is closer to how most people write message-
passing algorithms, and which serves as an example of a good policy for
anybody dealing with a lot of email. It’s not hard to show that this doesn’t
change what we can or can’t do in the system, because (in one direction) the
inbuf-less processes can always simulate an inbuf, and (in the other direction)
we can pretend that messages stored in the inbuf are only really delivered
when they are processed. We will still allow local computation events that
don’t require a message to be delivered, to allow processes to take action
spontaneously, but we won’t use them as much.

2.1 Basic message-passing model

We have a collection of n processes p1 . . . p2 , each of which has a state con-
sisting of a state from from state set Qi , together with an outbuf component
representing messages posted to be sent that have not yet been delivered.
Messages are point-to-point, with a single sender and recipient: if you want
broadcast, you have to pay for it. A configuration of the system consists
of a vector of states, one for each process. The configuration of the sys-
tem is updated by an event, which is either a delivery event (a message
is removed from some process’s outbuf, and the receiver gets to update its
state in response and possibly send one or more messages by adding them to

8
CHAPTER 2. MODEL 9

its own outbuf) or a computation event (some process updates its state
and possibly adds new messages to its outbuf). An execution segment
is a sequence of alternating configurations and events C0 , φ1 , C1 , φ2 , . . . , in
which each triple Ci φi+1 Ci+1 is consistent with the transition rules for the
event φi+1 , and the last element of the sequence (if any) is a configuration.
If the first configuration C0 is an initial configuration of the system, we
have an execution. A schedule is an execution with the configurations
removed.

2.1.1 Formal details

Let Q be the set of process states and M the set of messages.
Each process i has, in addition to its state statei ∈ Q, a variable outbuf i [j] ⊆
P M for each process j it can send messages to. We assume each process has
a transition function δ : Q × M ∪ {⊥} → QU P P × M that maps tuples
consisting of a message (or ⊥, for computation events) and a state to a new
state and a set of messages to be sent. An important feature of the transition
function is that the process’s behavior can’t depend on which of its previous
messages have been delivered or not. A delivery event del(i, j, m) removes
the message m from outbufi[j] and updates statej and outbuf j according to
δ(statej , m). A computation event comp(i) does the same thing, except that
it applies δ(statej , ⊥).
Some implicit features in this definition:

• A process can’t tell when its outgoing messages are delivered, because
the outbuf i variables aren’t included in the accessible state used as
input to the transition function.

• Processes are deterministic: The next action of each process depends

only on its current state, and not on extrinsic variables like the phase
of the moon, coin-flips, etc. We may wish to relax this condition later
by allowing coin-flips; to do so, we will need to extend the model to
incorporate probabilities.

• It is possible to determine the accessible state of a process by looking

only at events that involve that process. Specifically, given a schedule
S, define the restriction S|i to be the subsequence consisting of all
comp(i) and del(j, i, m) events (ranging over all possible j and m).
Since these are the only events that affect the accessible state of i,
and only the accessible state of i is needed to apply the transition
function, we can compute the accessible state of i looking only at
CHAPTER 2. MODEL 10

S|i. In particular, this means that i will have the same accessible
state after any two schedules S and S 0 where S|i = S 0 |i, and thus
will take the same actions in both schedules. This is the basis for
indistinguishability proofs (§3.2), a central technique in obtaining
lower bounds and impossibility results.

A curious feature of this particular model is that communication chan-

nels are not modeled separately from processes, but instead are baked into
processes as outbuf variables. This leads to some oddities like having to dis-
tinguish the accessible state of a process (which excludes the outbufs) from
the full state (which doesn’t). A different approach (taken, for example,
by [Lyn96]) would be to have separate automata representing processes and
communication channels. But since the resulting model produces essentially
the same executions, the exact details don’t really matter.

2.1.2 Network structure

It may be the case that not all processes can communicate directly; if so,
we impose a network structure in the form of a directed graph, where i
can send a message to j if and only if there is an edge from i to j in the
graph. Typically we assume that each process knows the identity of all its
neighbors.
For some problems (e.g., in peer-to-peer systems or other overlay net-
works) it may be natural to assume that there is a fully-connected un-
derlying network but that we have a dynamic network on top of it, where
processes can only send to other processes that they have obtained the ad-
dresses of in some way.

2.2 Asynchronous systems

In an asynchronous model, only minimal restrictions are placed on when
messages are delivered and when local computation occurs. A schedule is
said to be admissible if (a) there are infinitely many computation steps
for each process, and (b) every message is eventually delivered. (These are
fairness conditions.) The first condition (a) assumes that processes do not
explicitly terminate, which is the assumption used in [AW04]; an alternative,
which we will use when convenient, is to assume that every process either
has infinitely many computation steps or reaches an explicit halting state.
CHAPTER 2. MODEL 11

2.2.1 Example: client-server computing

Almost every distributed system in practical use is based on client-server
interactions. Here one process, the client, sends a request to a second
process, the server, which in turn sends back a response. We can model
this interaction using our asynchronous message-passing model by describing
what the transition functions for the client and the server look like: see
Algorithms 2.1 and 2.2.

1 initially do
2 send request to server
Algorithm 2.1: Client-server computation: client code

1 upon receiving request do

2 send response to client
Algorithm 2.2: Client-server computation: server code

The interpretation of Algorithm 2.1 is that the client sends request (by
adding it to its outbuf) in its very first computation event (after which it does
nothing). The interpretation of Algorithm 2.2 is that in any computation
event where the server observes request in its inbuf, it sends response.
We want to claim that the client eventually receives response in any
admissible execution. To prove this, observe that:
1. After finitely many steps, the client carries out a computation event.
This computation event puts request in its outbuf.
2. After finitely many more steps, a delivery event occurs that delivers
request to the server. This causes the server to send response.
3. After finitely many more steps, a delivery event delivers response to
the client, causing it to process response (and do nothing, given that
we haven’t included any code to handle this response).
Each step of the proof is justified by the constraints on admissible execu-
tions. If we could run for infinitely many steps without a particular process
doing a computation event or a particular message being delivered, we’d
violate those constraints.
Most of the time we will not attempt to prove the correctness of a pro-
tocol at quite this level of tedious detail. But if you are only interested in
CHAPTER 2. MODEL 12

distributed algorithms that people actually use, you have now seen a proof
of correctness for 99.9% of them, and do not need to read any further.

2.3 Synchronous systems

A synchronous message-passing system is exactly like an asynchronous
system, except we insist that the schedule consists of alternating phases in
which (a) every process executes a computation step, and (b) all messages
are delivered. The combination of a computation phase and a delivery phase
is called a round. Synchronous systems are effectively those in which all
processes execute in lock-step, and there is no timing uncertainty. This
makes protocols much easier to design, but makes them less resistant to
real-world timing oddities. Sometimes this can be dealt with by applying
a synchronizer (Chapter 8), which transforms synchronous protocols into
asynchronous protocols at a small cost in complexity.

2.4 Drawing message-passing executions

Though formally we can describe an execution in a message-passing system
as a long list of events, this doesn’t help much with visualizing the underlying
communication pattern. So it can sometimes be helpful to use a more visual
representation of a message-passing execution that shows how information
flows through the system.
A typical example is given in Figure 2.1. In this picture, time flows
from left to right, and each process is represented by a horizontal line. This
convention reflects the fact that processes have memory, so any information
available to a process at some time t is also available at all times t0 ≥ t.
Events are represented by marked points on these lines, and messages are
represented by diagonal lines between events. The resulting picture looks
like a collection of world lines as used in physics to illustrate the path
taken by various objects through spacetime.
Pictures like Figure 2.1 can be helpful for illustrating the various con-
straints we might put on message delivery. In Figure 2.1, the system is
completely asynchronous: messages can be delivered in any order, even if
sent between the same processes. If we run the same protocol under stronger
assumptions, we will get different communication patterns.
For example, Figure 2.2 shows an execution that is still asynchronous but
that assumes FIFO (first-in first-out) channels. A FIFO channel from some
process p to another process q guarantees that q receives messages in the
CHAPTER 2. MODEL 13

p3

p2

p1

Time →

Figure 2.1: Asynchronous message-passing execution. Time flows left-to-

right. Horizontal lines represent processes. Nodes represent events. Diago-
nal edges between events represent messages. In this execution, p1 executes
a computation event that sends messages to p2 and p3 . When p2 receives
this message, it sends messages to p1 and p3 . Later, p2 executes a com-
putation event that sends a second message to p1 . Because the system is
asynchronous, there is no guarantee that messages arrive in the same order
they are sent.

same order that p sends them (this can be simulated by a non-FIFO channel
by adding a sequence number to each message, and queuing messages at
the receiver until all previous messages have been processed).
If we go as far as to assume synchrony, we get the execution in Figure 2.3.
Now all messages take exactly one time unit to arrive, and computation
events follow each other in lockstep.

p3

p2

p1

Time →

Figure 2.2: Asynchronous message-passing execution with FIFO channels.

Multiple messages from one process to another are now guaranteed to be
delivered in the order they are sent.
CHAPTER 2. MODEL 14

p3

p2

p1

Time →

Figure 2.3: Synchronous message-passing execution. All messages are now

delivered in exactly one time unit, and computation events occur exactly
one time unit after the previous event.

2.5 Complexity measures

There is no explicit notion of time in the asynchronous model, but we can
define a time measure by adopting the rule that every message is delivered
and processed at most 1 time unit after it is sent. Formally, we assign time
0 to the first event, and assign the largest time we can to each subsequent
event, subject to the constraints that (a) no event is assigned a larger time
than any later event; (b) if a message m from i to j is created by an event at
time t, then the time for the delivery of m from i to j is no greater than j +1,
and (c) any computation step is assigned a time no later than the previous
event at the same process (or 0 if the process has no previous events). This
is consistent with an assumption that message propagation takes at most 1
time unit and that local computation takes 0 time units.
Another way to look at this is that it is a definition of a time unit in terms
of maximum message delay together with an assumption that message delays
dominate the cost of the computation. This last assumption is pretty much
always true for real-world networks with any non-trivial physical separation
between components, thanks to speed of light limitations.
An example of an execution annotated with times in this way is given in
Figure 2.4.
The time complexity of a protocol (that terminates) is the time of the
last event at any process.
Note that looking at step complexity, the number of computation
events involving either a particular process (individual step complexity)
or all processes (total step complexity) is not useful in the asynchronous
model, because a process may be scheduled to carry out arbitrarily many
computation steps without any of its incoming or outgoing messages being
CHAPTER 2. MODEL 15

p3
1 1
p2
1 1
p1
0 2 2
Time →

Figure 2.4: Asynchronous message-passing execution with times.

delivered, which probably means that it won’t be making any progress.

These complexity measures will be more useful when we look at shared-
memory models (Part II).
For a protocol that terminates, the message complexity is the total
number of messages sent. We can also look at message length in bits, to-
tal bits sent, etc., if these are useful for distinguishing our new improved
protocol from last year’s model.
For synchronous systems, time complexity becomes just the number of
rounds until a protocol finishes. Message complexity is still only loosely
connected to time complexity; for example, there are synchronous leader
election (Chapter 6) algorithms that, by virtue of grossly abusing the syn-
chrony assumption, have unbounded time complexity but very low message
complexity.
Chapter 3

Coordinated attack

(See also [Lyn96, §5.1].)

The Two Generals problem was the first widely-known distributed con-
sensus problem, described in 1978 by Jim Gray [Gra78, §5.8.3.3.1], although
the same problem previously appeared under a different name [AEH75].
The setup of the problem is that we have two generals on opposite sides
of an enemy army, who must choose whether to attack the army or retreat.
If only one general attacks, his troops will be slaughtered. So the generals
need to reach agreement on their strategy.
To complicate matters, the generals can only communicate by sending
messages by (unreliable) carrier pigeon. We also suppose that at some point
each general must make an irrevocable decision to attack or retreat. The
interesting property of the problem is that if carrier pigeons can become
lost, there is no protocol that guarantees agreement in all cases unless the
outcome is predetermined (e.g. the generals always attack no matter what
happens). The essential idea of the proof is that any protocol that does
guarantee agreement can be shortened by deleting the last message; iterating
this process eventually leaves a protocol with no messages.
Adding more generals turns this into the coordinated attack problem,
a variant of consensus; but it doesn’t make things any easier.

3.1 Formal description

To formalize this intuition, suppose that we have n ≥ 2 generals in a syn-
chronous system with unreliable channels—the set of messages received in
round i + 1 is always a subset of the set sent in round i, but it may be a
proper subset (even the empty set). Each general starts with an input 0

16
CHAPTER 3. COORDINATED ATTACK 17

(retreat) or 1 (attack) and must output 0 or 1 after some bounded number

of rounds. The requirements for the protocol are that, in all executions:

Agreement All processes output the same decision (0 or 1).

Validity If all processes have the same input x, and no messages are lost,
all processes produce output x. (If processes start with different inputs
or one or more messages are lost, processes can output 0 or 1 as long
as they all agree.)

Termination All processes terminate in a bounded number of rounds.1

Sadly, there is not protocol that satisfies all three conditions. We show
this in the next section.

3.2 Impossibility proof

To show coordinated attack is impossible,2 we use an indistinguishability
proof.
The basic idea of an indistinguishability proof is this:

• Execution A is indistinguishable from execution B for some process

p if p sees the same things (messages or operation results) in both
executions.

• If A is indistinguishable from B for p, then because p can’t tell which

of these two possible worlds it is in, it returns the same output in both.

So far, pretty dull. But now let’s consider a chain of hypothetical exe-
cutions A = A0 A1 . . . Ak = B, where each Ai is indistinguishable from Ai+1
for some process pi . Suppose also that we are trying to solve an agreement
1
Bounded means that there is a fixed upper bound on the length of any execution.
We could also demand merely that all processes terminate in a finite number of rounds.
In general, finite is a weaker requirement than bounded, but if the number of possible
outcomes at each step is finite (as they are in this case), they’re equivalent. The reason
is that if we build a tree of all configurations, each configuration has only finitely many
successors, and the length of each path is finite, then König’s lemma (see http://en.
wikipedia.org/wiki/Konig’s_lemma) says that there are only finitely many paths. So
we can take the length of the longest of these paths as our fixed bound. [BG97, Lemma
3.1]
2
Without making additional assumptions, always a caveat when discussing impossibil-
ity.
CHAPTER 3. COORDINATED ATTACK 18

task, where every process must output the same value. Then since pi out-
puts the same value in Ai and Ai+1 , every process outputs the same value
in Ai and Ai+1 . By induction on k, every process outputs the same value in
A and B, even though A and B may be very different executions.
This gives us a tool for proving impossibility results for agreement: show
that there is a path of indistinguishable executions between two executions
that are supposed to produce different output. Another way to picture
this: consider a graph whose nodes are all possible executions with an edge
between any two indistinguishable executions; then the set of output-0 exe-
cutions can’t be adjacent to the set of output-1 executions. If we prove the
graph is connected, we prove the output is the same for all executions.
For coordinated attack, we will show that no protocol satisfies all of
agreement, validity, and termination using an indistinguishability argument.
The key idea is to construct a path between the all-0-input and all-1-input
executions with no message loss via intermediate executions that are indis-
tinguishable to at least one process.
Let’s start with A = A0 being an execution in which all inputs are 1 and
all messages are delivered. We’ll build executions A1 , A2 , etc. by pruning
messages. Consider Ai and let m be some message that is delivered in the
last round in which any message is delivered. Construct Ai+1 by not deliv-
ering m. Observe that while Ai is distinguishable from Ai+1 by the recipient
of m, on the assumption that n ≥ 2 there is some other process that can’t
tell whether m was delivered or not (the recipient can’t let that other pro-
cess know, because no subsequent message it sends are delivered in either
execution). Continue until we reach an execution Ak in which all inputs are
1 and no messages are sent. Next, let Ak+1 through Ak+n be obtained by
changing one input at a time from 1 to 0; each such execution is indistin-
guishable from its predecessor by any process whose input didn’t change.
Finally, construct Ak+n through A2k+n by adding back messages in the re-
verse process used for A0 through Ak . This gets us to an execution Ak+n
in which all processes have input and no messages are lost. If agreement
holds, then the indistinguishability of adjacent executions to some process
means that the common output in A0 is the same as in A2k+n . But validity
requires that A0 outputs 1 and A2k+n outputs 0: so validity is violated.

3.3 Randomized coordinated attack

So we now know that we can’t solve the coordinated attack problem. But
maybe we want to solve it anyway. The solution is to change the problem.
CHAPTER 3. COORDINATED ATTACK 19

Randomized coordinated attack is like standard coordinated attack,

but with less coordination. Specifically, we’ll allow the processes to flip coins
to decide what to do, and assume that the communication pattern (which
messages get delivered in each round) is fixed and independent of the coin-
flips. This corresponds to assuming an oblivious adversary that can’t
see what is going on at all or perhaps a content-oblivious adversary
that can only see where messages are being sent but not the contents of the
messages. We’ll also relax the agreement property to only hold with some
high probability:

Randomized agreement For any adversary A, the probability that some

process decides 0 and some other process decides 1 given A is at most
.

Validity and termination are as before.

3.3.1 An algorithm
Here’s an algorithm that gives  = 1/r. (See [Lyn96, §5.2.2] for details
or [VL92] for the original version.) A simplifying assumption is that network
is complete, although a strongly-connected network with r greater than or
equal to the diameter also works.

• First part: tracking information levels

– Each process tracks its “information level,” initially 0. The state

of a process consists of a vector of (input, information-level) pairs
for all processes in the system. Initially this is (my-input, 0) for
itself and (⊥, −1) for everybody else.
– Every process sends its entire state to every other process in every
round.
– Upon receiving a message m, process i stores any inputs carried in
m and, for each process j, sets leveli [j] to max(leveli [j], levelm [j]).
It then sets its own information level to minj (leveli [j]) + 1.

• Second part: deciding the output

– Process 1 chooses a random key value uniformly in the range

[1, r].
– This key is distributed along with leveli [1], so that every process
with leveli [1] ≥ 0 knows the key.
CHAPTER 3. COORDINATED ATTACK 20

– A process decides 1 at round r if and only if it knows the key,

its information level is greater than or equal to the key, and all
inputs are 1.

3.3.2 Why it works

Termination Immediate from the algorithm.

Validity • If all inputs are 0, no process sees all 1 inputs (technically

requires an invariant that processes’ non-null views are consistent
with the inputs, but that’s not hard to prove.)
• If all inputs are 1 and no messages are lost, then the information
level of each process after k rounds is k (prove by induction) and
all processes learn the key and all inputs (immediate from first
round). So all processes decide 1.

Randomized Agreement • First prove a lemma: Define levelti [k] to

be the value of leveli [k] after t rounds. Then for all i, j, k, t, (1)
leveli [j]t ≥ levelj [j]t−1 and (2) leveli [k]t − levelj [k]t ≤ 1. As
always, the proof is by induction on rounds. Part (1) is easy and
boring so we’ll skip it. For part (2), we have:
– After 0 rounds, level0i [k] = level0j [k] = −1 if neither i nor j
equals k; if one of them is k, we have level0k [k] = 0, which is
still close enough.
– After t rounds, consider levelti [k] − levelt−1 i [k] and similarly
leveltj [k] − levelt−1
j [k]. It’s not hard to show that each can
jump by at most 1. If both deltas are +1 or both are 0,
there’s no change in the difference in views and we win from
the induction hypothesis. So the interesting case is when
leveli [k] stays the same and levelj [k] increases or vice versa.
– There are two ways for levelj [k] to increase:
∗ If j 6= k, then j received a message from some j 0 with
leveljt−1
0 [k] > levelj
t−1
[k]. From the induction hypothesis,
t−1 t−1
levelj 0 [k] ≤ leveli [k] + 1 = levelti [k]. So we are happy.
∗ If j = k, then j has leveltj [j] = 1 + mink6=j leveltj [k] ≤
1 + leveltj [i] ≤ 1 + levelti [i]. Again we are happy.
• Note that in the preceding, the key value didn’t figure in; so
everybody’s level at round r is independent of the key.
CHAPTER 3. COORDINATED ATTACK 21

• So now we have that levelri [i] is in {`, ` + 1}, where ` is some fixed
value uncorrelated with key. The only way to get some process
to decide 1 while others decide 0 is if ` + 1 ≥ key but ` < key.
(If ` = 0, a process at this level doesn’t know key, but it can still
reason that 0 < key since key is in [1, r].) This can only occur if
key = ` + 1, which occurs with probability at most 1/r since key
was chosen uniformly.

3.3.3 Almost-matching lower bound

The bound on the probability of disagreement in the previous algorithm is
almost tight. Varghese and Lynch [VL92] show that no synchronous algo-
1
rithm can get a probability of disagreement less than r+1 , using a stronger
validity condition that requires that the processes output 0 if any input is
0. This is a natural assumption for database commit, where we don’t want
to commit if any process wants to abort. We restate their result below:
Theorem 3.3.1. For any synchronous algorithm for randomized coordi-
nated attack that runs in r rounds that satisfies the additional condition
that all non-faulty processes decide 0 if any input is 0, Pr[disagreement] ≥
1/(r + 1).
Proof. Let  be the bound on the probability of disagreement. Define levelti [k]
as in the previous algorithm (whatever the real algorithm is doing). We’ll
show Pr[i decides 1] ≤  · (levelri [i] + 1), by induction on levelri [i].
• If levelri [i] = 0, the real execution is indistinguishable (to i) from an
execution in which some other process j starts with 0 and receives no
messages at all. In that execution, j must decide 0 or risk violating the
strong validity assumption. So i decides 1 with probability at most 
(from the disagreement bound).
• If levelri [i] = k > 0, the real execution is indistinguishable (to i) from
an execution in which some other process j only reaches level k − 1
and thereafter receives no messages. From the induction hypothesis,
Pr[j decides 1] ≤ k in that pruned execution, and so Pr[i decides 1] ≤
(k + 1) in the pruned execution. But by indistinguishability, we also
have Pr[i decides 1] ≤ (k + 1) in the original execution.
Now observe that in the all-1 input execution with no messages lost,
levelri [i] = r and Pr[i decides 1] = 1 (by validity). So 1 ≤ (r + 1), which
implies  ≥ 1/(r + 1).
Chapter 4

Broadcast and convergecast

Here we’ll describe protocols for propagating information throughout a net-

work from some central initiator and gathering information back to that
same initiator. We do this both because the algorithms are actually useful
and because they illustrate some of the issues that come up with keeping
time complexity down in an asynchronous message-passing system.

4.1 Flooding
Flooding is about the simplest of all distributed algorithms. It’s dumb and
expensive, but easy to implement, and gives you both a broadcast mecha-
nism and a way to build rooted spanning trees.
We’ll give a fairly simple presentation of flooding roughly following Chap-
ter 2 of [AW04].

4.1.1 Basic algorithm

The basic flooding algorithm is shown in Algorithm 4.1. The idea is that
when a process receives a message M , it forwards it to all of its neighbors
unless it has seen it before, which it tracks using a single bit seen-message.
Theorem 4.1.1. Every process receives M after at most D time and at
most |E| messages, where D is the diameter of the network and E is the set
of (directed) edges in the network.
Proof. Message complexity: Each process only sends M to its neighbors
once, so each edge carries at most one copy of M .
Time complexity: By induction on d(root, v), we’ll show that each v
receives M for the first time no later than time d(root, v) ≤ D. The base

22
CHAPTER 4. BROADCAST AND CONVERGECAST 23

1 initially do
2 if pid = root then
3 seen-message ← true
4 send M to all neighbors
5 else
6 seen-message ← false

7 upon receiving M do
8 if seen-message = false then
9 seen-message ← true
10 send M to all neighbors

Algorithm 4.1: Basic flooding algorithm

case is when v = root, d(root, v) = 0; here root receives message at time

0. For the induction step, Let d(root, v) = k > 0. Then v has a neighbor
u such that d(root, u) = k − 1. By the induction hypothesis, u receives M
for the first time no later than time k − 1. From the code, u then sends
M to all of its neighbors, including v; M arrives at v no later than time
(k − 1) + 1 = k.

Note that the time complexity proof also demonstrates correctness: every
process receives M at least once.
As written, this is a one-shot algorithm: you can’t broadcast a sec-
ond message even if you wanted to. The obvious fix is for each process
to remember which messages it has seen and only forward the new ones
(which costs memory) and/or to add a time-to-live (TTL) field on each
message that drops by one each time it is forwarded (which may cost ex-
tra messages and possibly prevents complete broadcast if the initial TTL
is too small). The latter method is what was used for searching in http:
//en.wikipedia.org/wiki/Gnutella, an early peer-to-peer system. An
interesting property of Gnutella was that since the application of flooding
was to search for huge (multiple MiB) files using tiny ( 100 byte) query mes-
sages, the actual bit complexity of the flooding algorithm was not especially
large relative to the bit complexity of sending any file that was found.
We can optimize the algorithm slightly by not sending M back to the
node it came from; this will slightly reduce the message complexity in many
cases but makes the proof a sentence or two longer. (It’s all a question of
what you want to optimize.)
CHAPTER 4. BROADCAST AND CONVERGECAST 24

4.1.2 Adding parent pointers

To build a spanning tree, modify Algorithm 4.1 by having each process
remember who it first received M from. The revised code is given as Algo-
rithm 4.2

1 initially do
2 if pid = root then
3 parent ← root
4 send M to all neighbors
5 else
6 parent ← ⊥

7 upon receiving M from p do

8 if parent = ⊥ then
9 parent ← p
10
11 send M to all neighbors

Algorithm 4.2: Flooding with parent pointers

We can easily prove that Algorithm 4.2 has the same termination prop-
erties as Algorithm 4.1 by observing that if we map parent to seen-message
by the rule ⊥ → false, anything else → true, then we have the same al-
gorithm. We would like one additional property, which is that when the
algorithm quiesces (has no outstanding messages), the set of parent point-
ers form a rooted spanning tree. For this we use induction on time:
Lemma 4.1.2. At any time during the execution of Algorithm 4.2, the
following invariant holds:
1. If u.parent 6= ⊥, then u.parent.parent 6= ⊥ and following parent point-
ers gives a path from u to root.

2. If there is a message M in transit from u to v, then u.parent 6= ⊥.

Proof. We have to show that any event preserves the invariant.
Delivery event M used to be in u.outbuf, now it’s in v.inbuf, but it’s still
in transit and u.parent is still not ⊥.1
1
This sort of extraneous special case is why I personally don’t like the split between
outbuf and inbuf used in [AW04], even though it makes defining the synchronous model
easier.
CHAPTER 4. BROADCAST AND CONVERGECAST 25

Computation event Let v receive M from u. There are two cases: if

v.parent is already non-null, the only state change is that M is no
longer in transit, so we don’t care about u.parent any more. If v.parent
is null, then

1. v.parent is set to u. This triggers the first case of the invariant.

From the induction hypothesis we have that u.parent 6= ⊥ and
that there exists a path from u to the root. Then v.parent.parent =
u.parent 6= ⊥ and the path from v → u → root gives the path
from v.
2. Message M is sent to all of v’s neighbors. Because M is now in
transit from v, we need v.parent 6= ⊥; but we just set it to u, so
we are happy.

At the end of the algorithm, the invariant shows that every process has
a path to the root, i.e., that the graph represented by the parent pointers is
connected. Since this graph has exactly |V | − 1 edges (if we don’t count the
self-loop at the root), it’s a tree.
Though we get a spanning tree at the end, we may not get a very good
spanning tree. For example, suppose our friend the adversary picks some
Hamiltonian path through the network and delivers messages along this
path very quickly while delaying all other messages for the full allowed 1
time unit. Then the resulting spanning tree will have depth |V | − 1, which
might be much worse than D. If we want the shallowest possible spanning
tree, we need to do something more sophisticated: see the discussion of
distributed breadth-first search in Chapter 5. However, we may be
happy with the tree we get from simple flooding: if the message delay on
each link is consistent, then it’s not hard to prove that we in fact get a
shortest-path tree. As a special case, flooding always produces a BFS tree
in the synchronous model.
Note also that while the algorithm works in a directed graph, the parent
pointers may not be very useful if links aren’t two-way.

4.1.3 Termination
See [AW04, Chapter 2] for further modifications that allow the processes to
detect termination. In a sense, each process can terminate as soon as it is
done sending M to all of its neighbors, but this still requires some mecha-
nism for clearing out the inbuf; by adding acknowledgments as described in
CHAPTER 4. BROADCAST AND CONVERGECAST 26

[AW04], we can terminate with the assurance that no further messages will
be received.

4.2 Convergecast
A convergecast is the inverse of broadcast: instead of a message propa-
gating down from a single root to all nodes, data is collected from outlying
nodes to the root. Typically some function is applied to the incoming data
at each node to summarize it, with the goal being that eventually the root
obtains this function of all the data in the entire system. (Examples would
be counting all the nodes or taking an average of input values at all the
nodes.)
A basic convergecast algorithm is given in Algorithm 4.3; it propagates
information up through a previously-computed spanning tree.

1 initially do
2 if I am a leaf then
3 send input to parent

4 upon receiving M from c do

5 append (c, M ) to buffer
6 if buffer contains messages from all my children then
7 v ← f (buffer, input)
8 if pid = root then
9 return v
10 else
11 send v to parent

Algorithm 4.3: Convergecast

The details of what is being computed depend on the choice of f :

• If input = 1 for all nodes and f is sum, then we count the number of
nodes in the system.

• If input is arbitrary and f is sum, then we get a total of all the input
values.

• Combining the above lets us compute averages, by dividing the total

of all the inputs by the node count.
CHAPTER 4. BROADCAST AND CONVERGECAST 27

• If f just concatenates its arguments, the root ends up with a vector

of all the input values.

Running time is bounded by the depth of the tree: we can prove by

induction that any node at height h (height is length of the longest path
from this node to some leaf) sends a message by time h at the latest. Mes-
sage complexity is exactly n − 1, where n is the number of nodes; this is
easily shown by observing that each node except the root sends exactly one
message.
Proving that convergecast returns the correct value is similarly done by
induction on depth: if each child of some node computes a correct value, then
that node will compute f applied to these values and its own input. What
the result of this computation is will, of course, depend on f ; it generally
makes the most sense when f represents some associative operation (as in
the examples above).

4.3 Flooding and convergecast together

A natural way to build the spanning tree used by convergecast is to run
flooding first. This also provides a mechanism for letting the leaves know
that they are leaves and initiating the protocol. The combined algorithm is
shown as Algorithm 4.4.
However, this may lead to very bad time complexity for the convergecast
stage. Consider a wheel-shaped network consisting of one central node p0
connected to nodes p1 , p2 , . . . , pn−1 , where each pi is also connected to pi+1 .
By carefully arranging for the pi pi+1 links to run much faster than the
p0 pi links, the adversary can make flooding build a tree that consists of
a single path p0 p1 p2 . . . pn−1 , even though the diameter of the network is
only 2. While it only takes 2 time units to build this tree (because every
node is only one hop away from the initiator), when we run convergecast
we suddenly find that the previously-speedy links are now running only at
the guaranteed ≤ 1 time unit per hop rate, meaning that convergecast takes
n − 1 time.
This may be less of an issue in real networks, where the latency of links
may be more uniform over time, meaning that a deep tree of fast links is
still likely to be fast when we reach the convergecast step. But in the worst
case we will need to be more clever about building the tree. We show how
to do this in Chapter 5.
CHAPTER 4. BROADCAST AND CONVERGECAST 28

1 initially do
2 children ← ∅
3 nonChildren ← ∅
4 if pid = root then
5 parent ← root
6 send init to all neighbors
7 else
8 parent ← ⊥

9 upon receiving init from p do

10 if parent = ⊥ then
11 parent ← p
12 send init to all neighbors
13 else
14 send nack to p

15 upon receiving nack from p do

16 nonChildren ← nonChildren ∪ {p}
17 as soon as children ∪ nonChildren includes all my neighbors do
18 v ← f (buffer, input)
19 if pid = root then
20 return v
21 else
22 send ack(v) to parent

23 upon receiving ack(v) from k do

24 add (k, v) to buffer
25 add k to children
Algorithm 4.4: Flooding and convergecast combined
Chapter 5

Distributed breadth-first
search

Here we describe some algorithms for building a breadth-first search

(BFS) tree in a network. All assume that there is a designated initia-
tor node that starts the algorithm. At the end of the execution, each node
except the initiator has a parent pointer and every node has a list of chil-
dren. These are consistent and define a BFS tree: nodes at distance k from
the initiator appear at level k of the tree.
In a synchronous network, flooding (§4.1) solves BFS; see [AW04, Lemma
2.8, page 21] or [Lyn96, §4.2]. So the interesting case is when the network
is asynchronous.
In an asynchronous network, the complication is that we can no longer
rely on synchronous communication to reach all nodes at distance d at the
same time. So instead we need to keep track of distances explicitly, or
possibly enforce some approximation to synchrony in the algorithm. (A
general version of this last approach is to apply a synchronizer to one of the
synchronous algorithms using a synchronizer; see Chapter 8.)
To keep things simple, we’ll drop the requirement that a parent learn
the IDs of its children, since this can be tacked on as a separate notification
protocol, in which each child just sends one message to its parent once it
figures out who its parent is.

5.1 Using explicit distances

This is a translation of the AsynchBFS automaton from [Lyn96, §15.4]. It’s
a very simple algorithm, closely related to Dijkstra’s algorithm for shortest

29
CHAPTER 5. DISTRIBUTED BREADTH-FIRST SEARCH 30

paths, but there is otherwise no particular reason to use it; it is dominated

by the O(D) time and O(DE) message complexity synchronizer-based al-
gorithm described in §5.3. (Here D is the diameter of the network, the
maximum distance between any two nodes.)
The idea is to run flooding with distances attached. Each node sets its
distance to 1 plus the smallest distance sent by its neighbors and its parent
to the neighbor supplying that smallest distance. A node notifies all its
neighbors of its new distance whenever its distance changes.
Pseudocode is given in Algorithm 5.1

1 initially do
2 if pid = initiator then
3 distance ← 0
4 send distance to all neighbors
5 else
6 distance ← ∞

7 upon receiving d from p do

8 if d + 1 < distance then
9 distance ← d + 1
10 parent ← p

Algorithm 5.1: AsynchBFS algorithm (from [Lyn96])

(See [Lyn96] for a precondition-effect description, which also includes

code for buffering outgoing messages.)
The claim is that after at most O(V E) messages and O(D) time, all
distance values are equal to the length of the shortest path from the initiator
to the appropriate node. The proof is by showing the following:
Lemma 5.1.1. The variable distancep is always the length of some path
from initiator to p, and any message sent by p is also the length of some
path from initiator to p.
Proof. The second part follows from the first; any message sent equals p’s
current value of distance. For the first part, suppose p updates its distance;
then it sets it to one more than the length of some path from initiator to p0 ,
which is the length of that same path extended by adding the pp0 edge.
We also need a liveness argument that says that distancep = d(initiator, p)
no later than time d(initiator, p). Note that we can’t detect when distance
stabilizes to the correct value without a lot of additional work.
CHAPTER 5. DISTRIBUTED BREADTH-FIRST SEARCH 31

In [Lyn96], there’s an extra |V | term in the time complexity that comes

from message pile-ups, since the model used there only allows one incoming
message to be processed per time units (the model in [AW04] doesn’t have
this restriction). The trick to arranging this to happen often is to build a
graph where node 1 is connected to nodes 2 and 3, node 2 to 3 and 4, node
3 to 4 and 5, etc. This allows us to quickly generate many paths of distinct
lengths from node 1 to node k, which produces k outgoing messages from
node k. It may be that a more clever analysis can avoid this blowup, by
showing that it only happens in a few places.

5.2 Using layering

This approach is used in the LayeredBFS algorithm in [Lyn96], which is due
to Gallager [Gal82].
Here we run a sequence of up to |V | instances of the simple algorithm
with a distance bound on each: instead of sending out just 0, the initiator
sends out (0, bound), where bound is initially 1 and increases at each phase.
A process only sends out its improved distance if it is less than bound.
Each phase of the algorithm constructs a partial BFS tree that contains
only those nodes within distance bound of the root. This tree is used to
report back to the root when the phase is complete. For the following phase,
notification of the increase in bound increase is distributed only through the
partial BFS tree constructed so far. With some effort, it is possible to
prove that in a bidirectional network that this approach guarantees that
each edge is only probed once with a new distance (since distance-1 nodes
are recruited before distance-2 nodes and so on), and the bound-update and
acknowledgment messages contribute at most |V | messages per phase. So
we get O(E + V D) total messages. But the time complexity is bad: O(D2 )
in the worst case.

5.3 Using local synchronization

The reason the layering algorithm takes so long is that at each phase we
have to phone all the way back up the tree to the initiator to get permission
to go on to the next phase. We need to do this to make sure that a node
is only recruited into the tree once: otherwise we can get pile-ups on the
channels as in the simple algorithm. But we don’t necessarily need to do
this globally. Instead, we’ll require each node at distance d to delay sending
out a recruiting message until it has confirmed that none of its neighbors
CHAPTER 5. DISTRIBUTED BREADTH-FIRST SEARCH 32

will be sending it a smaller distance. We do this by having two classes of

messages:1

• exactly(d): “I know that my distance is d.”

• more-than(d): “I know that my distance is > d.”

The rules for sending these messages for a non-initiator are:

1. I can send exactly(d) as soon as I have received exactly(d − 1) from at

least one neighbor and more-than(d − 2) from all neighbors.

2. I can send more-than(d) if d = 0 or as soon as I have received more-than(d−

1) from all neighbors.

The initiator sends exactly(0) to all neighbors at the start of the protocol
(these are the only messages the initiator sends).
My distance will be the unique distance that I am allowed to send in an
exactly(d) messages. Note that this algorithm terminates in the sense that
every node learns its distance at some finite time.
If you read the discussion of synchronizers in Chapter 8, this algorithm
essentially corresponds to building the alpha synchronizer into the syn-
chronous BFS algorithm, just as the layered model builds in the beta syn-
chronizer. See [AW04, §11.3.2] for a discussion of BFS using synchronizers.
The original approach of applying synchronizers to get BFS is due to Awer-
buch [Awe85].
We now show correctness. Under the assumption that local computation
takes zero time and message delivery takes at most 1 time unit, we’ll show
that if d(initiator, p) = d, (a) p sends more-than(d0 ) for any d0 < d by time
d0 , (b) p sends exactly(d) by time d, (c) p never sends more-than(d0 ) for any
d0 ≥ d, and (d) p never sends exactly(d0 ) for any d0 6= d. For parts (c) and
(d) we use induction on d0 ; for (a) and (b), induction on time. This is not
terribly surprising: (c) and (d) are safety properties, so we don’t need to
talk about time. But (a) and (b) are liveness properties so time comes in.
Let’s start with (c) and (d). The base case is that the initiator never
sends any more-than messages at all, and so never sends more-than(0), and
any non-initiator never sends exactly(0). For larger d0 , observe that if a
non-initiator p sends more-than(d0 ) for d0 ≥ d, it must first have received
1
In an earlier version of these notes, these messages where called distance(d) and
not-distance(d); the more self-explanatory exactly and more-than terminology is taken from
[BDLP08].
CHAPTER 5. DISTRIBUTED BREADTH-FIRST SEARCH 33

more-than(d0 − 1) from all neighbors, including some neighbor p0 at distance

d−1. But the induction hypothesis tells us that p0 can’t send more-than(d0 −
1) for d0 − 1 ≥ d − 1. Similarly, to send exactly(d0 ) for d0 < d, p must first
have received exactly(d0 − 1) from some neighbor p0 , but again p0 must be at
distance at least d−1 from the initiator and so can’t send this message either.
In the other direction, to send exactly(d0 ) for d0 > d, p must first receive
more-than(d0 − 2) from this closer neighbor p0 , but then d0 − 2 > d − 2 ≥ d − 1
so more-than(d0 − 2) is not sent by p0 .
Now for (a) and (b). The base case is that the initiator sends exactly(0)
to all nodes at time 0, giving (a), and there is no more-than(d0 ) with d0 < 0
for it to send, giving (b) vacuously; and any non-initiator sends more-than(0)
immediately. At time t + 1, we have that (a) more-than(t) was sent by any
node at distance t + 1 or greater by time t and (b) exactly(t) was sent by
any node at distance t by time t; so for any node at distance t + 2 we
send more-than(t + 1) no later than time t + 1 (because we already received
more-than(t) from all our neighbors) and for any node at distance t + 1 we
send exactly(t + 1) no later than time t + 1 (because we received all the
preconditions for doing so by this time).
Message complexity: A node at distance d sends more-than(d0 ) for all
0 < d0 < d and exactly(d) and no other messages. So we have message
complexity bounded by |E| · D in the worst case. Note that this is gives a
bound of O(DE), which is slightly worse than the O(E + DV ) bound for
the layered algorithm.
Time complexity: It’s immediate from (a) and (b) that all messages that
are sent are sent by time D, and indeed that any node p learns its distance
at time d(initiator, p). So we have optimal time complexity, at the cost of
higher message complexity. I don’t know if this trade-off is necessary, or if
a more sophisticated algorithm could optimize both.
Our time proof assumes that messages don’t pile up on edges, or that
such pile-ups don’t affect delivery time (this is the default assumption used
in [AW04]). A more sophisticated proof could remove this assumption.
One downside of this algorithm is that it has to be started simultane-
ously at all nodes. Alternatively, we could trigger “time 0” at each node
by a broadcast from the initiator, using the usual asynchronous broadcast
algorithm; this would give us a BFS tree in O(|E| · D) messages (since the
O(|E|) messages of the broadcast disappear into the constant) and 2D time.
The analysis of time goes through as before, except that the starting time 0
becomes the time at which the last node in the system is woken up by the
broadcast. Further optimizations are possible; see, for example, the paper
of Boulinier et al. [BDLP08], which shows how to run the same algorithm
CHAPTER 5. DISTRIBUTED BREADTH-FIRST SEARCH 34

with constant-size messages.

Chapter 6

Leader election

See [AW04, Chapter 3] or [Lyn96, Chapter 3] for details.

The basic idea of leader election is that we want a single process to
declare itself leader and the others to declare themselves non-leaders. The
non-leaders may or may not learn the identity of the leader as part of the pro-
tocol; if not, we can always add an extra phase where the leader broadcasts
its identity to the others. Traditionally, leader election has been used as a
way to study the effects of symmetry, and many leader election algorithms
are designed for networks in the form of a ring.
A classic result of Angluin [Ang80] shows that leader election in a ring
is impossible if the processes do not start with distinct identities. The proof
is that if everybody is in the same state at every step, they all put on the
crown at the same time. We discuss this result in more detail in §6.1.
With ordered identities, a simple algorithm due to Le Lann [LL77] and
Chang and Roberts [CR79] solves the problem in O(n) time with O(n2 )
messages: I send out my own id clockwise and forward any id bigger than
mine. If I get my id back, I win. This works with a unidirectional ring,
doesn’t require synchrony, and never produces multiple leaders. See §6.2.1
for more details.
On a bidirectional ring we can get O(n log n) messages and O(n) time
with power-of-2 probing, using an algorithm of Hirschberg and Sinclair [HS80].
This is described in §6.2.2.
An evil trick: if we have synchronized starting, known n, and known id
space, we can have process with id i wait until round i · n to start sending
its id around, and have everybody else drop out when they receive it; this
way only one process (the one with smallest id) ever starts a message and
only n messages are sent [FL87]. But the running time can be pretty bad.

35
CHAPTER 6. LEADER ELECTION 36

For general networks, we can apply the same basic strategy as in Le

Lann-Chang-Roberts by having each process initiate a broadcast/convergecast
algorithm that succeeds only if the initiator has the smallest id. This is de-
scribed in more detail in §6.3.
Some additional algorithms for the asynchronous ring are given in §§6.2.3
and 6.2.4. Lower bounds are shown in §6.4.

6.1 Symmetry
A system exhibits symmetry if we can permute the nodes without changing
the behavior of the system. More formally, we can define a symmetry as an
equivalence relation on processes, where we have the additional properties
that all processes in the same equivalence class run the same code; and
whenever p is equivalent to p0 , each neighbor q of p is equivalent to the
corresponding neighbor q 0 of p0 .
An example of a network with a lot of symmetries would be an anony-
mous ring, which is a network in the form of a cycle (the ring part) in
which every process runs the same code (the anonymous part). In this case
all nodes are equivalent. If we have a line, then we might or might not have
any non-trivial symmetries: if each node has a sense of direction that
tells it which neighbor is to the left and which is to the right, then we can
identify each node uniquely by its distance from the left edge. But if the
nodes don’t have a sense of direction, we can flip the line over and pair up
nodes that map to each other.1
Symmetries are convenient for proving impossibility results, as observed
by Angluin [Ang80]. The underlying theme is that without some mecha-
nism for symmetry breaking, a message-passing system escape from a
symmetric initial configuration. The following lemma holds for determin-
istic systems, basically those in which processes can’t flip coins:

Lemma 6.1.1. A symmetric deterministic message-passing system that

starts in an initial configuration in which equivalent processes have the same
state has a synchronous execution in which equivalent processes continue to
have the same state.

Proof. Easy induction on rounds: if in some round p and p0 are equivalent

and have the same state, and all their neighbors are equivalent and have the
1
Typically, this does not mean that the nodes can’t tell their neighbors apart. But it
does mean that if we swap the labels for all the neighbors (corresponding to flipping the
entire line from left to right), we get the same executions.
CHAPTER 6. LEADER ELECTION 37

same state, then p and p0 receive the same messages from their neighbors
and can proceed to the same state (including outgoing messages) in the next
round.

An immediate corollary is that you can’t do leader election in an anony-

mous system with a symmetry that puts each node in a non-trivial equiva-
lence class, because as soon as I stick my hand up to declare I’m the leader,
so do all my equivalence-class buddies.
With randomization, Lemma 6.1.1 doesn’t directly apply, since we can
break symmetry by having my coin-flips come up differently from yours. It
does show that we can’t guarantee convergence to a single leader in any fixed
amount of time (because otherwise we could just fix all the coin flips to get
a deterministic algorithm). Depending on what the processes know about
the size of the system, it may still be possible to show that a randomized
algorithm necessarily fails in some cases.2
A more direct way to break symmetry is to assume that all processes
have identities; now processes can break symmetry by just declaring that
the one with the smaller or larger identity wins. This approach is taken in
the algorithms in the following sections.

6.2 Leader election in rings

Here we’ll describe some basic leader election algorithms for rings. Histor-
ically, rings were the first networks in which leader election was studied,
because they are the simplest networks whose symmetry makes the problem
difficult, and because of the connection to token-ring networks, a method
for congestion control in local-area networks that is no longer used much.

6.2.1 The Le-Lann-Chang-Roberts algorithm

This is about the simplest leader election algorithm there is. It works in
a unidirectional ring, where messages can only travel clockwise.3 The
algorithms works does not require synchrony, but we’ll assume synchrony to
make it easier to follow.
2
Specifically, if the processes don’t know the size of the ring, we can imagine a ring
of size 2n in which the first n processes happen to get exactly the same coin-flips as the
second n processes for long enough that two matching processes, one in each region, both
think they have won the fight in a ring of size n and declare themself to be the leader.
3
We’ll see later in §6.2.3 that the distinction between unidirectional rings and bidirec-
tional rings is not a big deal, but for now let’s imagine that having a unidirectional ring
is a serious hardship.
CHAPTER 6. LEADER ELECTION 38

Formally, we’ll let the state space for each process i consist of two vari-
ables: leader, initially 0, which is set to 1 if i decides it’s a leader; and maxId,
the largest id seen so far. We assume that i denotes i’s position rather than
its id, which we’ll write as idi . We will also treat all positions as values mod
n, to simplify the arithmetic.
Code for the LCR algorithm is given in Algorithm 6.1.

1 initially do
2 leader ← 0
3 maxId ← idi
4 send idi to clockwise neighbor
5 upon receiving j do
6 if j = idi then
7 leader ← 1
8 if j > maxId then
9 maxId ← j
10 send j to clockwise neighbor

Algorithm 6.1: LCR leader election

6.2.1.1 Proof of correctness for synchronous executions

By induction on the round number k. The induction hypothesis is that
in round k, each process i’s leader bit is 0, its maxId value is equal to the
largest id in the range (i − k) . . . i, and that it sends idi−k if and only if idi−k
is the largest id in the range (i − k) . . . i. The base case is that when k = 0,
maxId = idi is the largest id in i . . . i, and i sends idi . For the induction step,
observe that in round k − 1, i − 1 sends id(i−1)−(k−1) = idi−k if and only if it
is the largest in the range (i − k) . . . (i − 1), and that i adopts it as the new
value of maxId and sends it just in case it is larger than the previous largest
value in (i − k + 1) . . . (i − 1), i.e., if it is the largest value in (i − k) . . . i.
Finally, in round n − 1, i − 1 sends idi−N = idi if and only if i is the
largest id in (i − n + 1) . . . i, the whole state space. So i receives idi and sets
leaderi = 1 if and only if it has the maximum id.

6.2.1.2 Performance
It’s immediate from the correctness proof that the protocols terminates after
exactly n rounds.
CHAPTER 6. LEADER ELECTION 39

To count message traffic, observe that each process sends at most 1

message per round, for a total of O(n2 ) messages. This is a tight bound
since if the ids are in decreasing order n, n − 1, n − 2, . . . 1, then no messages
get eaten until they hit n.

6.2.2 The Hirschberg-Sinclair algorithm

[[[ needs better intro and citation: the trick here is to only probe
2k away in phase k, gets O(n log n) messages and O(n) time. ]]]
Basically the same as for LCR but both the protocol and the invari-
ant get much messier. To specify the protocol, it may help to think of
messages as mobile agents and the state of each process as being of the
form (local-state, {agents I’m carrying}). Then the sending rule for a pro-
cess becomes ship any agents in whatever direction they want to go and the
transition rule is accept any incoming agents and update their state in terms
of their own internal transition rules. An agent state for LCR will be some-
thing like (original-sender, direction, hop-count, max-seen) where direction
is R or L depending on which way the agent is going, hop-count is initially
2k when the agent is sent and drops by 1 each time the agent moves, and
max-seen is the biggest id of any node the agent has visited. An agent turns
around (switches direction) when hop-count reaches 0.
To prove this works, we can mostly ignore the early phases (though we
have to show that the max-id node doesn’t drop out early, which is not
too hard). The last phase involves any surviving node probing all the way
around the ring, so it will declare itself leader only when it receives its own
agent from the left. That exactly one node does so is immediate from the
same argument for LCR.
Complexity analysis is mildly painful but basically comes down to the
fact that any node that sends a message 2k hops had to be a winner at
phase 2k−1 , which means that it is the largest of some group of 2k−1 ids.
Thus the 2k -hop senders are spaced at least 2k−1 away from each other and
there are at most n/2k−1 of them. Summing up over all dlg ne phases, we
Pdlg ne Pdlg ne
get k=0 2k n/2k−1 = O(n log n) messages and k=0 2k = O(n) time.

6.2.3 Peterson’s algorithm for the unidirectional ring

This algorithm is due to Peterson [Pet82] and assumes an asynchronous,
unidirectional ring. It gets O(n log n) message complexity in all executions.
The basic idea (2-way communication version): Start with n candidate
leaders. In each of at most lg n asynchronous phases, each candidate probes
CHAPTER 6. LEADER ELECTION 40

its nearest neighbors to the left and right; if its ID is larger than the IDs of
both neighbors, it survives to the next phase. Non-candidates act as relays
passing messages between candidates. As in Hirschberg and Sinclair (§6.2.2),
the probing operations in each phase take O(n) messages, and at least half
of the candidates drop out in each phase. The last surviving candidate wins
when it finds that it’s its own neighbor.
To make this work in a 1-way ring, we have to simulate 2-way communi-
cation by moving the candidates clockwise around the ring to catch up with
their unsendable counterclockwise messages. Peterson’s algorithm does this
with a two-hop approach that is inspired by the 2-way case above; in each
phase k, a candidate effectively moves two positions to the right, allowing it
to look at the ids of three phase-k candidates before deciding to continue in
phase k + 1 or not. Here is a very high-level description; it assumes that we
can buffer and ignore incoming messages from the later phases until we get
to the right phase, and that we can execute sends immediately upon receiv-
ing messages. Doing this formally in terms of I/O automata or the model of
§2.1 means that we have to build explicit internal buffers into our processes,
which we can easily do but won’t do here (see [Lyn96, pp. 483–484] for the
right way to do this.)
We can use a similar trick to transform any bidirectional-ring algorithm
into a unidirectional-ring algorithm: alternative between phases where we
send a message right, then send a virtual process right to pick up any left-
going messages deposited for us. The problem with this trick is that it
requires two messages per process per phase, which gives us a total message
complexity of O(n2 ) if we start with an O(n)-time algorithm. Peterson’s
algorithm avoids this by only propagating the surviving candidates.
Pseudocode for Peterson’s algorithm is given in Algorithm 6.2.
Note: the phase arguments in the probe messages are useless if one has
FIFO channels, which is why [Lyn96] doesn’t use them. Note also that the
algorithm does not elect the process with the highest ID, but the process
that is carrying the sole surviving candidate in the last phase.
Proof of correctness is essentially the same as for the 2-way algorithm.
For any pair of adjacent candidates, at most one of their current IDs survives
to the next phase. So we get a sole survivor after lg n phases. Each process
sends or relays at most 2 messages per phases, so we get at most 2n lg n
total messages.
CHAPTER 6. LEADER ELECTION 41

1 procedure candidate()
2 phase ← 0
3 current ← pid
4 while true do
5 send probe(phase, current)
6 wait for probe(phase, x)
7 id2 ← x
8 send probe(phase, current)
9 wait for probe(phase, x)
10 id3 ← x
11 if id2 = current then
12 I am the leader!
13 return
14 else if id2 > current and id2 > id3 do
15 current ← id2
16 phase ← phase + 1
17 else
18 switch to relay()

19 procedure relay()
20 upon receiving probe(p, i) do
21 send probe(p, i)

Algorithm 6.2: Peterson’s leader-election algorithm

CHAPTER 6. LEADER ELECTION 42

6.2.4 A simple randomized O(n log n)-message algorithm

An alternative to running a more sophisticated algorithm is to reduce the
average cost of LCR using randomization. The presentation here follows the
average-case analysis done by Chang and Roberts [CR79].
Run LCR where each id is constructed by prepending a long random
bit-string to the real id. This gives uniqueness (since the real id’s act as
tie-breakers) and something very close to a random permutation on the
constructed id’s. When we have unique random id’s, a simple argument
shows that the i-th largest id only propagates an expected n/i hops, giving
a total of O(nHn ) = O(n log n) hops.4 Unique random id’s occur with high
probability provided the range of the random sequence is  n2 .
The downside of this algorithm compared to Peterson’s is that knowledge
of n is required to pick random id’s from a large enough range. It also has
higher bit complexity since Peterson’s algorithm is sending only IDs (in the
official version) without any random padding.

6.3 Leader election in general networks

For general networks, a simple approach is to have each node initiate a
breadth-first-search and convergecast, with nodes refusing to participate in
the protocol for any initiator with a lower id. It follows that only the node
with the maximum id can finish its protocol; this node becomes the leader.
If messages from parallel broadcasts are combined, it’s possible to keep the
message complexity of this algorithm down to O(DE).
More sophisticated algorithms reduce the message complexity by coalesc-
ing local neighborhoods similar to what happens in the Hirschberg-Sinclair
and Peterson algorithms. A noteworthy example is an O(n log n) message-
complexity algorithm of Afek and Gafni [AG91], who also show an Ω(n log n)
lower bound on message complexity for any synchronous algorithm in a com-
plete network.

6.4 Lower bounds

Here we present two classic Ω(log n) lower bounds on message complexity
for leader election in the ring. The first, due to Burns [Bur80], assumes
4
Alternatively, we could consider the average-case complexity of the algorithm when
we assume all n! orderings of the ids are equally likely; this also gives O(n log n) expected
message complexity [CR79].
CHAPTER 6. LEADER ELECTION 43

that the system is asynchronous and that the algorithm is uniform: it

does not depend on the size of the ring. The second, due to Frederickson
and Lynch [FL87], allows a synchronous system and relaxes the uniformity
assumption, but requires that the algorithm can’t do anything to ids but
copy and compare them.

6.4.1 Lower bound on asynchronous message complexity

Here we describe a lower bound for uniform asynchronous leader election in
the ring. The description here is based on [AW04, §3.3.3]; a slightly different
presentation can also be found in [Lyn96, §15.1.4]. The original result is due
to Burns [Bur80]. We assume the system is deterministic.
The basic idea is to construct a bad execution in which n processes
send lots of messages recursively, by first constructing two bad (n/2)-process
executions and pasting them together in a way that generates many extra
messages. If the pasting step produces Θ(n) additional messages, we get a
recurrence T (n) ≥ 2T (n/2) + Θ(n) for the total message traffic, which has
solution T (n) = Ω(n log n).
We’ll assume that all processes are trying to learn the identity of the
process with the smallest id. This is a slightly stronger problem that mere
leader election, but it can be solved with at most an additional 2n messages
once we actually elect a leader. So if we get a lower bound of f (n) messages
on this problem, we immediately get a lower bound of f (n) − 2n on leader
election.
To construct the bad execution, we consider “open executions” on rings
of size n where no message is delivered across some edge (these will be partial
executions, because otherwise the guarantee of eventual delivery kicks in).
Because no message is delivered across this edge, the processes can’t tell if
there is really a single edge there or some enormous unexplored fragment of
a much larger ring. Our induction hypothesis will show that a line of n/2
processes can be made to send at least T (n/2) messages in an open execution
(before seeing any messages across the open edge); we’ll then show that a
linear number of additional messages can be generated by pasting two such
executions together end-to-end, while still getting an open execution with n
processes.
In the base case, we let n = 2. Somebody has to send a message even-
tually, giving T (2) ≥ 1.
For larger n, suppose that we have two open executions on n/2 processes
that each send at least T (n/2) messages. Break the open edges in both
executions and paste the resulting lines together to get a ring of size n;
CHAPTER 6. LEADER ELECTION 44

similarly paste the schedules σ1 and σ2 of the two executions together to

get a combined schedule σ1 σ2 with at least 2T (n/2) messages. Note that in
the combined schedule no messages are passed between the two sides, so the
processes continue to behave as they did in their separate executions.
Let e and e0 be the edges we used to past together the two rings. Extend
σ1 σ2 by the longest possible suffix σ3 in which no messages are delivered
across e and e0 . Since σ3 is as long as possible, after σ1 σ2 σ3 , there are no
messages waiting to be delivered across any edge except e and e0 and all
processes are quiescent—they will send no additional messages until they
receive one.
Now consider the processes in the half of the ring with the larger min-
imum id. Because each process must learn the minimum id in the other
half of the ring, each of these processes must receive a message in some
complete execution, giving an additional n/2 − 2 messages (since two of the
processes might receive undelivered messages on e or e0 that we’ve already
counted). But to prove our induction hypothesis, we need to keep one of
e or e0 open. Consider some execution σ1 σ2 σ3 σ4 in which all messages de-
layed on both e and e0 are delivered, and partition the n/2 process on the
losing side into two groups based on whether the first message they get is
triggered by opening e or e0 . One of these two groups must contain at least
half the n/2 − 2 processes who receive new messages, meaning that there is
an execution σ1 σ2 σ3 σ40 in which we open up only one edge and still get an
additional (n/2 − 2)/2 = Θ(n) messages. This concludes the proof.

6.4.2 Lower bound for comparison-based algorithms

Here we give an Ω(n log n) lower bound on messages for synchronous-start
comparison-based algorithms in bidirectional synchronous rings. For full
details see [Lyn96, §3.6], [AW04, §3.4.2], or the original JACM paper by
Frederickson and Lynch [FL87].
Basic ideas:

• Two fragments i . . . i+k and j . . . j +k of a ring are order-equivalent

provided idi+a > idi+b if and only if idj+a > idj+b for b = 0 . . . k.

• An algorithm is comparison-based if it can’t do anything to IDs but

copy them and test for <. The state of such an algorithm is modeled
by some non-ID state together with a big bag of IDs, messages have a
pile of IDs attached to them, etc. Two states/messages are equivalent
under some mapping of IDs if you can translate the first to the second
by running all IDs through the mapping.
CHAPTER 6. LEADER ELECTION 45

An equivalent version uses an explicit equivalence relation between

processes. Let executions of p1 and p2 be similar if both processes
send messages in the same direction(s) in the same rounds and both
processes declare themselves leader (or not) at the same round. Then
an algorithm is comparison-based based if order-equivalent rings yield
similar executions for corresponding processes. This can be turned
into the explicit-copying-IDs model by replacing the original protocol
with a full-information protocol in which each message is replaced
by the ID and a complete history of the sending process (including all
messages it has every received).

• Define an active round as a round in which at least 1 message is

sent. Claim: actions of i after k active rounds depends up to an order-
equivalent mapping of ids only on the order-equivalence class of ids
in i − k . . . i + k (the k-neighborhood of i). Proof: by induction on
k. Suppose i and j have order-equivalent (k − 1)-neighborhoods; then
after k − 1 active rounds they have equivalent states by the induc-
tion hypothesis. In inactive rounds, i and j both receive no messages
and update their states in the same way. In active rounds, i and j
receive order-equivalent messages and update their states in an order-
equivalent way.

• If we have an order of ids with a lot of order-equivalent k-neighborhoods,

then after k active rounds if one process sends a message, so do a lot
of other ones.

Now we just need to build a ring with a lot of order-equivalent neighbor-

hoods. For n a power of 2 we can use the bit-reversal ring, e.g., id sequence
000, 100, 010, 110, 001, 101, 011, 111 (in binary) when n = 8. Figure 6.1 gives
a picture of what this looks like for n = 32.
For n not a power of 2 we look up Frederickson and Lynch [FL87] or At-
tiya et al. [ASW88]. In either case we get Ω(n/k) order-equivalent members
of each equivalence class after k active rounds, giving Ω(n/k) messages per
active round, which sums to Ω(n log n).
For non-comparison-based algorithms we can still prove Ω(n log n) mes-
sages for time-bounded algorithms, but it requires techniques from Ram-
sey theory, the branch of combinatorics that studies when large enough
structures inevitably contain substructures with certain properties.5 Here
5
The classic example is Ramsey’s Theorem, which says that if you color the edges
of a complete graph red or blue, while trying to avoid having any subsets of k vertices
CHAPTER 6. LEADER ELECTION 46

35

30

25

20

15

10

0
-5 0 5 10 15 20 25 30 35

Figure 6.1: Labels in the bit-reversal ring with n = 32

“time-bounded” means that the running time can’t depend on the size of
the ID space. See [AW04, §3.4.2] or [Lyn96, §3.7] for the textbook version,
or [FL87, §7] for the original result.
The intuition is that for any fixed protocol, if the ID space is large
enough, then there exists a subset of the ID space where the protocol
acts like a comparison-based protocol. So the existence of an O(f (n))-
message time-bounded protocol implies the existence of an O(f (n))-message
comparison-based protocol, and from the previous lower bound we know
f (n) is Ω(n log n). Note that time-boundedness is necessary: we can’t prove
the lower bound for non-time-bounded algorithms because of the i · n trick.

with all edges between them the same color, you will no longer be able to once the graph
is large enough (for any fixed k). See [GRS90] for much more on the subject of Ramsey
theory.
Chapter 7

Logical clocks

Logical clocks assign a timestamp to all events in an asynchronous message-

passing system that simulates real time, thereby allowing timing-based al-
gorithms to run despite asynchrony. In general, they don’t have anything to
do with clock synchronization or wall-clock time; instead, they provide nu-
merical values that increase over time and are consistent with the observable
behavior of the system. In particular, messages are never delivered before
they are sent, when time is measured using the logical clock.

7.1 Causal ordering

The underlying notion of a logical clock is causal ordering, a partial order
on events that describes when one event e provably occurs before some other
event e0 .
For the purpose of defining casual ordering and logical clocks, we will
assume that a schedule consists of send events and receive events, which
correspond to some process sending a single message or receiving a single
message, respectively.
Given two schedules S and S 0 , call S and S 0 similar if S|p = S 0 |p for all
processes p; in other words, S and S 0 are similar if they are indistinguishable
by all participants. We can define a causal ordering on the events of some
schedule S implicitly by considering all schedules S 0 similar to S, and declare
that e < e0 if e precedes e0 in all such S. But it is usually more useful to
make this ordering explicit.
Following [AW04, §6.1.1] (and ultimately [Lam78]), define the happens-
before relation ⇒S on a schedule S to consist of:
1. All pairs (e, e0 ) where e precedes e0 in S and e and e0 are events of the

47
CHAPTER 7. LOGICAL CLOCKS 48

same process.
2. All pairs (e, e0 ) where e is a send event and e0 is the receive event for
the same message.
3. All pairs (e, e0 ) where there exists a third event e00 such that e ⇒S e00
and e00 ⇒S e0 . (In other words, we take the transitive closure of the
relation defined by the previous two cases.)
It is not terribly hard to show that this gives a partial order; the main
observation is that if e ⇒S e0 , then e precedes e0 in S. So ⇒S is a subset of
the total order <S given by the order of events in S.
A causal shuffle S 0 of a schedule S is a permutation of S that is consis-
tent with the happens-before relation on S; that is, if e happens-before e0 in
S, then e precedes e0 in S 0 . The importance of the happens-before relation
follows from the following lemma, which says that the causal shuffles of S
are precisely the schedules S 0 that are similar to S.
Lemma 7.1.1. Let S 0 be a permutation of the events in S. Then the fol-
lowing two statements are equivalent:
1. S 0 is a causal shuffle of S.
2. S 0 is the schedule of an execution fragment of a message-passing system
with S|p = S 0 |p for all S 0 .
Proof. (1 ⇒ 2). We need to show both similarity and that S 0 corresponds
to some execution fragment. We’ll show similarity first. Pick some p; then
every event at p in S also occurs in S 0 , and they must occur in the same order
by the first case of the definition of the happens-before relation. This gets
us halfway to showing S 0 is the schedule of some execution fragment, since
it says that any events initiated by p are consistent with p’s programming.
To get the rest of the way, observe that any other events are receive events.
For each receive event e0 in S, there must be some matching send event e
also in S; thus e and e0 are both in S 0 and occur in the right order by the
second case of the definition of happens-before.
(2 ⇒ 1). First observe that since every event e in S 0 occurs at some
process p, if S 0 |p = S|p for all p, then there is a one-to-one correspondence
between events in S 0 and S, and thus S 0 is a permutation of S. Now we
need to show that S 0 is consistent with ⇒S . Let e ⇒S e0 . There are three
cases.
1. e and e0 are events of the same process p and e <S e0 . But then e <S 0 e0
because S|p = S 0 |p.
CHAPTER 7. LOGICAL CLOCKS 49

2. e is a send event and e0 is the corresponding receive event. Then

e <S 0 e0 because S 0 is the schedule of an execution fragment.
3. e ⇒S e0 by transitivity. Then each step in the chain connecting e to
e0 uses one of the previous cases, and e <S 0 e0 by transitivity of <S 0 .

What this means: if I tell you ⇒S , then you know everything there is
to know about the order of events in S that you can deduce from reports
from each process together with the fact that messages don’t travel back in
time. But ⇒S is a pretty big relation (Θ(|S|2 ) bits with a naive encoding),
and seems to require global knowledge of <S to compute. So we can ask if
there is some simpler, easily computable description that works almost as
well. This is where logical clocks come in.

7.2 Implementations
The basic idea of a logical clock is to compute a timestamp for each event,
so that comparing timestamps gives information about ⇒S . Note that these
timestamps need not be totally ordered. In general, we will have a relation
<L between timestamps such that e ⇒S e0 implies e <L e0 , but it may be
that there are some pairs of events that are ordered by the logical clock
despite being incomparable in the happens-before relation.
Examples of logical clocks that use small timestamps but add extra or-
dering are Lamport clocks [Lam78], discussed in §7.2.1; and Neiger-Toueg-
Welch clocks [NT87, Wel87], discussed in §7.2.2. These both assign integer
timestamps to events and may order events that are not causally related.
The main difference between them is that Lamport clocks do not alter the
underlying execution, but may allow arbitrarily large jumps in the logical
clock values; while Neiger-Toueg-Welch clocks guarantee small increments
at the cost of possibly delaying parts of the system.1 A more restricted type
of logical clock are vector clock [Fid91, Mat93], discussed in §7.2.3, which
use n-dimensional vectors of integers to capture ⇒S exactly, at the cost of
much higher overhead.

7.2.1 Lamport clock

Lamport’s logical clock [Lam78] runs on top of any other message-passing
protocol, adding additional state at each process and additional content to
1
This makes them similar to synchronizers, which we will discuss in Chapter 8.
CHAPTER 7. LOGICAL CLOCKS 50

the messages (which is invisible to the underlying protocol). Every process

maintains a local variable clock. When a process sends a message or executes
an internal step, it sets clock ← clock + 1 and assigns the resulting value
as the clock value of the event. If it sends a message, it piggybacks the
resulting clock value on the message. When a process receives a message
with timestamp t, it sets clock ← max(clock, t) + 1; the resulting clock value
is taken as the time of receipt of the message. (To make life easier, we
assume messages are received one at a time.)

Theorem 7.2.1. If we order all events by clock value, we get an execution

of the underlying protocol that is locally indistinguishable from the original
execution.

Proof. Let e <L e0 if e has a lower clock value than e0 . If e and e0 are two
events of the same process, then e <L e0 . If e and e0 are send and receive
events of the same message, then again e <L e0 . So for any events e, e0 , if
e ⇒S e0 , then e <L e0 . Now apply Lemma 7.1.1.

7.2.2 Neiger-Toueg-Welch clock

Lamport’s clock has the advantage of requiring no changes in the behavior
of the underlying protocol, but has the disadvantage that clocks are entirely
under the control of the logical-clock protocol and may as a result make
huge jumps when a message is received. If this is unacceptable—perhaps
the protocol needs to do some unskippable maintenance task every 1000
clock ticks—then an alternative approach due to Neiger and Toueg [NT87]
and Welch [Wel87] can be used.
Method: Each process maintains its own variable clock, which it incre-
ments whenever it feels like it. To break ties, the process extends the clock
value to hclock, id, eventCounti where eventCount is a count of send and re-
ceive events (and possibly local computation steps). As in Lamport’s clock,
each message in the underlying protocol is timestamped with the current
extended clock value. Because the protocol can’t change the clock values on
its own, when a message is received with a timestamp later than the current
extended clock value, its delivery is delayed until clock exceeds the message
timestamp, at which point the receive event is assigned the extended clock
value of the time of delivery.

Theorem 7.2.2. If we order all events by clock value, we get an execution

of the underlying protocol that is locally indistinguishable from the original
execution.
CHAPTER 7. LOGICAL CLOCKS 51

Proof. Again, we have that (a) all events at the same process occur in in-
creasing order (since the event count rises even if the clock value doesn’t,
and we assume that the clock value doesn’t drop) and (b) all receive events
occur later than the corresponding send event (since we force them to). So
Lemma 7.1.1 applies.

The advantage of the Neiger-Toueg-Welch clock is that it doesn’t im-

pose any assumptions on the clock values, so it is possible to make clock be
a real-time clock at each process and nonetheless have a causally-consistent
ordering of timestamps even if the local clocks are not perfectly synchro-
nized. If some process’s clock is too far off, it will have trouble getting its
messages delivered quickly (if its clock is ahead) or receiving messages (if its
clock is behind)—the net effect is to add a round-trip delay to that process
equal to the difference between its clock and the clock of its correspon-
dent. But the protocol works well when the processes’ clocks are closely
synchronized, which has become a plausible assumption in the last 10-15
years thanks to the Network Time Protocol, cheap GPS receivers, and clock
synchronization mechanisms built into most cellular phone networks.2

7.2.3 Vector clocks

Logical clocks give a superset of the happens-before relation: if e ⇒S e0 ,
then e <L e0 (or conversely, if e 6<L e0 , then it is not the case that e ⇒S e0 ).
This is good enough for most applications, but what if we want to compute
⇒S exactly?
Here we can use a vector clock, invented independently by Fidge [Fid91]
and Mattern [Mat93]. Instead of a single clock value, each event is stamped
with a vector of values, one for each process. When a process executes a
local event or a send event, it increments only its own component xp of the
vector. When it receives a message, it increments xp and sets each xq to the
max of its previous value and the value of xq piggybacked on the message.
We define VC(e) ≤ VC(e0 ), where VC(e) is the value of the vector clock for
e, if VC(e)i ≤ VC(e0 )i for all i.

Theorem 7.2.3. Fix a schedule S; then for any e, e0 , V C(e) < V C(e0 ) if
and only if e ⇒S e0 .
2
As I write this, my computer reports that its clock is an estimated 289 microseconds
off from the timeserver it is synchronized to, which is less than a tenth of the round-trip
delay to machines on the same local-area network and a tiny fraction of the round-trip
delay to machines elsewhere, including the timeserver machine.
CHAPTER 7. LOGICAL CLOCKS 52

Proof. The if part follows immediately from the update rules for the vector
clock. For the only if part, suppose e does not happen-before e0 . Then e and
e0 are events of distinct processes p and p0 . For VC(e) < VC(e0 ) to hold, we
must have VC(e)p < VC(e0 )p ; but this can occur only if the value of VC(e)p
is propagated to p0 by some sequence of messages starting at p and ending
at p0 at or before e0 occurs. In this case we have e ⇒S e0 .

7.3 Applications
7.3.1 Consistent snapshots
A consistent snapshot of a message-passing computation is a description
of the states of the processes (and possibly messages in transit, but we
can reduce this down to just states by keeping logs of messages sent and
received) that gives the global configuration at some instant of a schedule
that is a consistent reordering of the real schedule (a consistent cut in
the terminology of [AW04, §6.1.2]. Without shutting down the protocol
before taking a snapshot this is the about the best we can hope for in a
message-passing system.
Logical time can be used to obtain consistent snapshots: pick some logi-
cal time and have each process record its state at this time (i.e. immediately
after its last step before the time or immediately before its first step after
the time). We have already argued that logical time gives a consistent re-
ordering of the original schedule, so the set of values recorded is just the
configuration at the end of an appropriate prefix of this reordering. In other
words, it’s a consistent snapshot.
If we aren’t building logical clocks anyway, there is a simpler consistent
snapshot algorithm due to Chandy and Lamport [CL85]. Here some central
initiator broadcasts a snap message, and each process records its state and
immediately forwards the snap message to all neighbors when it first receives
a snap message. To show that the resulting configuration is a configuration
of some consistent reordering, observe that (with FIFO channels) no process
receives a message before receiving snap that was sent after the sender sent
snap: thus causality is not violated by lining up all the pre-snap operations
before all the post-snap ones.
The full Chandy-Lamport algorithm adds a second marker message that
is used to sweep messages in transit out of the communications channels,
which avoids the need to keep logs if we want to reconstruct what messages
are in transit (this can also be done with the logical clock version). The
idea is that when a process records its state after receiving the snap mes-
CHAPTER 7. LOGICAL CLOCKS 53

sage, it issues a marker message on each outgoing channel. For incoming

channels, the process all records all messages received between the snapshot
and receiving a marker message on that channel (or nothing if it receives
marker before receiving snap). A process only reports its value when it has
received a marker on each channel. The marker and snap messages can also
be combined if the broadcast algorithm for snap resends it on all channels
anyway, and a further optimization is often to piggyback both on messages
of the underlying protocol if the underlying protocol is chatty enough.
Note that Chandy-Lamport is equivalent to the logical-time snapshot
using Lamport clocks, if the snap message is treated as a message with a
very large timestamp. For Neiger-Toueg-Welch clocks, we get an algorithm
where processes spontaneously decide to take snapshots (since Neiger-Toueg-
Welch clocks aren’t under the control of the snapshot algorithm) and delay
post-snapshot messages until the local snapshot has been taken. This can
be implemented as in Chandy-Lamport by separating pre-snapshot messages
from post-snapshot messages with a marker message, and essentially turns
into Chandy-Lamport if we insist that a process advance its clock to the
snapshot time when it receives a marker.

7.3.1.1 Property testing

Consistent snapshots are in principle useful for debugging (since one can
gather a consistent state of the system without being able to talk to every
process simultaneously), and in practice are mostly used for detecting stable
properties of the system. Here a stable property is some predicate on
global configurations that remains true in any successor to a configuration
in which it is true, or (bending the notion of properties a bit) functions
on configurations whose values don’t change as the protocol runs. Typical
examples are quiescence and its evil twin, deadlock. More exotic examples
include total money supply in a banking system that cannot create or destroy
money, or the fact that every process has cast an irrevocable vote in favor
of some proposal or advanced its Neiger-Toueg-Welch-style clock past some
threshold.
The reason we can test such properties using consistent snapshot is that
when the snapshot terminates with value C in some configuration C 0 , even
though C may never have occurred during the actual execution of the pro-
tocol, there is an execution which leads from C to C 0 . So if P holds in C,
stability means that it holds in C 0 .
Naturally, if P doesn’t hold in C, we can’t say much. So in this case we
re-run the snapshot protocol and hope we win next time. If P eventually
CHAPTER 7. LOGICAL CLOCKS 54

holds, we will eventually start the snapshot protocol after it holds and obtain
a configuration (which again may not correspond to any global configuration
that actually occurs) in which P holds.

7.3.2 Replicated state machines

[[[ this either needs more detail or needs to be deferred until we
get to Paxos ]]]
The main application for suggested by Lamport in his logical-clocks pa-
per [Lam78] was building a replicated state machine. In his construction,
any process can at any time issue an operation on the object by broadcast-
ing it with an attached timestamp. When a process receives an operation,
it buffers it in a priority queue ordered by increasing timestamp. It can ap-
ply the first operation in the queue only when it can detect that no earlier
operation will arrive, which it can do if it sees a message from every other
process with a later timestamp (or after a timeout, if we have some sort of
clock synchronization guarantee). It is not terribly hard to show that this
guarantees that every replica gets the same sequence of operations applied
to it, and that these operations are applied in an order consistent with the
processes’ ability to determine the actual order in which they were proposed.
Furthermore, if the processes spam each other regularly with their current
clock values, each operation will take effect after at most two message delays
(with Lamport clocks) if the clocks are not very well synchronized and af-
ter approximately one message delay (with Lamport or Neiger-Toueg-Welch
clocks) if they are. A process can also execute read operations on its own
copy immediately without notifying other processes (if it is willing to give
up linearizability for sequential consistency).
However, this particular construction assumes no failures, so for poorly-
synchronized clocks or systems in which sequentially-consistent reads are not
good enough, replicated state machines are no better than simply keeping
one copy of the object on a single process and having all operations go
through that process: 2 message delays + 2 messages per operation for the
single copy beats 2 message delays + many messages for full replication. But
replicated state machines take less time under good conditions, and when
augmented with more powerful tools like consensus or atomic broadcast are
the basis of most fault-tolerant implementations of general shared-memory
objects. [[[ We will see more examples of this idea [put in refs] ]]]
Chapter 8

Synchronizers

Synchronizers simulate an execution of a failure-free synchronous system

in a failure-free asynchronous system. See [AW04, Chapter 11] or [Lyn96,
Chapter 16] for a detailed (and rigorous) presentation.

8.1 Definitions
Formally, a synchronizer sits between the underlying network and the pro-
cesses and does one of two things:

• A global synchronizer guarantees that no process receives a message

for round r until all processes have sent their messages for round r.

• A local synchronizer guarantees that no process receives a message

for round r until all of that process’s neighbors have sent their messages
for round r.

In both cases the synchronizer packages all the incoming round r mes-
sages m for a single process together and delivers them as a single action
recv(p, m, r). Similarly, a process is required to hand over all of its outgoing
round-r messages to the synchronizer as a single action send(p, m, r)—this
prevents a process from changing its mind and sending an extra round-r
message or two. It is easy to see that the global synchronizer produces ex-
ecutions that are effectively indistinguishable from synchronous executions,
assuming that a synchronous execution is allowed to have some variability
in exactly when within a given round each process does its thing. The local
synchronizer only guarantees an execution that is locally indistinguishable
from an execution of the global synchronizer: an individual process can’t

55
CHAPTER 8. SYNCHRONIZERS 56

tell the difference, but comparing actions at different (especially widely sep-
arated) processes may reveal some process finishing round r + 1 while others
are still stuck in round r or earlier. Whether this is good enough depends
on what you want: it’s bad for coordinating simultaneous missile launches,
but may be just fine for adapting a synchronous message-passing algorithm
(e.g. for distributed breadth-first search as described in Chapter 5) to an
asynchronous system, if we only care about the final states of the processes
and not when precisely those states are reached.
Formally, the relation between global and local synchronization is de-
scribed by the following lemma:

Lemma 8.1.1. For any schedule S of a locally synchronous execution, there

is a schedule S 0 of a globally synchronous execution such that S|p = S 0 |p for
all processes p.

Proof. Essentially, we use the same happens-before relation as in Chap-

ter 7, and the fact that if a schedule S 0 is a causal shuffle of another schedule
S (i.e., a permutation of T that preserves causality), then S 0 |p = S|p for all
p (Lemma 7.1.1).
Given a schedule S, consider a schedule S 0 in which the events are ordered
first by increasing round and then by putting all sends before receives. This
ordering is consistent with ⇒S , so it’s a causal shuffle of S and S 0 |p = S|p.
But it’s globally synchronized, because no round-r operations at all happen
before a round-(r − 1) operation.

8.2 Implementations
These all implement at least a local synchronizer (the beta synchronizer is
global). The names were chosen by their inventor, Baruch Awerbuch [Awe85].
The main difference between them is the mechanism used to determine
when round-r messages have been delivered.
In the alpha synchronizer, every node sends a message to every neigh-
bor in every round (possibly a dummy message if the underlying protocol
doesn’t send a message); this allows the receiver to detect when it’s gotten
all its round-r messages (because it expects to get a message from every
neighbor) but may produce huge blow-ups in message complexity in a dense
graph.
In the beta synchronizer, messages are acknowledged by their receivers
(doubling the message complexity), so the senders can detect when all of
their messages are delivered. But now we need a centralized mechanism to
CHAPTER 8. SYNCHRONIZERS 57

collect this information from the senders and distribute it to the receivers,
since any particular receiver doesn’t know which potential senders to wait
for. This blows up time complexity, as we essentially end up building a
global synchronizer with a central leader.
The gamma synchronizer combines the two approaches at different lev-
els to obtain a trade-off between messages and time that depends on the
structure of the graph and how the protocol is organized.
Details of each synchronizer are given below.

8.2.1 The alpha synchronizer

The alpha synchronizer uses local information to construct a local synchro-
nizer. In round r, the synchronizer at p sends p’s message (tagged with the
round number) to each neighbor p0 or noMsg(r) if it has no messages. When
it collects a message or noMsg from each neighbor for round r, it delivers
all the messages. It’s easy to see that this satisfies the local synchronization
specification.
This produces no change in time but may drastically increase message
complexity because of all the extra noMsg messages flying around. For a
synchronous protocol that runs in T rounds with M messages, the same
protocol running with the alpha synchronizer will run in T time units, but
the message complexity may go up to M + T · |E| messages.

8.2.2 The beta synchronizer

The beta synchronizer centralizes detection of message delivery using a
rooted directed spanning tree (previously constructed). When p0 receives
a round-r message from p, it responds with ack(r). When p collects an ack
for all the messages it sent plus an OK from all of its children, it sends OK to
its parent. When the root has all the ack and OK messages it is expecting, it
broadcasts go. Receiving go makes p deliver the queued round-r messages.
This works because in order for the root to issue go, every round-r mes-
sage has to have gotten an acknowledgment, which means that all round-r
messages are waiting in the receivers’ buffers to be delivered. For the beta
synchronizer, message complexity increases slightly from M to 2M +2(n−1),
but time complexity goes up by a factor proportional to the depth of the
tree.
CHAPTER 8. SYNCHRONIZERS 58

8.2.3 The gamma synchronizer

The gamma synchronizer combines the alpha and beta synchronizers to try
to get low blowups on both time complexity and message complexity. The
essential idea is to cover the graph with a spanning forest and run beta
within each tree and alpha between trees. Specifically:

• Every message in the underlying protocol gets acked (including mes-

sages that pass between trees).

• When a process has collected all of its outstanding round-r acks, it

sends OK up its tree.

• When the root of a tree gets all acks and OK, it sends ready to the
roots of all adjacent trees (and itself). Two trees are adjacent if any
of their members are adjacent.

• When the root collects ready from itself and all adjacent roots, it broad-
casts go through its own tree.

As in the alpha synchronizer, we can show that no root issues go unless it

and all its neighbors issue ready, which happens only after both all nodes in
the root’s tree and all their neighbors (some of whom might be in adjacent
trees) have received acks for all messages. This means that when a node
receives go it can safely deliver its bucket of messages.
Message complexity is comparable to the beta synchronizer assuming
there aren’t too many adjacent trees: 2M messages for sends and acks,
plus O(n) messages for in-tree communication, plus O(Eroots ) messages for
root-to-root communication. Time complexity per synchronous round is
proportional to the depth of the trees: this includes both the time for in-
tree communication, and the time for root-to-root communication, which
might need to be routed through leaves.
In a particularly nice graph, the gamma synchronizer can give costs
comparable to the costs of the original synchronous algorithm. An example
in [Lyn96] is a ring of k-cliques, where we build a tree in each clique and get
O(1) time blowup and O(n) added messages. This is compared to O(n/k)
time blowup for beta and O(k) message blowup (or worse) for alpha. Other
graphs may favor tuning the size of the trees in the forest toward the alpha
or beta ends of the spectrum, e.g., if the whole graph is a clique (and we
didn’t worry about contention issues), we might as well just use beta and
get O(1) time blowup and O(n) added messages.
CHAPTER 8. SYNCHRONIZERS 59

8.3 Applications
See [AW04, §11.3.2] or [Lyn96, §16.5]. The one we have seen is distributed
breadth-first search, where the two asynchronous algorithms we described
in Chapter 5 were essentially the synchronous algorithms with the beta and
alpha synchronizers embedded in them. But what synchronizers give us
in general is the ability to forget about problems resulting from asynchrony
provided we can assume no failures (which may be a very strong assumption)
and are willing to accept a bit of overhead.

8.4 Limitations of synchronizers

Here we show some lower bounds on synchronizers, justifying our previous
claim that failures are trouble and showing that global synchronizers are
necessarily slow in a high-diameter network.

8.4.1 Impossibility with crash failures

[[[ this made more sense before we moved it before synchronous
agreement and FLP ]]]
The synchronizers above all fail badly if some process crashes. In the α
synchronizer, the system slowly shuts down as a wave of waiting propagates
out from the dead process. In the β synchronizer, the root never gives the
green light for the next round. The γ synchronizer, true to its hybrid nature,
fails in a way that is a hybrid of these two disasters.
This is unavoidable in the basic asynchronous model. Suppose that we
had a synchronizer that could tolerate crash failures (here, the process that
crashed in the asynchronous model would also appear to crash in the simu-
lated synchronous model, but everybody else would keep going). Then we
could use this fault-tolerant synchronizer to turn either of the synchronous
agreement protocols from Chapter 9 into an asynchronous protocol tolerat-
ing arbitrarily many crash failures. But this contradicts the FLP impossi-
bility result from Chapter 11.
We’ll see more examples of this trick of showing that a particular simu-
lation is impossible because it would allow us to violate impossibility results
later, especially when we start looking at the strength of shared-memory
objects in Chapter 18.
CHAPTER 8. SYNCHRONIZERS 60

8.4.2 Unavoidable slowdown with global synchronization

The session problem gives a lower bound on the speed of a global syn-
chronizer, or more generally on any protocol that tries to approximate syn-
chrony in a certain sense. Recall that in a global synchronizer, our goal is
to produce a simulation that looks synchronous from the outside; that is,
that looks synchronous to an observer that can see the entire schedule. In
contrast, a local synchronizer produces a simulation that looks synchronous
from the inside—the resulting execution is indistinguishable from a syn-
chronous execution to any of the processes, but an outside observer can
see that different processes execute different rounds at different times. The
global synchronizer we’ve seen takes more time than a local synchronizer;
the session problem shows that this is necessary.
In our description, we will mostly follow [AW04, §6.2.2].
A solution to the session problem is an asynchronous protocol in which
each process repeatedly executes some special action. Our goal is to guar-
antee that these special actions group into s sessions, where a session is an
interval of time in which every process executes at least one special action.
We also want the protocol to terminate: this means that in every execution,
every process executes a finite number of special actions.
A synchronous system can solve this problem trivially in s rounds: each
process executes one special action per round. For an asynchronous system,
a lower bound of Attiya and Mavronicolas [AM94] (based on an earlier
bound of Arjomandi, Fischer, and Lynch [AFL83], who defined the problem
in a slightly different communication model), shows that if the diameter of
the network is D, there is no solution to the s-session problem that takes
(s − 1)D time or less in the worst case. The argument is based on reordering
events in any such execution to produce fewer than s sessions, using the
happens-before relation from Chapter 7.
We now give an outline of the proof that this is expensive. (See [AW04,
§6.2.2] for the real proof.)
Fix some algorithm A for solving the s-session problem, and suppose
that its worst-case time complexity is (s − 1)D or less. Consider some
synchronous execution of A (that is, one where the adversary scheduler
happens to arrange the schedule to be synchronous) that takes (s − 1)D
rounds or less. Divide this execution into two segments: an initial segment
β that includes all rounds with special actions, and a suffix δ that includes
any extra rounds where the algorithm is still floundering around. We will
mostly ignore δ, but we have to leave it in to allow for the possibility that
whatever is happening there is important for the algorithm to work (say, to
CHAPTER 8. SYNCHRONIZERS 61

detect termination).
We now want to perform a causal shuffle on β that leaves it with only
s − 1 sessions. The first step is to chop β into at most s − 1 segments
β1 , β2 , . . . of at most D rounds each. Because the diameter of the network
is D, there exist processes p0 and p1 such that no chain of messages starting
at p0 within some segment reaches p1 before the end of the segment. It
follows that for any events e0 of p0 and e1 of p1 in the same segment βi , it
is not the case that e0 ⇒βδ e1 . So there exists a causal shuffle of βi that
puts all events of p0 after all events of p1 . By a symmetrical argument, we
can similarly put all events of p1 after all events of p0 . In both cases the
resulting schedule is indistinguishable by all processes from the original.
So now we apply these shuffles to each of the segments βi in alternating
order: p0 goes first in the even-numbered segments and p1 goes first in the
odd-numbered segments, yielding a sequence of shuffled segments βi0 . This
has the effect of putting the p0 events together, as in this example with
(s − 1) = 4:

βδ|(p0 , p1 ) = β1 β2 β3 β4 δ|(p0 , p1 )
= β10 β20 β30 β40 δ|(p0 , p1 )
= (p1 p0 )(p0 p1 )(p1 p0 )(p0 p1 )δ
= p1 (p0 p0 )(p1 p1 )(p0 p0 )p1 δ

(here each p0 , p1 stands in for a sequence of events of each process).

[[[ this notation is very confusing ]]]
Now let’s count sessions. We can’t end a session until we reach a point
where both processes have taken at least one step since the end of the last
session. If we mark with a slash the earliest places where this can happen,
we get a picture like this:

p1 p0 /p0 p1 /p1 p0 /p0 p1 /p1 δ.

We have at most s − 1 sessions! This concludes the proof.

Chapter 9

Synchronous agreement

Here we’ll consider synchronous agreement algorithm with stopping failures,

where a process stops dead at some point, sending and receiving no further
messages. We’ll also consider Byzantine failures, where a process deviates
from its programming by sending arbitrary messages, but mostly just to see
how crash-failure algorithms hold up; for algorithms designed specifically for
a Byzantine model, see Chapter 10.
If the model has communication failures instead, we have the coordinated
attack problem from Chapter 3.

9.1 Problem definition

We use the usual synchronous model with n processes with binary inputs
and binary outputs. Up to f processes may fail at some point; when a
process fails, one or one or more of its outgoing messages are lost in the
round of failure and all outgoing messages are lost thereafter.
There are two variants on the problem, depending on whether we want
a useful algorithm (and so want strong conditions to make our algorithm
more useful) or a lower bound (and so want weak conditions to make our
lower bound more general). For algorithms, we will ask for these conditions
to hold:

Agreement All non-faulty processes decide the same value.

Validity If all processes start with the same input, all non-faulty processes
decide it.

Termination All non-faulty processes eventually decide.

62
CHAPTER 9. SYNCHRONOUS AGREEMENT 63

For lower bounds, we’ll replace validity with non-triviality (often called
validity in the literature):

Non-triviality There exist failure-free executions A and B that produce

different outputs.

Non-triviality follows from validity but doesn’t imply validity; for exam-
ple, a non-trivial algorithm might have the property that if all non-faulty
processes start with the same input, they all decide something else. We’ll
start by using non-triviality, agreement, and termination to show a lower
bound on the number of rounds needed to solve the problem.

9.2 Lower bound on rounds

Here we show that synchronous agreement requires at least f + 1 rounds
if f processes can fail. This proof is modeled on the one in [Lyn96, §6.7]
and works backwards from the final state; for a proof of the same result
that works in the opposite direction, see [AW04, §5.1.4]. The original result
(stated for Byzantine failures) is due to Dolev and Strong [DS83], based on
a more complicated proof due to Fischer and Lynch [FL82]; see the chapter
notes for Chapter 5 of [AW04] for more discussion of the history.
Like the similar proof for coordinated attack (§3.2), the proof uses an
indistinguishability argument. But we have to construct a more complicated
chain of intermediate executions.
A crash failure at process i means that (a) in some round r, some or
all of the messages sent by i are not delivered, and (b) in subsequent rounds,
no messages sent by i are delivered. The intuition is that i keels over dead
in the middle of generating its outgoing messages for a round. Otherwise i
behaves perfectly correctly. A process that crashes at some point during an
execution is called faulty
We will show that if up to f processes can crash, and there are at least
f + 2 processes, then at least f + 1 rounds are needed (in some execution)
for any algorithm that satisfies agreement, termination, and non-triviality.
In particular, we will show that if all executions run in f or fewer rounds,
then the indistinguishability graph is connected; this implies non-triviality
doesn’t hold, because (as in §3.2), two adjacent states must decide the same
value because of the agreement property.1
1
The same argument works with even a weaker version of non-triviality that omits the
requirement that A and B are failure-free, but we’ll keep things simple.
CHAPTER 9. SYNCHRONOUS AGREEMENT 64

Now for the proof. To simplify the argument, let’s assume that all ex-
ecutions terminate in exactly f rounds (we can always have processes send
pointless chitchat to pad out short executions) and that every processes
sends a message to every other process in every round where it has not
crashed (more pointless chitchat). Formally, this means we have a sequence
of rounds 0, 1, 2, . . . , f −1 where each process sends a message to every other
process (assuming no crashes), and a final round f where all processes decide
on a value (without sending any additional messages).
We now want to take any two executions A and B and show that both
produce the same output. To do this, we’ll transform A’s inputs into B’s
inputs one process at a time, crashing processes to hide the changes. The
problem is that just crashing the process whose input changed might change
the decision value—so we have to crash later witnesses carefully to maintain
indistinguishability all the way across the chain.
Let’s say that a process p crashes fully in round r if it crashes in round
r and no round-r messages from p are delivered. The communication
pattern of an execution describes which messages are delivered between
processes without considering their contents—in particular, it tells us which
processes crash and what other processes they manage to talk to in the
round in which they crash.
With these definitions, we can state and prove a rather complicated
induction hypothesis:
Lemma 9.2.1. For any f -round protocol with n ≥ f + 2 process permitting
up to f crash failures; any process p; and any execution A in which at
most one processes crashes per round in rounds 0 . . . r − 1, p crashes fully in
round r + 1, and no other processes crash; there is a sequence of executions
A = A0 A1 . . . Ak such that each Ai is indistinguishable from Ai+1 by some
process, each Ai has at most one crash per round, and the communication
pattern in Ak is identical to A except that p crashes fully in round r.
Proof. By induction on f − r. If r = f , we just crash p in round r and
nobody else notices. For r < f , first crash p in round r instead of r + 1, but
deliver all of its round-r messages anyway (this is needed to make space for
some other process to crash in round r + 1). Then choose some message m
sent by p in round r, and let p0 be the recipient of m. We will show that we
can produce a chain of indistinguishable executions between any execution
in which m is delivered and the corresponding execution in which it is not.
If r = f − 1, this is easy; only p0 knows whether m has been delivered,
and since n ≥ f +2, there exists another non-faulty p00 that can’t distinguish
between these two executions, since p0 sends no messages in round f or later.
CHAPTER 9. SYNCHRONOUS AGREEMENT 65

If r < f − 1, we have to make sure p0 doesn’t tell anybody about the missing
message.
By the induction hypothesis, there is a sequence of executions starting
with A and ending with p0 crashing fully in round r + 1, such that each exe-
cution is indistinguishable from its predecessor. Now construct the sequence

A → (A with p0 crashing fully in r + 1)

→ (A with p0 crashing fully in r + 1 and m lost)
→ (A with m lost and p0 not crashing).

The first and last step apply the induction hypothesis; the middle one yields
indistinguishable executions since only p0 can tell the difference between m
arriving or not and its lips are sealed.
We’ve shown that we can remove one message through a sequence of
executions where each pair of adjacent executions is indistinguishable to
some process. Now paste together n − 1 such sequences (one per message)
to prove the lemma.

The rest of the proof: Crash some process fully in round 0 and then
change its input. Repeat until all inputs are changed.

9.3 Solutions
Here we give two solutions to synchronous agreement with crash failures.
The first, due to Dolev and Strong [DS83], is more practical but does not gen-
eralize well to Byzantine failures. The second is a variant on the exponential
information gathering algorithm of Pease, Shostak, and Lamport [PSL80],
which propagates enough information that it can in principle simulate any
other possible algorithm; it is mostly of interest because it can be used for
the Byzantine case as well.

9.3.1 Flooding
We’ll now show an algorithm that gets agreement, termination, and validity.
Validity here is stronger than the non-triviality condition used in the lower
bound, but the lower bound still applies: we can’t do this in less than f + 1
rounds.
So let’s do it in exactly f +1 rounds. There are two standard algorithms,
one of which generalizes to Byzantine processes under good conditions. We’ll
start with a simple approach based on flooding. This algorithm is described
CHAPTER 9. SYNCHRONOUS AGREEMENT 66

in more detail in [AW04, §5.1.3] or [Lyn96, §6.2.1]; the original is due to

Dolev and Strong [DS83].
Assumes very trustworthy processes. Each process keeps a set of (pro-
cess, input) pairs, initially just {(myId, myInput)}. At round r, I broadcast
my set to everybody and take the union of my set and all sets I receive.
At round f + 1, I decide on f (S), where f is some fixed function from sets
of process-input pairs to outputs that picks some input in S: for example,
f might take the input with the smallest process-id attached to it, take
the max of all known input values, or take the majority of all known input
values.

Lemma 9.3.1. After f + 1 rounds, all non-faulty processes have the same
set.

Proof. Let Sir be the set of process i after r rounds. What we’ll really show
is that if there are no failures in round k, then Sir = Sjr = Sik+1 for all i,
j, and r > k. To show this, observe that no faults in round k means that
all processes that are still alive at the start of round k send their message
to all other processes. Let L be the set of live processes in round k. At the
end of round k, for i in L we have Sik+1 = j∈L Sjk = S. Now we’ll consider
S

some round r = k + 1 + m and show by induction on m that Sik+m = S;

we already did m = 0, so for larger m notice that all messages are equal
to S and so Sik+1+m is the union of a whole bunch of S’s. So in particular
we have Sif +1 = S (since some failure-free round occurred in the preceding
f + 1 rounds) and everybody decides the same value f (S).

Flooding depends on being able to trust second-hand descriptions of

values; it may be that process 1 fails in round 0 so that only process 2 learns
its input. If process 2 can suddenly tell 3 (but nobody else) about the input
in round f + 1—or worse, tell a different value to 3 and 4—then we may
get disagreement. This remains true even if Byzantine processes can’t fake
inputs (e.g., because an input value is really a triple (i, v, signature(v)) using
an unforgeable digital signature)—the reason is that a Byzantine process
could horde some input (i, v, signature(v)) until the very last round and then
deliver it to only some of the non-faulty processes.

9.4 Exponential information gathering

The idea of exponential information gathering is that each process
will do a lot of gossiping, but now its state is no longer just a flat set of
CHAPTER 9. SYNCHRONOUS AGREEMENT 67

inputs, but a tree describing who it heard what from. We build this tree
out of pairs of the form (id-sequence, input) where id-sequence is a sequence
of intermediaries with no repetitions and input is some input. A process’s
state at each round is just a set of such pairs.
This is not really an improvement on flooding for crash failures, but it
can be used as a basis for building an algorithm for Byzantine agreement
(Chapter 10). Also useful as an example of a full-information algorithm,
in which every process records all that it knows about the execution; in
principle this allows the algorithm to simulate any other algorithm, which
can sometimes be useful for proving lower bounds.
See [AW04, §5.2.4] or [Lyn96, §6.2.3] for more details than we provide
here. The original exponential information-gathering algorithm (for Byzan-
tine processes) is due to Pease, Shostak, and Lamport [PSL80].
Initial state is (hi, myInput).
At round r, process i broadcasts all pairs (w, v) where |w| = r and i
does not appear in w (these restrictions make the algorithm slightly less
exponential). Upon receiving (w, v) from j, i adds (wj, v) to its list. If no
message arrives from j in round r, i adds (wj, ⊥) to its list for all non-
repeating w with |w| = r (this step can also be omitted).
A tree structure is obtained by letting w be the parent of wj for each j.
At round f + 1, apply some fixed decision rule to the set of all values
that appear in the tree (e.g. take the max, or decide on a default value v0
if there is more than one value in the tree). That this works follows pretty
much immediately from the fact that the set of node labels propagates just
as in the flooding algorithm (which is why EIG isn’t really an improvement).
But there are some complications from the messages that aren’t sent due to
the i-not-in-w restriction on sending. So we have to write out a real proof.
Below is basically just following the presentation in [Lyn96].
Let val(w, i) be the value v such that (w, v) appears in i’s list at the
end of round f + 1. We don’t worry about what earlier round it appears in
because we can compute that round as |w| + 1.

9.4.1 Basic invariants

• val(hi, i) = i’s input.

• Either val(wj, i) equals val(w, j) or val(wj, i) = ⊥ and j didn’t send a

message in round |w| + 1.

These are trivial.

CHAPTER 9. SYNCHRONOUS AGREEMENT 68

9.4.2 Stronger facts

• If val(xjy, i) 6= ⊥ then val(x, j) = val(xjy, i). Apply the invariant
inductively.

• If val(w, i) 6= ⊥ then it equals val(hi, j) for some j. Either w = hi and

we win or we can apply the previous fact to w = jy.

• If val(w, i) 6= ⊥ then there is some w0 that doesn’t contain i such that

val(w0 , i) = val(w, i). Let val(w, i) = v. If w doesn’t contain i we are
done, otherwise w = w0 iy for some w0 and y, and thus val(w0 , i) = v.

9.4.3 The payoff

Let Sir be the set of values in i’s list after r rounds. We’ll show that Sif +1 =
Sjf +1 for all non-faulty i and j. Let v be in Sif +1 . Then v = val(w, i) for
some w that doesn’t contain i (w here is really w0 from before). If |w| ≤ f ,
then i sends (w, v) to j at round |wi| and so val(wi, j) = v. Otherwise if
|w| = f + 1, w = xky for some non-faulty k, and from the first stronger fact
we have val(x, k) = v. Since k is non-faulty, it sends (x, v) to both i and j
in round |x| and we get val(xk, j) = v. We’ve just shown v in Sif +1 implies
v in Sjf +1 , and by symmetry the converse holds, so the sets are equal.
This is a lot of work to avoid sending messages that contain my own id!
However, by tacking on digital signatures, we can solve Byzantine agreement
in the case where f < n/2: see [Lyn96, §6.2.4] for details.

9.4.4 The real payoff

Run the same algorithm in a Byzantine system with n > 3f processes (treat-
ing bogus-looking messages as nulls), but compute the decision value by
taking recursive majorities of non-null values down the tree. Details are in
§10.2.1.

9.5 Variants
So far we have described binary consensus, since all inputs are 0 or 1. We
can also allow larger input sets. With crash failures, this allows a stronger
validity condition: the output must be equal to some input. Note that this
stronger condition doesn’t work if we have Byzantine failures. (Exercise:
why not?)
Chapter 10

Byzantine agreement

Like synchronous agreement (as in Chapter 9) except that we replace crash

failures with Byzantine failures, where a faulty process can ignore its
programming and send any messages it likes. Since we are operating under
a universal quantifier, this includes the case where the Byzantine processes
appear to be colluding with each other under the control of a centralized
adversary.

10.1 Lower bounds

We’ll start by looking at lower bounds.

10.1.1 Minimum number of rounds

We’ve already seen an f + 1 lower bound on rounds for crash failures (see
§9.2). This lower bound applies a fortiori to Byzantine failures, since Byzan-
tine failures can simulate crash failures.

10.1.2 Minimum number of processes

We can also show that we need n > 3f processes. For n = 3 and f = 1 the
intuition is that Byzantine B can play non-faulty A and C off against each
other, telling A that C is Byzantine and C that A is Byzantine. Since A is
telling C the same thing about B that B is saying about A, C can’t tell the
difference and doesn’t know who to believe. Unfortunately, this tragic soap
opera is not a real proof, since we haven’t actually shown that B can say
exactly the right thing to keep A and C from guessing that B is evil.

69
CHAPTER 10. BYZANTINE AGREEMENT 70

A0 B0 A0 B0

C1 C0

Č B1 A1

Figure 10.1: Three-process vs. six-process execution in Byzantine agreement

lower bound. Processes A0 and B0 in right-hand execution receive same mes-
sages as in left-hand three-process execution with Byzantine Č simulation
C0 through C1 . So validity forces them to decide 0. A similar argument
using Byzantine Ǎ shows the same for C0 .

The real proof:1 Consider an artificial execution where (non-Byzantine)

A, B, and C are duplicated and then placed in a ring A0 B0 C0 A1 B1 C1 ,
where the digits indicate inputs. We’ll still keep the same code for n = 3
on A0 , B0 , etc., but when A0 tries to send a message to what it thinks of
as just C we’ll send it to C1 while messages from B0 will instead go to C0 .
For any adjacent pair of processes (e.g. A0 and B0 ), the behavior of the rest
of the ring could be simulated by a single Byzantine process (e.g. C), so
each process in the 6-process ring behaves just as it does in some 3-process
execution with 1 Byzantine process. It follows that all of the processes
terminate and decide in the unholy 6-process Frankenexecution2 the same
value that they would in the corresponding 3-process Byzantine execution.
So what do they decide?
Given two processes with the same input, say, A0 and B0 , the giant exe-
cution is indistinguishable from an A0 B0 Č execution where Č is Byzantine
(see Figure 10.1. Validity says A0 and B0 must both decide 0. Since this
works for any pair of processes with the same input, we have each process
deciding its input. But now consider the execution of C0 A1 B̌, where B̌ is
Byzantine. In the big execution, we just proved that C0 decides 0 and A1
decides 1, but since the C0 A1 B execution is indistinguishable from the big
execution to C0 and A1 , they do the same thing here and violate agreement.
This shows that with n = 3 and f = 1, we can’t win. We can generalize
this to n = 3f . Suppose that there were an algorithm that solved Byzantine
agreement with n = 3f processes. Group the processes into groups of size f ,
1
The presentation here is based on [AW04, §5.2.3]. The original impossibility result
is due to Pease, Shostak, and Lamport [PSL80]. This particular proof is due to Fischer,
Lynch, and Merritt [FLM86].
2
Not a real word.
CHAPTER 10. BYZANTINE AGREEMENT 71

B0

B0 A0 D0

A0 D0 C0
C1
Č A1
B1
D1

Figure 10.2: Four-process vs. eight-process execution in Byzantine agree-

ment connectivity lower bound. Because Byzantine Č can simulate
C0 , D1 , B1 , A1 , and C1 , A0 , B0 and D0 must all decide 0 or risk violating
validity.

and let each of the n = 3 processes simulate one group, with everybody in
the group getting the same input, which can only make things easier. Then
we get a protocol for n = 3 and f = 1, an impossibility.

10.1.3 Minimum connectivity

So far, we’ve been assuming a complete communication graph. If the graph is
not complete, we may not be able to tolerate as many failures. In particular,
we need the connectivity of the graph (minimum number of nodes that must
be removed to split it into two components) to be at least 2f +1. See [Lyn96,
§6.5] for the full proof. The essential idea is that if we have an arbitrary
graph with a vertex cut of size k < 2f + 1, we can simulate it on a 4-process
graph where A is connected to B and C (but not D), B and C are connected
to each other, and D is connected only to B and C. Here B and C each
simulate half the processes in the size-k cut, A simulates all the processes
on one side of the cut and D all the processes on the other side. We then
construct an 8-process artificial execution with two non-faulty copies of each
of A, B, C, and D and argue that if one of B or C can be Byzantine then
the 8-process execution is indistinguishable to the remaining processes from
a normal 4-process execution. (See Figure 10.1.)
An argument similar to the n > 3f proof then shows we violate one
of validity or agreement: if we replacing C0 , C1 , and all the nodes on one
side of the C0 + C1 cut with a single Byzantine Č, we force the remaining
CHAPTER 10. BYZANTINE AGREEMENT 72

non-faulty nodes to decide their inputs or violate validity. But then doing
the same thing with B0 and B1 yields an execution that violates agreement.
Conversely, if we have connectivity 2f +1, then the processes can simulate
a general graph by sending each other messages along 2f + 1 predetermined
vertex-disjoint paths and taking the majority value as the correct message.
Since the f Byzantine processes can only corrupt one path each (assuming
the non-faulty processes are careful about who they forward messages from),
we get at least f + 1 good copies overwhelming the f bad copies. This
reduces the problem on a general graph with sufficiently high connectivity
to the problem on a complete graph, allowing Byzantine agreement to be
solved if the other lower bounds are met.

10.1.4 Weak Byzantine agreement

(Here we are following [Lyn96, §6.6]. The original result is due to Lam-
port [Lam83].)
Weak Byzantine agreement is like regular Byzantine agreement, but
validity is only required to hold if there are no faulty processes at all.3 If
there is a single faulty process, the non-faulty processes can output any value
regardless of their inputs (as long as they agree on it). Sadly, this weakening
doesn’t improve things much: even weak Byzantine agreement can be solved
only if n ≥ 3f + 1.
Proof: As in the strong Byzantine agreement case, we’ll construct a
many-process Frankenexecution to figure out a strategy for a single Byzan-
tine process in a 3-process execution. The difference is that now the number
of processes in our synthetic execution is much larger, since we want to
build an execution where at least some of our test subjects think they are
in a non-Byzantine environment. The trick is to build a very big, highly-
symmetric ring so that at least some of the processes are so far away from
the few points of asymmetry that might clue them in to their odd condition
that the protocol terminates before they notice.
Fix some protocol that allegedly solves weak Byzantine agreement, and
let r be the number of rounds for the protocol. Construct a ring of 6r pro-
cesses A01 B01 C01 A02 B02 C02 . . . A0r B0r C0r A10 B10 C10 . . . A1r B1r C1r , where each
Xij runs the code for process X in the 3-process protocol with input i. For
3
An alternative might be to weaken agreement or termination to apply only if there
are no non-faulty processes, but this makes the problem trivial. If we weaken agreement,
we can just have each process decide whatever process 1 tells it to, and if we weaken
termination, we can do more or less the same thing except that we only terminate if all
the other processes tell us they heard the same value from process 1.
CHAPTER 10. BYZANTINE AGREEMENT 73

each adjacent pair of processes, there is a 3-process Byzantine execution

which is indistinguishable from the 6r-process execution for that pair: since
agreement holds in all Byzantine executions, each adjacent pair decides the
same value in the big execution and so either everybody decides 0 or every-
body decides 1 in the big execution.
Now we’ll show that means that validity is violated in some no-failures 3-
process execution. We’ll extract this execution by looking at the execution
of processes A0r/2 B0r/2 C0r/2 . The argument is that up to round r, any
input-0 process that is at least r steps in the ring away from the nearest
1-input process acts like the corresponding process in the all-0 no-failures
3-process execution. Since A0,r/2 is 3r/2 > r hops away from A1r and
similarly for C0,r/2 , our 3 stooges all decide 0 by validity. But now repeat
the same argument for A1,r/2 B1,r/2 C1,r/2 and get 3 new stooges that all
decide 1. This means that somewhere in between we have two adjacent
processes where one decides 0 and one decides 1, violating agreement in the
corresponding 3-process execution where the rest of the ring is replaced by
a single Byzantine process. This concludes the proof.
This result is a little surprising: we might expect that weak Byzantine
agreement could be solved by allowing a process to return a default value
if it notices anything that might hint at a fault somewhere. But this would
allow a Byzantine process to create disagreement revealing its bad behavior
to just one other process in the very last round of an execution otherwise
headed for agreement on the non-default value. The chosen victim decides
the default value, but since it’s the last round, nobody else finds out. Even
if the algorithm is doing something more sophisticated, examining the 6r-
process execution will tell the Byzantine process exactly when and how to
start acting badly.

10.2 Upper bounds

Here we describe two upper bounds for Byzantine agreement, one of which
gets an optimal number of rounds at the cost of many large messages, and
the other of which gets smaller messages at the cost of more rounds. (We
are following §§5.2.4–5.2.5 of [AW04] in choosing these algorithms.) Neither
of these algorithms is state-of-the-art, but they demonstrate some of the
issues in solving Byzantine agreement without the sometimes-complicated
optimizations needed to get all the parameters of the algorithm down simul-
taneously.
CHAPTER 10. BYZANTINE AGREEMENT 74

10.2.1 Exponential information gathering gets n = 3f + 1

We’ll show that a variant of Exponential Information Gathering as defined in
§9.4 works with n ≥ 3f + 1 in f + 1 rounds. This is the same technique used
by Pease, Shostak, and Lamport [PSL80] to show that their impossibility
result is tight.
Recall EIG gives us at each node a set of pairs (path, value) where path
spans all sequences of 0 to n distinct ids and value is the input value for-
warded along that path. We write val(w, i) for the value stored in i’s list at
the end of the protocol that is associated with path w. Because we can’t
trust these val(w, i) values to be an accurate description of any process’s
input if there is a Byzantine process in w, each process computes for itself
replacement values val0 (w, i) that use majority voting to try to get a more
trustworthy picture of the original inputs.
Formally, we think of the set of paths as a tree where w is the parent of
wj for each path w and each id j not in w. To apply EIG in the Byzantine
model, ill-formed messages received from j are treated as missing messages,
but otherwise the data-collecting part of EIG proceeds as in the crash failure
model. However, we compute the decision value from the last-round values
recursively as follows. First replace any missing pair involving a path w with
|w| = f + 1 with (w, 0). Then for each path w, define val0 (w, i) to be the
majority value among val0 (wj, i) for all j, or val(w, i) if |w| = f + 1. Finally,
we have process i decide val0 (hi, i) (which it can compute locally from its
own stored values val(w, i)).
The val0 is a reconstruction of old values from later ones: as we move
up the tree from wj to w we are moving backwards in time, until in the
end we get the decision value val0 (hi, i) as a majority of reconstructed inputs
val0 (j, i). One way to think about this is that I don’t trust j to give me
the right value for wj—even when w = hi and j is just reporting its own
input—so instead a take a majority of values of wj that j allegedly reported
to other people. But since I don’t trust those other people either, I use the
same process recursively to construct those reports.

10.2.1.1 Proof of correctness

This is just a sketch of the proof from [Lyn96, §6.3.2]; essentially the same
argument appears in [AW04, §5.2.4].
We start with a basic observation that good processes send and record
values correctly:
CHAPTER 10. BYZANTINE AGREEMENT 75

Lemma 10.2.1. If i, j, and k are all non-faulty then for all w, val(wk, i) =
val(wk, j) = val(w, k).

Proof. Trivial: k announces the same value val(w, k) to both i and j.

More involved is this lemma, which says that when we reconstruct a

value for a trustworthy process at some level, we get the same value that
it sent us. In particular this will be used to show that the reconstructed
inputs val0 (j, i) are all equal to the real inputs for good processes.

Lemma 10.2.2. If j is non-faulty then val0 (wj, i) = val(wj, i) for all non-
faulty i and all w.

Proof. By induction on f + 1 − |wj|. If |wj| = f + 1, then val0 (wj, i) =

val(wj, i) = val(w, j) = val(wj, i). If |wj| < f + 1, then val(wj, k) = val(w, j)
for all non-faulty k. It follows that val(wjk, i) = val(w, j) for all non-faulty
i and k (that do no appear in w). The bad guys report at most f bad
values val(wj, k 0 ), but the good guys report at least n − f − |wj| good values
val(wj, k). Since n ≥ 3f + 1 and |wj| ≤ f , we have n − f − |wj| ≥ 3f + 1 −
f − f ≥ f + 1 good values, which are a majority.

We call a node w common val0 (w, i) = val0 (w, j) for all non-faulty i, j.
Lemma 10.2.2 says that wk is common if k is non-faulty. We can also show
that any node whose children are all common is also common, whether or
not the last process in its label is faulty.

Lemma 10.2.3. Let wk be common for all k. Then w is common.

Proof. Recall that, for |w| < f + 1, val0 (w, i) is the majority value among
all val0 (wk, i). If all wk are common, then val0 (wk, i) = val0 (wk, j) for all
non-faulty i and j, so i and j compute the same majority values and get
val0 (w, i) = val0 (w, j).

We can now prove the full result.

Theorem 10.2.4. Exponential information gathering using f + 1 rounds

in a synchronous Byzantine system with at most f faulty processes satisfies
validity and agreement, provided n ≥ 3f + 1.

Proof. Validity: Immediate application of Lemmas 10.2.1 and 10.2.2 when

w = hi. We have val0 (j, i) = val(j, i) = val(hi, j) for all non-faulty j and i,
which means that a majority of the val0 (j, i) values equal the common input
and thus so does val0 (hi, i).
CHAPTER 10. BYZANTINE AGREEMENT 76

Agreement: Observe that every path has a common node on it, since
a path travels through f + 1 nodes and one of them is good. If we then
suppose that the root is not common: by Lemma 10.2.3, it must have a
not-common child, that node must have a not-common child, etc. But this
constructs a path from the root to a leaf with no not-common nodes, which
we just proved can’t happen.

10.2.2 Phase king gets constant-size messages

The following algorithm, based on work of Berman, Garay, and Perry[BGP89],
achieves Byzantine agreement in 2(f + 1) rounds using constant-size mes-
sages, provided n ≥ 4f + 1. The description here is drawn from [AW04,
§5.2.5]. The original Berman-Garay-Perry paper gives somewhat better
bounds, but they’re more complicated.

10.2.2.1 The algorithm

The basic idea of the algorithm is that we avoid the recursive majority voting
of EIG by running a vote in each of f +1 phases through a phase king, some
process chosen in advance to run the phase. Since the number of phases
exceeds the number of faults, we eventually get a non-faulty phase king.
The algorithm is structured so that one non-faulty phase king is enough
to generate agreement and subsequent faulty phase kings can’t undo the
agreement.
Pseudocode appears in Algorithm 10.1. Each processes i maintains an
array pref i [j], where j ranges over all process ids. There are also utility
values majority, kingMajority and multiplicity for each process that are used
to keep track of what it hears from the other processes. Initially, pref i [i] is
just i’s input and pref i [j] = 0 for j 6= i.
The idea of the algorithm is that in each phase, everybody announces
their current preference (initially the inputs). If the majority of these pref-
erences is large enough (e.g. all inputs are the same), everybody adopts
the majority preference. Otherwise everybody adopts the preference of the
phase king. The majority rule means that once the processes agree, they
continue to agree despite bad phase kings. The phase king rule allows a
good phase king to end disagreement. By choosing a different king in each
phase, after f +1 phases, some king must be good. This intuitive description
is justified below.
CHAPTER 10. BYZANTINE AGREEMENT 77

1 pref i [i] = input

2 for j 6= i do pref i [j] = 0
3 for k ← 1 to f + 1 do
// First round of phase k
4 send pref i [i] to all processes (including myself)
5 pref i [j] ← vj , where vj is the value received from process j
6 majority ← majority value in pref i
7 multiplicity ← number of times majority appears in pref i
// Second round of phase k
8 if i = k then
// I am the phase king
9 send majority to all processes
10 receive kingMajority from phase king
11 if multiplicity > n/2 + f then
12 pref i [i] = majority
13 else
14 pref i [i] = kingMajority

15 return pref i [i]

Algorithm 10.1: Byzantine agreement: phase king
CHAPTER 10. BYZANTINE AGREEMENT 78

10.2.2.2 Proof of correctness

Termination is immediate from the algorithm.
For validity, suppose all inputs are v. We’ll show that all non-faulty i
have pref i [i] = v after every phase. In the first round of each phase, process
i receives at least n − f messages containing v; since n ≥ 4f + 1, we have
n − f ≥ 3f + 1 and n/2 + f ≤ (4f + 1)/2 + f = 3f + 1/2, and thus these
n − f messages exceed the n/2 + f threshold for adopting them as the new
preference. So all non-faulty processes ignore the phase king and stick with
v, eventually deciding v after round 2(f + 1).
For agreement, we’ll ignore all phases up to the first phase with a non-
faulty phase king. Let k be the first such phase, and assume that the pref
values are set arbitrarily at the start of this phase. We want to argue that
at the end of the phase, all non-faulty processes have the same preference.
There are two ways that a process can set its new preference in the second
round of the phase:

1. The process i observes a majority of more than n/2+f identical values

v and ignores the phase king. Of these values, more than n/2 of them
were sent by non-faulty processes. So the phase king also receives these
values (even if the faulty processes change their stories) and chooses
v as its majority value. Similarly, if any other process j observes a
majority of n/2 + f identical values, the two > n/2 non-faulty parts
of the majorities overlap, and so j also chooses v.

2. The process i takes its value from the phase king. We’ve already shown
that i then agrees with any j that sees a big majority; but since the
phase king is non-faulty, process i will agree with any process j that
also takes its new preference from the phase king.

This shows that after any phase with a non-faulty king, all processes
agree. The proof that the non-faulty processes continue to agree is the same
as for validity.

10.2.2.3 Performance of phase king

It’s not hard to see that this algorithm sends exactly (f +1)(n2 +n) messages
of 1 bit each (assuming 1-bit inputs). The cost is doubling the minimum
number of rounds and reducing the tolerance for Byzantine processes. As
mentioned earlier, a variant of phase-king with 3-round phases gets optimal
fault-tolerance with 3(f + 1) rounds (but 2-bit messages). Still better is
CHAPTER 10. BYZANTINE AGREEMENT 79

a rather complicated descendant of the EIG algorithm due to Garay and

Moses [GM98], which gets f + 1 rounds with n ≥ 3f + 1 while still having
polynomial message traffic.
Chapter 11

Impossibility of
asynchronous agreement

The Fischer-Lynch-Paterson (FLP) result [FLP85] says that you can’t do

agreement in an asynchronous message-passing system if even one crash
failure is allowed, unless you augment the basic model in some way, e.g.
by adding randomization or failure detectors. After its initial publication,
it was quickly generalized to other models including asynchronous shared
memory [LAA87], and indeed the presentation of the result in [Lyn96, §12.2]
is given for shared-memory first, with the original result appearing in [Lyn96,
§17.2.3] as a corollary of the ability of message passing to simulate shared
memory. In these notes, I’ll present the original result; the dependence on
the model is surprisingly limited, and so most of the proof is the same for
both shared memory (even strong versions of shared memory that support
e.g. atomic snapshots1 ) and message passing.
Section 5.3 of [AW04] gives a very different version of the proof, where
it is shown first for two processes in shared memory, then generalized to n
processes in shared memory by adapting the classic Borowsky-Gafni simu-
lation [BG93] to show that two processes with one failure can simulate n
processes with one failure. This is worth looking at (it’s an excellent exam-
ple of the power of simulation arguments, and BG simulation is useful in
many other contexts) but we will stick with the original argument, which
is simpler. We will look at this again when we consider BG simulation in
Chapter 27.
1
Chapter 19.

80
CHAPTER 11. IMPOSSIBILITY OF ASYNCHRONOUS AGREEMENT81

11.1 Agreement
Usual rules: agreement (all non-faulty processes decide the same value),
termination (all non-faulty processes eventually decide some value), va-
lidity (for each possible decision value, there an execution in which that
value is chosen). Validity can be tinkered with without affecting the proof
much.
To keep things simple, we assume the only two decision values are 0 and
1.

11.2 Failures
A failure is an internal action after which all send operations are disabled.
The adversary is allowed one failure per execution. Effectively, this means
that any group of n − 1 processes must eventually decide without waiting
for the n-th, because it might have failed.

11.3 Steps
The FLP paper uses a notion of steps that is slightly different from the
send and receive actions of the asynchronous message-passing model we’ve
been using. Essentially a step consists of receiving zero or more messages
followed by doing a finite number of sends. To fit it into the model we’ve
been using, we’ll define a step as either a pair (p, m), where p receives
message m and performs zero or more sends in response, or (p, ⊥), where
p receives nothing and performs zero or more sends. We assume that the
processes are deterministic, so the messages sent (if any) are determined by
p’s previous state and the message received. Note that these steps do not
correspond precisely to delivery and send events or even pairs of delivery
and send events, because what message gets sent in response to a particular
delivery may change as the result of delivering some other message; but this
won’t affect the proof.
The fairness condition essentially says that if (p, m) or (p, ⊥) is contin-
uously enabled it eventually happens. Since messages are not lost, once
(p, m) is enabled in some configuration C, it is enabled in all successor con-
figurations until it occurs; similarly (p, ⊥) is always enabled. So to ensure
fairness, we have to ensure that any non-faulty process eventually performs
any enabled step.
CHAPTER 11. IMPOSSIBILITY OF ASYNCHRONOUS AGREEMENT82

Comment on notation: I like writing the new configuration reached by

applying a step e to C like this: Ce. The FLP paper uses e(C).

11.4 Bivalence and univalence

The core of the FLP argument is a strategy allowing the adversary (who
controls scheduling) to steer the execution away from any configuration in
which the processes reach agreement. The guidepost for this strategy is
the notion of bivalence, where a configuration C is bivalent if there exist
traces T0 and T1 starting from C that lead to configurations CT0 and CT1
where all processes decide 0 and 1 respectively. A configuration that is not
bivalent is univalent, or more specifically 0-valent or 1-valent depending
on whether all executions starting in the configuration produce 0 or 1 as the
decision value. (Note that bivalence or univalence are the only possibilities
because of termination.) The important fact we will use about univalent
configurations is that any successor to an x-valent configuration is also x-
valent.
It’s clear that any configuration where some process has decided is not
bivalent, so if the adversary can keep the protocol in a bivalent configuration
forever, it can prevent the processes from ever deciding. The adversary’s
strategy is to start in an initial bivalent configuration C0 (which we must
prove exists) and then choose only bivalent successor configurations (which
we must prove is possible). A complication is that if the adversary is only
allowed one failure, it must eventually allow any message in transit to a
non-faulty process to be received and any non-faulty process to send its
outgoing messages, so we have to show that the policy of avoiding univalent
configurations doesn’t cause problems here.

11.5 Existence of an initial bivalent configuration

We can specify an initial configuration by specifying the inputs to all pro-
cesses. If one of these initial configurations is bivalent, we are done. Other-
wise, let C and C 0 be two initial configurations that differ only in the input
of one process p; by assumption, both C and C 0 are univalent. Consider two
executions starting with C and C 0 in which process p is faulty; we can ar-
range for these executions to be indistinguishable to all the other processes,
so both decide the same value x. It follows that both C and C 0 are x-valent.
But since any two initial configurations can be connected by some chain of
CHAPTER 11. IMPOSSIBILITY OF ASYNCHRONOUS AGREEMENT83

such indistinguishable configurations, we have that all initial configurations

are x-valent, which violations validity.

11.6 Staying in a bivalent configuration

Now start in a failure-free bivalent configuration C with some step e = (p, m)
or e = (p, ⊥) enabled in C. Let S be the set of configurations reachable
from C without doing e or failing any processes, and let e(S) be the set
of configurations of the form C 0 e where C 0 is in S. (Note that e is always
enabled in S, since once enabled the only way to get rid of it is to deliver
the message.) We want to show that e(S) contains a failure-free bivalent
configuration.
The proof is by contradiction: suppose that C 0 e is univalent for all C 0
in S. We will show first that there are C0 and C1 in S such that each Ci e
is i-valent. To do so, consider any pair of i-valent Ai reachable from C; if
Ai is in S, let Ci = Ai . If Ai is not in S, let Ci be the last configuration
before executing e on the path from C to Ai (Ci e is univalent in this case
by assumption).
So now we have C0 e and C1 e with Ci e i-valent in each case. We’ll now go
hunting for some configuration D in S and step e0 such that De is 0-valent
but De0 e is 1-valent (or vice versa); such a pair exists because S is connected
and so some step e0 crosses the boundary between the C 0 e = 0-valent and
the C 0 e = 1-valent regions.
By a case analysis on e and e0 we derive a contradiction:

1. Suppose e and e0 are steps of different processes p and p0 . Let both

steps go through in either order. Then Dee0 = De0 e, since in an
asynchronous system we can’t tell which process received its message
first. But De is 0-valent, which implies Dee0 is also 0-valent, which
contradicts De0 e being 1-valent.

2. Now suppose e and e0 are steps of the same process p. Again we let
both go through in either order. It is not the case now that Dee0 =
De0 e, since p knows which step happened first (and may have sent
messages telling the other processes). But now we consider some finite
sequence of steps e1 e2 . . . ek in which no message sent by p is delivered
and some process decides in Dee1 . . . ek (this occurs since the other
processes can’t distinguish Dee0 from the configuration in which p
died in D, and so have to decide without waiting for messages from
p). This execution fragment is indistinguishable to all processes except
CHAPTER 11. IMPOSSIBILITY OF ASYNCHRONOUS AGREEMENT84

p from De0 ee1 . . . ek , so the deciding process decides the same value i
in both executions. But Dee0 is 0-valent and De0 e is 1-valent, giving
a contradiction.

It follows that our assumption was false, and there is some reachable
bivalent configuration C 0 e.
Now to construct a fair execution that never decides, we start with a
bivalent configuration, choose the oldest enabled action and use the above
to make it happen while staying in a bivalent configuration, and repeat.

11.7 Generalization to other models

To apply the argument to another model, the main thing is to replace the
definition of a step and the resulting case analysis of 0-valent De0 e vs 1-valent
Dee0 to whatever steps are available in the other model. For example, in
asynchronous shared memory, if e and e0 are operations on different memory
locations, they commute (just like steps of different processes), and if they
are operations on the same location, either they commute (e.g. two reads)
or only one process can tell whether both happened (e.g. with a write and
a read, only the reader knows, and with two writes, only the first writer
knows). Killing the witness yields two indistinguishable configurations with
different valencies, a contradiction.
We are omitting a lot of details here. See [Lyn96, §12.2] for the real
proof, or Loui and Abu-Amara [LAA87] for the generalization to shared
memory, or Herlihy [Her91b] for similar arguments for a wide variety of
shared-memory primitives. We will see many of these latter arguments in
Chapter 18.
Chapter 12

Paxos

The Paxos algorithm for consensus in a message-passing system was first

described by Lamport in 1990 in a tech report that was widely considered
to be a joke (see http://research.microsoft.com/users/lamport/pubs/
pubs.html#lamport-paxos for Lamport’s description of the history). The
algorithm was finally published in 1998 [Lam98], and after the algorithm
continued to be ignored, Lamport finally gave up and translated the results
into readable English [Lam01]. It is now understood to be one of the most
efficient practical algorithms for achieving consensus in a message-passing
system with failure detectors, mechanisms that allow processes to give up
on other stalled processes after some amount of time (which can’t be done
in a normal asynchronous system because giving up can be made to happen
immediately by the adversary).
We will describe only the basic Paxos algorithm. The WikiPedia arti-
cle on Paxos (http://en.wikipedia.org/wiki/Paxos) gives a remarkably
good survey of subsequent developments and applications.

12.1 Motivation: replicated state machines

A replicated state machine is an object that is replicated across multiple
machines, with some mechanism for propagating operations on the object
to all replicas. Formally, we think of an object as having a set of states
Q, together with a transition relation δ that maps pairs of operations and
states to pairs of responses and states. Applying an operation to the object
means that we change the state to the output of δ and return the given
response. If we have failures, we will need some sort of consensus protocol
to coordinate which operations are applied and in what order. Paxos is a

85
CHAPTER 12. PAXOS 86

common choice, although other choices of consensus protocols will work too.

12.2 The Paxos algorithm

The algorithm runs in a message-passing model with asynchrony and less
than n/2 crash failures (but not Byzantine failures, at least in the original
algorithm). As always, we want to get agreement, validity, and termination.
The Paxos algorithm itself is mostly concerned with guaranteeing agreement
and validity while allowing for the possibility of termination if there is a long
enough interval in which no process restarts the protocol.
Processes are classified as proposers, accepters, and learners (a sin-
gle process may have all three roles). The idea is that a proposer attempts
to ratify a proposed decision value (from an arbitrary input set) by collect-
ing acceptances from a majority of the accepters, and this ratification is
observed by the learners. Agreement is enforced by guaranteeing that only
one proposal can get the votes of a majority of accepters, and validity follows
from only allowing input values to be proposed. The tricky part is ensuring
that we don’t get deadlock when there are more than two proposals or when
some of the processes fail. The intuition behind how this works is that any
proposer can effectively restart the protocol by issuing a new proposal (thus
dealing with lockups), and there is a procedure to release accepters from
their old votes if we can prove that the old votes were for a value that won’t
be getting a majority any time soon.
To organize this vote-release process, we attach a distinct proposal num-
ber to each proposal. The safety properties of the algorithm don’t depend
on anything but the proposal numbers being distinct, but since higher num-
bers override lower numbers, to make progress we’ll need them to increase
over time. The simplest way to do this in practice is to make the proposal
number be a timestamp with the proposer’s id appended to break ties. We
could also have the proposer poll the other processes for the most recent
proposal number they’ve seen and add 1 to it.
The revoting mechanism now works like this: before taking a vote, a
proposer tests the waters by sending a prepare(n) message to all accepters,
where n is the proposal number. An accepter responds to this with a promise
never to accept any proposal with a number less than n (so that old proposals
don’t suddenly get ratified) together with the highest-numbered proposal
that the accepter has accepted (so that the proposer can substitute this
value for its own, in case the previous value was in fact ratified). If the
proposer receives a response from a majority of the accepters, the proposer
CHAPTER 12. PAXOS 87

then does a second phase of voting where it sends accept(n, v) to all accepters
and wins if receives a majority of votes.
So for each proposal, the algorithm proceeds as follows:

1. The proposer sends a message prepare(n) to all accepters. (Sending

to only a majority of the accepters is enough, assuming they will all
respond.)

2. Each accepter compares n to the highest-numbered proposal for which

it has responded to a prepare message. If n is greater, it responds with
ack(n, v, nv ), where v is the highest-numbered proposal it has accepted
and nv is the number of that proposal (or ⊥ and 0 if there is no such
proposal). (An optimization at this point is to allow the accepter to
send back nack(n0 ) where n0 is some higher number to let the proposer
know that it’s doomed and should back off and try again—this keeps
a confused proposer who thinks it’s the future from locking up the
protocol until 2037.)

3. The proposer waits (possibly forever) to receive ack from a major-

ity of accepters. If any ack contained a value, it sets v to the most
recent (in proposal number ordering) value that it received. It then
sends accept(n, v) to all accepters (or just a majority). You should
think of accept as a demand (“Accept!”) rather than acquiescence (“I
accept”)—the accepters still need to choose whether to accept or not.

4. Upon receiving accept(n, v), an accepter accepts v unless it has already

received prepare(n0 ) for some n0 > n. If a majority of acceptors accept
the value of a given proposal, that value becomes the decision value of
the protocol.

Note that acceptance is a purely local phenomenon; additional messages

are needed to detect which if any proposals have been accepted by a majority
of accepters. Typically this involves a fourth round, where accepters send
accepted(n, v) to all learners (often just the original proposer).
There is no requirement that only a single proposal is sent out (indeed,
if proposers can fail we will need to send out more to jump-start the proto-
col). The protocol guarantees agreement and validity no matter how many
proposers there are and no matter how often they start.
CHAPTER 12. PAXOS 88

12.3 Informal analysis: how information flows be-

tween rounds
Call a round the collection of all messages labeled with some particular
proposal n. The structure of the algorithm simulates a sequential execution
in which higher-numbered rounds follow lower-numbered ones, even though
there is no guarantee that this is actually the case in a real execution.
When an acceptor sends ack(n, v, nv ), it is telling the round-n proposer
the last value preceding round n that it accepted. The rule that an acceptor
only acknowledges a proposal higher than any proposal it has previously ac-
knowledged prevents it from sending information “back in time”—the round
nv in an acknowledgment is always less than n. The rule that an acceptor
doesn’t accept any proposal earlier than a round it has acknowledged means
that the value v in an ack(n, v, nv ) message never goes out of date—there is
no possibility that an acceptor might retroactively accept some later value
in round n0 with nv < n0 < n. So the ack message values tell a consistent
story about the history of the protocol, even if the rounds execute out of
order.
The second trick is to use overlapping majorities to make sure that any
value that is accepted is not lost. If the only way to decide on a value in round
n is to get a majority of acceptors to accept it, and the only way to make
progress in round n0 is to get acknowledgments from a majority of acceptors,
these two majorities overlap. So in particular the overlapping process reports
the round-n proposal value to the proposer in round n0 , and we can show
by induction on n0 that this round-n proposal value becomes the proposal
value in all subsequent rounds that proceed past the acknowledgment stage.
So even though it may not be possible to detect that a decision has been
reached in round n (say, because some of the acceptors in the accepting
majority die without telling anybody what they did), no later round will be
able to choose a different value. This ultimately guarantees agreement.

12.4 Safety properties

We now present a more formal analysis of the Paxos protocol. We con-
sider only the safety properties of the protocol, corresponding to validity
and agreement; without additional assumptions, Paxos does not guarantee
termination.
Call a value chosen if it is accepted by a majority of accepters. The
safety properties of Paxos are:
CHAPTER 12. PAXOS 89

• No value is chosen unless it is first proposed. (This gives validity.)

• No two distinct values are both chosen. (This gives agreement.)

The first property is immediate from examination of the algorithm.

For the second property, we need some invariants. The intuition is that if
some value is chosen, then a majority of accepters have accepted it for some
proposal number n. Any proposal sent in an accept message with a higher
number n0 must be sent by a proposer that has seen an overlapping majority
respond to its prepare(n0 ) message. If we consider the process that overlaps,
this process must have accepted v before it received prepare(n0 ), since it can’t
accept afterwards, and unless it has accepted some other proposal since, it
responds with ack(n0 , v, n). If these are the only values that the proposer
receives with number n or greater, it chooses v as its new value.
Worrying about what happens in rounds between n and n0 is messy, so
we’ll use two formal invariants (taken more or less directly from Lamport’s
paper):

Invariant 1 An accepter accepts a proposal numbered n if and only if it

has not responded to a prepare message with a number n0 > n.

Invariant 2 For any v and n, if a proposal with value v and number n has
been issued (by sending accept messages), then there is a majority of
accepters S such that either (a) no accepter in S has accepted any
proposal numbered less than n, or (b) v is the value of the highest-
numbered proposal among all proposals numbered less than n accepted
by at least one accepter in S.

The proof of the first invariant is immediate from the rule for issuing
acks.
The proof of the second invariant follows from the first invariant and the
proposer’s rule for issuing proposals: it can only do so after receiving ack
from a majority of accepters—call this set S—and the value it issues is either
the proposal’s initial value if all responses are ack(n, ⊥, 0), or the maximum
value sent in by accepters in S if some responses are ack(n, v, nv ). In the
first case we have case (a) of the invariant: nobody accepted any proposals
numbered less than n before responding, and they can’t afterwards. In the
second case we have case (b): the maximum response value is the maximum-
numbered accepted value within S at the time of each response, and again no
new values numbered less than n will be accepted afterwards. Amazingly,
none of this depends on the temporal ordering of different proposals or
CHAPTER 12. PAXOS 90

messages: the accepters enforce that their acks are good for all time by
refusing to change their mind about earlier rounds later.
So now we suppose that some value v is eventually accepted by a majority
T with number n. Then we can show by induction on proposal number that
all proposals issued with higher numbers have the same value (even if they
were issued earlier). For any proposal accept(v 0 , n0 ) with n0 > n, there is
a majority S (which must overlap with T ) for which either case (a) holds
(a contradiction—once the overlapping accepter finally accepts, it violates
the requirement that no proposal less than n0 has been accepted) or case
(b) holds (in which case by the induction hypothesis v 0 is the value of some
earlier proposal with number n0 ≥ n, implying v 0 = v).

12.5 Learning the results

Somebody has to find out that a majority accepted a proposal in order to
get a decision value out. The usual way to do this is to have a fourth round
of messages where the accepters send chose(v, n) to some designated learner
(usually just the original proposer), which can then notify everybody else if
it doesn’t fail first. If the designated learner does fail first, we can restart by
issuing a new proposal (which will get replaced by the previous successful
proposal because of the safety properties).

12.6 Liveness properties

We’d like the protocol to terminate eventually. Suppose there is a single
proposer, and that it survives long enough to collect a majority of acks
and to send out accepts to a majority of the accepters. If everybody else
cooperates, we get termination in 3 message delays.
If there are multiple proposers, then they can step on each other. For
example, it’s enough to have two carefully-synchronized proposers alternate
sending out prepare messages to prevent any accepter from every accepting
(since an accepter promises not to accept accept(n, v) once it has responded
to prepare(n + 1)). The solution is to ensure that there is eventually some
interval during which there is exactly one proposer who doesn’t fail. One
way to do this is to use exponential random backoff (as popularized by
Ethernet): when a proposer decides it’s not going to win a round (e.g. by
receiving a nack or by waiting long enough to realize it won’t be getting
any more acks soon), it picks some increasingly large random delay before
CHAPTER 12. PAXOS 91

starting a new round; thus two or more will eventually start far enough
apart in time that one will get done without interference.
A more abstract solution is to assume some sort of weak leader election
mechanism, which tells each accepter who the “legitimate” proposer is at
each time. The accepters then discard messages from illegitimate proposers,
which prevents conflict at the cost of possibly preventing progress. Progress
is however obtained if the mechanism eventually reaches a state where a
majority of the accepters bow to the same non-faulty proposer long enough
for the proposal to go through.
Such a weak leader election method is an example of a more general
class of mechanisms known as failure detectors, in which each process
gets hints about what other processes are faulty that eventually converge to
reality. The particular failure detector in this case is known as the Ω failure
detector; there are other still weaker ones that we will talk about later that
can also be used to solve consensus. We will discuss failure detectors in
detail in Chapter 13.
Chapter 13

Failure detectors

Failure detectors were proposed by Chandra and Toueg [CT96] as a mech-

anism for solving consensus in an asynchronous message-passing system with
crash failures by distinguishing between slow processes and dead processes.
The basic idea is that each process has attached to it a failure detector mod-
ule that continuously outputs an estimate of which processes in the system
have failed. The output need not be correct; indeed, the main contribution
of Chandra and Toueg’s paper (and a companion paper by Chandra, Hadzi-
lacos, and Toueg [CHT96]) is characterizing just how bogus the output of a
failure detector can be and still be useful.
We will mostly follow Chandra and Toueg in these notes; see the paper
for the full technical details.
To emphasize that the output of a failure detector is merely a hint at
the actual state of the world, a failure detector (or the process it’s attached
to) is said to suspect a process at time t if it outputs failed at that time.
Failure detectors can then be classified based on when their suspicions are
correct.
We use the usual asynchronous message-passing model, and in particular
assume that non-faulty processes execute infinitely often, get all their mes-
sages delivered, etc. From time to time we will need to talk about time, and
unless we are clearly talking about real time this just means any steadily
increasing count (e.g., of total events), and will be used only to describe the
ordering of events.

92
CHAPTER 13. FAILURE DETECTORS 93

13.1 How to build a failure detector

Failure detectors are only interesting if you can actually build them. In
a fully asynchronous system, you can’t (this follows from the FLP result
and the existence of failure-detector-based consensus protocols). But with
timeouts, it’s not hard: have each process ping each other process from
time to time, and suspect the other process if it doesn’t respond to the ping
within twice the maximum round-trip time for any previous ping. Assuming
that ping packets are never lost and there is an (unknown) upper bound on
message delay, this gives what is known as an eventually perfect failure
detector: once the max round-trip times rise enough and enough time has
elapsed for the live processes to give up on the dead ones, all and only dead
processes are suspected.

13.2 Classification of failure detectors

Chandra and Toueg define eight classes of failure detectors, based on when
they suspect faulty processes and non-faulty processes. Suspicion of faulty
processes comes under the heading of completeness; of non-faulty pro-
cesses, accuracy.

13.2.1 Degrees of completeness

Strong completeness Every faulty process is eventually permanently sus-
pected by every non-faulty process.

Weak completeness Every faulty process is eventually permanently sus-

pected by some non-faulty process.

There are two temporal logic operators embedded in these statements:

“eventually permanently” means that there is some time t0 such that for
all times t ≥ t0 , the process is suspected. Note that completeness says
nothing about suspecting non-faulty processes: a paranoid failure detector
that permanently suspects everybody has strong completeness.

13.2.2 Degrees of accuracy

These describe what happens with non-faulty processes, and with faulty
processes that haven’t crashed yet.

Strong accuracy No process is suspected (by anybody) before it crashes.

CHAPTER 13. FAILURE DETECTORS 94

Weak accuracy Some non-faulty process is never suspected.

Eventual strong accuracy After some initial period of confusion, no pro-

cess is suspected before it crashes. This can be simplified to say that
no non-faulty process is suspected after some time, since we can take
end of the initial period of chaos as the time at which the last crash
occurs.

Eventual weak accuracy After some initial period of confusion, some

non-faulty process is never suspected.

Note that “strong” and “weak” mean different things for accuracy vs
completeness: for accuracy, we are quantifying over suspects, and for com-
pleteness, we are quantifying over suspectors. Even a weakly-accurate failure
detector guarantees that all processes trust the one visibly good process.

13.2.3 Boosting completeness

It turns out that any weakly-complete failure detector can be boosted to give
strong completeness. Recall that the difference between weak completeness
and strong completeness is that with weak completeness, somebody suspects
a dead process, while with strong completeness, everybody suspects it. So
to boost completeness we need to spread the suspicion around a bit. On
the other hand, we don’t want to break accuracy in the process, so there
needs to be some way to undo a premature rumor of somebody’s death.
The simplest way to do this is to let the alleged corpse speak for itself: I
will suspect you from the moment somebody else reports you dead until the
moment you tell me otherwise.
Pseudocode is given in Algorithm 13.1.

1 initially do
2 suspects ← ∅
3 while true do
4 Let S be the set of all processes my weak detector suspects.
5 Send S to all processes.
6 upon receiving S from q do
7 suspects ← (suspects ∪ p) \ {q}
Algorithm 13.1: Boosting completeness
CHAPTER 13. FAILURE DETECTORS 95

It’s not hard to see that this boosts completeness: if p crashes, some-
body’s weak detector eventually suspects it, this process tells everybody
else, and p never contradicts it. So eventually everybody suspects p.
What is slightly trickier is showing that it preserves accuracy. The es-
sential idea is this: if there is some good-guy process p that everybody trusts
forever (as in weak accuracy), then nobody ever reports p as suspect—this
also covers strong accuracy since the only difference is that now every non-
faulty process falls into this category. For eventual weak accuracy, wait for
everybody to stop suspecting p, wait for every message ratting out p to be
delivered, and then wait for p to send a message to everybody. Now every-
body trusts p, and nobody every suspects p again. Eventual strong accuracy
is again similar.
This will justify ignoring the weakly-complete classes.

13.2.4 Failure detector classes

Two degrees of completeness times four degrees of accuracy gives eight
classes of failure detectors, each of which gets its own name. But since
we can boost weak completeness to strong completeness, we can use this as
an excuse to consider only the strongly-complete classes.

P (perfect) Strongly complete and strongly accurate: non-faulty processes

are never suspected; faulty processes are eventually suspected by ev-
erybody. Easily achieved in synchronous systems.

S (strong) Strongly complete and weakly accurate. The name is mislead-

ing if we’ve already forgotten about weak completeness, but the corre-
sponding W (weak) class is only weakly complete and weakly accurate,
so it’s the strong completeness that the S is referring to.

♦P (eventually perfect) Strongly complete and eventually strongly ac-

curate.

♦S (eventually strong) Strongly complete and eventually weakly accu-

rate.

Jumping to the punch line: P can simulate any of the others, S and
♦P can both simulate ♦S but can’t simulate P or each other, and ♦S can’t
simulate any of the others (See Figure 13.1—we’ll prove all of this later.)
Thus ♦S is the weakest class of failure detectors in this list. However, ♦S is
strong enough to solve consensus, and in fact any failure detector (whatever
CHAPTER 13. FAILURE DETECTORS 96

S ♦P

♦S

Figure 13.1: Partial order of failure detector classes. Higher classes can
simulate lower classes.

its properties) that can solve consensus is strong enough to simulate ♦S

(this is the result in the Chandra-Hadzilacos-Toueg paper [CHT96])—this
makes ♦S the “weakest failure detector for solving consensus” as advertised.
Continuing our tour through Chandra and Toueg [CT96], we’ll show the
simulation results and that ♦S can solve consensus, but we’ll skip the rather
involved proof of ♦S’s special role from Chandra-Hadzilacos-Toueg.

13.3 Consensus with S

With the strong failure detector S, we can solve consensus for any number
of failures.
In this model, the failure detectors as applied to most processes are
completely useless. However, there is some non-faulty process c that nobody
every suspects, and this is enough to solve consensus with as many as n − 1
failures.
The basic idea of the protocol: There are three phases. In the first phase,
the processes gossip about input values for n − 1 asynchronous rounds. In
the second, they exchange all the values they’ve seen and prune out any
that are not universally known. In the third, each process decides on the
lowest-id input that hasn’t been pruned (minimum input also works since
at this point everybody has the same view of the inputs).
Pseudocode is given in Algorithm 13.2
In phase 1, each process p maintains two partial functions Vp and δp ,
where Vp lists all the input values hq, vq i that p has ever seen and δp lists
only those input values seen in the most recent of n−1 asynchronous rounds.
Both Vp and δp are initialized to {hp, vp i}. In round i, p sends (i, δp ) to all
processes. It then collects hi, δq i from each q that it doesn’t suspect and sets
δp to q δq \ Vp (where q ranges over the processes from which p received a
S
CHAPTER 13. FAILURE DETECTORS 97

1 Vp ← {hp, vp i}
2 δp ← {hp, vp i}
// Phase 1
3 for i ← 1 to n − 1 do
4 Send hi, δp i to all processes.
5 Wait to receive hi, δq i from all q I do not suspect.
S 
6 δp ← q δq \ Vp
S 
7 Vp ← q δq ∪ Vp
// Phase 2
8 Send hn, δp i to all processes.
9 Wait to receive hn, δq i from all q I do not suspect.
T 
10 Vp ← q Vq ∩ Vp
// Phase 3
11 return some input from Vp chosen via a consistent rule.
Algorithm 13.2: Consensus with a strong failure detector

message in round i) and sets Vp to Vp ∪ δp . In the next round, it repeats

the process. Note that each pair hq, vq i is only sent by a particular process
p the first round after p learns it: so any value that is still kicking around
in round n − 1 had to go through n − 1 processes.
In phase 2, each process p sends hn, Vp i, waits to receive hn, Vq i from
every process it does not suspect, and sets Vp to the intersection of Vp and
all received Vq . At the end of this phase all Vp values will in fact be equal,
as we will show.
In phase 3, everybody picks some input from their Vp vector according
to a consistent rule.

13.3.1 Proof of correctness

Let c be a non-faulty process that nobody every suspects.
The first observation is that the protocol satisfies validity, since every
Vp contains vc after round 1 and each Vp can only contain input values by
examination of the protocol. Whatever it may do to the other values, taking
intersections in phase 2 still leaves vc , so all processes pick some input value
from a nonempty list in phase 3.
To get termination we have to prove that nobody ever waits forever for
a message it wants; this basically comes down to showing that the first non-
CHAPTER 13. FAILURE DETECTORS 98

faulty process that gets stuck eventually is informed by the S-detector that
the process it is waiting for is dead.
For agreement, we must show that in phase 3, every Vp is equal; in
particular, we’ll show that every Vp = Vc . First it is necessary to show that
at the end of phase 1, Vc ⊆ Vp for all p. This is done by considering two
cases:

1. If hq, vq i ∈ Vc and c learns hq, vq i before round n − 1, then c sends

hq, vq i to p no later than round n − 1, p waits for it (since nobody ever
suspects c), and adds it to Vp .

2. If hq, vq i ∈ Vc and c learns hq, vq i only in round n − 1, then hq, vq i was

previously sent through n − 1 other processes, i.e., all of them. Each
process p 6= c thus added hq, vq i to Vp before sending it and again
hq, vq i is in Vp .

(The missing case where hq, vq i isn’t in Vc we don’t care about.)

But now phase 2 knocks out any extra elements in Vp , since Vp gets set
to Vp ∩ Vc ∩ (some other Vq ’s that are supersets of Vc ). It follows that, at
the end of phase 2, Vp = Vc for all p. Finally, in phase 3, everybody applies
the same selection rule to these identical sets and we get agreement.

13.4 Consensus with ♦S and f < n/2

The consensus protocol for S depends on some process c never being sus-
pected; if c is suspected during the entire (finite) execution of the proto-
col—as can happen with ♦S—then it is possible that no process will wait
to hear from c (or anybody else) and the processes will all decide their own
inputs. So to solve consensus with ♦S we will need to assume fewer than
n/2 failures, allowing any process to wait to hear from a majority no matter
what lies its failure detector is telling it.
The resulting protocol, known as the Chandra-Toueg consensus pro-
tocol, is structurally similar to the consensus protocol in Paxos.1 The differ-
ence is that instead of proposers blindly showing up, the protocol is divided
into rounds with a rotating coordinator pi in each round r with r = i
(mod n). The termination proof is based on showing that in any round
where the coordinator is not faulty and nobody suspects it, the protocol
finishes.
1
See Chapter 12.
CHAPTER 13. FAILURE DETECTORS 99

The consensus protocol uses as a subroutine a protocol for reliable

broadcast, which guarantees that any message that is sent is either re-
ceived by no non-faulty processes or exactly once by all non-faulty processes.
Pseudocode for reliable broadcast is given as Algorithm 13.3. It’s easy to
see that if a process p is non-faulty and receives m, then the fact that p is
non-faulty means that is successfully sends m to everybody else, and that
the other non-faulty processes also receive the message at least once and
deliver it.

1 procedure broadcast(m)
2 send m to all processes.
3 upon receiving m do
4 if I haven’t seen m before then
5 send m to all processes
6 deliver m to myself

Algorithm 13.3: Reliable broadcast

Here’s a sketch of the actual consensus protocol:

• Each process keeps track of a preference (initially its own input) and a
timestamp, the round number in which it last updated its preference.

• The processes go through a sequence of asynchronous rounds, each

divided into four phases:

1. All processes send (round, preference, timestamp) to the coordi-

nator for the round.
2. The coordinator waits to hear from a majority of the processes
(possibly including itself). The coordinator sets its own prefer-
ence to some preference with the largest timestamp of those it
receives and sends (round, preference) to all processes.
3. Each process waits for the new proposal from the coordinator
or for the failure detector to suspect the coordinator. If it re-
ceives a new preference, it adopts it as its own, sets timestamp
to the current round, and sends (round, ack) to the coordinator.
Otherwise, it sends (round, nack) to the coordinator.
4. The coordinator waits to receive ack or nack from a majority of
processes. If it receives ack from a majority, it announces the
CHAPTER 13. FAILURE DETECTORS 100

current preference as the protocol decision value using reliable

broadcast.
• Any process that receives a value in a reliable broadcast decides on it
immediately.
Pseudocode is in Algorithm 19.

1 preference ← input
2 timestamp ← 0
3 for round ← 1 . . . ∞ do
4 Send hround, preference, timestampi to coordinator
5 if I am the coordinator then
6 Wait to receive hround, preference, timestampi from majority of
processes.
7 Set preference to value with largest timestamp.
8 Send hround, preferencei to all processes.
9 Wait to receive round, preference0 from coordinator or to suspect
coordinator.
10 if I received round, preference0 then
11 preference ← preference0
12 timestamp ← round
13 Send ack(round) to coordinator.
14 else
15 Send nack(round) to coordinator.
16 if I am the coordinator then
17 Wait to receive ack(round) or nack(round) from a majority of
processes.
18 if I received no nack(round) messages then
19 Broadcast preference using reliable broadcast.

13.4.1 Proof of correctness

For validity, observe that the decision value is an estimate and all estimates
start out as inputs.
For termination, observe that no process gets stuck in phase 1, 2, or 4,
because either it isn’t waiting or it is waiting for a majority of non-faulty
processes who all sent messages unless they have already decided (this is
CHAPTER 13. FAILURE DETECTORS 101

why we need the nacks in phase 3). The loophole here is that processes
that decide stop participating in the protocol; but because any non-faulty
process retransmits the decision value in the reliable broadcast, if a process
is waiting for a response from a non-faulty process that already terminated,
eventually it will get the reliable broadcast instead and terminate itself.
In phase 3, a process might get stuck waiting for a dead coordinator, but
the strong completeness of ♦S means that it suspects the dead coordinator
eventually and escapes. So at worst we do finitely many rounds.
Now suppose that after some time t there is a process c that is never
suspected by any process. Then in the next round in which c is the co-
ordinator, in phase 3 all surviving processes wait for c and respond with
ack, c decides on the current estimate, and triggers the reliable broadcast
protocol to ensure everybody else decides on the same value. Since reli-
able broadcast guarantees that everybody receives the message, everybody
decides this value or some value previously broadcast—but in either case
everybody decides.
Agreement is the tricky part. It’s possible that two coordinators both
initiate a reliable broadcast and some processes choose the value from the
first and some the value from the second. But in this case the first coordi-
nator collected acks from a majority of processes in some round r, and all
subsequent coordinators collected estimates from an overlapping majority
of processes in some round r0 > r. By applying the same induction argu-
ment as for Paxos, we get that all subsequent coordinators choose the same
estimate as the first coordinator, and so we get agreement.

13.5 f < n/2 is still required even with ♦P

We can show that with a majority of failures, we’re in trouble with just ♦P
(and thus with ♦S, which is trivially simulated by ♦P ). The reason is that
♦P can lie to us for some long initial interval of the protocol, and consensus
is required to terminate eventually despite these lies. So the usual partition
argument works: start half of the processes with input 0, half with 1, and
run both halves independently with ♦P suspecting the other half until the
processes in both halves decide on their common inputs. We can now make
♦P happy by letting it stop suspecting the processes, but it’s too late.
CHAPTER 13. FAILURE DETECTORS 102

13.6 Relationships among the classes

It’s easy to see that P simulates S and ♦P simulates ♦S without modifica-
tion. It’s also immediate that P simulates ♦P and S simulates ♦S (make
“eventually” be “now”), which gives a diamond-shaped lattice structure be-
tween the classes. What is trickier is to show that this structure doesn’t
collapse: ♦P can’t simulate S, S can’t simulate ♦P , and ♦S can’t simulate
any of the other classes.
First let’s observe that ♦P can’t simulate S: if it could, we would get a
consensus protocol for f ≥ n/2 failures, which we can’t do. It follows that
♦P also can’t simulate P (because P can simulate S).
To show that S can’t simulate ♦P , choose some non-faulty victim process
v and consider an execution in which S periodically suspects v (which it is
allowed to do as long as there is some other non-faulty process it never
suspects). If the ♦P -simulator ever responds to this by refusing to suspect
v, there is an execution in which v really is dead, and the simulator violates
strong completeness. But if not, we violate eventual strong accuracy. Note
that this also implies S can’t simulate P , since P can simulate ♦P . It also
shows that ♦S can’t simulate either of ♦P or P .
We are left with showing ♦S can’t simulate S. Consider a system where
p’s ♦S detector suspects q but not r from the start of the execution, and
similarly r’s ♦S detector also suspects q but not p. Run p and r in isolation
until they give up and decide that q is in fact dead (which they must do
eventually by strong completeness, since this run is indistinguishable from
one in which q is faulty). Then wake up q and crash p and r. Since q is the
only non-faulty process, we’ve violated weak accuracy.
Chandra and Toueg [CT96] give as an example of a natural problem
that can be solved only with P the problem of terminating reliable
broadcast, in which a single leader process attempts to send a message
and all other processes eventually agree on the message if the leader is non-
faulty but must terminate after finite time with a default no message return
value if the leader is faulty.2 The process is solvable using P by just having
each process either wait for the message or for P to suspect the leader, which
can only occur if the leader does in fact crash. If the leader is dead, the
processes must eventually decide on no message; this separates P from ♦S
and ♦P since we can then wake up the leader and let it send its message.
But it also separates P from S, since we can have the S-detector only be
2
This is a slight weakening of the problem, which however still separates P from the
other classes. For the real problem see Chandra and Toueg [CT96].
CHAPTER 13. FAILURE DETECTORS 103

accurate for non-leaders. For other similar problems see the paper.
Chapter 14

Quorum systems

14.1 Basics
In the past few chapters, we’ve seen many protocols that depend on the fact
that if I talk to more than n/2 processes and you talk to more than n/2 pro-
cesses, the two groups overlap. This is a special case of a quorum system,
a family of subsets of the set of processes with the property that any two
subsets in the family overlap. By choosing an appropriate family, we may
be able to achieve lower load on each system member, higher availability,
defense against Byzantine faults, etc.
The exciting thing from a theoretical perspective is that these turn a
systems problem into a combinatorial problem: this means we can ask com-
binatorialists how to solve it.

14.2 Simple quorum systems

• Majority and weighted majorities

• Specialized read/write systems where write quorum is a column and

read quorum a row of some grid.

• Dynamic quorum systems: get more than half of the most recent copy.

• Crumbling walls [PW97b, PW97a]: optimal small-quorum system for

good choice of wall sizes.

104
CHAPTER 14. QUORUM SYSTEMS 105

14.3 Goals
• Minimize quorum size.

• Minimize load, defined as the minimum over all access strategies

(probability distributions on quorums) of the maximum over all servers
of probability it gets hit.

• Maximize capacity, defined as the maximum number of quorum ac-

cesses per time unit in the limit if each quorum access ties up a quorum
member for 1 time unit (but we are allowed to stagger a quorum access
over multiple time units).

• Maximize fault-tolerance: minimum number of server failures that

blocks all quorums. Note that for standard quorum systems this is
directly opposed to minimizing quorum size, since killing the smallest
quorum stops us dead.

• Minimize failure probability = probability that every quorum con-

tains at least one bad server, assuming each server fails with indepen-
dent probability.

Naor and Wool [NW98] describe trade-offs between these goals (some of
these were previously known, see the paper for citations):

• capacity = 1/load; this is obtained by selecting the quorums indepen-

dently at random according to the load-minimizing distribution. In
particular this means we can forget about capacity and just concen-
trate on minimizing load.

• load ≥ max(c/n, 1/c) where c is the minimum quorum size. The first
case is obvious: if every access hits c nodes, spreading them out as
evenly as possible still hits each node c/n of the time. The second is
trickier: Naor and Wool prove it using LP duality, but the argument
essentially says that if we have some quorum Q of size c, then since
every other quorum Q0 intersects Q in at least one place, we can show
that every Q0 adds at least 1 unit of load in total to the c members of
Q. So if we pick a random quorum Q0 , the average load added to all of
Q is at least 1, so the average load added to some particular element
of Q is at least 1/|Q| = 1/c. Combining the two cases, we can’t hope
√
to get load better than 1/ n, and to get this load we need quorums
√
of size at least n.
CHAPTER 14. QUORUM SYSTEMS 106

Figure 14.1: Figure 2 from [NW98]. Solid lines are G(3); dashed lines are
G∗ (3).

• failure probability is at least p when p > 1/2 (and optimal system is to

just pick a single leader in this case), failure probability can be made
exponentially small in size of smallest quorum when p < 1/2 (with
many quorums). These results are due to Peleg and Wool [PW95].

14.4 Paths system

This is an optimal-load system from Naor and Wool [NW98] with exponen-
tially low failure probability, based on percolation theory.
The basic idea is to build a d×d mesh-like graph where a quorum consists
of the union of a top-to-bottom path (TB path) and a left-to-right path (LR
√ √
path); this gives quorum size O( n) and load O(1/ n). Note that the TB
and LR paths are not necessarily direct: they may wander around for a
while in order to get where they are going, especially if there are a lot of
√
failures to avoid. But the smallest quorums will have size 2d + 1 = O( n).
The actual mesh is a little more complicated. Figure 14.1 reproduces
the picture of the d = 3 case from the Naor and Wool paper.
Each server corresponds to a pair of intersecting edges, one from the
CHAPTER 14. QUORUM SYSTEMS 107

G(d) grid and one from the G∗ (d) grid (the star indicates that G∗ (d) is the
dual graph1 of G(d). A quorum consists of a set of servers that produce an
LR path in G(d) and a TB path in G∗ (d). Quorums intersect, because any
LR path in G(d) must cross some TB path in G∗ (d) at some server (in fact,
each pair of quorums intersects in at least two places). The total number of
√
elements n is (d + 1)2 and the minimum size of a quorum is 2d + 1 = Θ( n).
The symmetry of the mesh gives that there exists a LR path in the
mesh if and only if there does not exist a TB path in its complement, the
graph that has an edge only if the mesh doesn’t. For a mesh with failure
probability p < 1/2, the complement is a mesh with failure probability
q = 1 − p > 1/2. Using results in percolation theory, it can be shown that
for failure probability q > 1/2, the probability that there exists a left-to-
right path is exponentially small in d (formally, for each p there is a constant
φ(p) such that Pr[∃LR path] ≤ exp(−φ(p)d)). We then have

Pr[∃(live quorum)] = Pr[∃(TB path) ∧ ∃(LR path)]

= Pr[¬∃(LR path in complement) ∨ ¬∃(TB path in complement)]
≤ Pr[¬∃(LR path in complement)] + Pr[¬∃(TB path in complement)]
≤ 2 exp(−φ(1 − p)d)
√
= 2 exp(−Θ( n)).

So the failure probability of this system is exponentially small for any fixed
p < 1/2.
See the paper [NW98] for more details.

14.5 Byzantine quorum systems

Standard quorum systems are great when you only have crash failures, but
with Byzantine failures you have to worry about finding a quorum that
includes a Byzantine serve who lies about the data. For this purpose you
need something stronger. Following Malkhi and Reiter [MR98] and Malkhi et
al. [MRWW01], one can define:

• A b-disseminating quorum system guarantees |Q1 ∩ Q2 | ≥ b +

1 for all quorums Q1 and Q2 . This guarantees that if I update a
1
See http://en.wikipedia.org/wiki/Dual_graph; the basic idea is that the dual of a
graph G embedded in the plane has a vertex for each region of G, and an edge connecting
each pair of vertices corresponding to adjacent regions, where a region is a subset of the
plane that is bounded by edges of G.
CHAPTER 14. QUORUM SYSTEMS 108

quorum Q1 and you update a quorum Q2 , and there are at most b

Byzantine processes, then there is some non-Byzantine process in both
our quorums. Mostly useful if data is “self-verifying,” e.g. signed with
digital signatures that the Byzantine processes can’t forge. Otherwise,
I can’t tell which of the allegedly most recent data values is the right
one since the Byzantine processes lie.

• A b-masking quorum system guarantees |Q1 ∩ Q2 | ≥ 2b + 1 for

all quorums Q1 and Q2 . (In other words, it’s the same as a 2b-
disseminating quorum system.) This allows me to defeat the Byzantine
processes through voting: given 2b + 1 overlapping servers, if I want
the most recent value of the data I take the one with the most recent
timestamp that appears on at least b + 1 servers, which the Byzantine
guys can’t fake.

An additional requirement in both cases is that for any set of servers B

with |B| ≤ b, there is some quorum Q such that Q ∩ B = ∅. This prevents
the Byzantine processes from stopping the system by simply refusing to
participate.
Note: these definitions are based on the assumption that there is some
fixed bound on the number of Byzantine processes. Malkhi and Reiter [MR98]
give more complicated definitions for the case where one has an arbitrary
family {B} of potential Byzantine sets. The definitions above are actually
simplified versions from [MRWW01].
The simplest way to build a b-disseminating quorum system is to use
supermajorities of size at least (n + b + 1)/2; the overlap between any two
such supermajorities is at least (n + b + 1) − n = b + 1. This gives a load of
1
substantially more
p than 2 . There are better constructions that knock the
load down to Θ( b/n); see [MRWW01].
For more on this topic in general, see the survey by by Merideth and
Reiter [MR10].

14.6 Probabilistic quorum systems

The problem with all standard (or strict) quorum systems is that we need
big quorums to get high fault tolerance, since the adversary can always stop
us by knocking out our smallest quorum. A probabilistic quorum sys-
tem or more specifically an -intersecting quorum system [MRWW01]
improves the fault-tolerance by relaxing the requirements. For such a sys-
tem we have not only a set system Q, but also a probability distribu-
CHAPTER 14. QUORUM SYSTEMS 109

tion w supplied by the quorum system designer, with the property that
Pr[Q1 ∩ Q2 = ∅] ≤  when Q1 and Q2 are chosen independently according
to their weights.

14.6.1 Example
√
Let a quorum be any set of size k n for some k and let all quorums be
chosen uniformly at random. Pick some quorum Q1 ; what is the probability
that a random Q2 does not intersect Q1 ? Imagine we choose the elements
of Q2 one at a time. The chance that the first element x1 of Q2 misses Q1
√ √
is exactly (n − k n)/n = 1 − k/ n, and conditioning on x1 through xi−1
√
missing Q1 the probability that xi also misses it is (n − k n − i + 1)/(n −
√ √
i + 1) ≤ (n − k n)/n = 1 − k/√ n. So taking the√product over all i gives
√ k n √
Pr[all miss Q1 ] ≤ (1 − k/ n) ≤ exp(−k n)k/ n) = exp(−k 2 ). So by
setting k = Θ(ln 1/), we can get our desired -intersecting system.

14.6.2 Performance
Failure probabilities, if naively defined, can be made arbitrarily small: add
low-probability singleton quorums that are hardly ever picked unless massive
failures occur. But the resulting system is still -intersecting.
One way to look at this is that it points out a flaw in the -intersecting
definition: -intersecting quorums may cease to be -intersecting conditioned
on a particular failure pattern (e.g. when all the non-singleton quorums are
knocked out by massive failures). But Malkhi et al. [MRWW01] address the
problem in a different way, by considering only survival of high quality
quorums, where a particular quorum Q is δ-high-quality if Pr[Q1 ∩ Q2 =
√
∅|Q1 = Q] ≤ δ and high quality if it’s -high-quality. It’s not hard to show
that a random quorum is δ-high-quality with probability at least /δ, so
a high quality quorum is one that fails to intersect a random quorum with
√
probability at most  and a high quality quorum is picked with probability
√
at least 1 − .
We can also consider load; Malkhi et al. [MRWW01] show that essen-
tially the same bounds on load for strict quorum systems also hold for -
√
intersecting quorum systems: load(S) ≥ max((E(|Q|)/n, (1 − )2 / E(|Q|)),
where E(|Q|) is the expected size of a quorum. The left-hand branch of the
max is just the average load applied to a uniformly-chosen server. For the
right-hand side, pick some high quality quorum Q0 with size less than or
√
equal to (1 − ) E(|Q|) and consider the load applied to its most loaded
member by its nonempty intersection (which occurs with probability at least
CHAPTER 14. QUORUM SYSTEMS 110

√
1− ) with a random quorum.

14.7 Signed quorum systems

A further generalization of probabilistic quorum systems gives signed quo-
rum systems [Yu06]. In these systems, a quorum consists of some set of
positive members (servers you reached) and negative members (servers you
tried to reach but couldn’t). These allow O(1)-sized quorums while tol-
erating n − O(1) failures, under certain natural probabilistic assumptions.
Because the quorums are small, the load on some servers may be very high:
so these are most useful for fault-tolerance rather than load-balancing. See
the paper for more details.
 Part II

Shared memory

111
Chapter 15

Model

Basic shared-memory model. See also [AW04, §4.1].

The idea of shared memory is that instead of sending messages to each
other, processes communicate through a pool of shared objects. These
are typically registers supporting read and write operations, but fancier
objects corresponding to more sophisticated data structures or synchroniza-
tion primitives may also be included in the model.
It is usually assumed that the shared objects do not experience faults.
This means that the shared memory can be used as a tool to prevent par-
titions and other problems that can arise in message passing if the number
of faults get too high. As a result, for large numbers of processor failures,
shared memory is a more powerful model than message passing, although
we will see in Chapter 16 that both models can simulate each other provided
a majority of processes are non-faulty.

15.1 Atomic registers

An atomic register supports read and write operations; we think of these
as happening instantaneously, and think of operations of different processes
as interleaved in some sequence. Each read operation on a particular reg-
ister returns the value written by the last previous write operation. Write
operations return nothing.
A process is defined by giving, for each state, the operation that it would
like to do next, together with a transition function that specifies how the
state will be updated in response to the return value of that operation. A
configuration of the system consists of a vector of states for the processes
and a vector of value for the registers. A sequential execution consists of a

112
CHAPTER 15. MODEL 113

sequence of alternating configurations and operations C0 , π1 , C1 , π2 , C2 . . . ,

where in each triple Ci , πi+1 , Ci+1 , the configuration Ci+1 is the result of
applying πi+1 to configuration Ci . For read operations, this means that the
state of the reading process is updated according to its transition function.
For write operations, the state of the writing process is updated, and the
state of the written register is also updated.
Pseudocode for shared-memory protocols is usually written using stan-
dard pseudocode conventions, with the register operations appearing either
as explicit subroutine calls or implicitly as references to shared variables.
Sometimes this can lead to ambiguity; for example, in the code fragment

done ← leftDone ∧ rightDone,

it is clear that the operation write(done, −) happens after read(leftDone)

and read(rightDone), but it is not clear which of read(leftDone and read(rightDone)
happens first. When the order is important, we’ll write the sequence out
explicitly:

1 leftIsDone ← read(leftDone)
2 rightIsDone ← read(rightDone)
3 write(done, leftIsDone ∧ rightIsDone)

Here leftIsDone and rightIsDone are internal variables of the process, so

using them does not require read or write operations to the shared memory.

15.2 Single-writer versus multi-writer registers

One variation that does come up even with atomic registers is what processes
are allowed to read or write a particular register. A typical assumption is
that registers are single-writer multi-reader—there is only one process
that can write to the register (which simplifies implementation since we don’t
have to arbitrate which of two near-simultaneous writes gets in last and thus
leaves the long-term value), although it’s also common to assume multi-
writer multi-reader registers, which if not otherwise available can be built
from single-writer multi-reader registers using atomic snapshot (see Chap-
ter 19). Less common are single-reader single-writer registers, which act
much like message-passing channels except that the receiver has to make an
explicit effort to pick up its mail.
CHAPTER 15. MODEL 114

15.3 Fairness and crashes

From the perspective of a schedule, the fairness condition says that every
processes gets to perform an operation infinitely often, unless it enters either
a crashed or halting state where it invokes no further operations. (Note
that unlike in asynchronous message-passing, there is no way to wake up a
process once it stops doing operations, since the only way to detect that any
activity is happening is to read a register and notice it changed.) Because
the registers (at least in in multi-reader models) provide a permanent fault-
free record of past history, shared-memory systems are much less vulnerable
to crash failures than message-passing systems (though FLP1 still applies);
so in extreme cases, we may assume as many as n − 1 crash failures, which
makes the fairness condition very weak. The n − 1 crash failures case is
called the wait-free case—since no process can wait for any other process
to do anything—and has been extensively studied in the literature.
For historical reasons, work on shared-memory systems has tended to as-
sume crash failures rather than Byzantine failures—possibly because Byzan-
tine failures are easier to prevent when you have several processes sitting in
the same machine than when they are spread across the network, or possibly
because in multi-writer situations a Byzantine process can do much more
damage. But the model by itself doesn’t put any constraints on the kinds
of process failures that might occur.

15.4 Concurrent executions

Often, the operations on our shared objects will be implemented using lower-
level operations. When this happens, it no longer makes sense to assume that
the high-level operations occur one at a time—although an implementation
may try to give that impression to its users. To model to possibility of
concurrency between operations, we split an operation into an invocation
and response, corresponding roughly to a procedure call and its return.
The user is responsible for invoking the object; the object’s implementation
(or the shared memory system, if the object is taken as a primitive) is
responsible for responding. Typically we will imagine that an operation is
invoked at the moment it becomes pending, but there may be executions
in which that does not occur. The time between the invocation and the
response for an operation is the interval of the operation.
A concurrent execution is a sequence of invocations and responses,
1
See Chapter 11.
CHAPTER 15. MODEL 115

where after any prefix of the execution, every response corresponds to some
preceding invocation, and there is at most one invocation for each pro-
cess—always the last—that does not have a corresponding response. How
a concurrent execution may or may not relate to a sequential execution
depends on the consistency properties of the implementation, as described
below.

15.5 Consistency properties

Different shared-memory systems may provide various consistency prop-
erties, which describe how views of an object by different processes mesh
with each other. The strongest consistency property generally used is lin-
earizability [HW90], where an implementation of an object is linearizable
if, for any concurrent execution of the object, there is a sequential execu-
tion of the object with the same operations and return values, where the
(total) order of operations in the sequential execution is a linearization of
the (partial) order of operations in the concurrent execution.
Less formally, this means that if operation a finishes before operation
b starts in the concurrent execution, then a must come before b in the
sequential execution. An equivalent definition is that we can assign each
operation a linearization point somewhere between when its invocation
and response, and the sequential execution obtained by assuming that all
operations occur atomically at their linearization points is consistent with
the specification of the object. Using either definition, we are given a fair bit
of flexibility in how to order overlapping operations, which can sometimes
be exploited by clever implementations (or lower bounds).
A weaker condition is sequential consistency [Lam79]. This says that
for any concurrent execution of the object, there exists some sequential
execution that is indistinguishable to all processes; however, this sequential
execution might include operations that occur out of order from a global
perspective. For example, we could have an execution of an atomic register
where you write to it, then I read from it, but I get the initial value that
precedes your write. This is sequentially consistent but not linearizable.
Mostly we will ask any implementations we consider to be linearizable.
However, both linearizability and sequential consistency are much stronger
than the consistency conditions provided by real multiprocessors. For some
examples of weaker memory consistency rules, a good place to start might
be the dissertation of Jalal Y. Kawash [Kaw00].
CHAPTER 15. MODEL 116

15.6 Complexity measures

There are several complexity measures for shared-memory systems.
Time Assume that no process takes more than 1 time unit between opera-
tions (but some fast processes may take less). Assign the first operation
in the schedule time 1 and each subsequent operation the largest time
consistent with the bound. The time of the last operation is the time
complexity. This is also known as the big-step or round measure
because the time increases by 1 precisely when every non-faulty pro-
cess has taken at least one step, and a minimum interval during which
this occurs counts as a big step or a round.
Total work The total work or total step complexity is just the length
of the schedule, i.e. the number of operations. This doesn’t consider
how the work is divided among the processes, e.g. an O(n2 ) total work
protocol might dump all O(n2 ) operations on a single process and leave
the rest with almost nothing to do. There is usually not much of a
direct correspondence between total work and time. For example, any
algorithm that involves busy-waiting—where a process repeatedly
reads a register until it changes—may have unbounded total work
(because the busy-waiter might spin very fast) even though it runs in
bounded time (because the register gets written to as soon as some
slower process gets around to it). However, it is trivially the case that
the time complexity is never greater than the total work.
Per-process work The per-process work, individual work, per-process
step complexity, or individual step complexity measures the
maximum number of operations performed by any single process. Op-
timizing for per-process work produces more equitably distributed
workloads (or reveals inequitably distributed workloads). Like total
work, per-process work gives an upper bound on time, since each time
unit includes at least one operation from the longest-running process,
but time complexity might be much less than per-process work (e.g.
in the busy-waiting case above).
Remote memory references As we’ve seen, step complexity doesn’t make
much sense for processes that busy-wait. An alternative measure is
remote memory reference complexity or RMR complexity. This
measure charges one unit for write operations and the first read oper-
ation by each process following a write, but charges nothing for sub-
sequent read operations if there are no intervening writes (see §17.5
CHAPTER 15. MODEL 117

for details). In this measure, a busy-waiting operation is only charged

one unit. RMR complexity can be justified to a certain extent by the
cost structure of multi-processor caching [MCS91, And90].
Contention In multi-writer or multi-reader situations, it may be bad to
have too many processes pounding on the same register at once. The
contention measures the maximum number of pending operations on
any single register during the schedule (this is the simplest of several
definitions out there). A single-reader single-writer algorithm always
has contention at most 2, but achieving such low contention may be
harder for multi-reader multi-writer algorithms. Of course, the con-
tention is never worse that n, since we assume each process has at
most one pending operation at a time.
Space Just how big are those registers anyway? Much of the work in this
area assumes they are very big. But we can ask for the maximum
number of bits in any one register (width) or the total size (bit com-
plexity) or number (space complexity) of all registers, and will try
to minimize these quantities when possible. We can also look at the
size of the internal states of the processes for another measure of space
complexity.

15.7 Fancier registers

In addition to stock read-write registers, one can also imagine more tricked-
out registers that provide additional operations. These usually go by the
name of read-modify-write (RMW) registers, since the additional oper-
ations consist of reading the state, applying some function to it, and writing
the state back, all as a single atomic action. Examples of RMW registers
that have appeared in real machines at various times in the past include:
Test-and-set bits A test-and-set operation sets the bit to 1 and returns
the old value.
Fetch-and-add registers A fetch-and-add operation adds some incre-
ment (typically -1 or 1) to the register and returns the old value.
Compare-and-swap registers A compare-and-swap operation writes
a new value only if the previous value is equal to a supplied test value.
These are all designed to solve various forms of mutual exclusion or
locking, where we want at most one process at a time to work on some
shared data structure.
CHAPTER 15. MODEL 118

Some more exotic read-modify-write registers that have appeared in the

literature are

Fetch-and-cons Here the contents of the register is a linked list; a fetch-

and-cons adds a new head and returns the old list.

Sticky bits (or sticky registers) With a sticky bit or sticky regis-
ter [Plo89], once the initial empty value is overwritten, all further
writes fail. The writer is not notified that the write fails, but may
be able to detect this fact by reading the register in a subsequent
operation.

Bank accounts Replace the write operation with deposit, which adds a
non-negative amount to the state, and withdraw, which subtracts a
non-negative amount from the state provided the result would not go
below 0; otherwise, it has no effect.

These solve problems that are hard for ordinary read/write registers un-
der bad conditions. Note that they all have to return something in response
to an invocation.
There are also blocking objects like locks or semaphores, but these don’t
fit into the RMW framework.
We can also consider generic read-modify-write registers that can com-
pute arbitrary functions (passed as an argument to the read-modify-write
operation) in the modify step. Here we typically assume that the read-
modify-write operation returns the old value of the register. Generic read-
modify-write registers are not commonly found in hardware but can be easily
simulated (in the absence of failures) using mutual exclusion.2

2
See Chapter 17.
Chapter 16

Distributed shared memory

In distributed shared memory, our goal is to simulate a collection of

memory locations or registers, each of which supports a read operation
that returns the current state of the register and a write operation that
updates the state. Our implementation should be linearizable [HW90],
meaning that read and write operations appear to occur instantaneously
(atomically) at some point in between when the operation starts and the
operation finishes; equivalently, there should be some way to order all the
operations on the registers to obtain a sequential execution consistent
with the behavior of a real register (each read returns the value of the most
recent write) while preserving the observable partial order on operations
(where π1 precedes π2 if π1 finishes before π2 starts). Implicit in this defi-
nition is the assumption that implemented operations take place over some
interval, between an invocation that starts the operation and a response
that ends the operation and returns its value.1
In the absence of process failures, we can just assign each register to some
process, and implement both read and write operations by remote procedure
calls to the process (in fact, this works for arbitrary shared-memory ob-
jects). With process failures, we need to make enough copies of the register
that failures can’t destroy all of them. This creates an asymmetry between
simulations of message-passing from shared-memory and vice versa; in the
former case (discussed briefly in §16.1 below), a process that fails in the
underlying shared-memory system only means that the same process fails in
the simulated message-passing system. But in the other direction, not only
does the failure of a process in the underlying message-passing system mean
that the same process fails in the simulated shared-memory system, but the
1
More details on the shared-memory model are given in Chapter 15.

119
CHAPTER 16. DISTRIBUTED SHARED MEMORY 120

simulation collapses completely if a majority of processes fail.

16.1 Message passing from shared memory

We’ll start with the easy direction. We can build a reliable FIFO channel
from single-writer single-reader registers using polling. The naive approach
is that for each edge uv in the message-passing system, we create a (very big)
register ruv , and u writes the entire sequence of every message it has ever
sent to v to ruv every time it wants to do a new send. To receive messages,
v polls all of its incoming registers periodically and delivers any messages in
the histories that it hasn’t processed yet.2
The ludicrous register width can be reduced by adding in an acknowl-
edgment mechanism in a separate register ackvu ; the idea is that u will only
write one message at a time to ruv , and will queue subsequent messages un-
til v writes in ackvu that the message in ruv has been received. With some
tinkering, it is possible to knock ruv down to only three possible states (send-
ing 0, sending 1, and reset) and ackvu down to a single bit (value-received,
reset-received), but that’s probably overkill for most applications.
Process failures don’t affect any of these protocols, except that a dead
process stops sending and receiving.

16.2 The Attiya-Bar-Noy-Dolev algorithm

Here we show how to implement shared memory from message passing. We’ll
assume that our system is asynchronous, that the network is complete, and
that we are only dealing with f < n/2 crash failures. We’ll also assume we
only want to build single-writer registers, just to keep things simple; we can
extend to multi-writer registers later.
Here’s the algorithm, which is due to Attiya, Bar-Noy, and Dolev [ABND95];
see also [Lyn96, §17.1.3]. (Section 9.3 of [AW04] gives an equivalent algo-
rithm, but the details are buried in an implementation of totally-ordered
broadcast). We’ll make n copies of the register, one on each process. Each
process’s copy will hold a pair (value, timestamp) where timestamps are (un-
bounded) integer values. Initially, everybody starts with (⊥, 0). A process
updates its copy with new values (v, t) upon receiving write(v, t) from any
other process p, provided t is greater than the process’s current timestamp.
2
If we are really cheap about using registers, and are willing to accept even more
absurdity in the register size, we can just have u write every message it ever sends to ru ,
and have each v poll all the ru and filter out any messages intended for other processes.
CHAPTER 16. DISTRIBUTED SHARED MEMORY 121

It then responds to p with ack(v, t), whether or not it updated its local
copy. A process will also respond to a message read(u) with a response
ack(value, timestamp, u); here u is a nonce3 used to distinguish between dif-
ferent read operations so that a process can’t be confused by out-of-date
acknowledgments.
To write a value, the writer increments its timestamp, updates its value
and sends write(value, timestamp) to all other processes. The write opera-
tion terminates when the writer has received acknowledgments containing
the new timestamp value from a majority of processes.
To read a value, a reader does two steps:

1. It sends read(u) to all processes (where u is any value it hasn’t used

before) and waits to receive acknowledgments from a majority of the
processes. It takes the value v associated with the maximum time-
stamp t as its return value (no matter how many processes sent it).

2. It then sends write(v, t) to all processes, and waits for a response

ack(v, t) from a majority of the processes. Only then does it return.

(Any extra messages, messages with the wrong nonce, etc. are dis-
carded.)
Both reads and writes cost Θ(n) messages (Θ(1) per process).
Intuition: Nobody can return from a write or a read until they are sure
that subsequent reads will return the same (or a later) value. A process
can only be sure of this if it knows that the values collected by a read will
include at least one copy of the value written or read. But since majorities
overlap, if a majority of the processes have a current copy of v, then the
majority read quorum will include it. Sending write(v, t) to all processes
and waiting for acknowledgments from a majority is just a way of ensuring
that a majority do in fact have timestamps that are at least t.
If we omit the write stage of a read operation, we may violate lineariz-
ability. An example would be a situation where two values (1 and 2, say),
have been written to exactly one process each, with the rest still holding
the initial value ⊥. A reader that observes 1 and (n − 1)/2 copies of ⊥
will return 1, while a reader that observes 2 and (n − 1)/2 copies of ⊥ will
return 2. In the absence of the write stage, we could have an arbitrarily
long sequence of readers return 1, 2, 1, 2, . . . , all with no concurrency. This
3
A nonce is any value that is guaranteed to be used at most once (the term originally
comes from cryptography, which in turn got it from linguistics). In practice, a reader will
most likely generate a nonce by combining its process id with a local timestamp.
CHAPTER 16. DISTRIBUTED SHARED MEMORY 122

would not be consistent with any sequential execution in which 1 and 2 are
only written once.

16.3 Proof of linearizability

Our intuition may be strong, but we still need a proof the algorithm works.
In particular, we want to show that for any trace T of the ABD protocol,
there is an trace of an atomic register object that gives the same sequence
of invoke and response events. The usual way to do this is to find a lin-
earization of the read and write operations: a total order that extends the
observed order in T where π1 < π2 in T if and only if π1 ends before π2
starts. Sometimes it’s hard to construct such an order, but in this case it’s
easy: we can just use the timestamps associated with the values written or
read in each operation. Specifically, we define the timestamp of a write or
read operation as the timestamp used in the write(v, t) messages sent out
during the implementation of that operation, and we put π1 before π2 if:

1. π1 has a lower timestamp than π2 , or

2. π1 has the same timestamp as π2 , π1 is a write, and π2 is a read, or

3. π1 has the same timestamp as π2 and π1 <T π2 , or

4. none of the other cases applies, and we feel like putting π1 first.

The intent is that we pick some total ordering that is consistent with both
<T and the timestamp ordering (with writes before reads when timestamps
are equal). To make this work we have to show (a) that these two orderings
are in fact consistent, and (b) that the resulting ordering produces values
consistent with an atomic register: in particular, that each read returns the
value of the last preceding write.
Part (b) is easy: since timestamps only increase in response to writes,
each write is followed by precisely those reads with the same timestamp,
which are precisely those that returned the value written.
For part (a), suppose that π1 <T π2 . The first case is when π2 is a read.
Then before the end of π1 , a set S of more than n/2 processes send the π1
process an ack(v1, t1 ) message. Since local timestamps only increase, from
this point on any ack(v2 , t2 , u) message sent by a process in S has t2 ≥ t1 .
Let S 0 be the set of processes sending ack(v2 , t2 , u) messages processed by
π2 . Since |S| > n/2 and |S 0 | > n/2, we have S ∩ S 0 is nonempty and so S 0
includes a process that sent ack(v2 , t2 ) with t2 ≥ t1 . So π2 is serialized after
CHAPTER 16. DISTRIBUTED SHARED MEMORY 123

π1 . The second case is when π2 is a write; but then π1 returns a timestamp

that precedes the writer’s increment in π2 , and so again is serialized first.

16.4 Proof that f < n/2 is necessary

This is pretty much the standard partition argument that f < n/2 is neces-
sary to do anything useful in a message-passing system. Split the processes
into two sets S and S 0 of size n/2 each. Suppose the writer is in S. Consider
an execution where the writer does a write operation, but all messages be-
tween S and S 0 are delayed. Since the writer can’t tell if the S 0 processes are
slow or dead, it eventually returns. Now let some reader in S 0 attempt to
read the simulated register, again delaying all messages between S and S 0 ;
now the reader is forced to return some value without knowing whether the
S processes are slow or dead. If the reader doesn’t return the value written,
we lose. If by some miracle it does, then we lose in the execution where the
write didn’t happen and all the processes in S really were dead.

16.5 Multiple writers

So far we have assumed a single writer. The main advantage of this approach
is that we don’t have to do much to manage timestamps: the single writer
can just keep track of its own. With multiple writers we can use essentially
the same algorithm, but each write needs to perform an initial round of
gathering timestamps so that it can pick a new timestamp bigger than those
that have come before. We also extend the timestamps to be of the form
hcount, idi, lexicographically ordered, so that two timestamps with the same
count field are ordered by process id. The modified write algorithm is:
1. Send read(u) to all processes and wait to receive acknowledgments
from a majority of the processes.

2. Set my timestamp to t = (maxq countq + 1, id) where the max is taken

over all processes q that sent me an acknowledgment. Note that this
is a two-field timestamp that is compared lexicographically, with the
id field used only to prevent duplicate timestamps.

3. Send write(v, t) to all processes, and wait for a response ack(v, t) from
a majority of processes.
This increases the cost of a write by a constant factor, but in the end
we still have only a linear number of messages. The proof of linearizability
CHAPTER 16. DISTRIBUTED SHARED MEMORY 124

is essentially the same as for the single-writer algorithm, except now we

must consider the case of two write operations by different processes. Here
we have that if π1 <T π2 , then π1 gets acknowledgments of its write with
timestamp t1 from a majority of processes before π2 starts its initial phase
to compute count. Since π2 waits for acknowledgments from a majority of
processes as well, these majorities overlap, so π2 ’s timestamp t2 must exceed
t1. So the linearization ordering previously defined still works.

16.6 Other operations

The basic ABD framework can be extended to support other operations.
One such operation is a collect [SSW91], where we read n registers in
parallel with no guarantee that they are read at the same time. This can
trivially be implemented by running n copies of ABD in parallel, and can
be implemented with the same time and message complexity as ABD for a
single register by combining the messages from the parallel executions into
single (possibly very large) messages.
Chapter 17

Mutual exclusion

For full details see [AW04, Chapter 4] or [Lyn96, Chapter 10].

17.1 The problem

The goal is to share some critical resource between processes without more
than one using it at a time—this is the fundamental problem in time-sharing
systems.
The solution is to only allow access while in a specially-marked block of
code called a critical section, and only allow one process at a time to be
in a critical section.
A mutual exclusion protocol guarantees this, usually in an asyn-
chronous shared-memory model.
Formally: We want a process to cycle between states trying (trying
to get into critical section), critical (in critical section), exiting (cleaning
up so that other processes can enter their critical sections), and remain-
der (everything else—essentially just going about its non-critical business).
Only in the trying and exiting states does the process run the mutual ex-
clusion protocol to decide when to switch to the next state; in the critical
or remainder states it switches to the next state on its own.

17.2 Goals
(See also [AW04, §4.2], [Lyn96, §10.2].)
Core mutual exclusion requirements:

Mutual exclusion At most one process is in the critical state at a time.

125
CHAPTER 17. MUTUAL EXCLUSION 126

No deadlock (progress) If there is at least one process in a trying state,

then eventually some process enters a critical state; similarly for exit-
ing and remainder states.

Note that the protocol is not required to guarantee that processes leave
the critical or remainder state, but we generally have to insist that the
processes at least leave the critical state on their own to make progress.
Additional useful properties (not satisfied by all mutual exclusion pro-
tocols; see [Lyn96, §10.4)]:

No lockout (lockout-freedom): If there is a particular process in a trying

or exiting state, that process eventually leaves that state. This means
that I don’t starve because somebody else keeps jumping past me and
seizing the critical resource before I can.

Stronger versions of lockout-freedom include explicit time bounds (how

many rounds can go by before I get in) or bounded bypass (nobody gets
in more than k times before I do).

17.3 Mutual exclusion using strong primitives

See [AW04, §4.3] or [Lyn96, 10.9]. The idea is that we will use some sort of
read-modify-write register, where the RMW operation computes a new
value based on the old value of the register and writes it back as a single
atomic operation, usually returning the old value to the caller as well.

17.3.1 Test and set

A test-and-set operation does the following sequence of actions atomically:

1 oldValue ← read(bit)
2 write(bit, 1)
3 return oldValue

Typically there is also a second reset operation for setting the bit back
to zero. For some implementations, this reset operation may only be used
safely by the last process to get 0 from the test-and-set bit.
Because a test-and-set operation is atomic, if two processes both try to
perform test-and-set on the same bit, only one of them will see a return value
CHAPTER 17. MUTUAL EXCLUSION 127

of 0. This is not true if each process simply executes the above code on a
stock atomic register: there is an execution in which both processes read
0, then both write 1, then both return 0 to whatever called the non-atomic
test-and-set subroutine.
Test-and-set provides a trivial implementation of mutual exclusion, shown
in Algorithm 17.1.

1 while true do
// trying
2 while testAndSet(lock) = 1 do nothing
// critical
3 (do critical section stuff)
// exiting
4 reset(lock)
// remainder
5 (do remainder stuff)
Algorithm 17.1: Mutual exclusion using test-and-set

It is easy to see that this code provides mutual exclusion, as once one
process gets a 0 out of lock, no other can escape the inner while loop until
that process calls the reset operation in its exiting state. It also provides
progress (assuming the lock is initially set to 0); the only part of the code
that is not straight-line code (which gets executed eventually by the fairness
condition) is the inner loop, and if lock is 0, some process escapes it, while
if lock is 1, some process is in the region between the testAndSet call and
the reset call, and so it eventually gets to reset and lets the next process
in (or itself, if it is very fast).
The algorithm does not provide lockout-freedom: nothing prevents a
single fast process from scooping up the lock bit every time it goes through
the outer loop, while the other processes ineffectually grab at it just after it
is taken away. Lockout-freedom requires a more sophisticated turn-taking
strategy.

17.3.2 A lockout-free algorithm using an atomic queue

Basic idea: In the trying phase, each process enqueues itself on the end of a
shared queue (assumed to be an atomic operation). When a process comes
to the head of the queue, it enters the critical section, and when exiting it
dequeues itself. So the code would look something like Algorithm 17.2.
CHAPTER 17. MUTUAL EXCLUSION 128

Note that this requires a queue that supports a head operation. Not all
implementations of queues have this property.

1 while true do
// trying
2 enq(Q, myId)
3 while head(Q) 6= myId do nothing
// critical
4 (do critical section stuff)
// exiting
5 deq(Q)
// remainder
6 (do remainder stuff)
Algorithm 17.2: Mutual exclusion using a queue

Here the proof of mutual exclusion is that only the process whose id is at
the head of the queue can enter its critical section. Formally, we maintain an
invariant that any process whose program counter is between the inner while
loop and the call to deq(Q) must be at the head of the queue; this invariant
is easy to show because a process can’t leave the while loop unless the test
fails (i.e., it is already at the head of the queue), no enq operation changes
the head value (if the queue is nonempty), and the deq operation (which
does change the head value) can only be executed by a process already at
the head (from the invariant).
Deadlock-freedom follows from proving a similar invariant that every
element of the queue is the id of some process in the trying, critical, or
exiting states, so eventually the process at the head of the queue passes the
inner loop, executes its critical section, and dequeues its id.
Lockout-freedom follows from the fact that once a process is at position
k in the queue, every execution of a critical section reduces its position by 1;
when it reaches the front of the queue (after some finite number of critical
sections), it gets the critical section itself.

17.3.2.1 Reducing space complexity

Following [AW04, §4.3.2], we can give an implementation of this algorithm
using a single read-modify-write (RMW) register instead of a queue; this
drastically reduces the (shared) space needed by the algorithm. The reason
this works is because we don’t really need to keep track of the position of
CHAPTER 17. MUTUAL EXCLUSION 129

each process in the queue itself; instead, we can hand out numerical tickets
to each process and have the process take responsibility for remembering
where its place in line is.
The RMW register has two fields, first and last, both initially 0. In-
crementing last simulates an enqueue, while incrementing first simulates a
dequeue. The trick is that instead of testing if it is at the head of the queue,
a process simply remembers the value of the last field when it “enqueued”
itself, and waits for the first field to equal it.
Algorithm 17.3 shows the code from Algorithm 17.2 rewritten to use this
technique. The way to read the RMW operations is that the first argument
specifies the variable to update and the second specifies an expression for
computing the new value. Each RMW operation returns the old state of the
object, before the update.

1 while true do
// trying
2 position ← RMW(V, hV.first, V.last + 1i)
// enqueue
3 while RMW(V, V ).first 6= position.last do
4 nothing
// critical
5 (do critical section stuff)
// exiting
6 RMW(V, hV.first + 1, V.lasti)
// dequeue
// remainder
7 (do remainder stuff)
Algorithm 17.3: Mutual exclusion using read-modify-write

17.4 Mutual exclusion using only atomic registers

While mutual exclusion is easier using powerful primitives, we can also solve
the problem using only registers.

17.4.1 Peterson’s tournament algorithm

Algorithm 17.4 shows Peterson’s lockout-free mutual exclusion protocol for
two processes p0 and p1 [Pet81] (see also [AW04, §4.4.2] or [Lyn96, §10.5.1]).
CHAPTER 17. MUTUAL EXCLUSION 130

It uses only atomic registers.

shared data:
1 waiting, initially arbitrary
2 present[i] for i ∈ {0, 1}, initially 0
3 Code for process i:
4 while true do
// trying
5 present[i] ← 1
6 waiting ← i
7 while true do
8 if present[¬i] = 0 then break
9
10 if waiting 6= i then break
11

// critical
12 (do critical section stuff)
// exiting
13 present[i] = 0
// remainder
14 (do remainder stuff)
Algorithm 17.4: Peterson’s mutual exclusion algorithm for two pro-
cesses

This uses three bits to communicate: present[0] and present[1] indicate

which of p0 and p1 are participating, and waiting enforces turn-taking. The
protocol requires that waiting be multi-writer, but it’s OK for present[0] and
present[1] to be single-writer.
In the description of the protocol, we write Lines 8 and 10 as two separate
lines because they include two separate read operations, and the order of
these reads is important.

17.4.1.1 Correctness of Peterson’s protocol

Intuitively, let’s consider all the different ways that the entry code of the two
processes could interact. There are basically two things that each process
does: it sets its own present in Line 5 and grabs the waiting variable in
Line 6. Here’s a typical case where one process gets in first:

1. p0 sets present[0] ← 1
CHAPTER 17. MUTUAL EXCLUSION 131

2. p0 sets waiting ← 0

3. p0 reads present[1] = 0 and enters critical section

4. p1 sets present[1] ← 1

5. p1 sets waiting ← 1

6. p1 reads present[0] = 1 and waiting = 1 and loops

7. p0 sets present[0] ← 0

8. p1 reads present[0] = 0 and enters critical section

The idea is that if I see a 0 in your present variable, I know that you
aren’t playing, and can just go in.
Here’s a more interleaved execution where the waiting variable decides
the winner:

1. p0 sets present[0] ← 1

2. p0 sets waiting ← 0

3. p1 sets present[1] ← 1

4. p1 sets waiting ← 1

5. p0 reads present[1] = 1

6. p1 reads present[0] = 1

7. p0 reads waiting = 1 and enters critical section

8. p1 reads present[0] = 1 and waiting = 1 and loops

9. p0 sets present[0] ← 0

10. p1 reads present[0] = 0 and enters critical section

Note that it’s the process that set the waiting variable last (and thus
sees its own value) that stalls. This is necessary because the earlier process
might long since have entered the critical section.
Sadly, examples are not proofs, so to show that this works in general, we
need to formally verify each of mutual exclusion and lockout-freedom. Mu-
tual exclusion is a safety property, so we expect to prove it using invariants.
The proof in [Lyn96] is based on translating the pseudocode directly into
CHAPTER 17. MUTUAL EXCLUSION 132

automata (including explicit program counter variables); we’ll do essentially

the same proof but without doing the full translation to automata. Below,
we write that pi is at line k if it the operation in line k is enabled but has
not occurred yet.

Lemma 17.4.1. If present[i] = 0, then pi is at Line 5 or 14.

Proof. Immediate from the code.

Lemma 17.4.2. If pi is at Line 12, and p¬i is at Line 8, 10, or 12, then
waiting = ¬i.

Proof. We’ll do the case i = 0; the other case is symmetric. The proof is by
induction on the schedule. We need to check that any event that makes the
left-hand side of the invariant true or the right-hand side false also makes
the whole invariant true. The relevent events are:

• Transitions by p0 from Line 8 to Line 12. These occur only if present[1] =

0, implying p1 is at Line Line 5 or 14 by Lemma 17.4.1. In this case
the second part of the left-hand side is false.

• Transitions by p0 from Line 10 to Line 12. These occur only if waiting 6=

0, so the right-hand side is true.

• Transitions by p1 from Line 6 to Line 8. These set waiting to 1, making

the right-hand side true.

• Transitions that set waiting to 0. These are transitions by p0 from

Line 6 to Line 10, making the left-hand side false.

We can now read mutual exclusion directly off of Lemma 17.4.2: if both
p0 and p1 are at Line 12, then we get waiting = 1 and waiting = 0, a
contradiction.
To show progress, observe that the only place where both processes can
get stuck forever is in the loop at Lines 8 and 10. But then waiting isn’t
changing, and so some process i reads waiting = ¬i and leaves. To show
lockout-freedom, observe that if p0 is stuck in the loop while p1 enters the
critical section, then after p1 leaves it sets present[1] to 0 in Line 13 (which
lets p0 in if p0 reads present[1] in time), but even if it then sets present[1]
back to 1 in Line 5, it still sets waiting to 1 in Line 6, which lets p0 into
the critical section. With some more tinkering this argument shows that p1
CHAPTER 17. MUTUAL EXCLUSION 133

enters the critical section at most twice while p0 is in the trying state, giving
2-bounded bypass; see [Lyn96, Lemma 10.12]. With even more tinkering we
get a constant time bound on the waiting time for process i to enter the
critical section, assuming the other process never spends more than O(1)
time inside the critical section.

17.4.1.2 Generalization to n processes

(See also [AW04, §4.4.3].)
The easiest way to generalize Peterson’s two-process algorithm to n pro-
cesses is to organize a tournament in the form of log-depth binary tree; this
method was invented by Peterson and Fischer [PF77]. At each node of the
tree, the roles of the two processes are taken by the winners of the subtrees,
i.e., the processes who have entered their critical sections in the two-process
algorithms corresponding to the child nodes. The winner of the tournament
as a whole enters the real critical section, and afterwards walks back down
the tree unlocking all the nodes it won in reverse order. It’s easy to see that
this satisfies mutual exclusion, and not much harder to show that it satisfies
lockout-freedom—in the latter case, the essential idea is that if the winner
at some node reaches the root infinitely often then lockout-freedom at that
node means that the winner of each child node reaches the root infinitely
often.
The most natural way to implement the nodes is to have present[0] and
present[1] at each node be multi-writer variables that can be written to by
any process in the appropriate subtree. Because the present variables don’t
do much, we can also implement them as the OR of many single-writer
variables (this is what is done in [Lyn96, §10.5.3]), but there is no immediate
payoff to doing this since the waiting variables are still multi-writer.
Nice properties of this algorithm are that it uses only bits and that it’s
very fast: O(log n) time in the absence of contention.

17.4.2 Fast mutual exclusion

With a bit of extra work, we can reduce the no-contention cost of mutual
exclusion to O(1), while keeping whatever performance we previously had
in the high-contention case. The trick (due to Lamport [Lam87]) is to put
an object at the entrance to the protocol that diverts a solo process onto a
“fast path” that lets it bypass the n-process mutex that everybody else ends
up on.
Our presentation mostly follows [AW04][§4.4.5], which uses the splitter
CHAPTER 17. MUTUAL EXCLUSION 134

abstraction of Moir and Anderson [MA95] to separate out the mechanism

for diverting a lone process.1 Code for a splitter is given in Algorithm 17.5.

shared data:
1 atomic register race, big enough to hold an id, initially ⊥
2 atomic register door, big enough to hold a bit, initially open
3 procedure splitter(id)
4 race ← id
5 if door = closed then
6 return right
7 door ← closed
8 if race = id then
9 return stop
10 else
11 return down

Algorithm 17.5: Implementation of a splitter

A splitter assigns to each processes that arrives at it the value right,

down, or stop. The useful properties of splitters are that if at least one
process arrives at a splitter, then (a) at least one process returns right or
stop; and (b) at least one process returns down or stop; (c) at most one
process returns stop; and (d) any process that runs by itself returns stop.
The first two properties will be useful when we consider the problem of
renaming in Chapter 24; we will prove them there. The last two properties
are what we want for mutual exclusion.
The names of the variables race and door follow the presentation in
[AW04, §4.4.5]; Moir and Anderson [MA95], following Lamport [Lam87],
call these X and Y . As in [MA95], we separate out the right and down out-
comes—even though they are equivalent for mutex—because we will need
them later for other applications.
The intuition behind Algorithm 17.5 is that setting door to closed closes
the door to new entrants, and the last entrant to write its id to race wins
(it’s a slow race), assuming nobody else writes race and messes things up.
The added cost of the splitter is always O(1), since there are no loops.
To reset the splitter, write open to door. This allows new processes to
enter the splitter and possibly return stop.
1
Moir and Anderson call these things one-time building blocks, but the name split-
ter has become standard in subsequent work.
CHAPTER 17. MUTUAL EXCLUSION 135

Lemma 17.4.3. After each time that door is set to open, at most one
process running Algorithm 17.5 returns stop.

Proof. To simplify the argument, we assume that each process calls splitter
at most once.
Let t be some time at which door is set to open (−∞ in the case of the
initial value). Let St be the set of processes that read open from door after
time t and before the next time at which some process writes closed to door,
and that later return stop by reaching Line 9.
Then every process in St reads door before any process in St writes door.
It follows that every process in St writes race before any process in St reads
race. If some process p is not the last process in St to write race, it will not
see its own id, and will not return stop. But only one process can be the
last process in St to write race.2

Lemma 17.4.4. If a process runs Algorithm 17.5 by itself starting from a

configuration in which door = open, it returns stop.

Proof. Follows from examining a solo execution: the process sets race to id,
reads open from door, then reads id from race. This causes it to return stop
as claimed.

To turn this into an n-process mutex algorithm, we use the splitter to

separate out at most one process (the one that gets stop) onto a fast path
that bypasses the slow path taken by the rest of the processes. The slow-
path process first fight among themselves to get through an n-process mutex;
the winner then fights in a 2-process mutex with the process (if any) on the
fast path.
Releasing the mutex is the reverse of acquiring it. If I followed the fast
path, I release the 2-process mutex first then reset the splitter. If I followed
the slow path, I release the 2-process mutex first then the n-process mutex.
This gives mutual exclusion with O(1) cost for any process that arrives
before there is any contention (O(1) for the splitter plus O(1) for the 2-
process mutex).
A complication is that if nobody wins the splitter, there is no fast-path
process to reset it. If we don’t want to accept that the fast path just breaks
forever in this case, we have to include a mechanism for a slow-path process
to reset the splitter if it can be assured that there is no fast-path process
2
It’s worth noting that this last process still might not return stop, because some later
process—not in St —might overwrite race. This can happen even if nobody ever resets the
splitter.
CHAPTER 17. MUTUAL EXCLUSION 136

left in the system. The simplest way to do this is to have each process mark
a bit in an array to show it is present, and have each slow-path process,
while still holding all the mutexes, check on its way out if the door bit is
set and no processes claim to be present. If it sees all zeros (except for
itself) after seeing door = closed, it can safely conclude that there is no
fast-path process and reset the splitter itself. The argument then is that the
last slow-path process to leave will do this, re-enabling the fast path once
there is no contention again. This approach is taken implicitly in Lamport’s
original algorithm, which combines the splitter and the mutex algorithms
into a single miraculous blob.

17.4.3 Lamport’s Bakery algorithm

See [AW04, §4.4.1] or [Lyn96, §10.7].
This is a lockout-free mutual exclusion algorithm that uses only single-
writer registers (although some of the registers may end up holding arbitrar-
ily large values). Code for the Bakery algorithm is given as Algorithm 17.6.

shared data:
1 choosing[i], an atomic bit for each i, initially 0
2 number[i], an unbounded atomic register, initially 0
3 Code for process i:
4 while true do
// trying
5 choosing[i] ← 1
6 number[i] ← 1 + maxj6=i number[j]
7 choosing[i] ← 0
8 for j 6= i do
9 loop until choosing[j] = 0
10 loop until number[j] = 0 or hnumber[i], ii < hnumber[j], ji
// critical
11 (do critical section stuff)
// exiting
12 number[i] ← 0
// remainder
13 (do remainder stuff)
Algorithm 17.6: Lamport’s Bakery algorithm

Note that several of these lines are actually loops; this is obvious for
CHAPTER 17. MUTUAL EXCLUSION 137

Lines 9 and 10, but is also true for Line 6, which includes an implicit loop
to read all n − 1 values of number[j].
Intuition for mutual exclusion is that if you have a lower number than I
do, then I block waiting for you; for lockout-freedom, eventually I have the
smallest number. (There are some additional complications involving the
choosing bits that we are sweeping under the rug here.) For a real proof
see [AW04, §4.4.1] or [Lyn96, §10.7].
Selling point is a strong near-FIFO guarantee and the use of only single-
writer registers (which need not even be atomic—it’s enough that they re-
turn correct values when no write is in progress). Weak point is unbounded
registers.

17.4.4 Lower bound on the number of registers

There is a famous result due to Burns and Lynch [BL93] that any mutual
exclusion protocol using only read/write registers requires at least n of them.
Details are in [Lyn96, §10.8]. A slightly different version of the argument
is given in [AW04, 4.4.4]. The proof is another nice example of an indistin-
guishability proof, where we use the fact that if a group of processes can’t
tell the difference between two executions, they behave the same in both.
Assumptions: We have a protocol that guarantees mutual exclusion and
progress. Our base objects are all atomic registers.
Key idea: In order for some process p to enter the critical section, it has
to do at least one write to let the other processes know it is doing so. If
not, they can’t tell if p ever showed up at all, so eventually either some p0
will enter the critical section and violate mutual exclusion or (in the no-p
execution) nobody enters the critical section and we violate progress. Now
suppose we can park a process pi on each register ri with a pending write
to i; in this case we say that pi covers ri . If every register is so covered, we
can let p go ahead and do whatever writes it likes and then deliver all the
covering writes at once, wiping out anything p did. Now the other processes
again don’t know if p exists or not. So we can say something stronger: before
some process p can enter a critical section, it has to write to an uncovered
register.
The hard part is showing that we can cover all the registers without
letting p know that there are other processes waiting—if p can see that
other processes are waiting, it can just sit back and wait for them to go
through the critical section and make progress that way. So our goal is to
produce states in which (a) processes p1 . . . , pk (for some k) between them
cover k registers, and (b) the resulting configuration is indistinguishable
CHAPTER 17. MUTUAL EXCLUSION 138

from an idle configuration to pk+1 . . . pn .

Lemma 17.4.5. Starting from any idle configuration C, there exists an

execution in which only processes p1 . . . pk take steps that leads to a config-
uration C 0 such that (a) C 0 is indistinguishable by any of pk+1 . . . pn from
some idle configuration C 00 and (b) k registers are covered by p1 . . . pk in C 0 .

Proof. The proof is by induction on k. For k = 1, just run p1 until it is

about to do a write, let C 0 be the resulting configuration and let C 00 = C.
For larger k, the essential idea is that starting from C, we first run
to a configuration C1 where p1 . . . pk−1 cover k − 1 registers and C1 is in-
distinguishable from an idle configuration by the remaining processes, and
then run pk until it covers one more register. If we let p1 . . . pk−1 go, they
overwrite anything pk wrote. Unfortunately, they may not come back to
covering the same registers as before if we rerun the induction hypothesis
(and in particular might cover the same register that pk does). So we have
to look for a particular configuration C1 that not only covers k − 1 registers
but also has an extension that covers the same k − 1 registers.
Here’s how we find it: Start in C. Run the induction hypothesis to get
C1 ; here there is a set W1 of k − 1 registers covered in C1 . Now let processes
p1 through pk−1 do their pending writes, then each enter the critical section,
leave it, and finish, and rerun the induction hypothesis to get to a state C2 ,
indistinguishable from an idle configuration by pk and up, in which k − 1
registers in W2 are covered. Repeat to get sets W3 , W4 , etc. Since this
r

sequence is unbounded, and there are only k−1 distinct sets of registers to
cover (where r is the number of registers), eventually we have Wi = Wj for
some i 6= j. The configurations Ci and Cj are now our desired configurations
covering the same k − 1 registers.
Now that we have Ci and Cj , we run until we get to Ci . We now run pk
until it is about to write some register not covered by Ci (it must do so, or
otherwise we can wipe out all of its writes while it’s in the critical section
and then go on to violate mutual exclusion). Then we let the rest of p1
through pk−1 do all their writes (which immediately destroys any evidence
that pk ran at all) and run the execution that gets them to Cj . We now
have k − 1 registers covered by p1 through pk−1 and a k-th register covered
by pk , in a configuration that is indistinguishable from idle: this proves the
induction step.

The final result follows by the fact that when k = n we cover n registers;
this implies that there are n registers to cover.
CHAPTER 17. MUTUAL EXCLUSION 139

17.5 RMR complexity

It’s not hard to see that we can’t build a shared-memory mutex without
busy-waiting: any process that is waiting can’t detect that the critical sec-
tion is safe to enter without reading a register, but if that register tells it
that it should keep waiting, it is back where it started and has to read it
again. This makes our standard step-counting complexity measures useless
for describe the worst-case complexity of a mutual exclusion algorithm.
However, the same argument that suggests we can ignore local compu-
tation in a message-passing model suggests that we can ignore local oper-
ations on registers in a shared-memory model. Real multiprocessors have
memory hierarchies where memory that is close to the CPU (or one of the
CPUs) is generally much faster than memory that is more distant. This
suggests charging only for remote memory references, or RMRs, where
each register is local to one of the processes and only operations on non-
local are expensive. This has the advantage of more accurately modeling
real costs [MCS91, And90], and allowing us to build busy-waiting mutual
exclusion algorithms with costs we can actually analyze.
As usual, there is a bit of a divergence here between theory and practice.
Practically, we are interested in algorithms with good real-time performance,
and RMR complexity becomes a heuristic for choosing how to assign memory
locations. This gives rise to very efficient mutual exclusion algorithms for
real machines, of which the most widely used is the beautiful MCS algorithm
of Mellor-Crummey and Scott [MCS91]. Theoretically, we are interested in
the question of how efficiently we can solve mutual exclusion in our formal
model, and RMR complexity becomes just another complexity measure, one
that happens to allow busy-waiting on local variables.

17.5.1 Cache-coherence vs. distributed shared memory

The basic idea of RMR complexity is that a process doesn’t pay for opera-
tions on local registers. But what determines which operations are local?
In the cache-coherent model (CC for short), once a process reads a
register it retains a local copy as long as nobody updates it. So if I do a se-
quence of read operations with no intervening operations by other processes,
I may pay an RMR for the first one (if my cache is out of date), but the rest
are free. The assumption is that each process can cache registers, and there
is some cache-coherence protocol that guarantees that all the caches stay
up to date. We may or may not pay RMRs for write operations or other
read operations, depending on the details of the cache-coherence protocol,
CHAPTER 17. MUTUAL EXCLUSION 140

but for upper bounds it is safest to assume that we do.

In the distributed shared memory model (DSM), each register is
assigned permanently to a single process. Other processes can read or write
the register, but only the owner gets to do so without paying an RMR. Here
memory locations are nailed down to specific processes.
In general, we expect the cache-coherent model to be cheaper than the
distributed shared-memory model, if we ignore constant factors. The reason
is that if we run a DSM algorithm in a CC model, then the process p to
which a register r is assigned incurs an RMR only if some other process
q accesses p since p’s last access. But then we can amortize p’s RMR by
charging q double. Since q incurs an RMR in the CC model, this tells us that
we pay at most twice as many RMRs in DSM as in CC for any algorithm.
The converse is not true: there are (mildly exotic) problems for which
it is known that CC algorithms are asymptotically more efficient than DSM
algorithms [Gol11, DH04].

17.5.2 RMR complexity of Peterson’s algorithm

As a warm-up, let’s look at the RMR complexity of Peterson’s two-process
mutual exclusion algorithm (Algorithm 17.4). Acquiring the mutex re-
quires going through mostly straight-line code, except for the loop that
tests present[¬i] and waiting.
In the DSM model, spinning on present[¬i] is not a problem (we can
make it a local variable of process i). But waiting is trouble. Whichever
process we don’t assign it to will pay an RMR every time it looks at it. So
Peterson’s algorithm behaves badly by the RMR measure in this model.
Things are better in the CC model. Now process i may pay RMRs for its
first reads of present[¬i] and waiting, but any subsequent reads are free unless
process ¬i changes one of them. But any change to either of the variables
causes process i to leave the loop. It follows that process i pays at most 3
RMRs to get through the busy-waiting loop, giving an RMR complexity of
O(1).
RMR complexities for parts of a protocol that access different registers
add just like step complexities, so the Peterson-Fischer tree construction
described in §17.4.1.2 works here too. The result is O(log n) RMRs per
critical section access, but only in the CC model.
CHAPTER 17. MUTUAL EXCLUSION 141

17.5.3 Mutual exclusion in the DSM model

Yang and Anderson [YA95] give a mutual exclusion algorithm for the DSM
model that requires Θ(log n) RMRs to reach the critical section. This is now
known to be optimal for deterministic algorithms [AHW08]. The core of the
algorithm is a 2-process mutex similar to Peterson’s, with some tweaks so
that each process spins only on its own registers. Pseudocode is given in
Algorithm 17.7; this is adapted from [YA95, Figure 1].

1 C[side(i)] ← i
2 T ←i
3 P [i] ← 0
4 rival ← C[¬side(i)]
5 if rival 6= ⊥ and T = i then
6 if P [rival] = 0 then
7 P [rival] = 1
8 while P [i] = 0 do spin
9 if T = i then
10 while P [i] ≤ 1 do spin

// critical section goes here

11 C[side(i)] ← ⊥
12 rival ← T
13 if rival 6= i then
14 P [rival] ← 2
Algorithm 17.7: Yang-Anderson mutex for two processes

The algorithm is designed to be used in a tree construction where a

process with id in the range {1 . . . n/2} first fights with all other processes
in this range, and similarly for processes in the range {n/2 + 1 . . . n}. The
function side(i) is 0 for the first group of processes and 1 for the second.
The variables C[0] and C[1] are used to record which process is the winner
for each side, and also take the place of the present variables in Peterson’s
algorithm. Each process has its own variable P [i] that it spins on when
blocked; this variable is initially 0 and ranges over {0, 1, 2}; this is used to
signal a process that it is safe to proceed, and tests on P substitute for tests
on the non-local variables in Peterson’s algorithm. Finally, the variable T
is used (like waiting in Peterson’s algorithm) to break ties: when T = i, it’s
i’s turn to wait.
Initially, C[0] = C[1] = ⊥ and P [i] = 0 for all i.
CHAPTER 17. MUTUAL EXCLUSION 142

When I want to enter my critical section, I first set C[side(i)] so you can
find me; this also has the same effect as setting present[side(i)] in Peterson’s
algorithm. I then point T to myself and look for you. I’ll block if I see
C[¬side(i)] = 1 and T = i. This can occur in two ways: one is that I really
write T after you did, but the other is that you only wrote C[¬side(i)] but
haven’t written T yet. In the latter case, you will signal to me that T may
have changed by setting P [i] to 1. I have to check T again (because maybe
I really did write T later), and if it is still i, then I know that you are ahead
of me and will succeed in entering your critical section. In this case I can
safely spin on P [i] waiting for it to become 2, which signals that you have
left.
There is a proof that this actually works in [YA95], but it’s 27 pages
of very meticulously-demonstrated invariants (in fairness, this includes the
entire algorithm, including the tree parts that we omitted here). For intu-
ition, this is not much more helpful than having a program mechanically
check all the transitions, since the algorithm for two processes is effectively
finite-state if we ignore the issue with different processes i jumping into the
role of side(i).
A slightly less rigorous proof but more human-accessible proof would be
analogous to the proof of Peterson’s algorithm. We need to show two things:
first, that no two processes ever both enter the critical section, and second,
that no process gets stuck.
For the first part, consider two processes i and j, where side(i) = 0 and
side(j) = 1. We can’t have both i and j skip the loops, because whichever
one writes T last sees itself in T . Suppose that this is process i and that
j skips the loops. Then T = i and P [i] = 0 as long as j is in the critical
section, so i blocks. Alternatively, suppose i writes T last but does so after
j first reads T . Now i and j both enter the loops. But again i sees T = i on
its second test and blocks on the second loop until j sets P [i] to 2, which
doesn’t happen until after j finishes its critical section.
Now let us show that i doesn’t get stuck. Again we’ll assume that i
wrote T second.
If j skips the loops, then j sets P [i] = 2 on its way out as long as T = i;
this falsifies both loop tests. If this happens after i first sets P [i] to 0, only
i can set P [i] back to 0, so i escapes its first loop, and any j 0 that enters
from the 1 side will see P [i] = 2 before attempting to set P [i] to 1, so P [i]
remains at 2 until i comes back around again. If j sets P [i] to 2 before i
sets P [i] to 0 (or doesn’t set it at all because T = j, then C[side(j)] is set
to ⊥ before i reads it, so i skips the loops.
If j doesn’t skip the loops, then P [i] and P [j] are both set to 1 after i
CHAPTER 17. MUTUAL EXCLUSION 143

and j enter the loopy part. Because j waits for P [j] 6= 0, when it looks at
T the second time it will see T = i 6= j and will skip the second loop. This
causes it to eventually set P [i] to 2 or set C[side(j)] to ⊥ before i reads it
as in the previous case, so again i eventually reaches its critical section.
Since the only operations inside a loop are on local variables, the algo-
rithm has O(1) RMR complexity. For the full tree this becomes O(log n).

17.5.4 Lower bounds

For deterministic algorithms, there is a lower bound due to Attiya, Hendler,
and Woelfel [AHW08] that shows that any one-shot mutual exclusion algo-
rithm for n processes incurs Ω(n log n) total RMRs in either the CC or DSM
models (which implies that some single process incurs Ω(log n) RMRs). This
is based on an earlier breakthrough lower bound of Fan and Lynch [FL06]
that proved the same lower bound for the number of times a register changes
state. Both bounds are information-theoretic: a family of n! executions is
constructed containing all possible orders in which the processes enter the
critical section, and it is shown that each RMR or state change only con-
tributes O(1) bits to choosing between them.
For randomized algorithms, Hendler and Woelfel [HW11] have an al-
gorithm that uses O(log n/ log log n) expected RMRs against an adaptive
adversary and Bender and Gilbert [BG11] can do O(log2 log n) amortized
expected RMRs against an oblivious adversary. Both bounds beat the de-
terministic lower bound. The adaptive-adversary bound is tight, due to a
matching lower bound of Giakkoupis and Woelfel [GW12b] that holds even
for systems that provide compare and swap objects. No non-trivial lower
bound is currently known for an oblivious adversary.
Chapter 18

The wait-free hierarchy

In a shared memory model, it may be possible to solve some problems us-

ing wait-free protocols, in which any process can finish the protocol in a
bounded number of steps, no matter what the other processes are doing (see
Chapter 26 for more on this and some variants).
The wait-free hierarchy hrm classifies asynchronous shared-memory ob-
ject types T by consensus number, where a type T has consensus number
n if with objects of type T and atomic registers (all initialized to appropriate
values1 ) it is possible to solve wait-free consensus (i.e., agreement, validity,
wait-free termination) for n processes but not for n + 1 processes. The con-
sensus number of any type is at least 1, since 1-process consensus requires
no interaction, and may range up to ∞ for particularly powerful objects.
The wait-free hierarchy was suggested by work by Maurice Herlihy [Her91b]
that classified many common (and some uncommon) shared-memory objects
by consensus number, and showed that an unbounded collection of objects
with consensus number n together with atomic registers gives a wait-free
implementation of any object in an n-process system. Various subsequent
authors noticed that this did not give a robust hierarchy in the sense that
combining two types of objects with consensus number n could solve wait-
free consensus for larger n, and the hierarchy hrm was proposed by Prasad
1
The justification for assuming that the objects can be initialized to an arbitrary state
is a little tricky. The idea is that if we are trying to implement consensus from objects of
type T that are themselves implemented in terms of objects of type S, then it’s natural to
assume that we initialize our simulated type-T objects to whatever states are convenient.
Conversely, if we are using the ability of type-T objects to solve n-process consensus to
show that they can’t be implemented from type-S objects (which can’t solve n-process
consensus), then for both the type-T and type-S objects we want these claims to hold no
matter how they are initialized.

144
CHAPTER 18. THE WAIT-FREE HIERARCHY 145

Jayanti [Jay97] as a way of classifying objects that might be robust: an

object is at level n of the hrm hierarchy if having unboundedly many objects
plus unboundedly many registers solves n-process wait-free consensus but
not (n + 1)-process wait-free consensus.2
Whether or not the resulting hierarchy is in fact robust for arbitrary
deterministic objects is still open, but Eric Ruppert [Rup00] subsequently
showed that it is robust for RMW registers and objects with a read opera-
tion that returns the current state, and there is a paper by Borowsky, Gafni,
and Afek [BGA94] that sketches a proof based on a topological characteriza-
tion of computability3 that hrm is robust for deterministic objects that don’t
discriminate between processes (unlike, say, single-writer registers). So for
well-behaved shared-memory objects (deterministic, symmetrically accessi-
ble, with read operations, etc.), consensus number appears to give a real
classification that allows us to say for example that any collection of read-
write registers (consensus number 1), fetch-and-increments (2), test-and-set
bits (2), and queues (2) is not enough to build a compare-and-swap (∞).
We won’t attempt to do the full robustness proofs of Borowsky et al. [BGA94]
or Ruppert [Rup00] that let us get away with this. Instead, we’ll concen-
trate on Herlihy’s original results and show that specific objects have specific
consensus numbers when used in isolation. The procedure in each case will
be to show an upper bound on the consensus number using a variant of
Fischer-Lynch-Paterson (made easier because we are wait-free and don’t
have to worry about fairness) and then show a matching lower bound (for
non-trivial upper bounds) by exhibiting an n-process consensus protocol for
some n. Essentially everything below is taken from Herlihy’s paper [Her91b],
so reading that may make more sense than reading these notes.

18.1 Classification by consensus number

Here we show the position of various types in the wait-free hierarchy. The
quick description is shown in Table 18.1; more details (mostly adapted
from [Her91b]) are given below.
2
The r in hrm stands for the registers, the m for having many objects of the given type.
Jayanti [Jay97] also defines a hierarchy hr1 where you only get finitely many objects. The
h stands for “hierarchy,” or, more specifically, h(T ) stands for the level of the hierarchy
at which T appears [Jay11].
3
See Chapter 28.
CHAPTER 18. THE WAIT-FREE HIERARCHY 146

Consensus Defining Examples

number characteristic
1 Read with Registers, counters,
interfering generalized counters, max registers,
no-return atomic snapshots.
RMW.
2 Interfering Test-and-set, fetch-and-write,
RMW; queue- fetch-and-add, queues,
like structures. process-to-memory swap.
m m-process consensus objects.
2m − 2 Atomic m-register write.
∞ First write- Queue with peek, fetch-and-cons,
like operation sticky bits, compare-and-swap,
wins. memory-to-memory swap, memory-
to-memory copy.

Table 18.1: Position of various types in the wait-free hierarchy

18.1.1 Level 1: atomic registers, counters, other interfering

RMW registers that don’t return the old value
First observe that any type has consensus number at least 1, since 1-process
consensus is trivial.
We’ll argue that a large class of particularly weak objects has consensus
number exactly 1, by running FLP with 2 processes. Recall from Chap-
ter 11 that in the Fischer-Lynch-Paterson [FLP85] proof we classify states
as bivalent or univalent depending on whether both decision values are still
possible, and that with at least one failure we can always start in a bivalent
state (this doesn’t depend on what objects we are using, since it depends
only on having invisible inputs). Since the system is wait-free there is no
constraint on adversary scheduling, and so if any bivalent state has a biva-
lent successor we can just do it. So to solve consensus we have to reach a
bivalent configuration C that has only univalent successors, and in particu-
lar has a 0-valent and a 1-valent successor produced by applying operations
x and y of processes px and py .
Assuming objects don’t interact with each other behind the scenes, x
and y must be operations of the same object. Otherwise Cxy = Cyx and
we get a contradiction.
Now let’s suppose we are looking at atomic registers, and consider cases:
CHAPTER 18. THE WAIT-FREE HIERARCHY 147

• x and y are both reads, Then x and y commute: Cxy = Cyx, and we
get a contradiction.

• x is a read and y is a write. Then py can’t tell the difference between

Cyx and Cxy, so running py to completion gives the same decision
value from both Cyx and Cxy, another contradiction.

• x and y are both writes. Now py can’t tell the difference between Cxy
and Cy, so we get the same decision value for both, again contradicting
that Cx is 0-valent and Cy is 1-valent.

There’s a pattern to these cases that generalizes to other objects. Sup-

pose that an object has a read operation that returns its state and one or
more read-modify-write operations that don’t return anything (perhaps we
could call them “modify-write” operations). We’ll say that the MW opera-
tions are interfering if, for any two operations x and y, either:

• x and y commute: Cxy = Cyx.

• One of x and y overwrites the other: Cxy = Cy or Cyx = Cx.

Then no pair of read or modify-write operations can get us out of a

bivalent state, because (a) reads commute; (b) for a read and MW, the non-
reader can’t tell which operation happened first; (c) and for any two MW
operations, either they commute or the overwriter can’t detect that the first
operation happened. So any MW object with uninformative, interfering
MW operations has consensus number 1.
For example, consider a counter that supports operations read, incre-
ment, decrement, and write: a write overwrites any other operation, and
increments and decrements commute with each other, so the counter has
consensus number 1. The same applies to a generalized counter that sup-
ports an atomic x ← x + a operation; as long as this operation doesn’t
return the old value, it still commutes with other atomic increments. Max
registers (reads on which return the largest value previously written) also
have commutative updates, so they also have consensus number 1.

18.1.2 Level 2: interfering RMW objects that return the old

value, queues (without peek)
Suppose now that we have a RMW object that returns the old value, and
suppose that it is non-trivial in the sense that it has at least one RMW
operation where the embedded function f that determines the new value is
CHAPTER 18. THE WAIT-FREE HIERARCHY 148

not the identity (otherwise RMW is just read). Then there is some value v
such that f (v) 6= v. To solve two-process consensus, have each process pi first
write its preferred value to a register ri , then execute the non-trivial RMW
operation on the RMW object initialized to v. The first process to execute
its operation sees v and decides its own value. The second process sees f (v)
and decides the first process’s value (which it reads from the register). It
follows that non-trivial RMW object has consensus number at least 2.
In many cases, this is all we get. Suppose that the operations of some
RMW type T are non-interfering in a way analogous to the previous defini-
tion, where now we say that x and y commute if they leave the object in the
same state (regardless of what values are returned) and that y overwrites x
if the object is always in the same state after both x and xy (again regard-
less of what is returned). The two processes px and py that carry out x and
y know what happenened, but a third process pz doesn’t. So if we run pz
to completion we get the same decision value after both Cx and Cy, which
means that Cx and Cy can’t be 0-valent and 1-valent. It follows that no
collection of RMW registers with interfering operations can solve 3-process
consensus, and thus all such objects have consensus number 2. Examples
of these objects include test-and-set bits, fetch-and-add registers, and
swap registers that support an operation swap that writes a new value and
return the previous value.
There are some other objects with consensus number 2 that don’t fit this
pattern. Define a wait-free queue as an object with enqueue and dequeue
operations (like normal queues), where dequeue returns ⊥ if the queue is
empty (instead of blocking). To solve 2-process consensus with a wait-free
queue, initialize the queue with a single value (it doesn’t matter what the
value is). We can then treat the queue as a non-trivial RMW register where
a process wins if it successfully dequeues the initial value and loses if it gets
empty.
However, enqueue operations are non-interfering: if px enqueues vx and
py enqueues vy , then any third process can detect which happened first;
similarly we can distinguish enq(x)deq() from deq()enq(x). So to show we
can’t do three process consensus we do something sneakier: given a bivalent
state C with allegedly 0- and 1-valent successors Cenq(x) and Cenq(y),
consider both Cenq(x)enq(y) and Cenq(y)enq(x) and run px until it does
a deq() (which it must, because otherwise it can’t tell what to decide) and
then stop it. Now run py until it also does a deq() and then stop it. We’ve
now destroyed the evidence of the split and poor hapless pz is stuck. In the
case of Cdeq()enq(x) and Cenq(x)deq() on a non-empty queue we can kill
the initial dequeuer immediately and then kill whoever dequeues x or the
CHAPTER 18. THE WAIT-FREE HIERARCHY 149

value it replaced, and if the queue is empty only the dequeuer knows. In
either case we reach indistinguishable states after killing only 2 witnesses,
and the queue has consensus number at most 2.
Similar arguments work on stacks, deques, and so forth—these all have
consensus number exactly 2.

18.1.3 Level ∞: objects where first write wins

These are objects that can solve consensus for any number of processes.
Here are a bunch of level-∞ objects:

Queue with peek Has operations enq(x) and peek(), which returns the
first value enqueued. (Maybe also deq(), but we don’t need it for
consensus). Protocol is to enqueue my input and then peek and return
the first value in the queue.

Fetch-and-cons Returns old cdr and adds new car on to the head of a list.
Use preceding protocol where peek() = tail(car :: cdr).

Sticky bit Has a write operation that has no effect unless register is in
the initial ⊥ state. Whether the write succeeds or fails, it returns
nothing. The consensus protocol is to write my input and then return
result of a read.

Compare-and-swap Has CAS(old, new) operation that writes new only if

previous value is old. Use it to build a sticky bit.

Load-linked/store-conditional Like compare-and-swap split into two op-

erations. The operation reads a memory location and marks it. The
operation succeeds only if the location has not been changed since the
preceding load-linked by the same process. Can be used to build a
sticky bit.

Memory-to-memory swap Has swap(ri , rj ) operation that atomically swaps

contents of ri with rj , as well as the usual read and write operations
for all registers. Use to implement fetch-and-cons. Alternatively, use
two registers input[i] and victory[i] for each process i, where victory[i]
is initialized to 0, and a single central register prize, initialized to 1.
To execute consensus, write your input to input[i], then swap victory[i]
with prize. The winning value is obtained by scanning all the victory
registers for the one that contains a 1, then returning the correspond-
ing input value.)
CHAPTER 18. THE WAIT-FREE HIERARCHY 150

Memory-to-memory copy Has a copy(ri , rj ) operation that copies ri to

rj atomically. Use the same trick as for memory-to-memory swap,
where a process copies prize to victory[i]. But now we have a process
follow up by writing 0 to prize. As soon as this happens, the victory
values are now fixed; take the leftmost 1 as the winner.4
Herlihy [Her91b] gives a slightly more complicated version of this pro-
cedure, where there is a separate prize[i] register for each i, and after
doing its copy a process writes 0 to all of the prize registers. This
shows that memory-to-memory copy solves consensus for arbitrarily
many processes even if we insist that copy operations can never over-
lap. The same trick also works for memory-to-memory swap, since we
can treat a memory-to-memory swap as a memory-to-memory copy
given that we don’t care what value it puts in the prize[i] register.

18.1.4 Level 2m − 2: simultaneous m-register write

Here we have a (large) collection of atomic registers augmented by an m-
register write operation that performs all the writes simultaneously. The
intuition for why this is helpful is that if p1 writes r1 and rshared while p2
writes r2 and rshared then any process can look at the state of r1 , r2 and
rshared and tell which write happened first. Code for this procedure is given
in Algorithm 18.1; note that up to 4 reads may be necessary to determine
the winner because of timing issues.
The workings of Algorithm 18.1 are straightforward:

• If the process reads r1 = r2 = ⊥, then we don’t care which went first,

because the reader (or somebody else) already won.

• If the process reads r1 = 1 and then r2 = ⊥, then p1 went first.

• If the process reads r2 = 2 and then r1 = ⊥, then p2 went first. (This

requires at least one more read after checking the first case.)

• Otherwise the process saw r1 = 1 and r2 = 2. Now read rshared : if it’s

1, p2 went first; if it’s 2, p1 went first.

Algorithm 18.1 requires 2-register writes, and will give us a protocol for 2
processes (since the reader above has to participate somewhere to make the
first case work). For m processes, we can do the same thing with m-register
writes. We have a register rpq = rqp for each pair of distinct processes p
4
Or use any other rule that all processes apply consistently.
CHAPTER 18. THE WAIT-FREE HIERARCHY 151

1 v1 ← r1
2 v2 ← r2
3 if v1 = v2 = ⊥ then
4 return no winner
5 if v1 = 1 and v2 = ⊥ then
// p1 went first
6 return 1
// read r1 again
7 v10 ← r1
8 if v2 = 2 and v10 = ⊥ then
// p2 went first
9 return 2
// both p1 and p2 wrote
10 if rshared = 1 then
11 return 2
12 else
13 return 1
Algorithm 18.1: Determining the winner of a race between 2-register
writes. The assumption is that p1 and p2 each wrote their own ids to ri
and rshared simultaneously. This code can be executed by any process
(including but not limited to p1 or p2 ) to determine which of these
2-register writes happened first.
CHAPTER 18. THE WAIT-FREE HIERARCHY 152

and q, plus a register rpp for each p; this gives a total of m 2

2 + m = O(m )
registers. All registers are initialized to ⊥. Process p then writes its initial
preference to some single-writer register pref p and then simultaneously writes
p to rpq for all q (including rpp ). It then attempts to figure out the first writer
by applying the above test for each q to rpq (standing in for rshared ), rpp (r1 )
and rqq (r2 ). If it won against all the other processes, it decides its own
value. If not, it repeats the test recursively for some p0 that beat it until
it finds a process that beat everybody, and returns its value. So m-register
writes solve m-process wait-free consensus.
A further tweak gets 2m−2: run two copies of an (m−1)–process protocol
using separate arrays of registers to decide a winner for each group. Then
add a second phase where each process has one register sp , in which each
process p from group 1 writes the winning id for its group simultaneously
into sp and sq for each q in the other group. To figure out who won in the
end, build a graph of all victories, where there is an edge from p to q if
and only if p beat q in phase 1 or p’s id was written before q’s id in phase
2. The winner is the (unique) process with at least one outgoing edge and
no incoming edges, which will be the process that won its own group (by
writing first) and whose value was written first in phase 2.

18.1.4.1 Matching impossibility result

It might seem that the technique used to boost from m-process consensus to
(2m − 2)-process consensus could be repeated to get up to at least Θ(m2 ),
but this turns out not to be the case. The essential idea is to show that in
order to escape bivalence, we have to get to a configuration C where every
process is about to do an m-register write leading to a univalent configura-
tion (since reads don’t help for the usual reasons, and normal writes can be
simulated by m-register writes with an extra m − 1 dummy registers), and
then argue that these writes can’t overlap too much. So suppose we are in
such a configuration, and suppose that Cx is 0-valent and Cy is 1-valent, and
we also have many other operations z1 . . . zk that lead to univalent states.
Following Herlihy [Her91b], we argue in two steps:

1. There is some register that is written to by x alone out of all the

pending operations. Proof: Suppose not. Then the 0-valent configu-
ration Cxyz1 . . . zk is indistinguishable from the 1-valent configuration
Cyz1 . . . zk by any process except px , and we’re in trouble.

2. There is some register that is written to by x and y but not by any of

the zi . Proof:: Suppose not. The each register is written by at most
CHAPTER 18. THE WAIT-FREE HIERARCHY 153

one of x and y, making it useless for telling which went first; or it is

overwritten by some zi , hiding the value that tells which went first.
So Cxyz1 . . . zk is indistinguishable from Cyxz1 . . . zk for any process
other than px and py , and we’re still in trouble.

Now suppose we have 2m − 1 processes. The first part says that each of
the pending operations (x, y, all of the zi ) writes to 1 single-writer register
and at least k two-writer registers where k is the number of processes leading
to a different univalent value. This gives k + 1 total registers simultaneously
written by this operation. Now observe that with 2m − 1 process, there is
some set of m processes whose operations all lead to a b-valent state; so
for any process to get to a (¬b)-valent state, it must write m + 1 registers
simultaneously. It follows that with only m simultaneous writes we can only
do (2m − 2)-consensus.

18.1.5 Level m: m-process consensus objects

An m-process consensus object has a single consensus operation that, the
first m times it is called, returns the input value in the first operation, and
thereafter returns only ⊥. Clearly this solves m-process consensus. To show
that it doesn’t solve (m + 1)-process consensus even when augmented with
registers, run a bivalent initial configuration to a configuration C where any
further operation yields a univalent state. By an argument similar to the m-
register write case, we can show that the pending operations in C must all be
consensus operations on the same consensus object (anything else commutes
or overwrites). Now run Cxyz1 . . . zk and Cyxz1 . . . zk , where x and y lead
to 0-valent and 1-valent states, and observe that pk can’t distinguish the
resulting configurations because all it got was ⊥. (Note: this works even if
the consensus object isn’t in its initial state, since we know that before x or
y the configuration is still bivalent.)
So the m-process consensus object has consensus number m. This shows
that hrm is nonempty at each level.
A natural question at this point is whether the inability of m-process
consensus objects to solve (m + 1)-process consensus implies robustness of
the hierarchy. One might consider the following argument: given any object
at level m, we can simulate it with an m-process consensus object, and
since we can’t combine m-process consensus objects to boost the consensus
number, we can’t combine any objects they can simulate either. The problem
here is that while m-process consensus objects can simulate any object in a
system with m processes (see below), it may be that some objects can do
CHAPTER 18. THE WAIT-FREE HIERARCHY 154

more in a system with m + 1 objects while still not solving (m + 1)-process

consensus. A simple way to see this would be to imagine a variant of the m-
process consensus object that doesn’t fail completely after m operations; for
example, it might return one of the first two inputs given to it instead of ⊥.
This doesn’t help with solving consensus, but it might (or might not) make
it too powerful to implement using standard m-process consensus objects.

18.2 Universality of consensus

Universality of consensus says that any type that can implement n-
process consensus can, together with atomic registers, give a wait-free im-
plementation of any object in a system with n processes. That consensus
is universal was shown by Herlihy [Her91b] and Plotkin [Plo89]. Both of
these papers spend a lot of effort on making sure that both the cost of each
operation and the amount of space used is bounded. But if we ignore these
constraints, the same result can be shown using a mechanism similar to the
replicated state machines of §12.1.
Here the processes repeatedly use consensus to decide between candidate
histories of the simulated object, and a process successfully completes an
operation when its operation (tagged to distinguish it from other similar
operations) appears in a winning history. A round structure avoids too
much confusion.
Details are given in Algorithm 18.2.
There are some subtleties to this algorithm. The first time that a process
calls consensus (on c[r]), it may supply a dummy input; the idea is that it
is only using the consensus object to obtain the agreed-upon history from a
round it missed. It’s safe to do this, because no process writes r to its round
register until c[r] is complete, so the dummy input can’t be accidentally
chosen as the correct value.
It’s not hard to see that whatever hr+1 is chosen in c[r+1] is an extension
of hr (it is constructed by appending operations to hr ), and that all processes
agree on it (by the agreement property of the consensus object c[r + 1]. So
this gives us an increasing sequence of consistent histories. We also need
to show that these histories are linearizable. The obvious linearization is
just the most recent version of hr . Suppose some call to apply(π1 ) finishes
before a call to apply(π2 ) starts. Then π1 is contained in some hr when
apply(π1 ) finishes, and since π2 can only enter h by being appended at the
end, we get π1 linearized before π2 .
Finally, we need to show termination. The algorithm is written with a
CHAPTER 18. THE WAIT-FREE HIERARCHY 155

1 procedure apply(π)
// announce my intended operation
2 op[i] ← π
3 while true do
// find a recent round
4 r ← maxj round[j]
// obtain the history as of that round
5 if hr = ⊥ then
6 hr ← consensus(c[r], ⊥)
7 if π ∈ hr then
8 return value π returns in hr
// else attempt to advance
9 h0 ← hr
10 for each j do
11 if op[j] 6∈ h0 then
12 append op[j] to h0

13 hr+1 ← consensus(c[r + 1], h0 )

14 round[i] ← r + 1

Algorithm 18.2: A universal construction based on consensus

CHAPTER 18. THE WAIT-FREE HIERARCHY 156

loop, so in principle it could run forever. But we can argue that no process
after executes the loop more than twice. The reason is that a process p puts
its operation in op[p] before it calculates r; so any process that writes r0 > r
to round sees p’s operation before the next round. It follows that p’s value
gets included in the history no later than round r + 2. (We’ll see this sort
of thing again when we do atomic snapshots in Chapter 19.)
Building a consistent shared history is easier with some particular objects
that solve consensus. For example, a fetch-and-cons object that supplies
an operation that pushes a new head onto a linked list and returns the old
head trivially implements the common history above without the need for
helping. One way to implement fetch-and-cons is with a swap object; to
add a new element to the list, create a cell with its next pointer pointing to
itself, then swap the next field with the head pointer for the entire list.
The solutions we’ve described here have a number of deficiencies that
make them impractical in a real system (even more so than many of the
algorithms we’ve described). If we store entire histories in a register, the
register will need to be very, very wide. If we store entire histories as a linked
list, it will take an unbounded amount of time to read the list. For solutions
to these problems, see [AW04, 15.3] or the papers of Herlihy [Her91b] and
Plotkin [Plo89].
Chapter 19

Atomic snapshots

We’ve seen in the previous chapter that there are a lot of things we can’t
make wait-free with just registers. But there are a lot of things we can.
Atomic snapshots are a tool that let us do a lot of these things easily.
An atomic snapshot object acts like a collection of n single-writer
multi-reader atomic registers with a special snapshot operation that returns
(what appears to be) the state of all n registers at the same time. This is
easy without failures: we simply lock the whole register file, read them all,
and unlock them to let all the starving writers in. But it gets harder if
we want a protocol that is wait-free, where any process can finish its own
snapshot or write even if all the others lock up.
We’ll give the usual sketchy description of a couple of snapshot algo-
rithms. More details on early snapshot results can be found in [AW04,
§10.3] or [Lyn96, §13.3]. There is also a reasonably recent survey by Fich
on upper and lower bounds for the problem [Fic05].

19.1 The basic trick: two identical collects equals

a snapshot
Let’s tag any value written with a sequence number, so that each value
written has a seqno field attached to it that increases over time. We can
now detect if a new write has occurred between two reads of the same
variable. Suppose now that we repeatedly perform collects—reads of all
n registers—until two successive collects return exactly the same vector of
values and sequence numbers. We can then conclude that precisely these
values were present in the registers at some time in between the two collects.
This gives us a very simple algorithm for snapshot. Unfortunately, it doesn’t

157
CHAPTER 19. ATOMIC SNAPSHOTS 158

terminate if there are a lot of writers around.1 So we need some way to slow
the writers down, or at least get them to do snapshots for us.

19.2 The Gang of Six algorithm

This is the approach taken by Afek and his five illustrious co-authors [AAD+ 93]
(see also [AW04, §10.3] or [Lyn96, §13.3.2]): before a process can write to
its register, it first has to complete a snapshot and leave the results behind
with its write.2 This means that if some slow process (including a slow
writer, since now writers need to do snapshots too) is prevented from doing
the two-collect snapshot because too much writing is going on, eventually
it can just grab and return some pre-packaged snapshot gathered by one of
the many successful writers.
Specifically, if a process executing a single snapshot operation σ sees
values written by a single process i with three different sequence numbers
s1 , s2 and s3 , then it can be assured that the snapshot σ3 gathered with
sequence number s3 started no earlier than s2 was written (and thus no
earlier than σ started, since σ read s1 after it started) and ended no later
than σ ended (because σ saw it). It follows that the snapshot can safely
return σ3 , since that represents the value of the registers at some time inside
σ3 ’s interval, which is contained completely within σ’s interval.
So a snapshot repeatedly does collects until either (a) it gets two identical
collects, in which case it can return the results (a direct scan, or (b) it sees
three different values from the same process, in which case it can take the
snapshot collected by the second write (an indirect scan). See pseudocode
in Algorithm 19.1.
Amazingly, despite the fact that updates keep coming and everybody is
trying to do snapshots all the time, a snapshot operation of a single process
is guaranteed to terminate after at most n + 1 collects. The reason is that
1
This isn’t always a problem, since there may be external factors that keep the writers
from showing up too much. Maurice Herlihy and I got away with using exactly this
snapshot algorithm in an ancient, pre-snapshot paper on randomized consensus [AH90a].
2
The algorithm is usually called the AADGMS algorithm by people who can remember
all the names—or at least the initials—of the team of superheroes who came up with it
(Afek, Attiya, Dolev, Gafni, Merritt, and Shavit) and Gang of Six by people who can’t.
Historically, this was one of three independent solutions to the problem that appeared at
about the same time; a similar algorithm for composite registers was given by James
Anderson [And94] and a somewhat different algorithm for consistent scan was given
by Aspnes and Herlihy [AH90b]. The Afek et al. algorithm had the advantage of using
bounded registers (in its full version), and so it and its name for atomic snapshot prevailed
over its competitors.
CHAPTER 19. ATOMIC SNAPSHOTS 159

in order to prevent case (a) from holding, the adversary has to supply at
least one new value in each collect after the first. But it can only supply one
new value for each of the n − 1 processes that aren’t doing collects before
case (b) is triggered (it’s triggered by the first process that shows up with
a second new value). Adding up all the collects gives 1 + (n − 1) + 1 =
n + 1 collects before one of the cases holds. Since each collect takes n − 1
read operations (assuming the process is smart enough not to read its own
register), a snapshot operation terminates after at most n2 − 1 reads.

1 procedure updatei (A, v)

2 s ← scan(A)
3 A[i] ← hA[i].count + 1, v, si
4 procedure scan(A)
5 initial ← collect(A)
6 previous ← initial while true do
7 s ← scan(A)
8 if s = previous then
// Two identical collects
9 return s
10 else if ∃j : s[j].count ≥ initial[j].count + 2 do
// Three different counts from j
11 return s[j].snapshot
12 else
// Nothing useful, try again
13 previous ← s

Algorithm 19.1: Snapshot of [AAD+ 93] using unbounded registers

For a write operation, a process first performs a snapshot, then writes

the new value, the new sequence number, and the result of the snapshot
to its register (these are very wide registers). The total cost is n2 − 1 read
operations for the snapshot plus 1 write operation.

19.2.1 Linearizability
We now need to argue that the snapshot vectors returned by the Afek et al.
algorithm really work, that is, that between each matching invoke-snapshot
and respond-snapshot there was some actual time where the registers in
the array contained precisely the values returned in the respond-snapshot
CHAPTER 19. ATOMIC SNAPSHOTS 160

action. We do so by assigning a linearization point to each snapshot vec-

tor, a time at which it appears in the registers (which for correctness of the
protocol had better lie within the interval between the snapshot invocation
and response). For snapshots obtained through case (a), take any time be-
tween the two collects. For snapshots obtained through case (b), take the
serialization point already assigned to the snapshot vector provided by the
third write. In the latter case we argue by induction on termination times
that the linearization point lies inside the snapshot’s interval.
Note that this means that all snapshots were ultimately collected by two
successive collects returning identical values, since any case-(b) snapshot
sits on top of a finite regression of case-(b) snapshots that must end with
a case-(a) snapshot. In practice what this means is that if there are many
writers, eventually all of them will stall waiting for a case-(a) snapshot to
complete, which happens because all the writers are stuck. So effectively
the process of requiring writers to do snapshots first almost gives us a form
of locking, but without the vulnerability to failures of a real lock. (In fact,
crash failures just mean that there are fewer writers to screw things up,
allowing snapshots to finish faster.)

19.2.2 Using bounded registers

The simple version of the Afek et al. algorithm requires unbounded registers
(since sequence numbers may grow forever). One of the reasons why this
algorithm required so many smart people was to get rid of this assumption:
the paper describes a (rather elaborate) mechanism for recycling sequence
numbers that prevents unbounded growth (see also [Lyn96, 13.3.3]). In
practice, unbounded registers are probably not really an issue once one has
accepted very large registers, but getting rid of them is an interesting theo-
retical problem.
It turns out that with a little cleverness we can drop the sequence num-
bers entirely. The idea is that we just need a mechanism to detect when
somebody has done a lot of writes while a snapshot is in progress. A naive
approach would be to have sequence numbers wrap around mod m for some
small constant modulus m; this fails because if enough snapshots happen
between two of my collects, I may notice none of them because all the se-
quence numbers wrapped around all the way. But we can augment mod-m
sequence numbers with a second handshaking mechanism that detects when
a large enough number of snapshots have occurred; this acts like the guard
bit on an automobile odometer, than signals when the odometer has over-
flowed to prevent odometer fraud by just running the odometer forward an
CHAPTER 19. ATOMIC SNAPSHOTS 161

extra million miles or so.

The result is the full version of Afek et al. [AAD+ 93]. (Our presentation
here follows [AW04, 10.3].) The key mechanism for detecting odometer fraud
is a handshake, a pair of single-writer bits used by two processes to signal
each other that they have done something. Call the processes S (for same)
and D (for different), and supposed we have handshake bits hS and hD . We
then provide operations tryHandshake (signal that something is happening)
and checkHandshake (check if something happened) for each process; these
operations are asymmetric. The code is:

tryHandshake(S): hS ← hD (make the two bits the same)

tryHandshake(D): hD ← ¬hS (make the two bits different)

checkHandshake(S): return hS 6= hD (return true if D changed its bit)

checkHandshake(D): return hS = hD (return true if S changed its bit)

The intent is that checkHandshake returns true if the other process

called tryHandshake after I did. The situation is a bit messy, however,
since tryHandshake involves two register operations (reading the other bit
and then writing my own). So in fact we have to look at the ordering of
these read and write events. Let’s assume that checkHandshake is called
by S (so it returns true if and only if it sees different values). Then we have
two cases:

1. checkHandshake(S) returns true. Then S reads a different value in

hD from the value it read during its previous call to tryHandshake(S).
It follows that D executed a write as part of a tryHandshake(D)
operation in between S’s previous read and its current read.

2. checkHandshake(S) returns false. Then S reads the same value in hD

as it read previously. This does not necessarily mean that D didn’t
write hD during this interval—it is possible that D is just very out
of date, and did a write that didn’t change the register value—but it
does mean that D didn’t perform both a read and a write since S’s
previous read.

How do we use this in a snapshot algorithm? The idea is that before

performing my two collects, I will execute tryHandshake on my end of a
pair of handshake bits for every other process. After performing my two
collects, I’ll execute checkHandshake. I will also assume each update (after
CHAPTER 19. ATOMIC SNAPSHOTS 162

performing a snapshot) toggles a mod-2 sequence number bit on the value

stored in its segment of the snapshot array. The hope is that between the
toggle and the handshake, I detect any changes. (See [AW04, Algorithm 30]
for the actual code.)
Does this work? Let’s look at cases:

1. The toggle bit for some process q is unchanged between the two snap-
shots taken by p. Since the bit is toggled with each update, this means
that an even number of updates to q 0 s segment occurred during the
interval between p’s writes. If this even number is 0, we are happy: no
updates means no call to tryHandshake by q, which means we don’t
see any change in q’s segment, which is good, because there wasn’t
any. If this even number is 2 or more, then we observe that each of
these events precedes the following one:

• p’s call to tryHandshake.

• p’s first read.
• q’s first write.
• q’s call to tryHandshake at the start of its second scan.
• q’s second write.
• p’s second read.
• p’s call to checkHandshake.

It follows that q both reads and writes the handshake bits in between
p’s calls to tryHandshake and checkHandshake, so p correctly sees
that q has updated its segment.

2. The toggle bit for q has changed. Then q did an odd number of updates
(i.e., at least one), and p correctly detects this fact.

What does p do with this information? Each time it sees that q has done
a scan, it updates a count for q. If the count reaches 3, then p can determine
that q’s last scanned value is from a scan that is contained completely within
the time interval of p’s scan. Either this is a direct scan, where q actually
performs two collects with no changes between them, or it’s an indirect
scan, where q got its value from some other scan completely contained
within q’s scan. In the first case p is immediately happy; in the second,
we observe that this other scan is also contained within the interval of p’s
scan, and so (after chasing down a chain of at most n − 1 indirect scans) we
eventually reach a direct scan contained within it that provided the actual
CHAPTER 19. ATOMIC SNAPSHOTS 163

value. In either case p returns the value of pair of adjacent collects with
no changes between them that occurred during the execution of its scan
operation, which gives us linearizability.

19.3 Faster snapshots using lattice agreement

The Afek et al. algorithm and its contemporaries all require O(n2 ) opera-
tions for each snapshot. It is possible to get this bound down to O(n) using
a more clever algorithm, [IMCT94] which is the best we can reasonably hope
for in the worst case given that (a) even a collect (which doesn’t guarantee
anything about linearizability) requires Θ(n) operations when implemented
in the obvious way, and (b) there is a linear lower bound, due to Jayanti,
Tan, and Toueg [JTT00], on a large class of wait-free objects that includes
snapshot.3
The first step, due to Attiya, Herlihy, and Rachman [AHR95], is a re-
duction to a related problem called lattice agreement.

19.3.1 Lattice agreement

A lattice is a partial order in which every pair of elements x, y has a least
upper bound x ∨ y called the join of x and y and a greatest lower bound
x ∧ y called the meet of x and y. For example, we can make a lattice out
of sets by letting join be union and meet be intersection; or we can make a
lattice out of integers by making join be max and meet be min.
In the lattice agreement problem, each process starts with an input xi
and produces an output yi , where both are elements of some lattice. The
requirements of the problem are:

Comparability For all i, j, yi ≤ yj or yj ≤ yi .

Downward validity For all i, xi ≤ yi .

Upward validity For all i, yi ≤ x1 ∨ x2 ∨ x3 ∨ . . . ∨ xn .

These requirements are analogous to the requirements for consensus.

Comparability acts like agreement: the views returned by the lattice-agreement
protocol are totally ordered. Downward validity says that each process will
include its own input in its output. Upward validity acts like validity: an
output can’t include anything that didn’t show up in some input.
3
But see §21.5 for a faster alternative if we allow either randomization or limits on the
number of times the array is updated.
CHAPTER 19. ATOMIC SNAPSHOTS 164

For the snapshot algorithm, we also demand wait-freedom: each pro-

cess terminates after a bounded number of its own steps, even if other pro-
cesses fail.
Note that if we are really picky, we can observe that we don’t actually
need meets; a semi-lattice that provides only joins is enough. In practice
we almost always end up with a full-blown lattice, because (a) we are working
with finite sets, and (b) we generally want to include a bottom element ⊥
that is less than all the other elements, to represent the “empty” state of
our data structure. But any finite join-semi-lattice with a bottom element
turns out to be a lattice, since we can define x ∧ y as the join of all elements
z such that z ≤ x and z ≤ y. We don’t use the fact that we are in a lattice
anywhere, but it does save us two syllables not to have to say “semi-lattice
agreement.”

19.3.2 Connection to vector clocks

The first step in reducing snapshot to lattice agreement is to have each writer
generates a sequence of increasing timestamps r1 , r2 , . . . , and a snapshot
corresponds to some vector of timestamps [t1 , t2 . . . tn ], where ti indicates
the most recent write by pi that is included in the snapshot (in other words,
we are using vector clocks again; see §7.2.3). Now define v ≤ v 0 if vi ≤ vi0
for all i; the resulting partial order is a lattice, and in particular we can
compute x ∨ y by the rule (x ∨ y)i = xi ∨ yi .
Suppose now that we have a bunch of snapshots that satisfy the compa-
rability requirement; i.e., they are totally ordered. Then we can construct a
sequential execution by ordering the snapshots in increasing order with each
update operation placed before the first snapshot that includes it. This se-
quential execution is not necessarily a linearization of the original execution,
and a single lattice agreement object won’t support more than one opera-
tion for each process, but the idea is that we can nonetheless use lattice
agreement objects to enforce comparability between concurrent executions
of snapshot, while doing some other tricks (exploiting, among other things,
the validity properties of the lattice agreement objects) to get linearizability
over the full execution.

19.3.3 The full reduction

The Attiya-Herlihy-Rachman algorithm is given as Algorithm 19.2. It uses
an array of registers Ri to hold round numbers (timestamps); an array Si
to hold values to scan; an unboundedly humongous array Vir to hold views
CHAPTER 19. ATOMIC SNAPSHOTS 165

obtained by each process in some round; and a collection of lattice-agreement

objects LAr , one for each round.

1 procedure scan()
// First attempt
2 Ri ← r ← max(R1 . . . Rn , Ri + 1)
3 collect ← read(S1 . . . Sn )
4 view ← LAr (collect)
5 if max(R1 . . . Rn ) > Ri then
// Fall through to second attempt
6 else
7 Vir ← view
8 return Vir
// Second attempt
9 Ri ← r ← max(R1 . . . Rn , Ri + 1)
10 collect ← read(S1 . . . Sn )
11 view ← LAr (collect)
12 if max(R1 . . . Rn ) > Ri then
13 Vir ← some nonempty Vjr
14 return Vir
15 else
16 Vir ← view
17 returnVir

Algorithm 19.2: Lattice agreement snapshot

The algorithm makes two attempts to obtain a snapshot. In both cases,

the algorithm advances to the most recent round it sees (or its previous round
plus one, if nobody else has reached this round yet), attempts a collect, and
then runs lattice-agreement to try to get a consistent view. If after getting
its first view it finds that some other process has already advanced to a
later round, it makes a second attempt at a new, higher round r0 and uses
some view that it obtains in this second round, either directly from lattice
agreement, or (if it discovers that it has again fallen behind), it uses an
indirect view from some speedier process.
The reason why I throw away my view if I find out you have advanced
to a later round is not because the view is bad for me but because it’s bad
for you: I might have included some late values in my view that you didn’t
see, breaking consistency between rounds. But I don’t have to do this more
CHAPTER 19. ATOMIC SNAPSHOTS 166

than once; if the same thing happens on my second attempt, I can use an
indirect view as in [AAD+ 93], knowing that it is safe to do so because any
collect that went into this indirect view started after I did.
The update operation is the usual update-and-scan procedure; for com-
pleteness this is given as Algorithm 19.3. To make it easier to reason about
the algorithm, we assume that an update returns the result of the embedded
scan.

1 procedure updatei (v)

2 Si ← (Si .seqno + 1, v)
3 return scan()
Algorithm 19.3: Update for lattice agreement snapshot

19.3.4 Why this works

We need to show three facts:

1. All views returned by the scan operation are comparable; that is, there
exists a total order on the set of views (which can be extended to a
total order on scan operations by breaking ties using the execution
order).

2. The view returned by an update operation includes the update (this

implies that future views will also include the update, giving the cor-
rect behavior for snapshot).

3. The total order on views respects the execution order: if π1 and π2 are
scan operations that return v1 and v2 , then scan1 <S scan2 implies
view1 ≤ view2 . (This gives us linearization.)

Let’s start with comparability. First observe that any view returned is
either a direct view (obtained from LAr ) or an indirect view (obtained from
Vjr for some other process j). In the latter case, following the chain of
indirect views eventually reaches some direct view. So all views returned for
a given round are ultimately outputs of LAr and thus satisfy comparability.
But what happens with views from different rounds? The lattice-agreement
objects only operate within each round, so we need to ensure that any view
returned in round r is included in any subsequent rounds. This is where
checking round numbers after calling LAr comes in.
CHAPTER 19. ATOMIC SNAPSHOTS 167

Suppose some process i returns a direct view; that is, it sees no higher
round number in either its first attempt or its second attempt. Then at
the time it starts checking the round number in Line 5 or 12, no process
has yet written a round number higher than the round number of i’s view
(otherwise i would have seen it). So no process with a higher round number
has yet executed the corresponding collect operation. When such a process
does so, it obtains values that are at least as current as those fed into LAr ,
and i’s round-r view is less than or equal to the vector of these values by
upward validity of LAr and thus less than or equal to the vector of values
returned by LAr0 for r0 > r by upward validity. So we have comparability
of all direct views, which implies comparability of all indirect views as well.
To show that each view returned by scan includes the preceding update,
we observe that either a process returns its first-try scan (which includes
the update by downward validity) or it returns the results of a scan in the
second-try round (which includes the update by downward validity in the
later round, since any collect in the second-try round starts after the update
occurs). So no updates are missed.
Now let’s consider two scan operations π1 and π2 where π1 precedes π2
in the execution. We want to show that, for the views v1 and v2 that these
scans return, v1 ≤ v2 . From the comparability property, the only way this
can fail is if v2 < v1 ; that is, there is some update included in v2 that is
not included in v1 . But this can’t happen; if π2 starts after π1 finishes, it
starts after any update π1 sees is already in one of the Sj registers, and so
π2 will include this update in its initial collect. (Slightly more formally, if s
is the contents of the registers at some time between when π1 finishes and
π2 starts, then v1 ≤ s by upward validity and s ≤ v2 by downward validity
of the appropriate LA objects.)

19.3.5 Implementing lattice agreement

There are several known algorithms for implementing lattice agreement, in-
cluding the original algorithm of Attiya, Herlihy, and Rachman [AHR95] and
an adaptive algorithm of Attiya and Fouren [AF01]. The best of them (as-
suming multi-writer registers) is Inoue et al.’s linear-time lattice agreement
protocol [IMCT94].
The intuition behind this protocol is to implement lattice agreement
using divide-and-conquer. The processes are organized into a tree, with
each leaf in the tree corresponding to some process’s input. Internal nodes
of the tree hold data structures that will report increasingly large subsets
of the inputs under them as they become available. At each internal node,
CHAPTER 19. ATOMIC SNAPSHOTS 168

a double-collect snapshot is used to ensure that the value stored at that

node is always the union of two values that appear in its children at the
same time. This is used to guarantee that, so long as each child stores an
increasing sequence of sets of inputs, the parent does so also.
Each process ascends the tree updating nodes as it goes to ensure that
its value is included in the final result. A rather clever data structure is
used to ensure that out-of-date smaller sets don’t overwrite larger ones at
any node, and the cost of using this data structure and carrying out the
double-collect snapshot at a node with m leaves below it is shown to be
O(m). So the total cost of a snapshot is O(n + n/2 + n/4 + . . . 1) = O(n),
giving the linear time bound.
Let’s now look at the details of this protocol. There are two main com-
ponents: the Union algorithm used to compute a new value for each node
of the tree, and the ReadSet and WriteSet operations used to store the
data in the node. These are both rather specialized algorithms and depend
on the details of the other, so it is not trivial to describe them in isolation
from each other; but with a little effort we can describe exactly what each
component demands from the other, and show that it gets it.
The Union algorithm does the usual two-collects-without change trick to
get the values of the children and then stores the result. In slightly more
detail:

1. Perform ReadSet on both children. This returns a set of leaf values.

2. Perform ReadSet on both children again.

3. If the values obtained are the same in both collects, call WriteSet on
the current node to store the union of the two sets and proceed to the
parent node. Otherwise repeat the preceding step.

The requirement of the Union algorithm is that calling ReadSet on a

given node returns a non-decreasing sequence of sets of values; that is, if
ReadSet returns some set S at a particular time and later returns S 0 , then
S ⊆ S 0 . We also require that the set returned by ReadSet is a superset
of any set written by a WriteSet that precedes it, and that it is equal to
some such set. This last property only works if we guarantee that the values
stored by WriteSet are all comparable (which is shown by induction on the
behavior of Union at lower levels of the tree).
Suppose that all these conditions hold; we want to show that the values
written by successive calls to Union are all comparable, that is, for any
values S, S 0 written by union we have S ⊆ S 0 or S 0 ⊆ S. Observe that
CHAPTER 19. ATOMIC SNAPSHOTS 169

S = L ∪ R and S 0 = L0 ∪ R0 where L, R and L0 , R0 are sets read from

the children. Suppose that the Union operation producing S completes its
snapshot before the operation producing S 0 . Then L ⊆ L0 (by the induction
hypothesis) and R ⊆ R0 , giving S ⊆ S 0 .
We now show how to implement the ReadSet and WriteSet operations.
The main thing we want to avoid is the possibility that some large set
gets overwritten by a smaller, older one. The solution is to have m registers
a[1 . . . m], and write a set of size s to every register in a[1 . . . s] (each register
gets a copy of the entire set). Because register a[s] gets only sets of size s or
larger, there is no possibility that our set is overwritten by a smaller one. If
we are clever about how we organize this, we can guarantee that the total
cost of all calls to ReadSet by a particular process is O(m), as is the cost of
the single call to WriteSet in Union.
Pseudocode for both is given as Algorithm 19.4. This is a simplified
version of the original algorithm from [IMCT94], which does the writes in
increasing order and thus forces readers to finish incomplete writes that they
observe, as in Attiya-Bar-Noy-Dolev [ABND95] (see also Chapter 16).

shared data: array a[1 . . . m] of sets, initially ∅

local data: index p, initially 0

1 procedure WriteSet(S)
2 for i ← |S| down to 1 do
3 a[i] ← S

4 procedure ReadSet()
// update p to last nonempty position
5 while true do
6 s ← a[p]
7 if p = m or a[p + 1] = ∅ then
8 break
9 else
10 p←p+1

11 return s
Algorithm 19.4: Increasing set data structure

Naively, one might think that we could just write directly to a[|S|] and
skip the previous ones, but this makes it harder for a reader to detect that
CHAPTER 19. ATOMIC SNAPSHOTS 170

a[|S|] is occupied. By writing all the previous registers, we make it easy to

tell if there is a set of size |S| or bigger in the sequence, and so a reader can
start at the beginning and scan forward until it reaches an empty register,
secure in the knowledge that no larger value has been written.4 Since we
want to guarantee that no reader every spends more that O(m) operations
on an array of m registers (even if it does multiple calls to ReadSet), we also
have it remember the last location read in each call to ReadSet and start
there again on its next call. For WriteSet, because we only call it once, we
don’t have to be so clever, and can just have it write all |S| ≤ m registers.
We need to show linearizability. We’ll do so by assigning a specific lin-
earization point to each high-level operation. Linearize each call to ReadSet
at the last time that it reads a[p]. Linearize each call to WriteSet(S) at the
first time at which a[|S|] = S and a[i] 6= ∅ for every i < |S| (in other words,
at the first time that some reader might be able to find and return S); if
there is no such time, linearize the call at the time at which it returns. Since
every linearization point is inside its call’s interval, this gives a linearization
that is consistent with the actual execution. But we have to argue that it
is also consistent with a sequential execution, which means that we need
to show that every ReadSet operation returns the largest set among those
whose corresponding WriteSet operations are linearized earlier.
Let R be a call to ReadSet and W a call to WriteSet(S). If R returns S,
then at the time that R reads S from a[|S|], we have that (a) every register
a[i] with i < |S| is non-empty (otherwise R would have stopped earlier),
and (b) |S| = m or a[|S| + 1] = ∅ (as otherwise R would have kept going
after later reading a[|S| + 1]. From the rule for when WriteSet calls are
linearized, we see that the linearization point of W precedes this time and
that the linearization point of any call to WriteSet with a larger set follows
it. So the return value of R is consistent.
The payoff: unless we do more updates than snapshots, don’t want to
assume multi-writer registers, are worried about unbounded space, have a
beef with huge registers, or care about constant factors, it costs no more
time to do a snapshot than a collect. So in theory we can get away with
assuming snapshots pretty much wherever we need them.
4
This trick of reading in one direction and writing in another dates back to a paper by
Lamport from 1977 [Lam77].
CHAPTER 19. ATOMIC SNAPSHOTS 171

19.4 Practical snapshots using LL/SC

Though atomic registers are enough for snapshots, it is possible to get
a much more efficient snapshot algorithm using stronger synchronization
primitives. An algorithm of Riany, Shavit, and Touitou [RST01] uses load-
linked/store-conditional objects to build an atomic snapshot protocol
with linear-time snapshots and constant-time updates using small registers.
We’ll give a sketch of this algorithm here.
The RST algorithm involves two basic ideas: the first is a snapshot algo-
rithm for a single scanner (i.e., only one process can do snapshots) in which
each updater maintains two copies of its segment, a high copy (that may
be more recent than the current scan) and a low copy (that is guaranteed
to be no more recent than the current scan). The idea is that when a scan
is in progress, updaters ensure that the values in memory at the start of
the scan are not overwritten before the scan is completed, by copying them
to the low registers, while the high registers allow new values to be written
without waiting for the scan to complete. Unbounded sequence numbers,
generated by the scanner, are used to tell which values are recent or not.
As long as there is only one scanner, nothing needs to be done to ensure
that all scans are consistent. But extending the algorithm to multiple scan-
ners is tricky. A simple approach would be to keep a separate low register for
each concurrent scan—however, this would require up to n low registers and
greatly increase the cost of an update. Instead, the authors devise a mecha-
nism, called a coordinated collect, that allows the scanners collectively to
implement a sequence of virtual scans that do not overlap. Each virtual scan
is implemented using the single-scanner algorithm, with its output written
to a common view array that is protected from inconsistent updates using
LL/SC operations. A scanner participates in virtual scans until it obtains a
virtual scan that is useful to it (this means that the virtual scan has to take
place entirely within the interval of the process’s actual scan operation); the
simplest way to arrange this is to have each scanner perform two virtual
scans and return the value obtained by the second one.
The paper puts a fair bit of work into ensuring that only O(n) view arrays
are needed, which requires handling some extra special cases where partic-
ularly slow processes don’t manage to grab a view before it is reallocated
for a later virtual scan. We avoid this complication by simply assuming an
unbounded collection of view arrays; see the paper for how to do this right.
A more recent paper by Fatourou and Kallimanis [FK07] gives improved
time and space complexity using the same basic technique.
CHAPTER 19. ATOMIC SNAPSHOTS 172

19.4.1 Details of the single-scanner snapshot

The single-scanner snapshot is implemented using a shared currSeq variable
(incremented by the scanner but used by all processes) and an array memory
of n snapshot segments, each of which is divided into a high and low com-
ponent consisting of a value and a timestamp. Initially, currSeq is 0, and all
memory locations are initialized to (⊥, 0). This part of the algorithm does
not require LL/SC.
A call to scan copies the first of memory[j].high or memory[j].low that
has a sequence number less than the current sequence number. Pseudocode
is given as Algorithm 19.5.

1 procedure scan()
2 currSeq ← currSeq + 1
3 for j ← 0 to n − 1 do
4 h ← memory[j].high
5 if h.seq < currSeq then
6 view[j] ← h.value
7 else
8 view[j] ← memory[j].low.value

Algorithm 19.5: Single-scanner snapshot: scan

The update operation for process i cooperates by copying memory[i].high

to memory[i].low if it’s old.
The update operation always writes its value to memory[i].high, but pre-
serves the previous value in memory[i].low if its sequence number indicates
that it may have been present at the start of the most recent call to scan.
This means that scan can get the old value if the new value is too recent.
Pseudocode is given in Algorithm 19.6.

1 procedure update()
2 seq ← currSeq
3 h ← memory[i].high
4 if h.seq 6= seq then
5 memory[i].low ← h
6 memory[i].high ← (value, seq)
Algorithm 19.6: Single-scanner snapshot: update
CHAPTER 19. ATOMIC SNAPSHOTS 173

To show this actually works, we need to show that there is a linearization

of the scans and updates that has each scan return precisely those values
whose corresponding updates are linearized before it. The ordering is based
on when each scan operation S increments currSeq and when each update
operation U reads it; specifically:
• If U reads currSeq after S increments it, then S < U .
• If U reads currSeq before S increments it and S reads memory[i].high
(where i is the process carrying out U ) before U writes it, then S < U .
• If U reads currSeq before S increments it, but S reads memory[i].high
after U writes it, then U < S.
Updates are ordered based on intervening scans (i.e., U1 < U2 if U1 < S
and S < U2 by the above rules), or by the order in which they read currSeq
if there is no intervening scan.
To show this is a linearization, we need first to show that it extends the
ordering between operations in the original schedule. Each of the above rules
has π1 < π2 only if some low-level operation of π1 precedes some low-level
operation of π2 , with the exception of the transitive ordering of two update
events with an intervening scan. But in this last case we observe that if
U1 < S, then U1 writes memory[i].high before S reads it, so if U1 precedes
U2 in the actual execution, U2 must write memory[i].high after S reads it,
implying S < U2 .
Now we show that the values returned by scan are consistent with the
linearization ordering; that, is, for each i, scan copies to view[i] the value
in the last update by process i in the linearization. Examining the code for
scan, we see that a scan operation S takes memory[i].high if its sequence
number is less than currSeq, i.e. if the update operation U that wrote
it read currSeq before S incremented it and wrote memory[i].high before
S read it; this gives U < S. Alternatively, if scan takes memory[i].low,
then memory[i].low was copied by some update operation U 0 from the value
written to memory[i].high by some update U that read currSeq before S
incremented it. Here U 0 must have written memory[i].high before S read
it (otherwise S would have taken the old value left by U ) and since U
precedes U 0 (being an operation of the same process) it must therefor also
have written memory[i].high before S read it. So again we get the first case
of the linearization ordering and U < S.
So far we have shown only that S obtains values that were linearized
before it, but not that it ignores values that were linearized after it. So now
let’s consider some U with S < U . Then one of two cases holds:
CHAPTER 19. ATOMIC SNAPSHOTS 174

• U reads currSeq after S increments it. Then U writes a sequence

number in memory[i].high that is greater than or equal to the currSeq
value used by S; so S returns memory[i].low instead, which can’t have
a sequence number equal to currSeq and thus can’t be U ’s value either.

• U reads currSeq before S increments it but writes memory[i].high after

S reads it. Now S won’t return U ’s value from memory[i].high (it didn’t
read it), and won’t get it from memory[i].low either (because the value
that is in memory[i].high will have seq < currSeq, and so S will take
that instead).

So in either case, if S < U , then S doesn’t return U ’s value. This

concludes the proof of correctness.

19.4.2 Extension to multiple scanners

See the paper for details.
The essential idea: view now represents a virtual scan viewr generated
cooperatively by all the scanners working together in some asynchronous
round r. To avoid conflicts, we update viewr using LL/SC or compare-and-
swap (so that only the first scanner to write wins), and pretend that reads
of memory[i] by losers didn’t happen. When viewr is full, start a new virtual
scan and advance to the next round (and thus the next viewr+1 ).

19.5 Applications
Here we describe a few things we can do with snapshots.

19.5.1 Multi-writer registers from single-writer registers

One application of atomic snapshot is building multi-writer registers from
single-writer registers. The idea is straightforward: to perform a write, a
process does a snapshot to obtain the maximum sequence number, tags its
own value with this sequence number plus one, and then writes it. A read
consists of a snapshot followed by returning the value associated with the
largest sequence number (breaking ties by process id). (See [Lyn96, §13.5]
for a proof that this actually works.) This requires using a snapshot that
doesn’t use multi-writer registers, and turns out to be overkill in practice;
there are simpler algorithms that give O(n) cost for reads and writes based
on timestamps (see [AW04, 10.2.3]).
CHAPTER 19. ATOMIC SNAPSHOTS 175

With additional work, it is even possible to eliminate the requirement

of multi-reader registers, and get a simulation of multi-writer multi-reader
registers that goes all the way down to single-writer single-read registers, or
even single-writer single-reader bits. See [AW04, §§10.2.1–10.2.2] or [Lyn96,
§13.4] for details.

19.5.2 Counters and accumulators

Given atomic snapshots, it’s easy to build a counter (supporting increment,
decrement, and read operations); or, in more generality, an accumulator
(supporting increments by arbitrary amounts); or, in even more general-
ity, an object supporting any collection of commutative update operations
(as long as these operations don’t return anything). The idea is that each
process stores in its segment the total of all operations it has performed
so far, and a read operation is implemented using a snapshot followed by
summing the results. This is a case where it is reasonable to consider multi-
writer registers in building the snapshot implementation, because there is
not necessarily any circularity in doing so.

19.5.3 Resilient snapshot objects

The previous examples can be generalized to objects with operations that
either read the current state of the object but don’t update it or update the
state but return nothing, provided the update operations either overwrite
each other (so that Cxy = Cy or Cyx = Cx) or commute (so that Cxy =
Cyx).
This was shown by Aspnes and Herlihy [AH90b] and improved on by
Anderson and Moir [AM93] by eliminating unbounded space usage (this
paper also defined the terms snapshot objects for those with separate read
and update operations and resilience for the property that all operations
commute or overwrite). The basic idea underneath both of these papers is to
use the multi-writer register construction given above, but break ties among
operations with the same sequence numbers by first placing overwritten
operations before overwriting operations and only then using process ids.
This almost shows that snapshots can implement any object with con-
sensus number 1 where update operations return nothing, because an object
that violates the commute-or-overwrite condition in some configuration has
consensus number at least 2 (see §18.1.2). It doesn’t quite work (as observed
in the Anderson-Moir paper), because the tie-breaking procedure assumes
a static ordering on which operations overwrite each other, so that given
CHAPTER 19. ATOMIC SNAPSHOTS 176

operations x and y where y overwrites x, y overwrites x in any configura-

tion. But there may be objects with dynamic ordering, where y overwrites
x in some configuration, x overwrites y in another, and perhaps even the
two operations commute in yet another. This prevents us from achieving
consensus, but also breaks the tie-breaking technique. So it may be possible
that there are objects with consensus number 1 and no-return updates that
we still can’t implement using only registers.
Chapter 20

Lower bounds on
perturbable objects

Being able to do snapshots in linear time means that we can build lineariz-
able counters, generalized counters, max registers, etc. in linear time, by
having each reader take a snapshot and combine the contributions of each
updater using the appropriate commutative and associative operation. A
natural question is whether we can do better by exploiting the particular
features of these objects.
Unfortunately, the Jayanti-Tan-Toueg [JTT00] lower bound for per-
turbable objects says each of these objects requires n − 1 space and n − 1
steps for a read operation in the worst case, for any solo-terminating imple-
mentation from historyless objects.1
Here perturbable means that the object has a particular property that
makes the proof work, essentially that the outcome of certain special exe-
cutions can be changed by stuffing lots of extra update operations in the
middle (see below for details). Solo-terminating means that a process
finishes its current operation in a finite number of steps if no other process
takes steps in between; it is a much weaker condition, for example, than
wait-freedom. Historyless objects are those for which any operation that
changes the state overwrites all previous operations (i.e., those for which
covering arguments work, as long as the covering processes never report
back what they say). Atomic registers are the typical example, while swap
objects (with a swap operation that writes a new state while returning the
old state) are the canonical example since they can implement any other
1
A caveat is that we may be able to make almost all read operations cheaper, although
we won’t be able to do anything about the space bound. See Chapter 21.

177
CHAPTER 20. LOWER BOUNDS ON PERTURBABLE OBJECTS 178

historyless object (and even have consensus number 2, showing that even
extra consensus power doesn’t necessarily help here).
Below is a sketch of the proof. See the original paper [JTT00] for more
details.

• Build executions of the form Λk Σk Π, where Λk is a preamble consist-

ing of various complete update operations and k incomplete update
operations, Σk delivers k delayed writes from the incomplete opera-
tions in Λk , and Π is a read operation whose first k reads are from
registers written in Σk .

– Induction hypothesis is that such an execution exists for each

k ≤ n − 1.
– Base case is Λ0 Σ0 = hi, covering 0 reads by Π.

• Now we look for a sequence of operations γ that change what Π re-

turns in Λk γΣk Π (the object is perturbable if such a sequence always
exists).

– For a max register, let γ include a bigger write than all the others.
– For a counter, let γ include at least n increments. The same
works for a mod-m counter if m is at least 2n.
∗ Why n increments? With fewer increments, we can make
Π return the same value by being sneaky about when the
partial increments represented in Σk are linearized.
– In contrast, historyless objects (including atomic registers) are
not perturbable: if Σk includes a write that sets the value of the
object, no set of operations inserted before it will change this
value. (This is good, because we know that it only takes one
atomic register to implement an atomic register.)

• Such a γ must write to some register not covered in Σk .

• Find a γ 0 that writes to the first uncovered register that Π looks at (if
none exists, the reader is wasting a step), truncate before that write,
and prepend the write to Σk .

– In more detail: let γ 0 = αβδ, where β is the first write by γ 0

to the first register read by Π that is not covered by Σk . Let
Λk+1 = Λk α and Σk+1 = βΣk . So now Λk+1 Σk+1 Π = Λk αβΣk Π
and in particular Σk+1 covers the first k + 1 registers read by Π.
CHAPTER 20. LOWER BOUNDS ON PERTURBABLE OBJECTS 179

– Note: γ 0 might be much longer than γ (this will be important

later, when we want to get around the JTT lower bound).

• Repeat until we’ve covered n − 1 registers. This implies that there

are at least n − 1 registers, and in the worst case a reader reads all of
them.
Chapter 21

Restricted-use objects

Here we are describing work by Aspnes, Attiya, and Censor [AAC09], plus
some extensions by Aspnes et al. [AACHE12] and Aspnes and Censor-
Hillel [ACH13]. The idea is to place restrictions on the size of objects that
would otherwise be subject to the Jayanti-Tan-Toueg bound [JTT00] (see
Chapter 20), in order to get cheap implementations.
The central object that is considered in this work is a max register,
for which read operation returns the largest value previously written, as
opposed to the last value previously written. So after writes of 0, 3, 5, 2, 6,
11, 7, 1, 9, a read operation will return 11.
These are perturbable objects in the sense of the Jayanti-Tan-Toueg
bound, so in the worst case a max-register read will have to read at least
n−1 distinct atomic registers, giving an n−1 lower bound on both individual
work and space. But we can get around this by considering bounded max
registers (which only hold values in some range 0 . . . m − 1); these are not
perturbable because once one hits its upper bound we can no longer insert
new operations to change the value returned by a read.

21.1 Implementing bounded max registers

For m = 1, the implementation is trivial: write does nothing and read always
returns 0.
For larger m, we’ll show how to paste together two max registers left and
right with m0 and m1 values together to get a max register r with m0 + m1
values. We’ll think of each value stored in the max register as a bit-vector,
with bit-vectors ordered lexicographically. In addition to left and right, we
will need a 1-bit atomic register switch used to choose between them. The

180
CHAPTER 21. RESTRICTED-USE OBJECTS 181

read procedure is straightforward and is shown in Algorithm 21.1; essentially

we just look at switch, read the appropriate register, and prepend the value
of switch to what we get.

1 procedure read(r)
2 if switch = 0 then
3 return 0(read(left))
4 else
5 return 1(read(right))

Algorithm 21.1: Max register read operation

For write operations, we have two somewhat asymmetrical cases depend-

ing on whether the value we are writing starts with a 0 bit or a 1 bit. These
are shown in Algorithm 21.2.

1 procedure write(r, 0x)

2 if switch = 0 then
3 write(left, x)

4 procedure write(r, 1x)

5 write(right, x)
6 switch ← 1
Algorithm 21.2: Max register write operations

The intuition is that the max register is really a big tree of switch vari-
ables, and we store a particular bit-vector in the max register by setting to 1
the switches needed to make read follow the path corresponding to that bit-
vector. The procedure for writing 0x tests switch first, because once switch
gets set to 1, any 0x values are smaller than the largest value, and we don’t
want them getting written to left where they might confuse particularly slow
readers into returning a value we can’t linearize. The procedure for writing
1x sets switch second, because (a) it doesn’t need to test switch, since 1x
always beats 0x, and (b) it’s not safe to send a reader down into right until
some value has actually been written there.
It’s easy to see that read and write operations both require exactly
one operation per bit of the value read or written. To show that we get
linearizability, we give an explicit linearization ordering (see the paper for a
full proof that this works):
CHAPTER 21. RESTRICTED-USE OBJECTS 182

1. All operations that read 0 from switch go in the first pile.

(a) Within this pile, we sort operations using the linearization order-
ing for left.

2. All operations that read 1 from switch or write 1 to switch go in the

second pile, which is ordered after the first pile.

(a) Within this pile, operations that touch right are ordered using
the linearization ordering for right. Operations that don’t (which
are the “do nothing” writes for 0x values) are placed consistently
with the actual execution order.

To show that this gives a valid linearization, we have to argue first that
any read operation returns the largest earlier write argument and that we
don’t put any non-concurrent operations out of order.
For the first part, any read in the 0 pile returns 0read(left), and read(left)
returns (assuming left is a linearizable max register) the largest value pre-
viously written to left, which will be the largest value linearized before the
read, or the all-0 vector if there is no such value. In either case we are
happy. Any read in the 1 pile returns 1read(right). Here we have to guard
against the possibility of getting an all-0 vector if no write operations lin-
earize before the read. But any write operation that writes 1x doesn’t set
switch to 1 until after it writes to right, so no read operation ever starts
read(right) until after at least one write to right has completed, implying
that that write to right linearizes before the read from right. So in this case
as well all the second-pile operations linearize.

21.2 Encoding the set of values

If we structure our max register as a balanced tree of depth k, we are essen-
tially encoding the values 0 . . . 2k − 1 in binary, and the cost of performing a
read or write operation on an m-valued register is exactly k = dlg me. But
if we are willing to build an unbalanced tree, any prefix code will work.
The paper describes a method of building a max register where the cost
of each operation that writes or reads a value v is O(log v). The essential
idea is to build a tree consisting of a rightward path with increasingly large
left subtrees hanging off of it, where each of these left subtrees is twice as big
as the previous. This means that after following a path encoded as 1k 0, we
hit a 2k -valued max register. The value returned after reading some v 0 from
this max register is v 0 + (2k − 1), where the 2k − 1 term takes into account
CHAPTER 21. RESTRICTED-USE OBJECTS 183

all the values represented by earlier max registers in the chain. Formally,
this is equivalent to encoding values using an Elias gamma code, tweaked
slightly by changing the prefixes from 0k 1 to 1k 0 to get the ordering right.

21.3 Unbounded max registers

While the unbalanced-tree construction could be used to get an unbounded
max register, it is possible that read operations might not terminate (if
enough writes keep setting 1 bits on the right path before the read gets to
them) and for very large values the cost even of terminating reads becomes
higher than what we can get out of a snapshot.
Here is the snapshot-based method: if each process writes its own contri-
bution to the max register to a single-writer register, then we can read the
max register by taking a snapshot and returning the maximum value. (It is
not hard to show that this is linearizable.) This gives an unbounded max
register with read and write cost O(n). So by choosing this in preference
to the balanced tree when m is large, the cost of either operation on a max
register is min (dlg me , O(n)).
We can combine this with the unbalanced tree by terminating the right
path with a snapshot-based max register. This gives a cost for reads and
writes of values v of O(min(log v, n)).

21.4 Lower bound

The min(dlg me , n − 1) cost of a max register read turns out to be ex-
actly optimal. Intuitively, we can show by a covering argument that once
some process attempts to write to a particular atomic register, then any
subsequent writes convey no additional information (because they can be
overwritten by the first delayed write)—so in effect, no algorithm can use
get more than one bit of information out of each atomic register.
For the lower bound proof, we consider solo-terminating executions in
which n − 1 writers do any number of max-register writes in some initial
prefix Λ, followed by a single max-register read Π by process pn . Let T (m, n)
be the optimal reader cost for executions with this structure with m values,
and let r be the first register read by process pn , assuming it is running an
algorithm optimized for this class of executions (we do not even require it
to be correct for other executions).
We are now going split up our set of values based on which will cause
a write to write to r. Let Sk be the set of all sequences of writes that only
CHAPTER 21. RESTRICTED-USE OBJECTS 184

write values ≤ k. Let t be the smallest value such that some execution in
St writes to r (there must be some such t, or our reader can omit reading r,
which contradicts the assumption that it is optimal).

Case 1 Since t is smallest, no execution in St−1 writes to r. If we restrict

writes to values ≤ t − 1, we can omit reading r, giving T (t, n) ≤
T (m, n) − 1 or T (m, n) ≥ T (t, n) + 1.

Case 2 Let α be some execution in St that writes to r.

• Split α as α0 δβ where δ is the first write to r by some process pi .

• Construct a new execution α0 η by letting all the max-register
writes except the one performing δ finish.
• Now consider any execution α0 ηγδ, where γ is any sequence of
max-register writes with values ≥ t that excludes pi and pn . Then
pn always sees the same value in r following these executions, but
otherwise (starting after α0 η) we have an (n − 1)-process max-
register with values t through m − 1.
• Omit the read of r again to get T (m, n) ≥ T (m − t, n − 1) + 1.

We’ve shown the recurrence T (m, n) ≥ mint (max(T (t, n), T (m−t, n)))+
1, with base cases T (1, n) = 0 and T (m, 1) = 0. The solution to this recur-
rence is exactly min(dlg me , n − 1), with is the same, except for a constant
factor on n, as the upper bound we got by choosing between a balanced
tree for small m and a snapshot for m ≥ 2n−1 . For small m, the recursive
split we get is also the same as in the tree-based algorithm: call the r reg-
ister switch and you can extract a tree from whatever algorithm somebody
gives you. So this says that the tree-based algorithm is (up to choice of the
tree) essentially the unique optimal bounded max register implementation
for m ≤ 2n−1 .
It is also possible to show lower bounds on randomized implementations
of max registers and other restricted-use objects. See [AAC09, AACHH12]
for examples.

21.5 Max-register snapshots

With some tinkering, it’s possible to extend the max-register construction
to get an array of max registers that supports snapshots. The description
in this section follows [AACHE12].
CHAPTER 21. RESTRICTED-USE OBJECTS 185

Formally, a max array is an object a that supports an operation write(a, i, v)

that sets a[i] ← max(v, a[i]), and an operation read(a) that returns a snap-
shot of all components of the array. The first step in building this beast is to
do it for two components. The resulting 2-component max array can then
be used as a building block for larger max arrays and for fast restricted-used
snapshots in general.
A k × ` max array a is one that permits values in the range 0 . . . k − 1 in
a[0] and 0 . . . ` − 1 in a[1]. We think of a[0] as the head of the max array and
a[1] as the tail. We’ll show how to construct such an object recursively from
smaller objects of the same type, analogous to the construction of an m-
valued max register (which we can think of as a m × 1 max array). The idea
is to split head into two pieces left and right as before, while representing tail
as a master copy stored in a max register at the top of the tree plus cached
copies at every internal node. These cached copies are updated by readers
at times carefully chosen to ensure linearizability.
The base of the construction is an `-valued max register r, used directly
as a 1×` max array; this is the case where the head component is trivial and
we only need to store a.tail = r. Here calling write(a, 0, v) does nothing,
while write(a, 1, v) maps to write(r, v), and read(a) returns h0, read(r)i.
For larger values of k, paste a k1 × ` max array left and a k2 × ` max
array right together to get a (k1 + k2 ) × ` max array. This construction uses
a switch variable as in the basic construction, along with an `-valued max
register tail that is used to store the value of a[1]. A call to write(a, 1, v)
operation writes tail directly, while write(a, 0, v) and read(a) follow the
structure of the corresponding operations for a simple max register, with
some extra work in read to make sure that the value in tail propagates into
left and right as needed to ensure the correct value is returned.
Pseudocode is given in Algorithm 21.3.
The individual step complexity of each operation is easily computed.
Assuming a balanced tree, write(a, 0, v) takes exactly lg k steps, while
write(a, 1, v) costs exactly lg ` steps; in both cases the cost is identical
to that of a max-register write. Read operations are more complicated. In
the worst case, we have two reads of a.tail and a write to a.right[1] at each
level, plus up to two operations on a.switch, for a total cost of at most
(3 lg k − 1)(lg ` + 2) = O(log k log `) steps.
In the special case where k = `, we get that writes cost the same number
of steps as in a single-component k-valued max register while the cost of
reads is squared.
CHAPTER 21. RESTRICTED-USE OBJECTS 186

1 procedure write(a, i, v)
2 if i = 0 then
3 if v < k1 then
4 if a.switch = 0 then
5 write(a.left, 0, v)
6 else
7 write(a.right, 0, v − k1 )
8 a.switch ← 1
9 else
10 write(a.tail, v)

11 procedure read(a)
12 x ← read(a.tail)
13 if a.switch = 0 then
14 write(a.left, 1, x)
15 return read(a.left)
16 else
17 x ← read(a.tail)
18 write(a.right, 1, x)
19 return hk1 , 0i + read(a.right)

Algorithm 21.3: Recursive construction of a 2-component max array

CHAPTER 21. RESTRICTED-USE OBJECTS 187

21.5.1 Linearizability
In broad outline, the proof of linearizability follows the proof for a simple
max register. But as with snapshots, we have to show that the ordering of
the head and tail components are consistent.
The key observation is the following lemma.

Lemma 21.5.1. Fix some execution of a max array a implemented as in

Algorithm 21.3. Suppose this execution contains a read(a) operation πleft
that returns vleft from a.left and a read(a) operation πright that returns vright
from a.right. Then vleft [1] ≤ vright [1].

Proof. Both vleft [1] and vright [1] are values that were previously written to
their respective max arrays by read(a) operations (such writes necessar-
ily exist because any process that reads a.left or a.right writes a.left[1] or
a.right[1] first). From examining the code, we have that any value written
to a.left[1] was read from a.tail before a.switch was set to 1, while any value
written to a.right[1] was read from a.tail after a.switch was set to 1. Since
max-register reads are non-decreasing, we have than any value written to
a.left[1] is less than or equal to any value written to a.right[1], proving the
claim.

The rest of the proof is tedious but straightforward: we linearize the

read(a) and write(a[0]) operations as in the max-register proof, then fit
the write(a[1]) operations in based on the tail values of the reads. The full
result is:

Theorem 21.5.2. If a.left and a.right are linearizable max arrays, and a.tail
is a linearizable max register, then Algorithm 21.3 implements a linearizable
max array.

It’s worth noting that the same unbalanced-tree construction used in

§§21.2 and 21.3 can be used here as well; this gives a cost of O(min(log v, n))
for writes and O(min(log v[0], n) · min(log v[1], n)) for reads, where v is the
value written or read.

21.5.2 Application to standard snapshots

To build an ordinary snapshot object from 2-component max arrays, we
construct a balanced binary tree in which each leaves holds a pointer to
an individual snapshot element and each internal node holds a pointer to
a partial snapshot containing all of the elements in the subtree of which
CHAPTER 21. RESTRICTED-USE OBJECTS 188

it is the root. The pointers themselves are non-decreasing indices into ar-
rays of values that consist of ordinary (although possibly very wide) atomic
registers.
When a process writes a new value to its component of the snapshot
object, it increases the pointer value in its leaf and then propagates the
new value up the tree by combining together partial snapshots at each step,
using 2-component max arrays to ensure linearizability. The resulting algo-
rithm is similar in many ways to the lattice agreement procedure of Inoue et
al. [IMCT94] (see §19.3.5), except that it uses a more contention-tolerant
snapshot algorithm than double collects and we allow processes to update
their values more than once. It is also similar to some constructions of
Jayanti [Jay02] for efficient computation of array aggregates (sum, min,
max, etc.) using LL/SC, the main difference being that because the index
values are non-decreasing, max arrays can substitute for LL/SC.
Each node in the tree except the root is represented by one component
of a 2-component max array that we can think of as being owned by its
parent, with the other component being the node’s sibling in the tree. To
propagate a value up the tree, at each level the process takes a snapshot
of the two children of the node and writes the sum of the indices to the
node’s component in its parent’s max array (or to an ordinary max register
if we are at the root). Before doing this last write, a process will combine
the partial snapshots from the two child nodes and write the result into
a separate array indexed by the sum. In this way any process that reads
the node’s component can obtain the corresponding partial snapshot in a
single register operation. At the root this means that the cost of obtaining
a complete snapshot is dominated by the cost of the max-register read, at
O(log v), where v is the number of updates ever performed.
A picture of this structure, adapted from [AACHE12], appears in Fig-
ure 21.1. The figure depicts an update in progress, with red values being the
new values written as part of the update. Only some of the tables associated
with the nodes are shown.
The cost of an update is dominated by the O(log n) max-array operations
needed to propagate the new value to the root. This takes O(log2 v log n)
steps.
The linearizability proof is trivial: linearize each update by the time at
which a snapshot containing its value is written to the root (which neces-
sarily occurs within the interval of the update, since we don’t let an update
finish until it has propagated its value to the top), and linearize reads by
when they read the root. This immediately gives us an O(log3 n) implemen-
tation—as long as we only want to use it polynomially many times—of any-
CHAPTER 21. RESTRICTED-USE OBJECTS 189

cms
5 cmr
bmr
br
ar
0 5 a

ms
0 0 3 3 mr
m

c
b 1 2
a

s
m r

Figure 21.1: Snapshot from max arrays [AACHE12]

thing we can build from snapshot, including counters, generalized counters,

and (by [AH90b, AM93]) any other object whose operations all commute
with or overwrite each other in a static pattern.
Randomization can eliminate the need to limit the number of times the
snapshot is used. The JTT bound still applies, so there will be occasional
expensive operations, but we can spread these out randomly so that any
particular operation has low expected cost. This gives a cost of O(log3 n)
expected steps for an unrestricted snapshot. See [ACH13] for details.
In some cases, further improvements are possible. The original max-
registers paper [AAC09] gives an implementation of counters using a similar
tree construction with only max registers that costs only O(log2 n) for incre-
ments; here the trick is to observe that counters that can only be incremented
by one are much easier to make linearizable, because there is no possibil-
ity of seeing an intermediate value that couldn’t be present in a sequential
execution.
Chapter 22

Common2

The common2 class, defined by Afek, Weisberger, and Weisman [AWW93]

consists of all read-modify-write objects where the modify functions either
(a) all commute with each other or (b) all overwrite each other. We can
think of it as the union of two simpler classes, the set of read-modify-
write objects where all update operations commute, called commuting
objects [AW99]; and the set of read-modify-write objects where all up-
dates produce a value that doesn’t depend on the previous state, called
historyless objects [FHS98]).
From §18.1.2, we know that both commuting objects and historyless
objects have consensus number at most 2, and that these objects have con-
sensus number exactly 2 provided they supply at least one non-trivial update
operation. The main result of Afek et al. [AWW93] is that commuting and
historyless objects can all be implemented from any object with consen-
sus number 2, even in systems with more than 2 processes. This gives a
completeness result analogous to completeness results in complexity the-
ory: any non-trivial common2 object can be used to implement any other
common2 object.
The main result in the paper has two parts, reflecting the two parts of
the common2 class: a proof that 2-process consensus plus registers is enough
to implement all commuting objects (which essentially comes down to build
a generalized fetch-and-add that returns an unordered list of all preceding
operations); and a proof that 2-process consensus plus registers is enough
to implement all overwriting objects (which is equivalent to showing that
we can implement swap objects). The construction of the generalized fetch-
and-add is pretty nasty, so we’ll concentrate on the implementation of swap
objects, limiting ourselves specifically to construction of single-use swap. For

190
CHAPTER 22. COMMON2 191

the remaining results, you’ll have to go to the paper itself [AWW93].

[[[ We may want to replace the swap part here. The swap
implementation from the Afek et al. paper from PODC 2011 is
cleaner. See also Guy Wertheim’s Master’s thesis. ]]]

22.1 Test-and-set and swap for two processes

The first step is to get test-and-set.
Algorithm 22.1 shows how to turn 2-process consensus into 2-process
test-and-set. The idea is that whoever wins the consensus protocol wins the
test-and-set. This is linearizable, because if I run TAS2 before you do, I win
the consensus protocol by validity.

1 procedure TAS2()
2 if Consensus2(myId) = myId then
3 return 0
4 else
5 return 1

Algorithm 22.1: Building 2-process TAS from 2-process consensus

Once we have test-and-set for two processes, we can easily get one-shot
swap for two processes. The trick is that a one-shot swap object always
returns ⊥ to the first process to access it and returns the other process’s value
to the second process. We can distinguish these two roles using test-and-set
and add a register to send the value across. Pseudocode is in Algorithm 22.2.

1 procedure swap(v)
2 a[myId] = v
3 if TAS2() = 0 then
4 return ⊥
5 else
6 return a[¬myId]

Algorithm 22.2: Two-process one-shot swap from TAS

CHAPTER 22. COMMON2 192

22.2 Building n-process TAS from 2-process TAS

To turn the TAS2 into full-blown n-process TAS, start by staging a tourna-
ment along the lines of [PF77] (§17.4.1.2). Each process walks up a tree of
nodes, and at each node it attempts to beat every process from the other
subtree using a TAS2 object (we can’t just have it fight one process, because
we don’t know which other process will have won the other subtree). A
process drops out if it ever sees a 1. We can easily show that at most one
process leaves each subtree with all zeros, including the whole tree itself.
Unfortunately, this process does not give a linearizable test-and-set ob-
ject. It is possible that p1 loses early to p2 , but then p3 starts (elsewhere
in the tree) after p1 finishes, and races to the top, beating out p2 . To avoid
this, we can follow [AWW93] and add a gate bit that locks out latecomers.1
The resulting construction looks something like Algorithm 22.3. This
gives a slightly different interface that straight TAS; instead of returning 0
for winning and 1 for losing, the algorithm returns ⊥ for winning and the id
of some process that beats you for losing. It’s not hard to see that this gives
a linearizable test-and-set after translating the values back to 0 and 1 (the
trick for linearizability is that any process that wins saw an empty gate, and
so started before any other process finished). It also sorts the processes into
a rooted tree, with each process linearizing after its parent (this latter claim
is a little trickier, but basically comes down to a loser linearizing after the
process that defeated it either on gate or on one of the TAS2 objects).

22.3 Single-use swap objects

Here we’ll show how to implement a single-use swap object, where each
process is only allowed to execute a single swap operation. The essential idea
is to explicitly string the processes into a sequence, where each process learns
the identity of the process ahead of it. This sequence gives the linearization
order and allows processes to compute their return values by reading the
input stored by their predecessor.
The algorithm proceeds in asynchronous rounds, with the participants
of each round organized into a tree using the compete procedure from Algo-
rithm 22.3. The winner at round k will attempt to thread itself behind some
process at round k − 1, starting with the process it lost to at that round (or
nobody if k = 1). In order for this to work, the round k − 1 process must be
1
The original version of this trick is from an earlier paper [AGTV92], where the gate
bit is implemented as an array of single-writer registers.
CHAPTER 22. COMMON2 193

1 procedure compete(i)
// check the gate
2 if gate 6= ⊥ then
3 return gate
4 gate ← i
// Do tournament, returning id of whoever I lose to
5 node ← leaf for i
6 while node 6= root do
7 for each j whose leaf is below sibling of node do
8 if TAS2(t[i, j]) = 1 then
9 return j

10 node ← node.parent
// I win!
11 return ⊥
Algorithm 22.3: Tournament algorithm with gate

locked down to round k − 1 (and thread itself behind some other process at
round k−1); this is done using a “trap” object implemented with a 2-process
swap. If the target process escapes by calling the trap object first, it will
leave behind the id of the process it lost to at round k − 1, allowing the
round-k winner to try again. If the round-k winner fails to trap anybody, it
will eventually thread itself behind the round-(k − 1) winner, who is stuck
at round k − 1.
Only those processes that are ancestors of the process that beat the
round-k winner may get trapped in round k − 1; everybody else will escape
and try again in a later round.
Pseudocode for the trap object is given in Algorithm 22.4. There are two
operations. The pass operation is called by the process trying to escape;
if it executes first, this process successfully escapes, but leaves behind the
identity of somebody else to try. The block operation locks the target down
so that pass fails. The shared data for a trap t consists of a two-process
swap object t[i, j] for each process i trying to block a process j. A utility
procedure passAll is included that calls pass on all potential blockers until
it fails.
It is not hard to see from the code that Algorithm 22.4 has the desired
properties: if the passer reaches the swap object first, it is not blocked but
leaves behind its value v for the passer; while if the blocker reaches the
CHAPTER 22. COMMON2 194

1 procedure block(t, j)
2 return swap(t[j, i], blocked)

3 procedure pass(t, j, v)
4 if swap(t[i, j], v) = blocked then
5 return false
6 else
7 return true

8 procedure passAll(t, v)
9 for j ← 1 to n do
10 if ¬pass(t, j, v) then return false
11

12 return true
Algorithm 22.4: Trap implementation from [AWW93]

object first, it obtains no value but successfully blocks the passer.

The full swap construction is given in Algorithm 22.5.
This just implements the blocking strategy described before, with the
main swap procedure implementing the round structure and the findValue
helper procedure implementing the walk up the tree.
It’s not hard to see that when all processes finish the protocol, they will
be neatly arranged in a chain with each process except the first obtaining
the value of its predecessor. Slightly less easy to see is that this ordering
will be consistent with the observed execution order, which is necessary for
linearizability. For a proof of linearizability, see the paper.
CHAPTER 22. COMMON2 195

1 procedure swap (v)

2 input[i] ← v
3 for k ← 1 to n do
// find our first target
4 t ← compete(tournament[k])
5 if t = ⊥ then
// I am the round-k winner
6 return findValue(k − 1, t0 )
7 else if ¬passAll(trap[k], t) do
// I did not escape
8 return findValue(k, t)
9 else
// I escaped, remember who I lost to
10 t0 ← t

11 procedure findValue(k, t)
12 if k = 0 then
13 return ⊥
14 else
15 repeat
16 x ← block(trap[k], t)
17 if x 6= ⊥ then t ← x
18 until x = ⊥
19 return input[t]

Algorithm 22.5: Single-use swap from [AWW93]

Chapter 23

Randomized consensus and

test-and-set

We’ve seen that we can’t solve consensus in an asynchronous system with

one crash failure [FLP85, LAA87], but that the problem becomes solvable
using failure detectors [CT96]. An alternative that also allows us to solve
consensus is to allow the processes to use randomization, by providing each
process with a local coin that can generate random values that are imme-
diately visible only to that process. The resulting randomized consensus
problem replaces the termination requirement with probabilistic ter-
mination: all processes terminate with probability 1. The agreement and
validity requirements remain the same.
In this chapter, we will describe how randomization interacts with the
adversary, give a bit of history of randomized consensus, and then concen-
trate on recent algorithms for randomized consensus and the closely-related
problem of randomized test-and-set. Much of the material in this chapter is
adapted from notes for a previous course on randomized algorithms [Asp11]
and my own recent papers [Asp12b, AE11, Asp12a].

23.1 Role of the adversary in randomized algo-

rithms
Because randomized processes are unpredictable, we need to become a little
more sophisticated in our handling of the adversary. As in previous asyn-
chronous protocols, we assume that the adversary has control over timing,
which we model by allowing the adversary to choose at each step which
process performs the next operation. But now the adversary may do so

196
CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET197

based on knowledge of the state of the protocol and its past evolution. How
much knowledge we give the adversary affects its power. Several classes of
adversaries have been considered in the literature; ranging from strongest
to weakest, we have:
1. An adaptive adversary. This adversary is a function from the state
of the system to the set of processes; it can see everything that has
happened so far (including coin-flips internal to processes that have not
yet been revealed to anybody else), but can’t predict the future. It’s
known that an adaptive adversary can force any randomized consensus
protocol to take Θ(n2 ) total steps [AC08]. The adaptive adversary is
also called a strong adversary following a foundational paper of
Abrahamson [Abr88].
2. An intermediate adversary or weak adversary [Abr88] is one
that limits the adversary’s ability to observe or control the system in
some way, without completely eliminating it. For example, a content-
oblivious adversary [Cha96] or value-oblivious adversary [Aum97]
is restricted from seeing the values contained in registers or pending
write operations and from observing the internal states of processes
directly. A location-oblivious adversary [Asp12b] can distinguish
between values and the types of pending operations, but can’t discrim-
inate between pending operations based one which register they are
operating on. These classes of adversaries are modeled by imposing
an equivalence relation on partial executions and insisting that the
adversary make the same choice of processes to go next in equivalent
situations. Typically they arise because somebody invented a consen-
sus protocol for the oblivious adversary below, and then looked for the
next most powerful adversary that still let the protocol work.
Weak adversaries often allow much faster consensus protocols than
adaptive adversaries. Each of the above adversaries permits consensus
to be achieved in O(log n) expected individual work using an appropri-
ate algorithm. But from a mathematical standpoint, weak adversaries
are a bit messy, and once you start combining algorithms designed for
different weak adversaries, it’s natural to move all the way down to
the weakest reasonable adversary, the oblivious adversary described
below.
3. A oblivious adversary has no ability to observe the system at all;
instead, it fixes a sequence of process ids in advance, and at each step
the next process in the sequence runs.
CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET198

We will describe below a protocol that guarantees O(log log n) ex-

pected individual work for an oblivious adversary. It is not known
whether this is optimal; in fact, is is consistent with the best known
lower bound (due to Attiya and Censor [AC08]) that consensus can be
solved in O(1) expected individual steps against an oblivious adver-
sary.

23.2 History
The use of randomization to solve consensus in an asynchronous system
with crash failures was proposed by Ben-Or et al.Ben-Or1983 for a message-
passing model. Chor, Israeli, and Li [CIL94] gave the first wait-free consen-
sus protocol for a shared-memory system, which assumed a particular kind
of weak adversary. Abrahamson [Abr88] defined strong and weak adver-
saries and gave the first wait-free consensus
 2  protocol for a strong adversary;
its expected step complexity was Θ 2n . After failing to show that expo-
nential time was necessary, Aspnes and Herlihy [AH90a] showed how to do
consensus in O(n4 ) total work, a value that was soon reduced to O(n2 log n)
by Bracha and Rachman [BR91]. This remained the best known bound for
the strong-adversary model until Attiya and Censor [AC08] showed match-
ing Θ(n2 ) upper and lower bounds for the problem; subsequent work [AC09]
showed that it was also possible to get an O(n) bound on individual work.
For weak adversaries, the best known upper bound on individual step
complexity was O(log n) for a long time [Cha96, Aum97, Asp12b], with
an O(n) bound on total step complexity for some models [Asp12b]. More
recent work has lowered the bound to O(log log n), under the assumption of
an oblivious adversary [Asp12a]. No non-trivial lower bound on expected
individual step complexity is known, although there is a known lower bound
on the distribution of of the individual step complexity [ACH10].

23.3 Reduction to simpler primitives

To show how to solve consensus using randomization, it helps to split the
problem in two: we will first see how to detect when we’ve achieved agree-
ment, and then look at how to achive agreement.
CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET199

23.3.1 Adopt-commit objects

Most known randomized consensus protocols have a round-based struc-
ture where we alternative between generating and detecting agreement.
Gafni [Gaf98] proposed adopt-commit protocols as a tool for detect-
ing agreement, and these protocols were later abstracted as adopt-commit
objects [MRRT08, AGGT09]. The version described here is largely taken
from [AE11], which shows bounds on the complexity of adopt-commit ob-
jects.
An adopt-commit object supports a single operation, AdoptCommit (u),
where u is an input from a set of m values. The result of this operation is an
output of the form (commit, v) or (adopt, v), where the second component
is a value from this set and the first component is a decision bit that
indicates whether the process should decide value v immediately or adopt it
as its preferred value in later rounds of the protocol.
The requirements for an adopt-commit object are the usual requirements
of validity and termination, plus:

1. Coherence. If the output of some operation is (commit, v), then every

output is either (adopt, v) or (commit, v).

2. Convergence. If all inputs are v, all outputs are (commit, v).

These last two requirement replace the agreement property of consensus.

They are also strictly weaker than consensus, which means that a consensus
object (with all its output labeled commit) is also an adopt-commit object.
The reason we like adopt-commit objects is that they allow the simple
consensus protocol shown in Algorithm 23.1.

1 preference ← input
2 for r ← 1 . . . ∞ do
3 (b, preference) ← AdoptCommit(AC[r], preference)
4 if b = commit then
5 return preference
6 else
7 do something to generate a new preference

Algorithm 23.1: Consensus using adopt-commit

The idea is that the adopt-commit takes care of ensuring that once some-
body returns a value (after receiving commit), everybody else who doesn’t
CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET200

return adopts the same value (follows from coherence). Conversely, if ev-
erybody already has the same value, everybody returns it (follows from
convergence). The only missing piece is the part where we try to shake all
the processes into agreement. For this we need a separate object called a
conciliator.

23.3.2 Conciliators
Conciliators are a weakened version of randomized consensus that replace
agreement with probabilistic agreement: it’s OK if the processes disagree
sometimes as long as they agree with constant probability despite interfer-
ence by the adversary. An algorithm that satisfies termination, validity, and
probabilistic agreement is called a conciliator.1
The important feature of conciliators is that if we plug a conciliator that
guarantees agreement with probability at least δ into Algorithm 23.1, then
on average we only have to execute the loop 1/δ times before every process
agrees. This gives an expected cost equal to 1/δ times the total cost of
AdoptCommit and the conciliator. Typically we will aim for constant δ.

23.4 Implementing an adopt-commit object

What’s nice about adopt-commit objects is that they can be implemented
deterministically. Here we’ll give a simple adopt-commit object for two
values, 0 and 1. Optimal (under certain assumptions) constructions of m-
valued adopt-commits can be found in [AE11].
Pseudocode is given in Algorithm 23.2.
Structurally, this is pretty similar to a splitter (see §17.4.2, except that
we use values instead of process ids.
We now show correctness. Termination and validity are trivial. For
coherence, observe that if I return (commit, v) I must have read a[¬v] = false
before any process with ¬v writes a[¬v]; it follows that all such processes
will see proposal 6= ⊥ and return (adopt, v). For convergence, observe that
if all processes have the same input v, they all write it to proposal and all
observe a[¬v] = false, causing them all to return (commit, v).
1
Warning: This name has not really caught on in the general theory-of-distributed-
computing community, and so far only appears in papers that have a particular researcher
as a co-author [Asp12a, AE11, Asp12b]. Unfortunately, there doesn’t seem to be a better
name for the same object that has caught on. So we are stuck with it for now.
CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET201

shared data: a[0], a[1], initially false; proposal, initially ⊥

1 procedure AdoptCommit(v)
2 a[v] ← 1
3 if proposal = ⊥ then
4 proposal ← v
5 else
6 v ← proposal
7 if a[¬v] = false then
8 return (commit, v)
9 else
10 return (adopt, v)

Algorithm 23.2: A 2-valued adopt-commit object

23.5 A one-register conciliator for an oblivious ad-

versary

shared data: register r, initially ⊥

1 k←0
2 while r = ⊥ do
2k
3 with probability 2n do
4 write v to r
5 else
6 do a dummy operation
7 k ←k+1
8 return r
Algorithm 23.3: Impatient first-mover conciliator from [Asp12b]

Algorithm 23.3 implements a conciliator using a single register; it works

against an oblivious adversary.2 This particular construction is taken from [Asp12b],
and is based on an earlier algorithm of Chor, Israeli, and Li [CIL94]. The
cost of this algorithm is expected O(n) total work and O(log n) individual
work. It’s not known whether it is possible to improve on this bound.
The basic idea is that processes alternate between reading a register r
2
Or any adversary dumb enough not to be able to block the write based on how the
coin-flip turned out.
CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET202

and (maybe) writing to the register; if a process reads a non-null value from
the register, it returns it. Any other process that reads the same non-null
value will agree with the first process; the only way that this can’t happen
is if some process writes a different value to the register before it notices the
first write.
The random choice of whether to write the register or not avoids this
problem. The idea is that even though the adversary can schedule a write
at a particular time, because it’s oblivious, it won’t be able to tell if the
process wrote (or was about to write) or did a no-op instead.
The basic version of this algorithm, due to Chor, Israeli, and Li [CIL94],
1
uses a fixed 2n probability of writing to the register. So once some process
writes to the register, the chance that any of the remaining n − 1 processes
write to it before noticing that it’s non-null is at most n−12n < 1/2. It’s also
not hard to see that this algorithm uses O(n) total operations, although it
may be that one single process running by itself has to go through the loop
2n times before it finally writes the register and escapes.
Using increasing probabilities avoids this problem, because any process
that executes the main loop dlg ne + 1 times will write the register. This
establishes the O(log n) per-process bound on operations. At the same time,
an O(n) bound on total operations still holds, since each write has at least
1
a 2n chance of succeeding. The price we pay for the improvement is that
we increase the chance that an initial value written to the register gets
overwritten by some high-probability write. But the intuition is that the
probabilities can’t grow too much, because the probability that I write on
my next write is close to the sum of the probabilities that I wrote on my
previous writes—suggesting that if I have a high probability of writing next
time, I should have done a write already.
Formalizing this intuition requires a little bit of work. Fix the schedule,
and let pi be the probability that the i-th write operation in this schedule
succeeds. Let t be the least value for which ti=1 pi ≥ 1/4. We’re going to
P

argue that with constant probability one of the first t writes succeeds, and
that the next n − 1 writes by different processes all fail.
The probability that none of the first t writes succeed is
t t
e−pi
Y Y
(1 − pi ) ≤
i=1 i=1
t
!
X
= exp pi
i=1
≤ e−1/4 .
CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET203

Now observe that if some process q writes at or before the t-th write,
then any process with a pending write either did no writes previously, or its
last write was among the first t − 1 writes, whose probabilities sum to less
1
than 1/4. In the first case, the process has a 2n chance of writing on its
P 1
next attempt. In the second, it has a i∈Sq pi + 2n chance of writing on its
next attempt, where Sq is the set of indices in 1 . . . t − 1 where q attempts
to write.
Summing up these probabilities over all processes gives a total of n−1
2n +
−1/4
i∈Sq pi ≤ 1/2+1/4 = 3/4. So with probabililty at least e
P P
q (1−3/4) =
e −1/4 /4, we get agreement.

23.6 Sifters
[[[ This is at least two papers out of date: there is the faster
sifter of Giakkoupis and Woelfel, and the O(log n)-space TAS of
Giakkoupis, Helmi, Higham, and Woelfel from STOC 2015. ]]]
A faster conciliator can be obtained using a sifter, which is a mechanism
for rapidly discarding processes using randomization [AA11] while keeping
at least one process around. The idea of a sifter is to have each process either
write a register (with low probability) or read it (with high probability); all
writers and all readers that see ⊥ continue to the next stage of the protocol,
while all readers who see a non-null value drop out. An appropriately-
√
tuned sifter will reduce n processes to at most 2 n processes on average; by
iterating this mechanism, the expected number of remaining processes can
be reduced to 1 +  after O(log log n + log(1/)) phases.
As with previous implementations of test-and-set (see Algorithm 22.3),
it’s often helpful to have a sifter return not only that a process lost but which
process it lost to. This gives the implementation shown in Algorithm 23.4.

1 procedure sifter(p, r)
2 with probability p do
3 r ← id
4 return ⊥
5 else
6 return r

Algorithm 23.4: A sifter

To use a sifter effectively, p should be tuned to match the number of

CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET204

processes that are likely to use it. This is because of the following lemma:

Lemma 23.6.1. Fix p, and let X processes executed a sifter with parameter
p. Let Y be the number of processes for which the sifter returns ⊥. Then
1
E [X | Y ] ≤ pX + . (23.6.1)
p

Proof. In order to return ⊥, a process must either (a) write to r, which

occurs with probability p, or (b) read r before any other process writes to
it. The expected number of writers, conditioned on X, is exactly pX. The
expected number of readers before the first write has a geometric distribution
truncated by X. Removing the truncation gives exactly p1 expected readers,
which is an upper bound on the correct value.

For n initial processes, the choice of p that minimizes the bound in

√
(23.6.1) is √1n , giving at most 2 n expected survivors. Iterating this process
√ q √
with optimal p at each step gives a sequence of at most n, 2 n, 2 2 n,
etc., expected survivors after each sifter. The twos are a little annoying, but
a straightforward induction bounds the expected survivors after i rounds by
−i
4 · n2 . In particular, we get at most 8 expected survivors after dlg lg ne
rounds.
At this point it makes sense to switch to a fixed p and a different anal-
ysis. For p = 1/2, the first process to access r always survives, and each
subsequent process survives with probability at most 3/4 (because it leaves
if the first process writes and it reads).
l So themnumber of “excess” processes
i
drops as (3/4) , and an additional log4/3 (7/) rounds are enough to reduce
the expected number of survivors from 1 + 7 to 1 +  for any fixed .3
It follows that

Theorem 23.6.2. An initial set of n processes can be reduced to 1 with

probability at least 1 −  using O(log log n + log(1/)) rounds of sifters.
l m
Proof. Let X be the number of survivors after dlg lg ne+ log4/3 (7/) rounds
of sifters, with probabilities tuned as described above. We’ve shown that
E [X] ≤ 1 + , so E [X − 1] ≤ . Since X − 1 ≥ 0, from Markov’s inequality
we have Pr [X ≥ 2] = Pr [X − 1 ≥ 1] ≤ E [X − 1] /1 ≤ .
3
This argument essentially follows the proof of [Asp12a, Theorem 2], which, because
of neglecting to subtract off a 1 at one point, ends up with 8/ instead of 7/.
CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET205

23.6.1 Test-and-set using sifters

Sifters were initially designed to be used for test-and-set. For this purpose,
we treat a return value of ⊥ as “keep going” and anything else as “leave
with value 1.” Using O(log log n) rounds of sifters, we can get down to one
process that hasn’t left with probability at least 1 − log−c n for any fixed
constant c. We then need a fall-back TAS to handle the log−c n chance that
we get more than one such survivor.
Alistarh and Aspnes [AA11] used the RatRace algorithm of Alistarh et
al. [AAG+ 10] for this purpose. This is an adaptive randomized test-and-set
built from splitters and two-process consensus objects that runs in O(log k)
expected time, where k is the number of processes that access the test-and-
set; a sketch of this algorithm is given in §24.5.2. If we want to avoid appeal-
ing to this algorithm, a somewhat simpler approach is to use an approach
similar to the Lamport’s fast-path mutual exclusion algorithm (described in
§17.4.2): any process that survives the splitters tries to rush to a two-process
TAS at the top of a tree of two-processes TASes by winning a splitter, and
if it doesn’t win the splitter, it enters at a leaf and pays O(log n) expected
steps. By setting  = 1/ log n, the overall expected cost of this final stage is
O(1).
This algorithm does not guarantee linearizability. I might lose a sifter
early on only to have a later process win all the sifters (say, by writing to
each one) and return 0. A gate bit as in Algorithm 22.3 solves this problem.
The full code is given in Algorithm 23.5.

23.6.2 Consensus using sifters

With some trickery, the sifter mechanism can be adapted to solve consensus,
still in O(log log n) expected individual work [Asp12a]. The main difficulty
is that a process can no longer drop out as soon as it knows that it lost:
it still needs to figure out who won, and possible help that winner over the
finish line.
The basic idea is that when a process p loses a sifter to some other process
q, p will act like a clone of q from that point on. In order to make this work,
each process writes down at the start of the protocol all of the coin-flips it
intends to use to decide whether to read or write at each round of sifting.
Together with its input, these coin-flips make up the process’s persona.
In analyzing the progress of the sifter, we count surviving personae (with
multiple copies of the same persona counting as one) instead of surviving
processes.
CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET206

1 if gate 6= ⊥ then
2 return 1
3 else
4 gate ← myId
l m
5 for i ← 1 . . . dlog log ne + log4/3 (7 log n) do
 −i+1

6 with probability min 1/2, 21−2 do
7 ri ← myId
8 else
9 w ← ri
10 if w 6= ⊥ then
11 return 1

12 if splitter() = stop then

13 return 0
14 else
15 return AWWTAS()
Algorithm 23.5: Test-and-set in O(log log n) expected time

Pseudocode for this algorithm is given in Algorithm 23.6. Note that the
loop body is essentially the same as the code in Algorithm 23.4, except that
the random choice is replaced by a lookup in persona.chooseWrite.
To show that this works, we need to argue that having multiple copies
of a persona around doesn’t change the behavior of the sifter. In each
round, we will call the first process with a given persona p to access ri
the representative of p, and argue that a persona survives round i in
this algorithm precisely when its representative would survive round i in
a corresponding test-and-set sifter with the schedule restricted only to the
representatives.
There are three cases:
1. The representative of p writes. Then at least one copy of p survives.
2. The representative of p reads a null value. Again at least one copy of
p survives.
3. The representative of p reads a non-null value. Then no copy of p
survives: all subsequent reads by processes carrying p also read a non-
null value and discard p, and since no process with p writes, no other
process adopts p.
CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET207

1 procedure conciliator(input)
l m
2 Let R = dlog log ne + log4/3 (7/)
3 Let chooseWrite be a vector of R independent random Boolean
variables with Pr[chooseWrite[i] = 1] = pi , where
−i+1 −i
pi = 21−2 (n)−2 for i ≤ dlog log ne and pi = 1/2 for larger i.
4 persona ← hinput, chooseWrite, myIdi
5 for i ← 1 . . . R do
6 if persona.chooseWrite[i] = 1 then
7 ri ← persona
8 else
9 v ← ri
10 if v 6= ⊥ then
11 persona ← v

12 return persona.input
Algorithm 23.6: Sifting conciliator (from [Asp12a])

From the preceding analysis for test-and-set, we have that after O(log log n+
log 1/) rounds with appropriate probabilities of writing, at most 1+ values
survive on average. This gives a probability of at most  of disagreement. By
alternating these conciliators with adopt-commit objects, we get agreement
in O(log log n + log m/ log log m) expected time, where m is the number of
possible input values.
I don’t think the O(log log n) part of this expression is optimal, but I
don’t know how to do better.

23.7 O(log∗ n) Randomized test-and-set

A more sophisticated sifter due to Giakkoupis and Woelfel [GW12a] removes
all but O(log n) processes, on average, using two operations for each pro-
cess. Iterating this sifter reduces the expected survivors to O(1) in O(log∗ n)
rounds. A particularly nice feature of the Giakkoupis-Woelfel algorithm is
that (if you don’t care about space) it doesn’t have any parameters that
require tuning to n: this means that exactly the same structure can be used
in each round. An unfortunate feature is that it’s not possible to guaran-
tee that every process that leaves learns the identity of a process that stays:
this means that it can’t adapted into a consensus protocol using the persona
CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET208

trick described in §23.6.2.

Pseudocode is given in Algorithm 23.7. In this simplified version, we
assume an infinitely long array A[1 . . . ], so that we don’t need to worry
about n. Truncating the array at log n also works, but the analysis requires
handling the last position as a special case, which I am too lazy to do here.

1 Choose r ∈ Z+ such that Pr [r = i] = 2−i

2 A[r] ← 1
3 if A[r + 1] = 0 then
4 stay
5 else
6 leave
Algorithm 23.7: Giakkoupis-Woelfel sifter [GW12a]

Lemma 23.7.1. In any execution of Algorithm 23.7 with an oblivious ad-

versary and n processes, at least one process stays, and the expected number
of processes that stay is O(log n).

Proof. For the first part, observe that any process that picks the largest
value of r among all processes will survive; since the number of processes is
finite, there is at least one such survivor.
For the second part, let Xi be the number of survivors with r = i. Then
E [Xi ] is bounded by n · 2−i , since no process survives with r = i without
first choosing r = i. But we can also argue that E [Xi ] ≤ 3 for any value of
n, by considering the sequence of write operations in the execution.
Because the adversary is oblivious, the location of these writes is uncor-
related with their ordering. If we assume that the adversary is trying to
maximize the number of survivors, its best strategy is to allow each process
to read immediately after writing, as delaying this read can only increase the
probability that A[r + 1] is nonzero. So in computing Xi , we are counting
the number of writes to A[i] before the first write to A[i + 1]. Let’s ignore
all writes to other registers; then the j-th write to either of A[i] or A[i + 1]
has a conditional probability of 2/3 of landing on A[i] and 1/3 on A[i + 1].
We are thus looking at a geometric distribution with parameter 1/3, which
has expectation 3.
CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET209

Combining these two bounds gives E [Xi ] ≤ min(3, 2−i ). So then

∞
min(3, n · 2−i )
X
E [survivors] ≤
i=1
= 3 lg n + O(1),

because once n · 2−i drops below 3, the remaining terms form a geometric
series.

Like square root, logarithm is concave, so Jensen’s inequality applies here

as well. So O(log∗ n) rounds of Algorithm 23.7 reduces us to an expected
constant number of survivors, which can then be fed to RatRace.
With an adaptive adversary, all of the sifter-based test-and-sets fail
badly: in this particular case, an adaptive adversary can sort the processes
in order of increasing write location so that every process survives. The best
known n-process test-and-set for an adaptive adversary is still a tree of 2-
process randomized test-and-sets, as in the Afek et al. [AWW93] algorithm
described in §22.2. Whether O(log n) expected steps is in fact necessary
is still open (as is the exact complexity of test-and-set with an oblivious
adversary).

23.8 Space bounds

√
A classic result of Fich, Herlihy, and Shavit [FHS98] shows that Ω( n) reg-
isters are needed to solve consensus even under the very weak requirement of
nondeterministic solo termination, which says that for every reachable
configuration and every process p, there exists some continuation of the exe-
cution in which the protocol terminates with only p running. The best known
upper bound is the trivial n—one single-writer register per process—since
any multi-writer register algorithm can be translated into a single-writer al-
gorithm and (assuming wide enough registers) multiple registers of a single
process can be combined into one.
There has been very little progress in closing the gap between these
two bounds since the original conference version of the FHS paper from
1993, although very recently, Giakkoupis et al. [GHHW13] have shown a
√
surprising O( n)-space algorithm for the closely related problem of leader
election, which is basically test-and-set without guaranteeing linearizability.
The main difference between leader election and consensus is that in consen-
sus every process learns the identity of the winner, instead of just whether
CHAPTER 23. RANDOMIZED CONSENSUS AND TEST-AND-SET210

it personally won or lost. It is not clear whether the techniques used for this
problem could carry across to consensus.
Chapter 24

Renaming

We will start by following the presentation in [AW04, §16.3]. This mostly de-
scribes results of the original paper of Attiya et al. [ABND+ 90] that defined
the renaming problem and gave a solution for message-passing; however, it’s
now more common to treat renaming in the context of shared-memory, so
we will follow Attiya and Welch’s translation of these results to a shared-
memory setting.

24.1 Renaming
In the renaming problem, we have n processes, each starts with a name
from some huge namespace, and we’d like to assign them each unique names
from a much smaller namespace. The main application is allowing us to run
algorithms that assume that the processes are given contiguous numbers,
e.g. the various collect or atomic snapshot algorithms in which each process
is assigned a unique register and we have to read all of the registers. With
renaming, instead of reading a huge pile of registers in order to find the few
that are actually used, we can map the processes down to a much smaller
set.
Formally, we have a decision problem where each process has input xi
(its original name) and output yi , with the requirements:

Termination Every nonfaulty process eventually decides.

Uniqueness If pi 6= pj , then yi 6= yj .

Anonymity The code executed by any process depends only on its input
xi : for any execution of processes p1 . . . pn with inputs x1 . . . xn , and

211
CHAPTER 24. RENAMING 212

any permutation π of [1 . . . n], there is a corresponding execution of

processes pπ(1) . . . pπ(n) with inputs x1 . . . xn in which pπ(i) performs
exactly the same operations as pi and obtains the same output yi .

The last condition is like non-triviality for consensus: it excludes algo-

rithms where pi just returns i in all executions. Typically we do not have
to do much to prove anonymity other than observing that all processes are
running the same code.
We will be considering renaming in a shared-memory system, where we
only have atomic registers to work with.

24.2 Performance
Conventions on counting processes:

• N = number of possible original names.

• n = maximum number of processes.

• k = number of processes that actually execute the algorithm.

Ideally, we’d like any performance measures we get to depend on k alone

if possible (giving an adaptive algorithm). Next best would be something
polynomial in n and k. Anything involving N is bad.
We’d also like to minimize the size of the output namespace. How well
we can do this depends on what assumptions we make. For deterministic
algorithms using only read-write registers, a lower bound due to Herlihy and
Shavit [HS99] shows that we can’t get fewer than 2n − 1 names for general
n.1 Our target thus will be exactly 2n − 1 output names if possible, or 2k − 1
if we are trying to be adaptive. For randomized algorithm, it is possible to
solve strong or tight renaming, where the size of the namespace is exactly
k; we’ll see how to do this in §24.5.
A small note on bounds: There is a lot of variation in the literature on
how bounds on the size of the output namespace are stated. The original
Herlihy-Shavit lower bound [HS99] says that there is no general renaming
algorithm that uses 2n names for n + 1 processes; in other words, any n-
process algorithm uses at least 2n − 1 names. Many subsequent papers
1
This lower bound was further refined by Castañeda and Rajsbaum [CR08], who show
that 2n − 2 (but no less!) is possible for certain special values of n; all of these lower
bounds make extensive use of combinatorial topology, so we won’t try to present them
here.
CHAPTER 24. RENAMING 213

discussing lower bounds on the namespace follow the approach of Herlihy and
Shavit and quote lower bounds that are generally 2 higher than the minimum
number of names needed for n processes. This requires a certain amount of
translation when comparing these lower bounds with upper bounds, which
use the more natural convention.

24.3 Order-preserving renaming

Before we jump into upper bounds, let’s do an easy lower bound from the
Attiya et al. paper [ABND+ 90]. This bound works on a variant of renaming
called order-preserving renaming, where we require that yi < yj when-
ever xi < xj . Unfortunately, this requires a very large output namespace:
with t failures, any asynchronous algorithm for order-preserving renaming
requires 2t (n − t + 1) − 1 possible output names. This lower bound applies
regardless of the model, as long as some processes may start after other
processes have already been assigned names.
For the wait-free case, we have t = n − 1, and the bound becomes just
n
2 −1. This is a simpler case than the general t-failure case, but the essential
idea is the same: if I’ve only seen a few of the processes, I need to leave room
for the others.

Theorem 24.3.1. There is no order-preserving renaming algorithm for n

processes using fewer than 2n − 1 names.

Proof. By induction on n. For n = 1, we use 21 − 1 = 1 names; this is the

base case. For larger n, suppose we use m names, and consider an execution
in which one process pn runs to completion first. This consumes one name
yn and leaves k names less than yn and m − k − 1 names greater than yn . By
setting all the inputs xi for i < n either less than xn or greater than xn , we
can force the remaining processes to choose from the remaining k or m−k−1
names. Applying the induction hypothesis, this gives k ≥ 2n−1 − 1 and
m−k−1 ≥ 2n−1 −1, so m = k+(m−k−1)+1 ≥ 2(2n−1 −1)+1 = 2n −1.

24.4 Deterministic renaming

In deterministic renaming, we can’t use randomization, and may or may
not have any primitives stronger than atomic registers. With just atomic
registers, we can only solve loose renaming; with test-and-set, we can solve
tight renaming. In this section, we describe some basic algorithms for de-
terministic renaming.
CHAPTER 24. RENAMING 214

24.4.1 Wait-free renaming with 2n − 1 names

Here we use Algorithm 55 from [AW04], which is an adaptation to shared
memory of the message-passing renaming algorithm of [ABND+ 90]. One
odd feature of the algorithm is that, as written, it is not anonymous: pro-
cesses communicate using an atomic snapshot object and use their process
ids to select which component of the snapshot array to write to. But if we
think of the process ids used in the algorithm as the inputs xi rather than
the actual process ids i, then everything works. The version given in Algo-
rithm 24.1 makes this substitution explicit, by treating the original name i
as the input.

1 procedure getName()
2 s←1
3 while true do
4 a[i] ← s
5 view ← snapshot(a)
6 if view[j] = s for some j then
7 r ← |{j : view[j] 6= ⊥ ∧ j ≤ i}|
8 s ← r-th positive integer not in
{view[j] : j 6= i ∧ view[j] = ⊥}
9 else
10 return s

Algorithm 24.1: Wait-free deterministic renaming

The array a holds proposed names for each process (indexed by the
original names), or ⊥ for processes that have not proposed a name yet. If a
process proposes a name and finds that no other process has proposed the
same name, it takes it; otherwise it chooses a new name by first computing
its rank r among the active processes and then choosing the r-th smallest
name that hasn’t been proposed by another process. Because the rank is at
most n and there are at most n − 1 names proposed by the other processes,
this always gives proposed names in the range [1 . . . 2n − 1]. But it remains
to show that the algorithm satisfies uniqueness and termination.
For uniqueness, consider two process with original names i and j. Sup-
pose that i and j both decide on s. Then i sees a view in which a[i] = s and
a[j] 6= s, after which it no longer updates a[i]. Similarly, j sees a view in
which a[j] = s and a[i] 6= s, after which it no longer updates a[j]. If i’s view
is obtained first, then j can’t see a[i] 6= s, but the same holds if j’s view is
CHAPTER 24. RENAMING 215

obtained first. So in either case we get a contradiction, proving uniqueness.

Termination is a bit trickier. Here we argue that no process can run
forever without picking a name, by showing that if we have a set of processes
that are doing this, the one with smallest original name eventually picks a
name. More formally, call a process trying if it runs for infinitely many steps
without choosing a name. Then in any execution with at least one trying
process, eventually we reach a configuration where all processes have either
finished or are trying. In some subsequent configuration, all the processes
have written to the a array at least once; from this point on, the set of non-
null positions in a—and thus the rank each process computes for itself—is
stable.
Starting from some such stable configuration, look at the trying process
i with the smallest original name, and suppose it has rank r. Let F =
{z1 < z2 . . . } be the set of “free names” that are not proposed in a by any of
the finished processes. Observe that no trying process j 6= i ever proposes
a name in {z1 . . . zr }, because any such process has rank greater than r.
This leaves zr open for i to claim, provided the other names in {z1 . . . zr }
eventually become free. But this will happen, because only trying processes
may have proposed these names (early on in the execution, when the finished
processes hadn’t finished yet), and the trying processes eventually propose
new names that are not in this range. So eventually process i proposes zr ,
sees no conflict, and finishes, contradicting the assumption that it is trying.
Note that we haven’t proved any complexity bounds on this algorithm at
all, but we know that the snapshot alone takes at least Ω(N ) time and space.
Brodksy et al. [BEW11] cite a paper of Bar-Noy and Dolev [BND89] as giving
a shared-memory version of [ABND+ 90] with complexity O(n·4n ); they also
give algorithms and pointers to algorithms with much better complexity.

24.4.2 Long-lived renaming

In long-lived renaming a process can release a name for later use by other
processes (or the same process, if it happens to run choose-name again). Now
the bound on the number of names needed is 2k−1, where k is the maximum
number of concurrently active processes. Algorithm 24.1 can be converted
to a long-lived renaming algorithm by adding the releaseName procedure
given in Algorithm 24.2. This just erases the process’s proposed name, so
that some other process can claim it.
Here the termination requirement is weakened slightly, to say that some
process always makes progress in getName. It may be, however, that there
is some process that never successfully obtains a name, because it keeps
CHAPTER 24. RENAMING 216

1 procedure releaseName()
2 a[i] ← ⊥
Algorithm 24.2: Releasing a name

getting stepped on by other processes zipping in and out of getName and

releaseName.

24.4.3 Renaming without snapshots

Moir and Anderson [MA95] give a renaming protocol that is somewhat easier
to understand and doesn’t require taking snapshots over huge arrays. A
downside is that the basic version requires k(k + 1)/2 names to handle k
active processes.

24.4.3.1 Splitters
The Moir-Anderson renaming protocol uses a network of splitters, which
we last saw providing a fast path for mutual exclusion in §17.4.2. Each
splitter is a widget, built from a pair of atomic registers, that assigns to
each processes that arrives at it the value right, down, or stop. As discussed
previously, the useful properties of splitters are that if at least one process
arrives at a splitter, then (a) at least one process returns right or stop; and
(b) at least one process returns down or stop; (c) at most one process returns
stop; and (d) any process that runs by itself returns stop.
We proved the last two properties in §17.4.2; we’ll prove the first two
here. Another way of describing these properties is that of all the processes
that arrive at a splitter, some process doesn’t go down and some process
doesn’t go right. By arranging splitters in a grid, this property guarantees
that every row or column that gets at least one process gets to keep it—which
means that with k processes, no process reaches row k + 1 or column k + 1.
Algorithm 24.3 gives the implementation of a splitter (it’s identical to
Algorithm 17.5, but it will be convenient to have another copy here).

Lemma 24.4.1. If at least one process completes the splitter, at least one
process returns stop or right.

Proof. Suppose no process returns right; then every process sees open in
door, which means that every process writes its id to race before any process
closes the door. Some process writes its id last: this process will see its own
id in race and return stop.
CHAPTER 24. RENAMING 217

shared data:
1 atomic register race, big enough to hold an id, initially ⊥
2 atomic register door, big enough to hold a bit, initially open
3 procedure splitter(id)
4 race ← id
5 if door = closed then
6 return right
7 door ← closed
8 if race = id then
9 return stop
10 else
11 return down

Algorithm 24.3: Implementation of a splitter

Lemma 24.4.2. If at least one process completes the splitter, at least one
process returns stop or down.
Proof. First observe that if no process ever writes to door, then no process
completes the splitter, because the only way a process can finish the splitter
without writing to door is if it sees closed when it reads door (which must
have been written by some other process). So if at least one process finishes,
at least one process writes to door. Let p be any such process. From the
code, having written door, it has already passed up the chance to return
right; thus it either returns stop or down.

24.4.3.2 Splitters in a grid

Now build an m-by-m triangular grid of splitters, arranged as rows 0 . . . m−1
and columns 0 . . . m − 1, where a splitter appears in each position (r, c)
with r + c ≤ m − 1 (see Figure 24.1 for an example; this figure is taken
from [Asp10]). Assign a distinct name to each of the m


2 splitters in this
grid. To obtain a name, a process starts at (r, c) = (0, 0), and repeatedly
executes the splitter at its current position (r, c). If the splitter returns right,
it moves to (r, c + 1); if down, it moves to (r + 1, c); if stop, it stops, and
returns the name of its current splitter. This gives each name to at most
one process (by Lemma 17.4.3); we also have to show that if at most m
processes enter the grid, every process stops at some splitter.
The argument for this is simple. Suppose some process p leaves the
grid on one of the 2m output wires. Look at the path it takes to get there
CHAPTER 24. RENAMING 218

Figure 24.1: A 6 × 6 Moir-Anderson grid

(see Figure 24.2, also taken from [Asp10]). Each splitter on this path must
handle at least two processes (or p would have stopped at that splitter, by
Lemma 17.4.4). So some other process leaves on the other output wire, either
right or down. If we draw a path from each of these wires that continues right
or down to the end of the grid, then along each of these m disjoint paths
either some splitter stops a process, or some process reaches a final output
wire, each of which is at a distinct splitter. But this gives m processes in
addition to p, for a total of m + 1 processes. It follows that:

Theorem 24.4.3. An m × m Moir-Anderson grid solves renaming for up

to m processes.

The time complexity of the algorithm is O(m): Each process spends at

most 4 operations on each splitter, and no process goes through more than
2m splitters. In general, any splitter network will take at least n steps to
stop n processes, because the adversary can run them all together in a horde
that drops only one process at each splitter.
If we don’t know k in advance, we can still guarantee names of size O(k 2
)
by carefully arranging them so that each k-by-k subgrid contains the first k2
names. This gives an adaptive renaming algorithm (although the namespace
size is pretty high). We still have to choose our grid to be large enough for
the largest k we might actually encounter; the resulting space complexity is
O(n2 ).
CHAPTER 24. RENAMING 219

Figure 24.2: Path taken by a single process through a 6 × 6 Moir-Anderson

grid (heavy path), and the 6 disjoint paths it spawns (dashed paths).

With a slightly more clever arrangement of the splitters, it is possible to

reduce the space complexity to O(n3/2 ) [Asp10]. Whether further reductions
are possible is an open problem. Note however that linear time complexity
makes splitter networks uncompetitive with much faster randomized algo-
rithms (as we’ll see in §24.5), so this may not be a very important open
problem.

24.4.4 Getting to 2n − 1 names in polynomial space

From before, we have an algorithm that will get 2n − 1 names for n pro-
cesses out of N possible processes when run using O(N ) space (for the
enormous snapshots). To turn this into a bounded-space algorithm, run
Moir-Anderson first to get down to Θ(k 2 ) names, then run the previous
algorithm (in Θ(n2 ) space) using these new names as the original names.
Since we didn’t prove anything about time complexity of the humongous-
snapshot algorithm, we can’t say much about the time complexity of this
combined one. Moir and Anderson suggest instead using an O(N k 2 ) algo-
rithm of Borowsky and Gafni to get O(k 4 ) time for the combined algorithm.
This is close to the best known: a later paper by Afek and Merritt [AM99]
holds the current record for deterministic adaptive renaming into 2k − 1
names at O(k 2 ) individual steps. On the lower bound side, it is known that
Ω(k) is a lower bound on the individual steps of any renaming protocol with
CHAPTER 24. RENAMING 220

a polynomial output namespace [AAGG11].

24.4.5 Renaming with test-and-set

Moir and Anderson give a simple renaming algorithm based on test-and-set
that is strong (k processes are assigned exactly the names 1 . . . k), adaptive
(the time complexity to acquire a name is O(k)), and long-lived, which
means that a process can release its name and the name will be available to
processes that arrive later. In fact, the resulting algorithm gives long-lived
strong renaming, meaning that the set of names in use will always be no
larger than the set of processes that have started to acquire a name and not
yet finished releasing one; this is a little stronger than just saying that the
algorithm is strong and that it is long-lived separately.
The algorithm is simple: we have a line of test-and-set bits T [1] . . . T [n].
To acquire a name, a process starts at T [1] and attempts to win each test-
and-set until it succeeds; whichever T [i] it wins gives it name i. To release
a name, a process releases the test-and-set.
Without the releases, the same mechanism gives fetch-and-increment [AWW93].
Fetch-and-increment by itself solves tight renaming (although not long-lived
renaming, since there is no way to release a name).

24.5 Randomized renaming

With randomization, we can beat both the 2k − 1 lower bound on the size of
the output namespace from [HS99] and the Ω(k) lower bound on individual
work from [AAGG11], achieving strong renaming with O(log k) expected
individual work [AACH+ 11].
The basic idea is that we can use randomization for load balancing,
where we avoid the problem of having an army of processes marching to-
gether with only a few peeling off at a time (as in splitter networks) by
having the processes split up based on random choices. For example, if each
process generates a random name consisting of 2 dlg ne bits, then it is rea-
sonably likely that every process gets a unique name in a namespace of size
O(n2 ) (we can’t hope for less than O(n2 ) because of the birthday para-
dox). But we want all processes to be guaranteed to have unique names, so
we need some more machinery.
We also need the processes to have initial names; if they don’t, there
is always some nonzero probability that two identical processes will flip
their coins in exactly the same way and end up with the same name.
CHAPTER 24. RENAMING 221

This observation was formalized by Buhrman, Panconesi, Silvestri, and

Vitányi [BPSV06].

24.5.1 Randomized splitters

Attiya, Kuhn, Plaxton, Wattenhofer, and Wattenhofer [AKP+ 06] suggested
the use of randomized splitters in the context of another problem (adap-
tive collect) that is closely related to renaming.
A randomized splitter is just like a regular splitter, except that if a
process doesn’t stop it flips a coin to decide whether to go right or down.
Randomized splitters are nice because they usually split better than de-
terministic splitters: if k processes√reach a randomized splitter, with high
probability no more than k/2 + O( k log k) will leave on either output wire.
It’s not hard to show that a binary tree of these things of depth 2 dlg ne
stops all but a constant expected number of processes on average;2 processes
that don’t stop can be dropped into a backup renaming algorithm (Moir-
Anderson, for example) only a constant increase in expected individual work.
Furthermore, the binary tree of randomized splitters is adaptive; if only
k processes show up, we only need O(log k) levels levels on average to split
them up. This gives renaming into a namespace with expected size O(k 2 )
in O(log k) expected individual steps.

24.5.2 Randomized test-and-set plus sampling

Subsequent work by Alistarh et al. [AAG+ 10] showed how some of the same
ideas could be used to get strong renaming, where the output namespace has
size exactly n (note this is not adaptive; another result in the same paper
gives adaptive renaming, but it’s not strong). There are two pieces to this
result: an implementation of randomized test-and-set called RatRace, and
a sampling procedure for getting names called ReShuffle.
The RatRace protocol implements a randomized test-and-set with O(log k)
expected individual work. The essential idea is to use a tree of randomized
splitters to assign names, then have processes walk back up the same tree
attempting to win a 3-process randomized test-and-set at each node (there
are 3 processes, because in addition to the winners of each subtree, we may
also have a process that stopped on that node in the renaming step); this
2
The proof is to consider the expected number of pairs

of processes that flip their coins
the same way for all 2 dlg ne steps. This is at most n2 n−2 < 1/2, so on average at most 1
process escapes the tree, giving (by symmetry) at most a 1/n chance that any particular
process escapes. Making the tree deeper can give any polynomial fraction of escapees
while still keeping O(log n) layers.
CHAPTER 24. RENAMING 222

test-and-set is just a very small binary tree of 2-process test-and-sets im-

plemented using the algorithm of Tromp and Vitányi [TV02]. A gate bit is
added at the top as in the test-and-set protocol of Afek et al. [AGTV92] to
get linearizability.
Once we have test-and-set, we could get strong renaming using a linear
array of test-and-sets as suggested by Moir and Anderson [MA95], but it’s
more efficient to use the randomization to spread the processes out. In
the ReShuffle protocol, each process chooses a name in the range [1 . . . n]
uniformly at random, and attempts to win a test-and-set guarding that
name. If it doesn’t work, it tries again. Alistarh et al. show that this method
produces unique names for everybody in O(n log4 n) total steps with high
probability. The individual step complexity of this algorithm, however, is
not very good: there is likely to be some unlucky process that needs Ω(n)
probes (at an expected cost of Θ(log n) steps each) to find an empty slot.

24.5.3 Renaming with sorting networks

A later paper by Alistarh et al. [AACH+ 11] reduces the cost of renaming
still further, getting O(log k) expected individual step complexity for acquir-
ing a name. The resulting algorithm is both adaptive and strong: with k
processes, only names 1 through k are used. We’ll describe the non-adaptive
version here.
The basic idea is to build a sorting network out of test-and-sets; the re-
sulting structure, called a renaming network, routes each process through
a sequence of test-and-sets to a unique output wire. Unlike a splitter net-
work, a renaming network uses the stronger properties of test-and-set to
guarantee that (once the dust settles) only the lowest-numbered output wires
are chosen; this gives strong renaming.

24.5.3.1 Sorting networks

A sorting network is a kind of parallel sorting algorithm that proceeds in
synchronous rounds, where in each round the elements of an array at certain
fixed positions are paired off and swapped if they are out of order. The
difference between a sorting network and a standard comparison-based sort
is that the choice of which positions to compare at each step is static, and
doesn’t depend on the outcome of previous comparisons; also, the only effect
of a comparison is possibly swapping the two values that were compared.
Sorting networks are drawn as in Figure 24.3. Each horizontal line or
wire corresponds to a position in the array. The vertical lines are com-
CHAPTER 24. RENAMING 223

Figure 24.3: A sorting network

parators that compare two values coming in from the left and swap the
larger value to the bottom. A network of comparators is a sorting network
if the sequences of output values is always sorted no matter what the order
of values on the inputs is.
The depth of a sorting network is the maximum number of comparators
on any path from an input to an output. The width is the number of wires;
equivalently, the number of values the network can sort. The sorting network
in Figure 24.3 has depth 3 and width 4.
Explicit constructions of sorting networks with width n and depth O(log2 n)
are known [Bat68]. It is also known that sorting networks with depth
O(log n) exist [AKS83], but no explicit construction of such a network is
known.

24.5.3.2 Renaming networks

To turn a sorting network into a renaming network, we replace the compara-
tors with test-and-set bits, and allow processes to walk through the network
asynchronously. This is similar to an earlier mechanism called a counting
network [AHS94], which used certain special classes of sorting networks as
counters, but here any sorting network works.
Each process starts on a separate input wire, and we maintain the in-
variant that at most one process ever traverses a wire. It follows that each
test-and-set bit is only used by two processes. The first process to reach the
test-and-set bit is sent out the lower output, while the second is sent out the
CHAPTER 24. RENAMING 224

upper output. If we imagine each process that participates in the protocol

as a one and each process that doesn’t as a zero, the test-and-set bit acts
as a comparator: if no processes show up on either input (two zeros), no
processes leave (two zeros again); if processes show up on both inputs (two
ones), processes leave on both (two ones again); and if only one process
ever shows up (a zero and a one), it leaves on the bottom output (zero and
one, sorted). Because the original sorting network sorts all the ones to the
bottom output wires, the corresponding renaming network sorts all the pro-
cesses that arrive to the bottom outputs. Label these outputs starting at 1
at the bottom to get renaming.
Since each test-and-set involves at most two processes, we can carry them
out in O(1) expected register operations using, for example, the protocol of
Tromp and Vitányi [TV02]. The expected cost for a process to acquire a
name is then O(log n) (using an AKS sorting network). A more complicated
construction in the Alistarh et al. paper shows how to make this adaptive,
giving an expected cost of O(log k) instead.
The use of test-and-sets to route processes to particular names is simi-
lar to the line of test-and-sets proposed by Moir and Anderson [MA95] as
described in §24.4.5. Some differences between that protocol and renaming
networks is that renaming networks do not by themselves give fetch-and-
increment (although Alistarh et al. show how to build fetch-and-increment
on top of renaming networks at a small additional cost), and renaming net-
works do not provide any mechanism for releasing names. The question of
whether it is possible to get cheap long-lived strong renaming is still open.

24.5.4 Randomized loose renaming

Loose renaming should be easier than strong renaming, and using a random-
ized algorithm it essentially reduces to randomized load balancing. A basic
approach is to use 2n names, and guard each with a test-and-set; because
less than half of the names are taken at any given time, each process gets a
name after O(1) tries and the most expensive renaming operation over all
n processes takes O(log n) expected steps.
A more sophisticated version of this strategy, which appears in [AAGW13],
uses n(1 + ) output names to get O(log log n) maximum steps. The intu-
ition for why this works is if n processes independently choose one of cn
names uniformly at random, then the expected
number of collisions—pairs
of processes that choose the same name—is n2 /cn, or about n/2c. This
may seem like only a constant-factor improvement, but if we instead look at
the ratio between the survivors n/2c and the number of allocated names cn,
CHAPTER 24. RENAMING 225

we have now moved from 1/c to 1/2c2 . The 2 gives us some room to reduce
the number of names in the next round, to cn/2, say, while still keeping a
1/c2 ratio of survivors to names.
So the actual renaming algorithm consists of allocating cn/2i names to
round i, and squaring the ratio of survivors to names in each rounds. It only
takes O(log log n) rounds to knock the ratio of survivors to names below 1/n,
so at this point it is likely that all processes will have finished. At the same
time, the sum over all rounds of the allocated names forms a geometric
series, so only O(n) names are needed altogether.
Swept under the carpet here is a lot of careful analysis of the probabili-
ties. Unlike what happens with sifters (see §23.6), Jensen’s inequality goes
the wrong way here, so some additional technical tricks are needed (see the
paper for details). But the result is that only O(log log n) rounds are to
assign every process a name with high probability, which is the best value
currently known.
There is a rather weak lower bound in the Alistarh et al. paper that
shows that Ω(log log n) steps are needed for some process in the worst case,
under the assumption that the renaming algorithm uses only test-and-set
objects and that a process acquires a name as soon as it wins some test-and-
set object. This does not give a lower bound on the problem in general, and
indeed the renaming-network based algorithms discussed previously do not
have this property. So the question of the exact complexity of randomized
loose renaming is still open.
Chapter 25

Software transactional
memory

Not covered in 2014. If you are interested in software transactional mem-

ory from a theoretical perspective, there is a good recent survey on this
material by Attiya [Att14], available at http: // www. eatcs. org/ images/
bulletin/ beatcs112. pdf .
Software transactional memory, or STM for short, goes back to
Shavit and Touitou [ST97] based on earlier proposals for hardware support
for transactions by Herlihy and Moss [HM93]. Recently very popular in
programming language circles. We’ll give a high-level description of the
Shavit and Touitou results; for full details see the actual paper.
We start with the basic idea of a transaction. In a transaction, I read a
bunch of registers and update their values, and all of these operations appear
to be atomic, in the sense that the transaction either happens completely or
not at all, and serializes with other transactions as if each occurred instan-
taneously. Our goal is to implement this with minimal hardware support,
and use it for everything.
Generally we only consider static transactions where the set of mem-
ory locations accessed is known in advance, as opposed to dynamic trans-
actions where it may vary depending on what we read (for example, maybe
we have to follow pointers through some data structure). Static transactions
are easier because we can treat them as multi-word read-modify-write.
Implementations are usually non-blocking: some infinite stream of
transactions succeed, but not necessarily yours. This excludes the sim-
plest method based on acquiring locks, since we have to keep going even if
a lock-holder crashes, but is weaker than wait-freedom since we can have

226
CHAPTER 25. SOFTWARE TRANSACTIONAL MEMORY 227

starvation.

25.1 Motivation
Some selling points for software transactional memory:
1. We get atomic operations without having to use our brains much.
Unlike hand-coded atomic snapshots, counters, queues, etc., we have
a universal construction that converts any sequential data structure
built on top of ordinary memory into a concurrent data structure.
This is useful since most programmers don’t have very big brains. We
also avoid burdening the programmer with having to remember to lock
things.

2. We can build large shared data structures with the possibility of con-
current access. For example, we can implement atomic snapshots so
that concurrent updates don’t interfere with each other, or an atomic
queue where enqueues and dequeues can happen concurrently so long
as the queue always has a few elements in it to separate the enqueuers
and dequeuers.

3. We can execute atomic operations that span multiple data structures,

even if the data structures weren’t originally designed to work together,
provided they are all implemented using the STM mechanism. This
is handy in classic database-like settings, as when we want to take $5
from my bank account and put it in yours.
On the other hand, we now have to deal with the possibility that oper-
ations may fail. There is a price to everything.

25.2 Basic approaches

• Locking (not non-blocking). Acquire either a single lock for all of
memory (doesn’t allow much concurrency) or a separate lock for each
memory location accessed. The second approach can lead to deadlock
if we aren’t careful, but we can prove that if every transaction acquires
locks in the same order (e.g., by increasing memory address), then
we never get stuck: we can order the processes by the highest lock
acquired, and somebody comes out on top. Note that acquiring locks
in increasing order means that I have to know which locks I want before
I acquire any of them, which may rule out dynamic transactions.
CHAPTER 25. SOFTWARE TRANSACTIONAL MEMORY 228

• Single-pointer compare-and-swap (called ”Herlihy’s method” in [ST97],

because of its earlier use for constructing concurrent data structures
by Herlihy [Her93]). All access to the data structure goes through
a pointer in a CAS. To execute a transaction, I make my own copy
of the data structure, update it, and then attempt to redirect the
pointer. Advantages: trivial to prove that the result is linearizable
(the pointer swing is an atomic action) and non-blocking (somebody
wins the CAS); also, the method allows dynamic transactions (since
you can do anything you want to your copy). Disadvantages: There’s
a high overhead of the many copies,1 and the single-pointer bottleneck
limits concurrency even when two transactions use disjoint parts of
memory.
• Multiword RMW: This is the approach suggested by Shavit and Touitou,
which most subsequent work follows. As usually implemented, it only
works for static transactions. The idea is that I write down what reg-
isters I plan to update and what I plan to do to them. I then attempt
to acquire all the registers. If I succeed, I update all the values, store
the old values, and go home. If I fail, it’s because somebody else al-
ready acquired one of the registers. Since I need to make sure that
somebody makes progress (I may be the only process left alive), I’ll
help that other process finish its transaction if possible. Advantages:
allows concurrency between disjoint transactions. Disadvantages: re-
quires implementing multi-word RMW—in particular, requires that
any process be able to understand and simulate any other process’s
transactions. Subsequent work often simplifies this to implementing
multi-word CAS, which is sufficient to do non-blocking multi-word
RMW since I can read all the registers I need (without any locking)
and then do a CAS to update them (which fails only if somebody else
succeeded).

25.3 Implementing multi-word RMW

We’ll give a sketchy description of Shavit and Touitou’s method [ST97],
which essentially follows the locking approach but allows other processes to
help dead ones so that locks are always released.
The synchronization primitive used is LL/SC: LL (load-linked) reads
a register and leaves our id attached to it, SC (store-conditional) writes a
1
This overhead can be reduced in many cases by sharing components, a subject that
has seen much work in the functional programming literature. See for example [Oka99].
CHAPTER 25. SOFTWARE TRANSACTIONAL MEMORY 229

register only if our id is still attached, and clears any other id’s that might
also be attached. It’s easy to build a 1-register CAS (CAS1) out of this,
though Shavit and Touitou exploit some additional power of LL/SC.

25.3.1 Overlapping LL/SC

The particular trick that gets used in the Shavit-Touitou protocol is to
use two overlapping LL/SC pairs to do a CAS-like update on one memory
location while checking that another memory location hasn’t changed. The
purpose of this is to allow multiple processes to work on the same transaction
(which requires the first CAS to avoid conflicts with other transactions) while
making sure that slow processes don’t cause trouble by trying to complete
transactions that have already finished (the second check).
To see this in action, suppose we have a register r that we want to do
a CAS on, while checking that a second register status is ⊥ (as opposed to
success or failure). If we execute the code fragment in Algorithm 25.1, it will
succeed only if nobody writes to status between its LL and SC and similarly
for r; if this occurs, then at the time of LL(r), we know that status = ⊥, and
we can linearize the write to r at this time if we restrict all access to r to
go through LL/SC.

1 if LL(status) = ⊥ then
2 if LL(r) = oldValue then
3 if SC(status, ⊥) = true then
4 SC(r, newValue)

Algorithm 25.1: Overlapping LL/SC

25.3.2 Representing a transaction

Transactions are represented by records rec. Each such record consists of a
status component that describes how far the transaction has gotten (needed
to coordinate cooperating processes), a version component that distinguishes
between versions that may reuse the same space (and that is used to shut
down the transaction when complete), a stable component that indicates
when the initialization is complete, an Op component that describes the
RMW to be performance, an array addresses[] of pointers to the arguments
to the RMW, and an array oldValues[] of old values at these addresses (for
the R part of the RWM). These are all initialized by the initiator of the
CHAPTER 25. SOFTWARE TRANSACTIONAL MEMORY 230

transaction, who will be the only process working on the transaction until
it starts acquiring locks.

25.3.3 Executing a transaction

Here we give an overview of a transaction execution:

1. Initialize the record rec for the transaction. (Only the initiator does
this.)

2. Attempt to acquire ownership of registers in addresses[]. See the

AcquireOwnerships code in the paper for details. The essential idea is
that we want to set the field owner[r] for each memory location r that
we need to lock; this is done using an overlapping LL/SC as described
above so that we only set owner[r] if (a) r is currently unowned, and
(b) nothing has happened to rec.status or rec.version. Ownership is
acquired in order of increasing memory address; if we fail to acquire
ownership for some r, our transaction fails. In case of failure, we set
rec.status to failure and release all the locks we’ve acquired (checking
rec.version in the middle of each LL/SC so we don’t release locks for
a later version using the same record). If we are the initiator of this
transaction, we will also go on to attempt to complete the transaction
that got in our way.

3. Do a LL on rec.status to see if AcquireOwnerships succeeded. If so,

update the memory, store the old results in oldValues, and release the
ownerships. If it failed, release ownership and help the next transaction
as described above.

Note that only an initiator helps; this avoids a long chain of helping
and limits the cost of each attempted transaction to the cost of doing two
full transactions, while (as shown below) still allowing some transaction to
finish.

25.3.4 Proof of linearizability

Intuition is:

• Linearizability follows from the linearizability of the locking protocol:

acquiring ownership is equivalent to grabbing a lock, and updates oc-
cur only when all registers are locked.
CHAPTER 25. SOFTWARE TRANSACTIONAL MEMORY 231

• Complications come from (a) two or more processes trying to complete

the same transaction and (b) some process trying to complete an old
transaction that has already terminated. For the first part we just
make sure that the processes don’t interfere with each other, e.g. I
am happy when trying to acquire a location if somebody else acquires
it for the same transaction. For the second part we have to check
rec.status and rec.version before doing just about anything. See the
pseudocode in the paper for details on how this is done.

25.3.5 Proof of non-blockingness

To show that the protocol is non-blocking we must show that if an un-
bounded number of transactions are attempted, one eventually succeeds.
First observe that in order to fail, a transaction must be blocked by another
transaction that acquired ownership of a higher-address location than it did;
eventually we run out of higher-address locations, so there is some trans-
action that doesn’t fail. Of course, this transaction may not succeed (e.g.,
if its initiator dies), but either (a) it blocks some other transaction, and
that transaction’s initiator will complete it or die trying, or (b) it blocks no
future transactions. In the second case we can repeat the argument for the
n − 1 surviving processes to show that some of them complete transactions,
ignoring the stalled transaction from case (b).

25.4 Improvements
One downside of the Shavit and Touitou protocol is that it uses LL/SC
very aggressively (e.g. with overlapping LL/SC operations) and uses non-
trivial (though bounded, if you ignore the ever-increasing version numbers)
amounts of extra space. Subsequent work has aimed at knocking these down;
for example a paper by Harris, Fraser, and Pratt [HFP02] builds multi-
register CAS out of single-register CAS with O(1) extra bits per register.
The proof of these later results can be quite involved; Harris et al, for
example, base their algorithm on an implementation of 2-register CAS whose
correctness has been verified only by machine (which may be a plus in some
views).
CHAPTER 25. SOFTWARE TRANSACTIONAL MEMORY 232

25.5 Limitations
There has been a lot of practical work on STM designed to reduce over-
head on real hardware, but there’s still a fair bit of overhead. On the
theory side, a lower bound of Attiya, Hillel, and Milani [AHM09] shows that
any STM system that guarantees non-interference between non-overlapping
RMW transactions has the undesirable property of making read-only trans-
actions as expensive as RMW transactions: this conflicts with the stated
goals of many practical STM implementations, where it is assumed that
most transactions will be read-only (and hopefully cheap). So there is quite
a bit of continuing research on finding the right trade-offs.
Chapter 26

Obstruction-freedom

Not covered in 2014.

The gold standard for shared-memory objects is wait-freedom: I can
finish my operation in a bounded number of steps no matter what anybody
else does. Like the gold standard in real life, this can be overly constraining.
So researchers have developed several weaker progress guarantees that are
nonetheless useful. The main ones are:

Lock-freedom An implementation is lock-free if infinitely many opera-

tions finish in any infinite execution. In simpler terms, somebody al-
ways makes progress, but maybe not you. (Also called non-blocking.)

Obstruction-freedom An implementation is obstruction-free if, start-

ing from any reachable configuration, any process can finish in a
bounded number of steps if all of the other processes stop. This defini-
tion was proposed in 2003 by Herlihy, Luchangco, and Moir [HLM03].
In lower bounds (e.g., the Jayanti-Tan-Toueg bound described in Chap-
ter 20) essentially the same property is often called solo-terminating.

Both of these properties exclude traditional lock-based algorithms, where

some process grabs a lock, updates the data structure, and then release the
lock; if this process halts, no more operations finish. Both properties are
also weaker than wait-freedom. It is not hard to show that lock-freedom is
a stronger condition that obstruction-freedom: given a lock-free implemen-
tation, if we can keep some single process running forever in isolation, we
get an infinite execution with only finitely many completed operations. So
we have a hierarchy: wait-free > lock-free > obstruction-free > locking.

233
CHAPTER 26. OBSTRUCTION-FREEDOM 234

26.1 Why build obstruction-free algorithms?

The pitch is similar to the pitch for building locking algorithms: an obstruction-
free algorithm might be simpler to design, implement, and reason about than
a more sophisticated algorithm with stronger properties. Unlike locking al-
gorithms, an obstruction-free algorithm won’t fail because some process dies
holding the lock; instead, it fails if more than one process runs the algorithm
at the same time. This possibility may be something we can avoid by build-
ing a contention manager, a high-level protocol that detects contention
and delays some processes to avoid it (say, using randomized exponential
back-off).

26.2 Examples
26.2.1 Lock-free implementations
Pretty much anything built using compare-and-swap or LL/SC ends up
being lock-free. A simple example would be a counter, where an increment
operation does

1 x ← LL(C)
2 SC(C, x + 1)

This is lock-free (the only way to prevent a store-conditional from suc-

ceeding is if some other store-conditional succeeds, giving infinitely many
successful increments) but not wait-free (I can starve). It’s also obstruction-
free, but since it’s already lock-free we don’t care about that.

26.2.2 Double-collect snapshots

Similarly, suppose we are doing atomic snapshots. We know that there
exist wait-free implementations of atomic snapshots, but they are subtle
and confusing. So we want to do something simpler, and hope that we at
least get obstruction-freedom.
If we do double-collects, that is, we have updates just write to a register
and have snapshots repeatedly collect until they get two collects in a row
with the same values, then any snapshot that finishes is correct (assuming
no updaters ever write the same value twice, which we can enforce with
nonces). This isn’t wait-free, because we can keep a snapshot going forever
CHAPTER 26. OBSTRUCTION-FREEDOM 235

by doing a lot of updates. It is lock-free, because we have to keep doing

updates to make this happen.
We can make this merely obstruction-free if we work hard (there is no rea-
son to do this, but it illustrates the difference between lock-freedom—good—and
obstruction-freedom—not so good). Suppose that every process keeps a
count of how many collects it has done in a register that is included in
other process’s collects (but not its own). Then two concurrent scans can
stall each other forever (the implementation is not lock-free), but if only one
is running it completes two collects in O(n) operations without seeing any
changes (it is obstruction-free).

26.2.3 Software transactional memory

Similar things happen with software transactional memory (see Chapter 25).
Suppose that I have an implementation of multiword compare-and-swap, and
I want to carry out a transaction. I read all the values I need, then execute an
MCAS operation that only updates if these values have not changed. The
resulting algorithm is lock-free (if my transaction fails, it’s because some
update succeeded). If however I am not very clever and allow some values
to get written outside of transactions, then I might only be obstruction-free.

26.2.4 Obstruction-free test-and-set

Algorithm 26.1 gives an implementation of 2-process test-and-set from atomic
registers that is obstruction-free; this demonstrates that obstruction-freedom
lets us evade the wait-free impossibility results implied by the consensus hi-
erarchy ([Her91b], discussed in Chapter 18).
The basic idea goes back to the racing counters technique used in
consensus protocols starting with Chor, Israeli, and Li [CIL94], and there is
some similarity to a classic randomized wait-free test-and-set due to Tromp
and Vitányi [TV02]. Each process keeps a position x in memory that it
also stores from time to time in its register a[i]. If a process gets 2 steps
ahead of the other process (as observed by comparing x to a[i − 1], it wins
the test-and-set; if a process falls one or more steps behind, it (eventually)
loses. To keep space down and guarantee termination in bounded time, all
values are tracked modulo 5.
Why this works: observe that whenever a process computes δ, x is equal
to a[i]; so δ is always an instantaneous snapshot of a[i] − a[1 − i]. If I observe
δ = 2 and return 0, your next read will either show you δ = −2 or δ = −1
(depending on whether you increment a[1 − i] after my read). In the latter
CHAPTER 26. OBSTRUCTION-FREEDOM 236

1 x←0
2 while true do
3 δ ← x − a[1 − i]
4 if δ = 2 (mod 5) then
5 return 0
6 else if δ = −1 (mod 5) do
7 return 1
8 else
9 x ← (x + 1) mod 5
10 a[i] ← x

Algorithm 26.1: Obstruction-free 2-process test-and-set

case, you return 1 immediately; in the former, you return after one more
increment (and more importantly, you can’t return 0). Alternatively, if I
ever observe δ = −1, your next read will show you either δ = 1 or δ = 2;
in either case, you will eventually return 0. (We chose 5 as a modulus
because this is the smallest value that makes the cases δ = 2 and δ = −2
distinguishable.)
We can even show that this is linearizable, by considering a solo execution
in which the lone process takes two steps and returns 0 (with two processes,
solo executions are the only interesting case for linearizability).
However, Algorithm 26.1 is not wait-free or even lock-free: if both pro-
cesses run in lockstep, they will see δ = 0 forever. But it is obstruction-free.
If I run by myself, then whatever value of δ I start with, I will see −1 or 2
after at most 6 operations.1
This gives an obstruction-free step complexity of 6, where the
obstruction-free step complexity is defined as the maximum number of op-
erations any process can take after all other processes stop. Note that our
usual wait-free measures of step complexity don’t make a lot of sense for
obstruction-free algorithms, as we can expect a sufficiently cruel adversary
to be able to run them up to whatever value he likes.
Building a tree of these objects as in §22.2 gives n-process test-and-set
with obstruction-free step complexity O(log n).
1
The worst case is where an increment by my fellow process leaves δ = −1 just before
my increment.
CHAPTER 26. OBSTRUCTION-FREEDOM 237

26.2.5 An obstruction-free deque

(We probably aren’t going to do this in class.)
So far we don’t have any good examples of why we would want to be
obstruction-free if our algorithm is based on CAS. So let’s describe the case
Herlihy et al. suggested.
A deque is a generalized queue that supports push and pop at both ends
(thus it can be used as either a queue or a stack, or both). A classic problem
in shared-memory objects is to build a deque where operations at one end of
the deque don’t interfere with operations at the other end. While there exist
lock-free implementation with this property, there is a particularly simple
implementation using CAS that is only obstruction-free.
Here’s the idea: we represent the deque as an infinitely-long array of
compare-and-swap registers (this is a simplification from the paper, which
gives a bounded implementation of a bounded deque). The middle of the
deque holds the actual contents. To the right of this region is an infinite
sequence of right null (RN) values, which are assumed never to appear as
a pushed value. To the left is a similar infinite sequence of left null (LN)
values. Some magical external mechanism (called an oracle in the paper)
allows processes to quickly find the first null value at either end of the non-
null region; the correctness of the protocol does not depend on the properties
of the oracle, except that it has to point to the right place at least some of
the time in a solo execution. We also assume that each cell holds a version
number whose only purpose is to detect when somebody has fiddled with
the cell while we aren’t looking (if we use LL/SC, we can drop this).
Code for rightPush and rightPop is given in Algorithm 26.2 (the code
for leftPush and leftPop is symmetric).
It’s easy to see that in a solo execution, if the oracle doesn’t lie, either
operation finishes and returns a plausible value after O(1) operations. So
the implementation is obstruction-free. But is it also correct?
To show that it is, we need to show that any execution leaves the deque
in a sane state, in particular that it preserves the invariant that the deque
consists of left-nulls followed by zero or more values followed by right-nulls,
and that the sequence of values in the queue is what it should be.
This requires a detailed case analysis of which operations interfere with
each other, which can be found in the original paper. But we can give some
intuition here. The two CAS operations in rightPush or rightPop succeed
only if neither register was modified between the preceding read and the
CAS. If both registers are unmodified at the time of the second CAS, then
the two CAS operations act like a single two-word CAS, which replaces the
CHAPTER 26. OBSTRUCTION-FREEDOM 238

1 procedure rightPush(v)
2 while true do
3 k ← oracle(right)
4 prev ← a[k − 1]
5 next ← a[k]
6 if prev.value 6= RN and next.value = RN then
7 if CAS(a[k − 1], prev, [prev.value, prev.version + 1]) then
8 if CAS(a[k], next, [v, next.version + 1]) then
9 we win, go home

10 procedure rightPop()
11 while true do
12 k ← oracle(right)
13 cur ← a[k − 1]
14 next ← a[k]
15 if cur.value 6= RN and next.value = RN then
16 if cur.value = LN and A[k − 1] = cur then
17 return empty
18 else if CAS(a[k], next, [RN, next.version + 1]) do
19 if CAS(a[k − 1], cur, [RN, cur.version + 1]) then
20 return cur.value

Algorithm 26.2: Obstruction-free deque

CHAPTER 26. OBSTRUCTION-FREEDOM 239

previous values (top, RN) with (top, value) in rightPush or (top, value) with
(top, RN) in rightPop; in either case the operation preserves the invariant.
So the only way we get into trouble is if, for example, a rightPush does a
CAS on a[k −1] (verifying that it is unmodified and incrementing the version
number), but then some other operation changes a[k − 1] before the CAS on
a[k]. If this other operation is also a rightPush, we are happy, because it
must have the same value for k (otherwise it would have failed when it saw
a non-null in a[k − 1]), and only one of the two right-pushes will succeed
in applying the CAS to a[k]. If the other operation is a rightPop, then it
can only change a[k − 1] after updating a[k]; but in this case the update to
a[k] prevents the original right-push from changing a[k]. With some more
tedious effort we can similarly show that any interference from leftPush or
leftPop either causes the interfering operation or the original operation to
fail. This covers 4 of the 16 cases we need to consider. The remaining cases
will be brushed under the carpet to avoid further suffering.

26.3 Boosting obstruction-freedom to wait-freedom

Since we didn’t cover this in 2014, this section has not been updated to
include some recent results [ACHS13, GHHW13].
Naturally, having an obstruction-free implementation of some object is
not very helpful if we can’t guarantee that some process eventually gets
its unobstructed solo execution. In general, we can’t expect to be able
to do this without additional assumptions; for example, if we could, we
could solve consensus using a long sequence of adopt-commit objects with
no randomization at all.2 So we need to make some sort of assumption
about timing, or find somebody else who has already figured out the right
assumption to make.
Those somebodies turn out to be Faith Ellen Fich, Victor Luchangco,
Mark Moir, and Nir Shavit, who give an algorithm for boosting obstruction-
freedom to wait-freedom [FLMS05]. The timing assumption is unknown-
bound semisynchrony, which means that in any execution there is some
maximum ratio R between the shortest and longest time interval between
any two consecutive steps of the same non-faulty process, but the processes
don’t know what this ratio is.3 In particular, if I can execute more than R
2
This fact was observed by Herlihy et al. [HLM03] in their original obstruction-free
paper; it also implies that there exists a universal obstruction-free implementation of
anything based on Herlihy’s universal construction.
3
This is a much older model, which goes back to a famous paper of Dwork, Lynch, and
CHAPTER 26. OBSTRUCTION-FREEDOM 240

steps without you doing anything, I can reasonably conclude that you are
dead—the semisynchrony assumption thus acts as a failure detector.
The fact that R is unknown might seem to be an impediment to using
this failure detector, but we can get around this. The idea is to start with
a small guess for R; if a process is suspected but then wakes up again, we
increment the guess. Eventually, the guessed value is larger than the correct
value, so no live process will be falsely suspected after this point. Formally,
this gives an eventually perfect (♦P ) failure detector, although the algorithm
does not specifically use the failure detector abstraction.
To arrange for a solo execution, when a process detects a conflict (be-
cause its operation didn’t finish quickly), it enters into a “panic mode” where
processes take turns trying to finish unmolested. A fetch-and-increment reg-
ister is used as a timestamp generator, and only the process with the smallest
timestamp gets to proceed. However, if this process is too sluggish, other
processes may give up and overwrite its low timestamp with ∞, temporarily
ending its turn. If the sluggish process is in fact alive, it can restore its low
timestamp and kill everybody else, allowing it to make progress until some
other process declares it dead again.
The simulation works because eventually the mechanism for detecting
dead processes stops suspecting live ones (using the technique described
above), so the live process with the winning timestamp finishes its operation
without interference. This allows the next process to proceed, and eventually
all live processes complete any operation they start, giving the wait-free
property.
The actual code is in Algorithm 26.3. It’s a rather long algorithm but
most of the details are just bookkeeping.
The preamble before entering PANIC mode is a fast-path computation
that allows a process that actually is running in isolation to skip testing
any timestamps or doing any extra work (except for the one register read of
PANIC). The assumption is that the constant B is set high enough that any
process generally will finish its operation in B steps without interference. If
there is interference, then the timestamp-based mechanism kicks in: we grab
a timestamp out of the convenient fetch-and-add register and start slugging
it out with the other processes.
(A side note: while the algorithm as presented in the paper assumes a
fetch-and-add register, any timestamp generator that delivers increasing val-
ues over time will work. So if we want to limit ourselves to atomic registers,
we could generate timestamps by taking snapshots of previous timestamps,
Stockmeyer [DLS88].
CHAPTER 26. OBSTRUCTION-FREEDOM 241

1 if ¬PANIC then
2 execute up to B steps of the underlying algorithm
3 if we are done then return
4 PANIC ← true // enter panic mode
5 myTimestamp ← fetchAndIncrement()
6 A[i] ← 1 // reset my activity counter
7 while true do
8 T [i] ← myTimestamp
9 minTimestamp ← myTimestamp; winner ← i
10 for j ← 1 . . . n, j 6= i do
11 otherTimestamp ← T [j]
12 if otherTimestamp < minTimestamp then
13 T [winner] ← ∞ // not looking so winning any more
14 minTimestamp ← otherTimestamp; winner ← j
15 else if otherTimestamp < ∞ do
16 T [j] ← ∞

17 if i = winner then
18 repeat
19 execute up to B steps of the underlying algorithm
20 if we are done then
21 T [i] ← ∞
22 PANIC ← false
23 return
24 else
25 A[i] ← A[i] + 1
26 PANIC ← true
27 until T [i] = ∞
28 repeat
29 a ← A[winner]
30 wait a steps
31 winnerTimestamp ← T [winner]
32 until a = A[winner] or winnerTimestamp 6= minTimestamp
33 if winnerTimestamp = minTimestamp then
34 T [winner] ← ∞ // kill winner for inactivity

Algorithm 26.3: Obstruction-freedom booster from [FLMS05]

CHAPTER 26. OBSTRUCTION-FREEDOM 242

adding 1, and appending process ids for tie-breaking.)

Once I have a timestamp, I try to knock all the higher-timestamp pro-
cesses out of the way (by writing ∞ to their timestamp registers). If I see a
smaller timestamp than my own, I’ll drop out myself (T [i] ← ∞), and fight
on behalf of its owner instead. At the end of the j loop, either I’ve decided
I am the winner, in which case I try to finish my operation (periodically
checking T [i] to see if I’ve been booted), or I’ve decided somebody else is
the winner, in which case I watch them closely and try to shut them down
if they are too slow (T [winner] ← ∞). I detect slow processes by inactivity
in A[winner]; similarly, I signal my own activity by incrementing A[i]. The
value in A[i] is also used as an increasing guess for the time between incre-
ments of A[i]; eventually this exceeds the R(B + O(1)) operations that I
execute between incrementing it.
We still need to prove that this all works. The essential idea is to show
that whatever process has the lowest timestamp finishes in a bounded num-
ber of steps. To do so, we need to show that other processes won’t be
fighting it in the underlying algorithm. Call a process active if it is in the
loop guarded by the “if i = winner” statement. Lemma 1 from the paper
states:
Lemma 26.3.1 ([FLMS05, Lemma 1]). If processes i and j are both active,
then T [i] = ∞ or T [j] = ∞.
Proof. Assume without loss of generality that i last set T [i] to myTimestamp
in the main loop after j last set T [j]. In order to reach the active loop, i
must read T [j]. Either T [j] = ∞ at this time (and we are done, since only j
can set T [j] < ∞), or T [j] is greater than i’s timestamp (or else i wouldn’t
think it’s the winner). In the second case, i sets T [j] = ∞ before entering
the active loop, and again the claim holds.

The next step is to show that if there is some process i with a minimum
timestamp that executes infinitely many operations, it increments A[i] in-
finitely often (thus eventually making the failure detector stop suspecting
it). This gives us Lemma 2 from the paper:
Lemma 26.3.2 ([FLMS05, Lemma 2]). Consider the set of all processes that
execute infinitely many operations without completing an operation. Suppose
this set is non-empty, and let i hold the minimum timestamp of all these
processes. Then i is not active infinitely often.
Proof. Suppose that from some time on, i is active forever, i.e., it never
leaves the active loop. Then T [i] < ∞ throughout this interval (or else i
CHAPTER 26. OBSTRUCTION-FREEDOM 243

leaves the loop), so for any active j, T [j] = ∞ by the preceding lemma. It
follows that any active T [j] leaves the active loop after B + O(1) steps of j
(and thus at most R(B + O(1)) steps of i). Can j re-enter? If j’s timestamp
is less than i’s, then j will set T [i] = ∞, contradicting our assumption. But
if j’s timestamp is greater than i’s, j will not decide it’s the winner and
will not re-enter the active loop. So now we have i alone in the active loop.
It may still be fighting with processes in the initial fast path, but since i
sets PANIC every time it goes through the loop, and no other process resets
PANIC (since no other process is active), no process enters the fast path after
some bounded number of i’s steps, and every process in the fast path leaves
after at most R(B + O(1)) of i’s steps. So eventually i is in the loop alone
forever—and obstruction-freedom means that it finishes its operation and
leaves. This contradicts our initial assumption that i is active forever.

So now we want to argue that our previous assumption that there exists
a bad process that runs forever without winning leads to a contradiction,
by showing that the particular i from Lemma 26.3.2 actually finishes (note
that Lemma 26.3.2 doesn’t quite do this—we only show that i finishes if it
stays active long enough, but maybe it doesn’t stay active).
Suppose i is as in Lemma 26.3.2. Then i leaves the active loop infinitely
often. So in particular it increments A[i] infinitely often. After some finite
number of steps, A[i] exceeds the limit R(B +O(1)) on how many steps some
other process can take between increments of A[i]. For each other process j,
either j has a lower timestamp than i, and thus finishes in a finite number of
steps (from the premise of the choice of i), or j has a higher timestamp than
i. Once we have cleared out all the lower-timestamp processes, we follow the
same logic as in the proof of Lemma 26.3.2 to show that eventually (a) i sets
T [i] < ∞ and PANIC = true, (b) each remaining j observes T [i] < ∞ and
PANIC = true and reaches the waiting loop, (c) all such j wait long enough
(since A[i] is now very big) that i can finish its operation. This contradicts
the assumption that i never finishes the operation and completes the proof.

26.3.1 Cost
If the parameters are badly tuned, the potential cost of this construction is
quite bad. For example, the slow increment process for A[i] means that the
time a process spends in the active loop even after it has defeated all other
processes can be as much as the square of the time it would normally take
to complete an operation alone—and every other process may pay R times
this cost waiting. This can be mitigated to some extent by setting B high
CHAPTER 26. OBSTRUCTION-FREEDOM 244

enough that a winning process is likely to finish in its first unmolested pass
through the loop (recall that it doesn’t detect that the other processes have
reset T [i] until after it makes its attempt to finish). An alternative might
be to double A[i] instead of incrementing it at each pass through the loop.
However, it is worth noting (as the authors do in the paper) that nothing
prevents the underlying algorithm from incorporating its own contention
management scheme to ensure that most operations complete in B steps
and PANIC mode is rarely entered. So we can think of the real function of
the construction as serving as a backstop to some more efficient heuristic
approach that doesn’t necessarily guarantee wait-free behavior in the worst
case.

26.4 Lower bounds for lock-free protocols

This material was not covered in 2014 and is out of date. In particular, the
results in the journal version of Ellen, Hendler, and Shavit’s paper [EHS12]
are a bit stronger than the conference version [FHS05]. This section still
mostly follows the conference version.
So far we have seen that obstruction-freedom buys us an escape from
the impossibility results that plague wait-free constructions, while still al-
lowing practical implementations of useful objects under plausible timing
assumptions. Yet all is not perfect: it is still possible to show non-trivial
lower bounds on the costs of these implementations in the right model. We
will present one of these lower bounds, the linear-contention lower bound of
Ellen, Hendler, and Shavit [EHS12].4 First we have to define what is meant
by contention.

26.4.1 Contention
A limitation of real shared-memory systems is that physics generally won’t
permit more than one process to do something useful to a shared object
at a time. This limitation is often ignored in computing the complexity of
a shared-memory distributed algorithm (and one can make arguments for
ignoring it in systems where communication costs dominate update costs
in the shared-memory implementation), but it is useful to recognize it if we
can’t prove lower bounds otherwise. Complexity measures that take the cost
of simultaneous access into account go by the name of contention.
4
The result first appeared in FOCS in 2005 [FHS05], with a small but easily fixed
bug in the definition of the class of objects the proof applies to. We’ll use the corrected
definition from the journal version.
CHAPTER 26. OBSTRUCTION-FREEDOM 245

The particular notion of contention used in the Ellen et al. paper is an

adaptation of the contention measure of Dwork, Herlihy, and Waarts [DHW97].
The idea is that if I access some shared object, I pay a price in memory
stalls for all the other processes that are trying to access it at the same time
but got in first. In the original definition, given an execution of the form
Aφ1 φ2 . . . φk φA0 , where all operations φi are applied to the same object as φ,
and the last operation in A is not, then φk incurs k memory stalls. Ellen et
al. modify this to only count sequences of non-trivial operations, where an
operation is non-trivial if it changes the state of the object in some states
(e.g., writes, increments, compare-and-swap—but not reads). Note that this
change only strengthens the bound they eventually prove, which shows that
in the worst case, obstruction-free implementations of operations on objects
in a certain class incur a linear number of memory stalls (possibly spread
across multiple base objects).

26.4.2 The class G

The Ellen et al. bound is designed to be as general as possible, so the
authors define a class G of objects to which it applies. As is often the case
in mathematics, the underlying meaning of G is “a reasonably large class
of objects for which this particular proof works,” but the formal definition
is given in terms of when certain operations of the implemented object are
affected by the presence or absence of other operations—or in other words,
when those other operations need to act on some base object in order to let
later operations know they occurred.
An object is in class G if it has some operation Op and initial state s
such that for any two processes p and q and every sequence of operations
AφA0 , where

1. φ is an instance of Op executed by p,

2. no operation in A or A0 is executed by p,

3. no operation in A0 is executed by q, and

4. no two operations in A0 are executed by the same process;

then there exists a sequence of operations Q by q such that for every sequence
HφH 0 where

1. HH 0 is an interleaving of Q and the sequences AA0 |r for each process

r,
CHAPTER 26. OBSTRUCTION-FREEDOM 246

2. H 0 contains no operations of q, and

3. no two operations in H 0 are executed by the same process;

then the return value of φ to p changes depending on whether it occurs after

Aφ or Hφ.
This is where “makes the proof work” starts looking like a much simpler
definition. The intuition is that deep in the guts of the proof, we are going
to be injecting some operations of q into an existing execution (hence adding
Q), and we want to do it in a way that forces q to operate on some object
that p is looking at (hence the need for Aφ to return a different value from
Hφ), without breaking anything else that is going on (all the rest of the
conditions). The reason for pulling all of these conditions out of the proof
into a separate definition is that we also want to be able to show that
particular classes of real objects satisfy the conditions required by the proof,
without having to put a lot of special cases into the proof itself.

Lemma 26.4.1. A mod-m fetch-and-increment object, with m ≥ n, is in

G.

Proof. This is a classic proof-by-unpacking-the-definition. Pick some exe-

cution AφA0 satisfying all the conditions, and let a be the number of fetch-
and-increments in A and a0 the number in A0 . Note a0 ≤ n − 2, since all
operations in A0 are by different processes.
Now let Q be a sequence of n − a0 − 1 fetch-and-increments by q, and let
HH 0 be an interleaving of Q and the sequences AA0 |r for each r, where H 0
includes no two operation of the same process and no operations at all of
q. Let h, h0 be the number of fetch-and-increments in H, H 0 , respectively.
Then h + h0 = a + a0 + (n − a0 − 1) = n + a − 1 and h0 ≤ n − 2 (since H 0
contains at most one fetch-and-increment for each process other than p and
q). This gives h ≥ (n + a + 1) − (n − 2) = a + 1 and h ≤ n + a − 1, and the
return value of φ after Hφ is somewhere in this range mod m. But none of
these values is equal to a mod m (that’s why we specified m ≥ n, although
as it turns out m ≥ n − 1 would have been enough), so we get a different
return value from Hφ than from Aφ.

As a corollary, we also get stock fetch-and-increment registers, since we

can build mod-m registers from them by taking the results mod m.
A second class of class-G objects is obtained from snapshot:
CHAPTER 26. OBSTRUCTION-FREEDOM 247

Lemma 26.4.2. Single-writer snapshot objects are in G.5

Proof. Let AφA0 be as in the definition, where φ is a scan operation. Let

Q consist of a single update operation by q that changes its segment. Then
in the interleaved sequence HH 0 , this update doesn’t appear in H 0 (it’s
forbidden), so it must be in H. Nobody can overwrite the result of the
update (single-writer!), so it follows that Hφ returns a different snapshot
from Aφ.

26.4.3 The lower bound proof

Theorem 26.4.3 ([EHS12, Theorem 5.2]). For any obstruction-free imple-
mentation of some object in class G from RMW base objects, there is an
execution in which some operation incurs n − 1 stalls.

We can’t do better than n − 1, because it is easy to come up with im-

plementations of counters (for example) that incur at most n − 1 stalls.
Curiously, we can even spread the stalls out in a fairly arbitrary way over
multiple objects, while still incurring at most n − 1 stalls. For example, a
counter implemented using a single counter (which is a RMW object) gets
exactly n−1 stalls if n−1 processes try to increment it at the same time, de-
laying the remaining process. At the other extreme, a counter implemented
by doing a collect over n − 1 single-writer registers (also RMW objects) gets
at least n − 1 stalls—distributed as one per register—if each register has a
write delivered to it while the reader waiting to read it during its collect. So
we have to allow for the possibility that stalls are concentrated or scattered
or something in between, as long as the total number adds up at least n − 1.
The proof supposes that the theorem is not true and then shows how to
boost an execution with a maximum number k < n−1 stalls to an execution
with k+1 stalls, giving a contradiction. (Alternatively, we can read the proof
as giving a mechanism for generating an (n − 1)-stall execution by repeated
boosting, starting from the empty execution.)
This is pretty much the usual trick: we assume that there is a class of
bad executions, then look for an extreme member of this class, and show
that it isn’t as extreme as we thought. In doing so, we can restrict our
attention to particularly convenient bad executions, so long as the existence
of some bad execution implies the existence of a convenient bad execution.
5
For the purposes of this lemma, “single-writer” means that each segment can be
written to by only one process, not that there is only one process that can execute update
operations.
CHAPTER 26. OBSTRUCTION-FREEDOM 248

Formally, the authors define a k-stall execution for process p as an exe-

cution Eσ1 . . . σi where E and σi are sequence of operations such that:

1. p does nothing in E,

2. Sets of processes Sj , j = 1 . . . i, whose union S = ij=1 Sj has size k,

S

are each covering objects Oj after E with pending non-trivial opera-

tions,

3. Each σj consists of p applying events by itself until it is about to apply

an event to Oj , after which each process in Sj accesses Oj , after which
p accesses Oj .

4. All processes not in S are idle after E,

5. p starts at most one operation of the implemented object in σ1 . . . σi ,

and

6. In every extension of E in which p and the processes in S don’t take

steps, no process applies a non-trivial event to any base object accessed
in σ1 . . . σi . (We will call this the weird condition below.)

So this definition includes both the fact that p incurs k stalls and some
other technical details that make the proof go through. The fact that p
incurs k stalls follows from observing that it incurs |Sj | stalls in each segment
σj , since all processes in Sj access Oj just before p does.
Note that the empty execution is a 0-stall execution (with i = 0) by the
definition. This shows that a k-stall execution exists for some k.
Note also that the weird condition is pretty strong: it claims not only
that there are no non-trivial operation on O1 . . . Oi in τ , but also that there
are no non-trivial operations on any objects accessed in σ1 . . . σi , which may
include many more objects accessed by p.6
We’ll now show that if a k-stall execution exists, for k ≤ n − 2, then
a (k + k 0 )-stall execution exists for some k 0 > 0. Iterating this process
eventually produces an (n − 1)-stall execution.
Start with some k-stall execution Eσ1 . . . σi . Extend this execution by
a sequence of operations σ in which p runs in isolation until it finishes its
operation φ (which it may start in σ if it hasn’t done so already), then each
process in S runs in isolation until it completes its operation. Now linearize
6
And here is where I screwed up in class on 2011-11-14, by writing the condition as
the weaker requirement that nobody touches O1 . . . Oi .
CHAPTER 26. OBSTRUCTION-FREEDOM 249

the high-level operations completed in Eσ1 . . . σi σ and factor them as AφA0

as in the definition of class G.
Let q be some process not equal to p or contained in any Sj (this is
where we use the assumption k ≤ n − 2). Then there is some sequence of
high-level operations Q of q such that Hφ does not return the same value as
Aφ for any interleaving HH 0 of Q with the sequences of operations in AA0
satisfying the conditions in the definition. We want to use this fact to shove
at least one more memory stall into Eσ1 . . . σi σ, without breaking any of
the other conditions that would make the resulting execution a (k + k 0 )-stall
execution.
Consider the extension τ of E where q runs alone until it finishes every
operation in Q. Then τ applies no nontrivial events to any base object
accessed in σ1 . . . σk , (from the weird condition on k-stall executions) and
the value of each of these base objects is the same after E and Eτ , and thus
is also the same after Eσ1 . . . σk and Eτ σ1 . . . σk .
Now let σ 0 be the extension of Eτ σ1 . . . σk defined analogously to σ: p
finishes, then each process in each Sj finishes. Let HφH 0 factor the lin-
earization of Eτ σ1 . . . σi σ 0 . Observe that HH 0 is an interleaving of Q and
the high-level operations in AA0 , that H 0 contains no operations by q (they
all finished in τ , before φ started), and that H 0 contains no two operations
by the same process (no new high-level operations start after φ finishes,
so there is at most one pending operation per process in S that can be
linearized after φ).
Now observe that q does some non-trivial operation in τ to some base
object accessed by p in σ. If not, then p sees the same responses in σ 0 and
in σ, and returns the same value, contradicting the definition of class G.
So does q’s operation in τ cause a stall in σ? Not necessarily: there
may be other operations in between. Instead, we’ll use the existence of q’s
operation to demonstrate the existence of at least one operation, possibly by
some other process we haven’t even encountered yet, that does cause a stall.
We do this by considering the set F of all finite extensions of E that are free
of p and S operations, and look for an operation that stalls p somewhere in
this infinitely large haystack.
Let Oi+1 be the first base object accessed by p in σ that is also accessed
by some non-trivial event in some sequence in F . We will show two things:
first, that Oi+1 exists, and second, that Oi+1 is distinct from the objects
O1 . . . Oi . The first part follows from the fact that τ is in F , and we have
just shown that τ contains a non-trivial operation (by q) on a base object
accessed by p in σ. For the second part, we use the weird condition on
k-stall executions again: since every extension of E in F is ({p} ∪ S)-free, no
CHAPTER 26. OBSTRUCTION-FREEDOM 250

process applies a non-trivial event to any base object accessed in σ1 . . . σi ,

which includes all the objects O1 . . . Oi .
You’ve probably guessed that we are going to put our stalls in on Oi+1 .
We choose some extension X from F that maximizes the number of processes
with simultaneous pending non-trivial operations on Oi+1 (we’ll call this set
of processes Si+1 and let |Si+1 | be the number k 0 > 0 we’ve been waiting for),
and let E 0 be the minimum prefix of X such that these pending operations
are still pending after EE 0 .
We now look at the properties of EE 0 . We have:

• EE 0 is p-free (follows from E being p-free and E 0 ∈ F , since everything

in F is p-free).

• Each process in Sj has a pending operation on Oj after EE 0 (it did

after E, and didn’t do anything in E 0 ).

This means that we can construct an execution EE 0 σ1 . . . σi σi+1 that

includes k + k 0 memory stalls, by sending in the same sequences σ1 . . . σi as
before, then appending a new sequence of events where (a) p does all of its
operations in σ up to its first operation on Oi+1 ; then (b) all the processes in
the set Si+1 of processes with pending events on Oi+1 execute their pending
events on Oi+1 ; then (c) p does its first access to Oi+1 from σ. Note that in
addition to giving us k + k 0 memory stalls, σi+1 also has the right structure
for a (k + k 0 )-stall execution. But there is one thing missing: we have to
show that the weird condition on further extensions still holds.
Specifically, letting S 0 = S∪Si+1 , we need to show that any ({p}∪S 0 )-free
extension α of EE 0 includes a non-trivial access to a base object accessed
in σ1 . . . σi+1 . Observe first that since α is ({p} ∪ S 0 )-free, then E 0 α is
({p} ∪ S)-free, and so it’s in F : so by the weird condition on Eσ1 . . . σi , E 0 α
doesn’t have any non-trivial accesses to any object with a non-trivial access
in σ1 . . . σi . So we only need to squint very closely at σi+1 to make sure it
doesn’t get any objects in there either.
Recall that σi+1 consists of (a) a sequence of accesses by p to objects
already accessed in σ1 . . . σi (already excluded); (b) an access of p to Oi+1 ;
and (c) a bunch of accesses by processes in Si+1 to Oi+1 . So we only need
to show that α includes no non-trivial accesses to Oi+1 . Suppose that it
does: then there is some process that eventually has a pending non-trivial
operation on Oi+1 somewhere in α. If we stop after this initial prefix α0 of α,
we get k 0 + 1 processes with pending operations on Oi+1 in EE 0 α0 . But then
E 0 α0 is an extension of E with k 0 + 1 processes with a simultaneous pending
operation on Oi+1 . This contradicts the choice of X to maximize k 0 . So
CHAPTER 26. OBSTRUCTION-FREEDOM 251

if our previous choice was in fact maximal, the weird condition still holds,
and we have just constructed a (k + k 0 )-stall execution. This concludes the
proof.

26.4.4 Consequences
We’ve just shown that counters and snapshots have (n − 1)-stall executions,
because they are in the class G. A further, rather messy argument (given
in the Ellen et al. paper) extends the result to stacks and queues, obtaining
a slightly weaker bound of n total stalls and operations for some process in
the worst case.7 In both cases, we can’t expect to get a sublinear worst-case
bound on time under the reasonable assumption that both a memory stall
and an actual operation takes at least one time unit. This puts an inherent
bound on how well we can handle hot spots for many practical objects, and
means that in an asynchronous system, we can’t solve contention at the
object level in the worst case (though we may be able to avoid it in our
applications).
But there might be a way out for some restricted classes of objects. We
saw in Chapter 21 that we could escape from the Jayanti-Tan-Toueg [JTT00]
lower bound by considering bounded objects. Something similar may hap-
pen here: the Fich-Herlihy-Shavit bound on fetch-and-increments requires
executions with n(n − 1)d + n increments to show n − 1 stalls for some fetch-
and-increment if each fetch-and-increment only touches d objects, and even
for d = log n this is already superpolynomial. The max-register construction
of a counter [AAC09] doesn’t help here, since everybody hits the switch bit
at the top of the max register, giving n − 1 stalls if they all hit it at the
same time. But there might be some better construction that avoids this.

26.4.5 More lower bounds

There are many more lower bounds one can prove on lock-free implemen-
tations, many of which are based on previous lower bounds for stronger
models. We won’t present these in class, but if you are interested, a good
place to start is [AGHK06].
7
This is out of date: Theorem 6.2 of [EHS12] gives a stronger result than what’s in
[FHS05].
CHAPTER 26. OBSTRUCTION-FREEDOM 252

26.5 Practical considerations

Also beyond the scope of what we can do, there is a paper by Fraser and
Harris [FH07] that gives some nice examples of the practical trade-offs in
choosing between multi-register CAS and various forms of software transac-
tional memory in implementing lock-free data structures.
Chapter 27

BG simulation

The Borowsky-Gafni simulation [BG93], or BG simulation for short, is

a deterministic, wait-free algorithm that allows t+1 processes to collectively
construct a simulated execution of a system of n > t processes of which t may
crash. For both the simulating and simulated system, the underlying shared-
memory primitives are atomic snapshots; these can be replaced by atomic
registers using any standard snapshot algorithm. The main consequence
of the BG simulation is that the question of what decision tasks can be
computed deterministically by an asynchronous shared-memory system that
tolerates t crash failures reduces to the question of what can be computed
by a wait-free system with exactly t + 1 processes. This is an easier problem,
and in principle can be determined exactly using the topological approach
described in Chapter 28.
The intuition for how this works is that the t + 1 simulating processes
solve a sequence of agreement problems to decide what the n simulated
processes are doing; these agreement problems are structured so that the
failure of a simulator stops at most one agreement. So if at most t of the
simulating processes can fail, only t simulated processes get stuck as well.
We’ll describe here a version of the BG simulation that appears in
a follow-up paper by Borowsky, Gafni, Lynch, and Rajsbaum [BGLR01].
This gives a more rigorous presentation of the mechanisms of the original
Borowsky-Gafni paper, and includes a few simplifications.

27.1 Safe agreement

The safe agreement mechanism performs agreement without running into
the FLP bound, by using termination condition: it is guaranteed to termi-

253
CHAPTER 27. BG SIMULATION 254

nate only if there are no failures by any process during an initial unsafe
section of its execution. Each process i starts the agreement protocol with a
proposei (v) event for its input value v. At some point during the execution
of the protocol, the process receives a notification safei , followed later (if
the protocol finishes) by a second notification agreei (v 0 ) for some output
value v 0 . It is guaranteed that the protocol terminates as long as all pro-
cesses continue to take steps until they receive the safe notification, and
that the usual validity (all outputs equal some input) and agreement (all
outputs equal each other) conditions hold. There is also a wait-free progress
condition that the safei notices do eventually arrive for any process that
doesn’t fail, no matter what the other processes do (so nobody gets stuck
in their unsafe section).
Pseudocode for a safe agreement object is given in Algorithm 27.1. This
is a translation of the description of the algorithim in [BGLR01], which is
specified at a lower level using I/O automata.

// proposei (v)
1 A[i] ← hv, ii
2 if snapshot(A) contains hj, 2i for some j 6= i then
// Back off
3 A[i] ← hv, 0i
4 else
// Advance
5 A[i] ← hv, 2i
// safei
6 repeat
7 s ← snapshot(A)
8 until s does not contain hj, 1i for any j
// agreei
9 return s[j].value where j is smallest index with s[j].level = 2
Algorithm 27.1: Safe agreement (adapted from [BGLR01])

The communication mechanism is a snapshot object containing a pair

A[i] = hvaluei , leveli i for each process i, initially h⊥, 0i. When a process
carries out proposei (v), it sets A[i] to hv, 1i, advancing to level 1. It then
looks around to see if anybody else is at level 2; if so, it backs off to 0, and
if not, it advances to 2. In either case it then spins until it sees a snapshot
with nobody at level 1, and agrees on the level-2 value with the smallest
index i.
CHAPTER 27. BG SIMULATION 255

The safei transition occurs when the process leaves level 1 (no matter
which way it goes). This satisfies the progress condition, since there is
no loop before this, and guarantees termination if all processes leave their
unsafe interval, because no process can then wait forever for the last 1 to
disappear.
To show agreement, observe that at least one process advances to level 2
(because the only way a process doesn’t is if some other process has already
advanced to level 2), so any process i that terminates observes a snapshot
s that contains at least one level-2 tuple and no level-1 tuples. This means
that any process j whose value is not already at level 2 in s can at worst
reach level 1 after s is taken. But then j sees a level-2 tuples and backs
off. It follows that any other process i0 that takes a later snapshot s0 that
includes no level-1 tuples sees the same level-2 tuples as i, and computes the
same return value. (Validity also holds, for the usual trivial reasons.)

27.2 The basic simulation algorithm

The basic BG simulation uses a single snapshot object A with t + 1 compo-
nents (one for each simulating process) and an infinite array of safe agree-
ment objects Sjr , where the output of Sjr represents the value sjr of the
r-th snapshot performed by simulated process j. Each component A[i] of A
is itself a vector of n components A[i][j], each of which is a tuple hv, ri rep-
resenting the value v that process i determines process j would have written
after taking its r-th snapshot.1
Each simulating process i cycles through all simulated processes j. Sim-
ulating one round of a particular process j involves four phases:

1. The process makes an initial guess for sjr by taking a snapshot of A and
taking the value with the largest round number for each component
A[−][k].

2. The process initiates the safe agreement protocol Sjr using this guess.
It continues to run Sjr until it leaves the unsafe interval.
1
The underlying assumption is that all simulated processes alternate between taking
snapshots and doing updates. This assumption is not very restrictive, because two snap-
shots with no intervening update are equivalent to two snapshots separated by an update
that doesn’t change anything, and two updates with no intervening snapshot can be re-
placed by just the second update, since the adversary could choose to schedule them
back-to-back anyway.
CHAPTER 27. BG SIMULATION 256

3. The process attempts to finish Sjr , by performing one iteration of the

loop from Algorithm 27.1. If this iteration doesn’t succeed, it moves on
to simulating j + 1 (but will come back to this phase for j eventually).

4. If Sjr terminates, the process computes a new value vjr for j to write
based on the simulated snapshot returned by Sjr , and updates A[i][j]
with hvjr , ri.

Actually implementing this while maintaining an abstraction barrier

around safe agreement is tricky. One approach might be to have each pro-
cess i manage a separate thread for each simulated process j, and wrap the
unsafe part of the safe agreement protocol inside a mutex just for threads
of i. This guarantees that i enters the unsafe part of any safe agreement
object on behalf of only one simulated j at a time, while preventing delays
in the safe part of Sjr from blocking it from finishing some other Sj 0 r0 .

27.3 Effect of failures

So now what happens if a simulating process i fails? This won’t stop any
other process i0 from taking snapshots on behalf of j, or from generating its
own values to put in A[i0 ][j]. What it may do is prevent some safe agreement
object Sjr from terminating. The termination property of Sjr means that
this can only occur if the failure occurs while i is in the unsafe interval for
Sjr —but since i is only in the unsafe interval for at most one Sjr at a time,
this stalls only one simulated process j. It doesn’t block any i0 , because any
other i0 is guaranteed to leave its own unsafe interval for Sjr after finitely
many steps, and though it may waste some effort waiting for Sjr to finish,
once it is in the safe interval it doesn’t actually wait for it before moving on
to other simulated j 0 .
It follows that each failure of a simulating process knocks out at most
one simulated process. So a wait-free system with t + 1 processes—and thus
at most t failures in the executions we care about—will produces at most t
failures inside the simulation.

27.4 Inputs and outputs

Two details not specified in the description above are how i determines
j’s initial input and how i determines its own outputs from the outputs
of the simulated processes. For the basic BG simulation, this is pretty
straightforwards: we use the safe agreement objects Sj0 to agree on j’s
CHAPTER 27. BG SIMULATION 257

input, after each i proposes its own input vector for all j based on its own
input to the simulator protocol. For outputs, i waits for at least n − t of the
simulated processes to finish, and computes its own output based on what
it sees.
One issue that arises here is that we can only use the simulation to
solve colorless tasks, which are decision problems where any process can
take the output of any other process without causing trouble.2 This works
for consensus or k-set agreement, but fails pretty badly for renaming. The
extended BG simulation, due to Gafni [Gaf09], solves this problem by
mapping each simulating process p to a specific simulated process qp , and
using a more sophisticated simulation algorithm to guarantee that qp doesn’t
crash unless p does. Details can be found in Gafni’s paper; there is also a
later paper by Imbs and Raynal [IR09] that simplifies some details of the
construction. Here, we will limit ourselves to the basic BG simulation.

27.5 Correctness of the simulation

To show that the simulation works, observe that we can extract a simulated
execution by applying the following rules:

1. The round-r write operation of j is represented by the first write tagged

with round r performed for j.

2. The round-r snapshot operation of j is represented by whichever snap-

shot operation wins Sjr .

The simulated execution then consists of a sequence of write and snap-

shot operations, with order of the operations determined by the order of
their representatives in the simulating execution, and the return values of
the snapshots determined by the return values of their representatives.
Because all processes that simulate a write for j in round r use the same
snapshots to compute the state of j, they all write the same value. So
the only way we get into trouble is if the writes included in our simulated
snapshots are inconsistent with the ordering of the simulated operations
defined above. Here the fact that each simulated snapshot corresponds to a
real snapshot makes everything work: when a process performs a snapshot
for Sjr , then it includes all the simulated write operations that happen
2
The term “colorless” here comes from use of colors to represent process ids in the topo-
logical approach described in Chapter 28. These colors aren’t really colors, but topologists
like coloring nodes better than assigning them ids.
CHAPTER 27. BG SIMULATION 258

before this snapshot, since the s-th write operation by k will be represented
in the snapshot if and only if the first instance of the s-th write operation
by k occurs before it. The only tricky bit is that process i’s snapshot for
Sjr might include some operations that can’t possibly be included in Sjr ,
like j’s round-r write or some other operation that depends on it. But this
can only occur if some other process finished Sjr before process i takes its
snapshot, in which case i’s snapshot will not win Sjr and will be discarded.

27.6 BG simulation and consensus

BG simulation was originally developed to attack k-set agreement, but (as
pointed out by Gafni [Gaf09]) it gives a particularly simple proof of the
impossibility of consensus with one faulty process. Suppose that we had
a consensus protocol that solved consensus for n > 1 processes with one
crash failure, using only atomic registers. Then we could use BG simulation
to get a wait-free consensus protocol for two processes. But it’s easy to
show that atomic registers can’t solve wait-free consensus, because (follow-
ing [LAA87]), we only need to do the last step of FLP that gets a contradic-
tion when moving from a bivalent C to 0-valent Cx or 1-valent Cy. We thus
avoid the complications that arise in the original FLP proof from having to
deal with fairness.
More generally, BG simulation means that increasing the number of
processes while keeping the same number of crash failures doesn’t let us
compute anything we couldn’t before. This gives a formal justification for
the slogan that the difference between distributed computing and parallel
computing is that in a distributed system, more processes can only make
things worse.
Chapter 28

Topological methods

Here we’ll describe some results applying topology to distributed computing,

mostly following a classic paper of Herlihy and Shavit [HS99]. This was one
of several papers [BG93, SZ00] that independently proved lower bounds on
k-set agreement [Cha93], which is a relaxation of consensus where we
require only that there are at most k distinct output values (consensus is
1-set agreement). These lower bounds had failed to succumb to simpler
techniques.

28.1 Basic idea

• Represent indistinguishability proofs using tools from topology.

• Typical indistinguishability proof:

– Show certain executions are indistinguishable to some process

(and thus that process produces same output in both executions).
– In general case, have a chain of schedules S1 , S2 , . . . , Sk such that
for each i there is some p with Si |p = Si+1 |p. The restriction
to p acts as an edge between points representing executions, and
we use the existence of a path of such edges as a proof that the
decision value in S1 is the same as in Sk , assuming all processes
must agree on the decision value.

• Topological version:

– Essentially the dual of the above: points are now individual

process states (or histories), and edges (and higher-dimensional

259
CHAPTER 28. TOPOLOGICAL METHODS 260

structures) represent consistent states of different processes (i.e.,

executions in which both states occur).
– Considering many possible states produces a simplicial com-
plex, a finite combinatorial structure used in topology to model
continuous surfaces.
– Properties of the simplicial complex resulting from some protocol
or problem specification can then be used to determine properties
of the underlying protocol or problem.
– Topologists know a lot of properties to look at.

28.2 k-set agreement

The motivating problem for much of this work was getting impossibility
results for k-set agreement, proposed by Chaudhuri [Cha93]. The k-set
agreement problem is similar to consensus, where each process starts with
an input and eventually returns a decision value that must be equal to some
process’s input, but the agreement condition is relaxed to require only that
the set of decision values include at most k values.
With k − 1 crash failures, it’s easy to build a k-set agreement algorithm:
wait until you seen n − k + 1 input values, then choose the smallest one you
see. This works because any value a process returns is necessarily among
the k smallest input values (including the k − 1 it didn’t see). Chaudhuri
conjectured that k-set agreement was not solvable with k failures, and gave
a proof of a partial result (analogous to the existence of an initial bivalent
configuration for consensus) based on Sperner’s Lemma [Spe28]. This is a
classic result in topology that says that certain colorings of the vertices of a
graph in the form of a triangle that has been divided into smaller triangles
necessarily contain a small triangle with three different colors on its corners.
This connection between k-set renaming and Sperner’s Lemma became the
basic idea behind each the three independent proofs of the conjecture that
appeared shortly thereafter [HS99, BG93, SZ00].
Our plan is to give a sufficient high-level description of the topologi-
cal approach that the connection between k-set agreement and Sperner’s
Lemma becomes obvious. It is possible to avoid this by approaching the
problem purely combinatorially, as is done in Section 16.3 of [AW04]. The
presentation there is obtained by starting with a topological argument and
getting rid of the topology (in fact, the proof in [AW04] contains a proof
of Sperner’s Lemma with the serial numbers filed off). The disadvantage of
this approach is that it obscures what is really going in and makes it harder
CHAPTER 28. TOPOLOGICAL METHODS 261

to obtain insight into how topological techniques might help for other prob-
lems. The advantage is that (unlike these notes) the resulting text includes
actual proofs instead of handwaving.

28.3 Representing distributed computations using

topology
Topology is the study of properties of shapes that are preserved by continu-
ous functions between their points that have continuous inverses, which get
the rather fancy name of homeomorphisms. A continuous function1 is one
that maps nearby points to nearby points. A homeomorphism is continuous
in both directions: this basically means that you can stretch and twist and
otherwise deform your object however you like, as long as you don’t tear
it (which would map nearby points on opposite sides of the tear to distant
points) or glue bits of it together (which turns into tearing when we look
at the inverse function). Topologists are particularly interested in showing
when there is no homeomorphism between two objects; the classic example
is that you can’t turn a sphere into a donut without damaging it, but you
can turn a donut into a coffee mug (with a handle).
Working with arbitrary objects embedded in umpteen-dimensional spaces
is messy, so topologists invented a finite way of describing certain well-
behaved objects combinatorially, by replacing ugly continuous objects like
spheres and coffee mugs with simpler objects pasted together in complex
ways. The simpler objects are simplexes, and the more complicated pasted-
together objects are called simplicial complexes. The nifty thing about
simplicial complexes is that they give a convenient tool for describing what
states or outputs of processes in a distributed algorithm are “compatible” in
some sense, and because topologists know a lot about simplicial complexes,
we can steal their tools to describe distributed algorithms.

28.3.1 Simplicial complexes and process states

The formal definition of a k-dimensional simplex is the convex closure of
(k + 1) points {x1 . . . xk+1 } in general position; the convex closure part
ai = 1 and each ai ≥ 0, and
P P
means the set of all points ai xi where
the general position part means that the xi are not all contained in some
1
Strictly speaking, a continuous function between metric spaces; there is an even more
general definition of continuity that holds for spaces that are too strange to have a con-
sistent notion of distance.
CHAPTER 28. TOPOLOGICAL METHODS 262

subspace of dimension (k − 1) or smaller (so that the simplex isn’t squashed

flat somehow). What this gives us is a body with (k + 1) corners and (k + 1)
faces, each of which is a (k − 1)-dimensional simplex (the base case is that
a 0-dimensional simplex is a point). Each face includes all but one of the
corners, and each corner is on all but one of the faces. So we have:

• 0-dimensional simplex: point.2

• 1-dimensional simplex: line segment with 2 endpoints (which are both

corners and faces).

• 2-dimensional simplex: triangle (3 corners with 3 1-dimensional sim-

plexes for sides).

• 3-dimensional simplex: tetrahedron (4 corners, 4 triangular faces).

• 4-dimensional simplex: 5 corners, 5 tetrahedral faces. It’s probably

best not to try to visualize this.

A simplicial complex is a bunch of simplexes stuck together; formally,

this means that we pretend that some of the corners (and any faces that
include them) of different simplexes are identical points. There are ways to
do this right using equivalence relations. But it’s easier to abstract out the
actual geometry and go straight to a combinatorial structure.
An (abstract) simplicial complex is just a collection of sets with the
property that if A is a subset of B, and B is in the complex, then A is also
in the complex (this means that if some simplex is included, so are all of
its faces, their faces, etc.). This combinatorial version is nice for reasoning
about simplicial complexes, but is not so good for drawing pictures.
The trick to using this for distributed computing problems is that we
are going to build simplicial complexes by letting points be process states
(or sometimes process inputs or outputs), each labeled with a process id,
and letting the sets that appear in the complex be those collections of
states/inputs/outputs that are compatible with each other in some sense.
For states, this means that they all appear in some global configuration
in some admissible execution of some system; for inputs and outputs, this
means that they are permitted combinations of inputs or outputs in the
specification of some problem.
2
For consistency, it’s sometimes convenient to define a point as having a single (−1)-
dimensional face defined to be the empty set. We won’t need to bother with this, since
0-dimensional simplicial complexes correspond to 1-process distributed systems, which are
amply covered in almost every other Computer Science class you have ever taken.
CHAPTER 28. TOPOLOGICAL METHODS 263

Example: For 2-process binary consensus with processes 0 and 1, the in-
put complex, which describes all possible combinations of inputs, consists
of the sets

{{}, {p0}, {q0}, {p1}, {q1}, {p0, q0}, {p0, q1}, {p1, q0}, {p1, q1}} ,

which we might draw like this:

p0 q0

q1 p1

Note that there are no edges from p0 to p1 or q0 to q1: we can’t have

two different states of the same process in the same global configuration.
The output complex, which describes the permitted outputs, is

{{}, {p0}, {q0}, {p1}, {q1}, {p0, q0}, {p1, q1}} .

As a picture, this omits two of the edges (1-dimensional simplexes) from the
input complex:
p0 q0

q1 p1

One thing to notice about this output complex is that it is not con-
nected: there is no path from the p0–q0 component to the q1–p1 compo-
nent.
Here is a simplicial complex describing the possible states of two pro-
cesses p and q, after each writes 1 to its own bit then reads the other process’s
bit. Each node in the picture is labeled by a sequence of process ids. The
first id in the sequence is the process whose view this node represents; any
other process ids are processes this first process sees (by seeing a 1 in the
other process’s register). So p is the view of process p running by itself,
while pq is the view of process p running in an execution where it reads q’s
register after q writes it.
p qp pq q
CHAPTER 28. TOPOLOGICAL METHODS 264

The edges express the constraint that if we both write before we read,
then if I don’t see your value you must see mine (which is why there is no
p–q edge), but all other combinations are possible. Note that this complex
is connected: there is a path between any two points.
Here’s a fancier version in which each process writes its input (and re-
members it), then reads the other process’s register (i.e., a one-round full-
information protocol). We now have final states that include the process’s
own id and input first, then the other process’s id and input if it is visible.
For example, p1 means p starts with 1 but sees a null and q0p1 means q
starts with 0 but sees p’s 1. The general rule is that two states are compat-
ible if p either sees nothing or q’s actual input and similarly for q, and that
at least one of p or q must see the other’s input. This gives the following
simplicial complex:

p0 q0p0 p0q0 q0

q1p0 p1q0

p0q1 q0p1

q1 p1q1 q1p1 p1

Again, the complex is connected.

The fact that this looks like four copies of the p–qp–pq–q complex pasted
into each edge of the input complex is not an accident: if we fix a pair of
inputs i and j, we get pi–qjpi–piqj–qj, and the corners are pasted together
because if p sees only p0 (say), it can’t tell if it’s in the p0/q0 execution or
the p0/q1 execution.
The same process occurs if we run a two-round protocol of this form,
where the input in the second round is the output from the first round.
Each round subdivides one edge from the previous round into three edges:
CHAPTER 28. TOPOLOGICAL METHODS 265

p−q

p − qp − pq − q

p − (qp)p − p(qp) − qp − (pq)(qp) − (qp)(pq) − pq − q(pq) − (pq)q − q

Here (pq)(qp) is the view of p after seeing pq in the first round and seeing
that q saw qp in the first round.

28.3.2 Subdivisions
In the simple write-then-read protocol above, we saw a single input edge turn
into 3 edges. Topologically, this is an example of a subdivision, where we
represent a simplex using several new simplexes pasted together that cover
exactly the same points.
Certain classes of protocols naturally yield subdivisions of the input
complex. The iterated immediate snapshot (IIS) model, defined by
Borowsky and Gafni [BG97], considers executions made up of a sequence
of rounds (the iterated part) where each round is made up of one or more
mini-rounds in which some subset of the processes all write out their cur-
rent views to their own registers and then take snapshots of all the registers
(the immediate snapshot part). The two-process protocols of the previous
section are special cases of this model.
Within each round, each process p obtains a view vp that contains the
previous-round views of some subset of the processes. We can represent the
views as a subset of the processes, which we will abbreviate in pictures by
putting the view owner first: pqr will be the view {p, q, r} as seen by p, while
qpr will be the same view as seen by q. The requirements on these views
are that (a) every process sees its own previous view: p ∈ vp for all p; (b)
all views are comparable: vp ⊆ vq or vq ⊆ vp ; and (c) if I see you, then I see
everything you see: q ∈ vp implies vq ⊆ vp . This last requirement is called
immediacy and follows from the assumption that writes and snapshots are
done in the same mini-round: if I see your write, then your snapshot takes
place no later than mine does.
The IIS model does not correspond exactly to a standard shared-memory
model (or even a standard shared-memory model augmented with cheap
snapshots). There are two reasons for this: standard snapshots don’t provide
CHAPTER 28. TOPOLOGICAL METHODS 266

immediacy, and standard snapshots allow processes to go back an perform

more than one snapshot on the same object. The first issue goes away if we
are looking at impossibility proofs, because the adversary can restrict itself
only to those executions that satisfy immediacy. The second issue is more
delicate, but Borowsky and Gafni demonstrate that any decision protocol
that runs in the standard model can be simulated in the IIS model, using a
variant of BG simulation.
For three processes, one round of immediate snapshots gives rise to the
simplicial complex depicted in Figure 28.1. The corners of the big triangle
are the solo views of processes that do their snapshots before anybody else
shows up. Along the edges of the big triangle are views corresponding to 2-
process executions, while in the middle are complete views of processes that
run late enough to see everything. Each little triangle corresponds to some
execution. For example, the triangle with corners p, qp, rpq corresponds to
a sequential execution where p sees nobody, q sees p, and r sees both p and
q. The triangle with corners pqr, qpr, and rpq is the maximally-concurrent
execution where all three processes write before all doing their snapshots:
here everybody sees everybody. It is not terribly hard to enumerate all
possible executions and verify that the picture includes all of them. In higher
dimension, the picture is more complicated, but we still get a subdivision
that preserves the original topological structure [BG97].
Figure 28.2 shows (part of) the next step of this process: here we have
done two iterations of immediate snapshot, and filled in the second-round
subdivisions for the p–qpr–rpq and pqr–qpr–rpq triangles. (Please imagine
similar subdivisions of all the other triangles that I was too lazy to fill in
by hand.) The structure is recursive, with each first-level triangle mapping
to an image of the entire first-level complex. As in the two-process case,
adjacent triangles overlap because the relevant processes don’t have enough
information; for example, the points on the qpr–rpq edge correspond to
views of q or r that don’t include p in round 2 and so can’t tell whether p
saw p or pqr in round 1.
The important feature of the round-2 complex (and the round-k complex
in general) is that it’s a triangulation of the original outer triangle: a
partition into little triangles where each corner aligns with corners of other
little triangles.
(Better pictures of this process in action can be found in Figures 25 and
26 of [HS99].)
CHAPTER 28. TOPOLOGICAL METHODS 267

rp qp

qpr rpq
pr pq

pqr

r q
qr rq

Figure 28.1: Subdivision corresponding to one round of immediate snapshot

CHAPTER 28. TOPOLOGICAL METHODS 268

rp qp

qpr rpq
pr pq

pqr

r q
qr rq

Figure 28.2: Subdivision corresponding to two rounds of immediate snapshot

CHAPTER 28. TOPOLOGICAL METHODS 269

2 1

1 1
1 3

2 3 3 3

Figure 28.3: An attempt at 2-set agreement

28.4 Impossibility of k-set agreement

Now let’s show that there is no way to do k-set agreement with n = k + 1
processes in the IIS model.
Suppose that after some fixed number of rounds, each process chooses
an output value. This output can only depend on the view of the process,
so is fixed for each vertex in the subdivision. Also, the validity condition
means that a process can only choose an output that it can see among the
inputs in its view. This means that at the corners of the outer triangle
(corresponding to views where the process thinks it’s alone), a process must
return its input, while along the outer edges (corresponding to views where
two processes may see each other but not the third), a process must return
one of the two inputs that appear in the corners incident to the edge. Internal
corners correspond to views that include—directly or indirectly—the inputs
of all processes, so these can be labeled arbitrarily. An example is given in
Figure 28.3, for a one-round protocol with three processes.
CHAPTER 28. TOPOLOGICAL METHODS 270

We now run into Sperner’s Lemma [Spe28], which says that, for any
subdivision of a simplex into smaller simplexes, if each corner of the original
simplex has a different color, and each corner that appears on some face of
the original simplex has a color equal to the color of one of the corners of
that face, then within the subdivision there are an odd number of simplexes
whose corners are all colored differently.3
How this applies to k-set agreement: Suppose we have n = k+1 processes
in a wait-free system (corresponding to allowing up to k failures). With
the cooperation of the adversary, we can restrict ourselves to executions
consisting of ` rounds of iterated immediate snapshot for some ` (termination
comes in here to show that ` is finite). This gives a subdivision of a simplex,
where each little simplex corresponds to some particular execution and each
corner some process’s view. Color all the corners of the little simplexes in
this subdivision with the output of the process holding the corresponding
view. Validity means that these colors satisfy the requirements of Sperner’s
Lemma. Sperner’s Lemma then says that some little simplex has all k + 1
colors, giving us a bad execution with more than k distinct output values.
The general result says that we can’t do k-set agreement with k failures
for any n > k. We haven’t proved this result, but it can be obtained from
the n = k + 1 version using a simulation of k + 1 processes with k failures
by n processes with k failures due to Borowsky and Gafni [BG93].

28.5 Simplicial maps and specifications

Let’s step back and look at consensus again.
3
The proof of Sperner’s Lemma is not hard, and is done by induction on the dimension
k. For k = 0, any subdivision consists of exactly one zero-dimensional simplex whose single
corner covers all k + 1 = 1 colors. For k + 1, suppose that the colors are {1, . . . , k + 1},
and construct a graph with a vertex for each little simplex in the subdivision and an
extra vertex for the region outside the big simplex. Put an edge in this graph between
each pair of regions that share a k-dimensional face with colors {1, . . . , k}. The induction
hypothesis tells us that there are an odd number of edges between the outer-region vertex
and simplexes on the {1, . . . , k}-colored face of the big simplex. The Handshaking Lemma
from graph theory says that the sum of the degrees of all the nodes in the graph is even.
But this can only happen if there are an even number of nodes with odd degree, implying
that the are are an odd number of simplexes in the subdivision with an odd number of
faces colored {1, . . . , k}.
Now suppose we have a simplex with at least one face f colored {1, . . . , k}. If the
opposite corner is colored c 6= k + 1, then it has exactly two faces colored {1, . . . , k}: f ,
and the face that replaces f ’s c-colored corner with the opposite corner. So the only way
to get an odd number of {1, . . . , k}-colored faces is to have all k + 1 colors. It follows that
there are an odd number of (k + 1)-colored simplexes.
CHAPTER 28. TOPOLOGICAL METHODS 271

One thing we could conclude from the fact that the output complex for
consensus was not connected but the ones describing our simple protocols
were was that we can’t solve consensus (non-trivially) using these protocols.
The reason is that to solve consensus using such a protocol, we would need
to have a mapping from states to outputs (this is just whatever rule tells
each process what to decide in each state) with the property that if some
collection of states are consistent, then the outputs they are mapped to are
consistent.
In simplical complex terms, this means that the mapping from states
to outputs is a simplicial map, a function f from points in one simplicial
complex C to points in another simplicial complex D such that for any
simplex A ∈ C, f (A) = {f (x)|x ∈ A} gives a simplex in D. (Recall that
consistency is represented by including a simplex, in both the state complex
and the output complex.) A mapping from states to outputs that satisfies
the consistency requirements encoded in the output complex s always a
simplicial map, with the additional requirement that it preserves process ids
(we don’t want process p to decide the output for process q). Conversely,
any id-preserving simplicial map gives an output function that satisfies the
consistency requirements.
Simplicial maps are examples of continuous functions, which have all
sorts of nice topological properties. One nice property is that a continuous
function can’t separate a connected space into disconnected components. We
can prove this directly for simplical maps: if there is a path of 1-simplexes
{x1 , x2 }, {x2 , x3 }, . . . {xk−1 , xk } from x1 to xk in C, and f : C → D is a
simplicial map, then there is a path of 1-simplexes {f (x1 ), f (x2 )}, . . . from
f (x1 ) to f (xk ). Since being connected just means that there is a path
between any two points,4 if C is connected we’ve just shown that f (C) is as
well.
Getting back to our consensus example, it doesn’t matter what simplicial
map f you pick to map process states to outputs; since the state complex C
is connected, so is f (C), so it lies entirely within one of the two connected
components of the output complex. This means in particular that everybody
always outputs 0 or 1: the protocol is trivial.

28.5.1 Mapping inputs to outputs

For general decision tasks, it’s not enough for the outputs to be consistent
with each other. They also have to be consistent with the inputs. This can
4
Technically, this is the definition of path-connected, which is the same as connected
for well-behaved topological spaces.
CHAPTER 28. TOPOLOGICAL METHODS 272

be expressed by a relation ∆ between input simplexes and output simplexes.

Formally, a decision task is modeled by a triple (I, O, ∆), where I is the
input complex, O is the output complex, and (A, B) ∈ ∆ if and only if B is
a permissible output given input I. Here there are no particular restrictions
on ∆ (for example, it doesn’t have to be a simplicial map or even a function),
but it probably doesn’t make sense to look at decision tasks unless there is
at least one permitted output simplex for each input simplex.

28.6 The asynchronous computability theorem

Given a decision task specified in this way, there is a topological characteriza-
tion of when it has a wait-free solution. This is given by the Asynchronous
Computability Theorem (Theorem 3.1 in [HS99]), which says:

Theorem 28.6.1. A decision task (I, O, ∆) has a wait-free protocol using

shared memory if and only if there exists a chromatic subdivision σ of I and
a color-preserving simplicial map µ : σ(I) → O such that for each simplex s
in σ(I), µ(S) ∈ ∆(carrier(S, I)).

To unpack this slightly, a chromatic subdivision is a subdivision where

each vertex is labeled by a process id (a color), and no simplex has two
vertices with the same color. A color-preserving simplicial map is a simplicial
map that preserves ids. The carrier of a simplex in a subdivision is whatever
original simplex it is part of. So the theorem says that I can only solve a
task if I can find a simplicial map from a subdivision of the input complex
to the output complex that doesn’t do anything strange to process ids and
that is consistent with ∆.
Looking just at the theorem, one might imagine that the proof consists
of showing that the protocol complex defined by the state complex after
running the protocol to completion is a subdivision of the input complex,
followed by the same argument we’ve seen already about mapping the state
complex to the output complex. This is almost right, but it’s complicated by
two inconvenient facts: (a) the state complex generally isn’t a subdivision of
the input complex, and (b) if we have a map from an arbitrary subdivision
of the input complex, it is not clear that there is a corresponding protocol
that produces this particular subdivision.
So instead the proof works like this:

Protocol implies map Even though we don’t get a subdivision with the
full protocol, there is a restricted set of executions that does give a
CHAPTER 28. TOPOLOGICAL METHODS 273

subdivision. So if the protocol works on this restricted set of execu-

tions, an appropriate map exists. There are two ways to prove this:
Herlihy and Shavit do so directly, by showing that this restricted set
of executions exists, and Borowksy and Gafni [BG97] do so indirectly,
by showing that the IIS model (which produces exactly the standard
chromatic subdivision used in the ACT proof) can simulate an ordi-
nary snapshot model. Both methods are a bit involved, so we will skip
over this part.

Map implies protocol This requires an algorithm. The idea here is that
that participating set algorithm, originally developed to solve k-set
agreement [BG93], produces precisely the standard chromatic subdivi-
sion used in the ACT proof. In particular, it can be used to solve the
problem of simplex agreement, the problem of getting the processes
to agree on a particular simplex contained within the subdivision of
their original common input simplex. This is a little easier to explain,
so we’ll do it.

28.6.1 The participating set protocol

Algorithm 28.1 depicts the participating set protocol; this first appeared
in [BG93], although the presentation here is heavily influenced by the version
in Elizabeth Borowsky’s dissertation [Bor95]. The shared data consists of
a snapshot object level, and processes start at a high level and float down
until they reach a level i such that there are already i processes at this level
or below. The set returned by a process consists of all processes it sees at
its own level or below, and it can be shown that this in fact implements a
one-shot immediate snapshot. Since immediate snapshots yield a standard
subdivision, this gives us what we want for converting a color-preserving
simplicial map to an actual protocol.

1 Initially, level[i] = n + 2 for all i.

2 repeat
3 level[i] ← level[i] − 1
4 v ← snapshot(level)
5 S ← {j | v[j] ≤ level[i]}
6 until |S| ≥ level[i]
7 |S| ≥ level[i] return S
Algorithm 28.1: Participating set
CHAPTER 28. TOPOLOGICAL METHODS 274

The following theorem shows that the return values from participating
set have all the properties we want for iterated immediate snapshot:

Theorem 28.6.2. Let Si be the output of the participating set algorithm for
process i. Then all of the following conditions hold:

1. For all i, i ∈ Si . (Self-containment.)

2. For all i, j, Si ⊆ Sj or Sj ⊆ Si . (Atomic snapshot.)

3. For all i, j, if i ∈ Sj , then Si ⊆ Sj . (Immediacy.)

Proof. Self-inclusion is trivial, but we will have to do some work for the
other two properties.
The first step is to show that Algorithm 28.1 neatly sorts the processes
out into levels, where each process that returns at level k returns precisely
the set of processes at level k and below.
For each process i, let Si be defined as above, let `i be the final value of
level[i] when i returns, and let Si0 = {j | `j ≤ Si }. Our goal is to show that
Si0 = Si , justifying the above claim.
Because no process ever increases its level, if process i observes level[j] ≤
`i in its last snapshot, then `j ≤ level[j] ≤ `i . So Si0 is a superset of Si . We
thus need to show only that no extra processes sneak in; in particular, we
will to show that Si = Si0 , by showing that both equal `i .
The first step is to show that |Si0 | ≥ |Si | ≥ `i . The first inequality follows
from the fact that Si0 ⊇ Si ; the second follows from the code (if not, i would
have stayed in the loop).
The second step is to show that |Si0 | ≤ `i . Suppose not; that is, suppose
that |Si0 | > `i . Then there are at least `i +1 processes with level `i or less, all
of which take a snapshot on level `i + 1. Let i0 be the last of these processes
to take a snapshot while on level `i + 1. Then i0 sees at least `i + 1 processes
at level `i + 1 or less and exits, contradicting the assumption that it reaches
level `i . So |Si0 | ≤ `i .
The atomic snapshot property follows immediately from the fact that if
`i ≤ `j , then `k ≤ `i implies `k ≤ `j , giving Si = Si0 ⊆ Sj0 = Sj . Similarly,
for immediacy we have that if i ∈ Sj , then `i ≤ `j , giving Si ≤ Sj by the
same argument.

The missing piece for turning this into IIS is that in Algorithm 28.1, I
only learn the identities of the processes I am supposed to include but not
their input values. This is easily dealt with by adding an extra register for
each process, to which it writes its input before executing participating set.
CHAPTER 28. TOPOLOGICAL METHODS 275

28.7 Proving impossibility results

To show something is impossible using the ACT, we need to show that there
is no color-preserving simplicial map from a subdivision of I to O satisfying
the conditions in ∆. This turns out to be equivalent to showing that there
is no continuous function from I to O with the same properties, because
any such simplicial map can be turned into a continuous function (on the
geometric version of I, which includes the intermediate points in addition
to the corners). Fortunately, topologists have many tools for proving non-
existence of continuous functions.

28.7.1 k-connectivity
Define the m-dimensional disk to be the set of all points at most 1 unit away
from the origin in Rm , and the m-dimensional sphere to be the surface of
the (m + 1)-dimensional disk (i.e., all points exactly 1 unit away from the
origin in Rm+1 ). Note that what we usually think of as a sphere (a solid
body), topologists call a disk, leaving the term sphere for just the outside
part.
An object is k-connected if any continuous image of an m-dimensional
sphere can be extended to a continuous image of an (m + 1)-dimensional
disk, for all m ≤ k.5 This is a roundabout way of saying that if we can
draw something that looks like a deformed sphere inside our object, we can
always include the inside as well: there are no holes that get in the way.
The punch line is that continuous functions preserve k-connectivity: if we
map an object with no holes into some other object, the image had better
not have any holes either.
Ordinary path-connectivity is the special case when k = 0; here, the
0-sphere consists of two points and the 1-disk is the path between them. So
0-connectivity says that for any two points, there is a path between them.
For 1-connectivity, if we draw a loop (a path that returns to its origin), we
can include the interior of the loop somewhere. One way to thinking about
this is to say that we can shrink the loop to a point without leaving the object
(the technical term for this is that the path is null-homotopic, where a
homotopy is a way to transform one thing continuously into another thing
over time and the null path sits on a single point). An object that is
1-connected is also called simply connected.
5
This definition is for the topological version of k-connectivity. It is not related in any
way to the definition of k-connectivity in graph theory, where a graph is k-connected if
there are k disjoint paths between any two points.
CHAPTER 28. TOPOLOGICAL METHODS 276

For 2-connectivity, we can’t contract a sphere (or box, or the surface of

a 2-simplex, or anything else that looks like a sphere) to a point.
The important thing about k-connectivity is that it is possible to prove
that any subdivision of a k-connected simplicial complex is also k-connected
(sort of obvious if you think about the pictures, but it can also be proved
formally), and that k-connectivity is preserved by simplicial maps (if not,
somewhere in the middle of all the k-simplexes representing our surface is a
(k + 1)-simplex in the domain that maps to a hole in the range, violating the
rule that simplicial maps map simplexes to simplexes). So a quick way to
show that the Asynchronous Computability Theorem implies that something
is not asynchronously computable is to show that the input complex is k-
connected and the output complex isn’t.

28.7.2 Impossibility proofs for specific problems

Here are some applications of the Asynchronous Computability Theorem
and k-connectivity:

Consensus There is no nontrivial wait-free consensus protocol for n ≥ 2

processes. Proof: The input complex is 1-connected, but the output
complex is not, and we need a map that covers the entire output
complex (by nontriviality).

k-set agreement There is no wait-free k-set agreement for n ≥ k + 1

processes. Proof: The output complex for k-set agreement is not k-
connected, because buried inside it are lots of (k + 1)-dimensional
holes corresponding to missing simplexes where all k + 1 processes
choose different values. But these holes aren’t present in the input
complex—it’s OK if everybody starts with different inputs—and the
validity requirements for k-set agreement force us to map the surfaces
of these non-holes around holes in the output complex. (This proof
actually turns into the Sperner’s Lemma proof if we fully expand the
claim about having to map the input complex around the hole.)

Renaming There is no wait-free renaming protocol with less than 2n − 1

output names for all n. The general proof of this requires showing that
with fewer names we get holes that are too big (and ultimately reduces
to Sperner’s Lemma); for the special case of n = 3 and m = 4, see
Figure 28.4, which shows how the output complex of renaming folds
up into the surface of a torus. This means that renaming for n = 3
CHAPTER 28. TOPOLOGICAL METHODS 277

b1 c2 a1 b2 c1 a2 b1

c3 a4 b3 c4 a3 b4 c3

a1 b2 c1 a2 b1 c2 a1

Figure 28.4: Output complex for renaming with n = 3, m = 4. Each vertex

is labeled by a process id (a, b, c) and a name (1, 2, 3, 4). Observe that the
left and right edges of the complex have the same sequence of labels, as do
the top and bottom edges; the complex thus folds up into a (twisted) torus.
(This is a poor imitation of part of [HS99, Figure 9].)

and m = 4 is exactly the same as trying to stretch a basketball into

an inner tube.
Chapter 29

Approximate agreement

Not covered in 2014.

The approximate agreement [DLP+ 86] or -agreement problem is
another relaxation of consensus where input and output values are real num-
bers, and a protocol is required to satisfy modified validity and agreement
conditions.
Let xi be the input of process i and yi its output. Then a protocol
satisfies approximate agreement if it satisfies:

Termination Every nonfaulty process eventually decides.

Validity Every process returns an output within the range of inputs. For-
mally, for all i, it holds that (minj xj ) ≤ yi ≤ (maxj xj ).

-agreement For all i and j, |i − j| ≤ .

Unlike consensus, approximate agreement has wait-free algorithms for

asynchronous shared memory, which we’ll see in §29.1). But a curious prop-
erty of approximate agreement is that it has no bounded wait-free algo-
rithms, even for two processes (see §29.2)

29.1 Algorithms for approximate agreement

Not only is approximate agreement solvable, it’s actually easily solvable, to
the point that there are many known algorithms for solving it.
We’ll use the algorithm of Moran [Mor95], mostly as presented in [AW04,

278
CHAPTER 29. APPROXIMATE AGREEMENT 279

Algorithm 54] but with a slight bug fix;1 pseudocode appears in Algo-
rithm 29.1.2
The algorithm carries out a sequence of asynchronous rounds in which
processes adopt new values, such that the spread of the vector of all values
Vr in round r, defined as spread Vr = max Vr − min Vr , drops by a factor of 2
per round. This is done by having each process choose a new value in each
round by taking the midpoint (average of min and max) of all the values it
sees in the previous round. Slow processes will jump to the maximum round
they see rather than propagating old values up from ancient rounds; this
is enough to guarantee that latecomer values that arrive after some process
writes in round 2 are ignored.
The algorithm uses a single snapshot object A to communicate, and each
process stores its initial input and a round number along with its current
preference. We assume that the initial values in this object all have round
number 0, and that log2 0 = −∞ (which avoids a special case in the termi-
nation test).

1 A[i] ← hxi , 1, xi i
2 repeat
3 hx01 , r1 , v1 i . . . hx0n , rn , vn i ← snapshot(A)
4 rmax ← maxj rj
5 v ← midpoint{vj | rj = rmax }
6 A[i] ← hxi , rmax + 1, vi
7 until rmax ≥ 2 and rmax ≥ log2 (spread({x0j })/)
8 return v
Algorithm 29.1: Approximate agreement

To show this works, we want to show that the midpoint operation guar-
antees that the spread shrinks by a factor of 2 in each round. Let Vr be the
set of all values v that are ever written to the snapshot object with round
1
The original algorithm from [AW04] does not include the test rmax ≥ 2. This allows
for bad executions in which process 1 writes its input of 0 in round 1 and takes a snapshot
that includes only its own input, after which process 2 runs the algorithm to completion
with input 1. Here process 2 will see 0 and 1 in round 1, and will write (1/2, 2, 1) to
A[2]; on subsequent iterations, it will see only the value 1/2 in the maximum round, and
after dlog2 (1/)e rounds it will decide on 1/2. But if we now wake process 1 up, it will
decided 0 immediately based on its snapshot, which includes only its own input and gives
spread(x) = 0. Adding the extra test prevents this from happening, as new values that
arrive after somebody writes round 2 will be ignored.
2
Showing that this particular algorithm works takes a lot of effort. If I were to do this
over, I’d probably go with a different algorithm due to Schenk [Sch95].
CHAPTER 29. APPROXIMATE AGREEMENT 280

number r. Let Ur ⊆ Vr be the set of values that are ever written to the snap-
shot object with round number r before some process writes a value with
round number r + 1 or greater; the intuition here is that Ur includes only
those values that might contribute to the computation of some round-(r + 1)
value.
Lemma 29.1.1. For all r for which Vr+1 is nonempty,
spread(Vr+1 ) ≤ spread(Ur )/2.
Proof. Let Uri be the set of round-r values observed by a process i in the
iteration in which it sees rmax = r in some iteration, if such an iteration
exists. Note that Uri ⊆ Ur , because if some value with round r + 1 or greater
is written before i’s snapshot, then i will compute a larger value for rmax .
Given two processes i and j, we can argue from the properties of snapshot
that either Uri ⊆ Urj or Urj ⊆ Uri . The reason is that if i’s snapshot comes
first, then j sees at least as many round-r values as i does, because the only
way for a round-r value to disappear is if it is replaced by a value in a later
round. But in this case, process j will compute a larger value for rmax and
will not get a view for round r. The same holds in reverse if j’s snapshot
comes first.
Observe that if Uri ⊆ Urj , then
midpoint(Uri ) − midpoint(Urj ) ≤ spread(Urj )/2.
This holds because midpoint(Uri ) lies within the interval min Urj , max Urj ,
 

and every point in this interval is within spread(Urj )/2 of midpoint(Urj ). The
same holds if Urj ⊆ Uri . So any two values written in round r + 1 are within
spread(Ur )/2 of each other.
In particular, the minimum and maximum values in Vr+1 are within
spread(Ur )/2 of each other, so spread(Vr+1 ) ≤ spread(Ur )/2.
Corollary 29.1.2. For all r ≥ 2 for which Vr is nonempty,
spread(Vr ) ≤ spread(U1 )/2r−1 .
Proof. By induction on r. For r = 2, this is just Lemma 29.1.1. For larger
r, use the fact that Ur−1 ⊆ Vr−1 and thus spread(Ur−1 ) ≤ spread(Vr−1 ) to
compute
spread(Vr ) ≤ spread(Ur−1 )/2
≤ spread(Vr−1 )/2
≤ (spread(U1 )/2r−2 )/2
= spread(U1 )/2r−1 .
CHAPTER 29. APPROXIMATE AGREEMENT 281

Let i be some process that finishes in the fewest number of rounds. Pro-
cess i can’t finish until it reaches round rmax +1, where rmax ≥ log2 (spread({x0j })/)
for a vector of input values x0 that it reads after some process writes round
2 or greater. We have spread({x0j }) ≥ spread(U1 ), because every value in
U1 is included in x0 . So rmax ≥ log2 (spread(U1 )/) and spread(Vrmax +1 ) ≤
spread(U1 )/2rmax ≤ spread(U1 )/(spread(U1 )/) = . Since any value re-
turned is either included in Vrmax +1 or some later Vr0 ⊆ Vrmax +1 , this gives
us that the spread of all the outputs is less than : Algorithm 29.1 solves
approximate agreement.
The cost of Algorithm 29.1 depends on the cost of the snapshot oper-
ations, on , and on the initial input spread D. For linear-cost snapshots,
this works out to O(n log(D/)).

29.2 Lower bound on step complexity

The dependence on D/ is necessary, at least for deterministic algorithms.
Here we give a lower bound due to Herlihy [Her91a], which shows that
any deterministic approximate agreement algorithm takes at least log3 (D/)
total steps even with just two processes.
Define the preference of a process in some configuration as the value it
will choose if it runs alone starting from this configuration. The preference
of a process p is well-defined because the process is deterministic; it also can
only change as a result of a write operation by another process q (because
no other operations are visible to p, and p’s own operations can’t change
its preference). The validity condition means that in an initial state, each
process’s preference is equal to its input.
Consider an execution with two processes p and q, where p starts with
preference p0 and q starts with preference q0 . Run p until it is about to
perform a write that would change q’s preference. Now run q until it is
about to change p’s preference. If p’s write no longer changes q’s preference,
start p again and repeat until both p and q have pending writes that will
change the other process’s preference. Let p1 and q1 be the new preferences
that result from these operations. The adversary can now choose between
running P only and getting to a configuration with preferences p0 and q1 ,
Q only and getting p1 and q0 , or both and getting p1 and q1 ; each of these
choices incurs at least one step. By the triangle inequality, |p0 − q0 | ≤
|p0 − q1 | + |q1 − p1 | + |p1 − q0 |, so at least on of these configurations has
a spread between preferences that is at least 1/3 of the initial spread. It
CHAPTER 29. APPROXIMATE AGREEMENT 282

follows that after k steps the best spread we can get is D/3k , requiring
k ≥ log3 (D/) steps to get -agreement.
Herlihy uses this result to show that there are decisions problems that
have wait-free but not bounded wait-free deterministic solutions using regis-
ters. Curiously, the lower bound says nothing about the dependence on the
number of processes; it is conceivable that there is an approximate agree-
ment protocol with running time that depends only on D/ and not n.
 Part III

Direct interaction

283
Chapter 30

Self-stabilization

[[[ classic self-stabilization results. These don’t really fit in either

message-passing or shared-memory because the guarded-command
approach allows simultaneous update of neighbors. ]]]

284
Chapter 31

Population protocols

[[[ import http://pine.cs.yale.edu/pinewiki/PopulationProtocols

[[[ pinewiki PopulationProtocols ]]] ]]]
[[[ update for more recent results ]]]

285
Chapter 32

Chemical reaction networks

[[[ say something about chemical reaction networks, maybe merge

with population protocols ]]]

286
Chapter 33

Mobile robots

[[[ say something about mobile robots ]]]

[[[ does ANTS fit in here? ]]]

287
Chapter 34

Self-assembly

[[[ say something about self-assembly ]]]

288
Appendix

289
Appendix A

Assignments

Assignments should be submitted using the drop box mechanism in class-

esv2. To use this, you will need to log into https://classesv2.yale.edu/
using your NetId, then click on the “Drop Box” tool in the list of tools on
the left. You should then be shown a folder with your name on it, to which
you can upload files using the “Add” pulldown menu.
Assignments should be submitted in PDF format. How you get your
assignment into this format is up to you. I personally recommend learning
LATEX if you haven’t already, but exporting to PDF from your favorite word
processor or scanning in legibly hand-written solutions are also sensible op-
tions. Assignments that are not submitted in PDF format may be treated
unjustly. I would also recommend giving your assignment an informative
filename to avoid confusion.
Questions about assignments can be sent to the instructor directly at
james.aspnes@gmail.com, or submitted to the course Piazza site at http:
//piazza.com/yale/spring2016/cpsc465.

A.1 Assignment 1: due Wednesday, 2016-02-17,

at 5:00pm
[[[ To be announced. ]]]

A.2 Assignment 2: due Wednesday, 2016-03-09,

at 5:00pm
[[[ To be announced. ]]]

290
APPENDIX A. ASSIGNMENTS 291

A.3 Assignment 3: due Wednesday, 2016-04-20,

at 5:00pm
[[[ To be announced. ]]]
Appendix B

Sample assignments from

Spring 2014

B.1 Assignment 1: due Wednesday, 2014-01-29, at

5:00pm
Bureaucratic part
Send me email! My address is james.aspnes@gmail.com.
In your message, include:

1. Your name.

2. Your status: whether you are an undergraduate, grad student, auditor,

etc.

3. Anything else you’d like to say.

(You will not be graded on the bureaucratic part, but you should do it
anyway.)

B.1.1 Counting evil processes

A connected bidirectional asynchronous network of n processes with iden-
tities has diameter D and may contain zero or more evil processes. Fortu-
nately, the evil processes, if they exist, are not Byzantine, fully conform to
RFC 3514 [Bel03], and will correctly execute any code we provide for them.
Suppose that all processes wake up at time 0 and start whatever protocol
we have given them. Suppose that each process initially knows whether it is

292
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 293

evil, and knows the identities of all of its neighbors. However, the processes
do not know the number of processes n or the diameter of the network D.
Give a protocol that allows every process to correctly return the number
of evil processes no later than time D. Your protocol should only return a
value once for each process (no converging to the correct answer after an
initial wrong guess).

Solution
There are a lot of ways to do this. Since the problem doesn’t ask about
message complexity, we’ll do it in a way that optimizes for algorithmic sim-
plicity.
At time 0, each process initiates a separate copy of the flooding algorithm
(Algorithm 4.1). The message hp, N (p), ei it distributes consists of its own
identity, the identities of all of its neighbors, and whether or not it is evil.
In addition to the data for the flooding protocol, each process tracks a
set I of all processes it has seen that initiated a protocol and a set N of all
processes that have been mentioned as neighbors. The initial values of these
sets for process p are {p} and N (p), the neighbors of p.
Upon receiving a message hq, N (q), ei, a process adds q to I and N (q) to
N . As soon as I = N , the process returns a count of all processes for which
e = true.
Termination by D: Follows from the same analysis as flooding. Any
process at distance d from p has p ∈ I by time d, so I is complete by time
D.
S
Correct answer: Observe that N = i∈I N (i) always. Suppose that
there is some process q that is not in I. Since the graph is connected, there
is a path from p to q. Let r be the last node in this path in I, and let s be
the following node. Then s ∈ N \ I and N 6= I. By contraposition, if I = N
then I contains all nodes in the network, and so the count returned at this
time is correct.

B.1.2 Avoiding expensive processes

Suppose that you have a bidirectional but not necessarily complete asyn-
chronous message-passing network represented by a graph G = (V, E) where
each node in V represents a process and each edge in E connects two pro-
cesses that can send messages to each other. Suppose further that each
process is assigned a weight 1 or 2. Starting at some initiator process, we’d
like to construct a shortest-path tree, where each process points to one of
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 294

its neighbors as its parent, and following the parent pointers always gives a
path of minimum total weight to the initiator.1
Give a protocol that solves this problem with reasonable time, message,
and bit complexity, and show that it works.

Solution
There’s an ambiguity in the definition of total weight: does it include the
weight of the initiator and/or the initial node in the path? But since these
values are the same for all paths to the initiator from a given process, they
don’t affect which is lightest.
If we don’t care about bit complexity, there is a trivial solution: Use an
existing BFS algorithm followed by convergecast to gather the entire struc-
ture of the network at the initiator, run your favorite single-source shortest-
path algorithm there, then broadcast the results. This has time complexity
O(D) and message complexity O(DE) if we use the BFS algorithm from
§5.3. But the last couple of messages in the convergecast are going to be
pretty big.
A solution by reduction: Suppose that we construct a new graph G0
where each weight-2 node u in G is replaced by a clique of nodes u1 , u2 , . . . uk ,
with each node in the clique attached to a different neighbor of u. We then
run any breadth-first search protocol of our choosing on G0 , where each
weight-2 node simulates all members of the corresponding clique. Because
any path that passes through a clique picks up an extra edge, each path in
the breadth-first search tree has a length exactly equal to the sum of the
weights of the nodes other than its endpoints.
A complication is that if I am simulating k nodes, between them they
may have more than one parent pointer. So we define u.parent to be ui .parent
where ui is a node at minimum distance from the initiator in G0 . We also
re-route any incoming pointers to uj 6= ui to point to ui instead. Because
ui was chosen to have minimum distance, this never increases the length of
any path, and the resulting modified tree is a still a shortest-path tree.
Adding nodes blows up |E 0 |, but we don’t need to actually send messages
between different nodes ui represented by the same process. So if we use the
§5.3 algorithm again, we only send up to D messages per real edge, giving
O(D) time and O(DE) messages.
1
Clarification added 2014-01-26: The actual number of hops is not relevant for the
construction of the shortest-path tree. By shortest path, we mean path of minimum total
weight.
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 295

If we don’t like reductions, we could also tweak one of our existing al-
gorithms. Gallager’s layered BFS (§5.2) is easily modified by changing the
depth bound for each round to a total-weight bound. The synchronizer-
based BFS can also be modified to work, but the details are messy.

B.2 Assignment 2: due Wednesday, 2014-02-12, at

5:00pm
B.2.1 Synchronous agreement with weak failures
Suppose that we modify the problem of synchronous agreement with crash
failures from Chapter 9 so that instead of crashing a process forever, the
adversary may jam some or all of its outgoing messages for a single round.
The adversary has limited batteries on its jamming equipment, and can
only cause f such one-round faults. There is no restriction on when these
one-round jamming faults occur: the adversary might jam f processes for
one round, one process for f rounds, or anything in between, so long as the
sum over all rounds of the number of processes jammed in each round is at
most f . For the purposes of agreement and validity, assume that a process
is non-faulty if it is never jammed.2
As a function of f and n, how many rounds does it take to reach agree-
ment in the worst case in this model, under the usual assumptions that
processes are deterministic and the algorithm must satisfy agreement, ter-
mination, and validity? Give the best upper and lower bounds that you
can.

Solution
√
The par solution for this is an Ω( f ) lower bound and O(f ) upper bound.
I don’t know if it is easy to do better than this.
For the lower bound, observe that the adversary can simulate an ordinary
crash failure by jamming a process in every round starting in the round it
crashes in. This means that in an r-round protocol, we can simulate k crash
failures with kr jamming faults. From the Dolev-Strong lower bound [DS83]
(see also Chapter 9), we know that there is no r-round protocol with k = r
crash failures faults, so there is no r-round protocol with r2 jamming faults.
2
Clarifications added 2014-02-10: We assume that processes don’t know that they are
being jammed or which messages are lost (unless the recipient manages to tell them that
a message was not delivered). As in the original model, we assume a complete network
and that all processes have known identities.
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 296

√ 
This gives a lower bound of f + 1 on the number of rounds needed to
solve synchronous agreement with f jamming faults.3
For the upper bound, have every process broadcast its input every round.
After f +1 rounds, there is at least one round in which no process is jammed,
so every process learns all the inputs and can take, say, the majority value.

B.2.2 Byzantine agreement with contiguous faults

Suppose that we restrict the adversary in Byzantine agreement to corrupting
a connected subgraph of the network; the idea is that the faulty nodes need
to coordinate, but can’t relay messages through the non-faulty nodes to do
so.
Assume the usual model for Byzantine agreement with a network in the
form of an m × m torus. This means that each node has a position (x, y)
in {0, . . . , m − 1} × {0, . . . , m − 1}, and its neighbors are the four nodes
(x + 1 mod m, y), (x − 1 mod m, y), (x, y + 1 mod m), and (x, y − 1 mod m).
For sufficiently large m,4 what is the largest number of faults f ; that
this system can tolerate and still solve Byzantine agreement?

Solution
The relevant bound here is the requirement that the network have enough
connectivity that the adversary can’t take over half of a vertex cut (see
§10.1.3). This is complicated slightly by the requirement that the faulty
nodes be contiguous.
The smallest vertex cut in a sufficiently large torus consists of the four
neighbors of a single node; however, these nodes are not connected. But we
can add a third node to connect two of them (see Figure B.1).
By adapting the usual lower bound we can use this construction to show
that f = 3 faults are enough to prevent agreement when m ≥ 3. The
question then is whether f = 2 faults is enough.
By a case analysis, we can show that any two nodes in a sufficiently
large torus are either adjacent themselves or can be connected by three
paths, where no two paths have adjacent vertices. Assume without loss of
generality that one of the nodes is at position (0, 0). Then any other node
is covered by one of the following cases:
3
Since Dolev-Strong only needs to crash one process per round, we don’t really need
the full r jamming faults for processes that crash late. This could be used to improve the
constant for this argument.
4
Problem modified 2014-02-03. In the original version, it asked to compute f for all
m, but there are some nasty special cases when m is small.
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 297

Figure B.1: Connected Byzantine nodes take over half a cut

1. Nodes adjacent to (0, 0). These can communicate directly.

2. Nodes at (0, i) or (i, 0). These cases are symmetric, so we’ll describe
the solution for (0, i). Run one path directly north: (0, 1), (0, 2), . . . , (0, i−
1). Similarly, run a second path south: (0, −1), (0, −2), . . . (0, i + 1).
For the third path, take two steps east and then run north and back
west: (1, 0), (2, 0), (2, 1), (2, 2), . . . , (2, i), (1, i). These paths are all
non-adjacent as long as m ≥ 4.

3. Nodes at (±1, i) or (i, ±1), where i is not −1, 0, or 1. Suppose the node
is at (1, i). Run one path east then north through (1, 0), (1, 1), . . . , (1, i−
1). The other two paths run south and west, with a sideways jog in the
middle as needed. This works for m sufficiently large to make room
for the sideways jogs.

4. Nodes at (±1, ±1) or (i, j) where neither of i or j is −1, 0, or 1. Now we

can run one path north then east, one east then north, one south then
west, and one west then south, creating four paths in a figure-eight
pattern centered on (0, 0).

B.3 Assignment 3: due Wednesday, 2014-02-26, at

5:00pm
B.3.1 Among the elect
The adversary has decided to be polite and notify each non-faulty processes
when he gives up crashing it. Specifically, we have the usual asynchronous
message-passing system with up to f faulty processes, but every non-faulty
process is eventually told that it is non-faulty. (Faulty processes are told
nothing.)
For what values of f can you solve consensus in this model?
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 298

Solution
We can tolerate f < n/2, but no more.
If f < n/2, the following algorithm works: Run Paxos, where each
process i waits to learn that it is non-faulty, then acts as a proposer for
proposal number i. The highest-numbered non-faulty process then carries
out a proposal round that succeeds because no higher proposal is ever issued,
and both the proposer (which is non-faulty) and a majority of accepters
participate.
If f ≥ n/2, partition the processes into two groups of size bn/2c, with
any leftover process crashing immediately. Make all of the processes in both
groups non-faulty, and tell each of them this at the start of the protocol.
Now do the usual partitioning argument: Run group 0 with inputs 0 with no
messages delivered from group 1 until all processes decide 0 (we can do this
because the processes can’t distinguish this execution from one in which
the group 1 processes are in fact faulty). Run group 1 similarly until all
processes decide 1. We have then violated agreement, assuming we didn’t
previously violate termination of validity.

B.3.2 Failure detectors on the cheap

Suppose we do not have the budget to equip all of our machines with failure
detectors. Instead, we order an eventually strong failure detector for k
machines, and the remaining n − k machines get fake failure detectors that
never suspect anybody. The choice of which machines get the real failure
detectors and which get the fake ones is under the control of the adversary.
This means that every faulty process is eventually permanently suspected
by every non-faulty process with a real failure detector, and there is at
least one non-faulty process that is eventually permanently not suspected
by anybody. Let’s call the resulting failure detector ♦Sk .
Let f be the number of actual failures. Under what conditions on f , k,
and n can you still solve consensus in the usual deterministic asynchronous
message-passing model using ♦Sk ?

Solution
First observe that ♦S can simulate ♦Sk for any k by having n − k processes
ignore the output of their failure detectors. So we need f < n/2 by the
usual lower bound on ♦S.
If f ≥ k, we are also in trouble. The f > k case is easy: If there exists
a consensus protocol for f > k, then we can transform it into a consensus
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 299

protocol for n − k processes and f − k failures, with no failure detectors

at all, by pretending that there are an extra k processes with real failure
detectors that crash immediately. The FLP impossibility result rules this
out.
If f = k, we have to be a little more careful. By immediately crashing
f − 1 processes with real failure detectors, we can reduce to the f = k = 1
case. Now the adversary runs the FLP strategy. If no processes crash, then
all n − k + 1 surviving process report no failures; if it becomes necessary to
crash a process, this becomes the one remaining process with the real failure
detector. In either case the adversary successfully prevents consensus.
So let f < k. Then we have weak completeness, because every faulty
process is eventually permanently suspected by at least k − f > 0 processes.
We also have weak accuracy, because it is still the case that some process
is eventually permanently never suspected by anybody. By boosting weak
completeness to strong completeness as described in §13.2.3, we can turn
out failure detector into ♦S, meaning we can solve consensus precisely when
f < min(k, n/2).

B.4 Assignment 4: due Wednesday, 2014-03-26, at

5:00pm
B.4.1 A global synchronizer with a global clock
Consider an asynchronous message-passing system with n processes in a
bidirectional ring with no failures. Suppose that the processes are equipped
with a global clock, which causes a local event to occur simultaneously at
each process every c time units, where as usual 1 is the maximum message
delay. We would like to use this global clock to build a global synchronizer.
Provided c is at least 1, a trivial approach is to have every process advance
to the next round whenever the clock pulse hits. This gives one synchronous
round every c time units.
Suppose that c is greater than 1 but still o(n). Is it possible to build
a global synchronizer in this model that runs more than a constant ratio
faster than this trivial global synchronizer in the worst case?

B.4.2 A message-passing counter

A counter is a shared object that support operations inc and read, where
read returns the number of previous inc operations.
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 300

Algorithm B.1 purports to implement a counter in an asynchronous

message-passing system subject to f < n/2 crash failures. In the algo-
rithm, each process i maintains a vector ci of contributions to the counter
from all the processes, as well as a nonce ri used to distinguish responses to
different read operations from each other. All of these values are initially
zero.
Show that the implemented counter is linearizable, or give an example
of an execution where it isn’t.

1 procedure inc
2 ci [i] ← ci [i] + 1
3 Send ci [i] to all processes.
4 Wait to receive ack(ci [i]) from a majority of processes.
5 upon receiving c from j do
6 ci [j] ← max(ci [j], c)
7 Send ack(c) to j.
8 procedure read
9 ri ← ri + 1
10 Send read(ri ) to all processes.
11 Wait to receive respond(ri , cj ) from a majority of processes j.
P
12 return k maxj cj [k]
13 upon receiving read(r) from j do
14 Send respond(r, ci ) to j
Algorithm B.1: Counter algorithm for Problem B.4.2.

B.5 Assignment 5: due Wednesday, 2014-04-09, at

5:00pm
B.5.1 A concurrency detector
Consider the following optimistic mutex-like object, which we will call a
concurrency detector. A concurrency detector supports two operations
for each process i, enteri and exiti . These operations come in pairs: a
process enters a critical section by executing enteri , and leaves by executing
exiti . The behavior of the object is undefined if a process calls enteri twice
without an intervening exiti , or calls exiti without first calling enteri .
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 301

Unlike mutex, a concurrency detector does not enforce that only one
process is in the critical section at a time; instead, exiti returns 1 if the
interval between it and the previous enteri overlaps with some interval
between a enterj and corresponding exitj for some j 6= i, and returns 0 if
there is no overlap.
Is there a deterministic linearizable wait-free implementation of a con-
currency detector from atomic registers? If there is, give an implementation.
If there is not, give an impossibility proof.

Solution
It is not possible to implement this object using atomic registers.
Suppose that there were such an implementation. Algorithm B.2 im-
plements two-process consensus using a two atomic registers and a single
concurrency detector, initialized to the state following enter1 .

1 procedure consensus1 (v)

2 r1 ← v
3 if exit1 () = 1 then
4 return r2
5 else
6 return v

7 procedure consensus2 (v)

8 r2 ← v
9 enter2 ()
10 if exit2 () = 1 then
11 return v
12 else
13 return r1

Algorithm B.2: Two-process consensus using the object from Prob-

lem B.5.1

Termination is immediate from the absence of loops in the code.

To show validity and termination, observe that one of two cases holds:

1. Process 1 executes exit1 before process 2 executes enter2 . In this

case there is no overlap between the interval formed by the implicit
enter1 and exit1 and the interval formed by enter2 and exit2 . So
the exit1 and exit2 operations both return 0, causing process 1 to
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 302

return its own value and process 2 to return the contents of r1 . These
will equal process 1’s value, because process 2’s read follows its call to
enter2 , which follows exit1 and thus process 1’s write to r1 .

2. Process 1 executes exit1 after process 2 executes enter2 . Now both

exit operations return 1, and so process 2 returns its own value while
process 1 returns the contents of r2 , which it reads after process 2
writes its value there.
In either case, both processes return the value of the first process to
access the concurrency detector, satisfying both agreement and validity.
This would give a consensus protocol for two processes implemented from
atomic registers, contradicting the impossibility result of Loui and Abu-
Amara [LAA87].

B.5.2 Two-writer sticky bits

A two-writer sticky bit is a sticky bit that can be read by any process,
but that can only be written to by two specific processes.
Suppose that you have an unlimited collection of two-writer sticky bits
for each pair of processes, plus as many ordinary atomic registers as you
need. What is the maximum number of processes for which you can solve
wait-free binary consensus?

Solution
If n = 2, then a two-writer sticky bit is equivalent to a sticky bit, so we can
solve consensus.
If n ≥ 3, suppose that we maneuver our processes as usual to a bivalent
configuration C with no bivalent successors. Then there are three pending
operations x, y, and z, that among them produce both 0-valent and 1-valent
configurations. Without loss of generality, suppose that Cx and Cy are both
0-valent and Cz is 1-valent. We now consider what operations these might
be.

1. If x and z apply to different objects, then Cxz = Czx must be both

0-valent and 1-valent, a contradiction. Similarly if y and z apply to
different objects. This shows that all three operations apply to the
same object O.

2. If O is a register, then the usual case analysis of Loui and Abu-

Amara [LAA87] gives us a contradiction.
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 303

3. If O is a two-writer sticky bit, then we can split cases further based

on z:

(a) If z is a read, then either:

i. At least one of x and y is a read. But then Cxz = Czx or
Cyz = Czy, and we are in trouble.
ii. Both x and y are writes. But then Czx (1-valent) is indis-
tinguishable from Cx (0-valent) by the two processes that
didn’t perform z: more trouble.
(b) If z is a write, then at least one of x or y is a read; suppose it’s
x. Then Cxz is indistinguishable from Cz by the two processes
that didn’t perform x.

Since we reach a contradiction in all cases, it must be that when n ≥ 3,

every bivalent configuration has a bivalent successor, which shows that we
can’t solve consensus in this case. The maximum value of n for which we
can solve consensus is 2.

B.6 Assignment 6: due Wednesday, 2014-04-23, at

5:00pm
B.6.1 A rotate register
Suppose that you are asked to implement a concurrent m-bit register that
supports in addition to the usual read and write operations a RotateLeft
operation that rotates all the bits to the left; this is equivalent to doing a
left shift (multiplying the value in the register by two) followed by replacing
the lowest-order bit with the previous highest-order bit.
For example, if the register contains 1101, and we do RotateLeft, it now
contains 1011.
Show that if m is sufficiently large as a function of the number of pro-
cesses n, Θ(n) steps per operation in the worst case are necessary and suffi-
cient to implement a linearizable wait-free m-bit shift register from atomic
registers.

Solution
The necessary part is easier, although we can’t use JTT (Chapter 20) di-
rectly because having write operations means that our rotate register is not
perturbable. Instead, we argue that if we initialize the register to 1, we
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 304

get a mod-m counter, where increment is implemented by RotateLeft and

read is implemented by taking the log of the actual value of the counter.
Letting m ≥ 2n gives the desired Ω(n) lower bound, since a mod-2n counter
is perturbable.
For sufficiency, we’ll show how to implement the rotate register using
snapshots. This is pretty much a standard application of known tech-
niques [AH90b, AM93], but it’s not a bad exercise to write it out.
Pseudocode for one possible solution is given in Algorithm B.3.
The register is implemented using a single snapshot array A. Each entry
in the snapshot array holds four values: a timestamp and process id indi-
cating which write the process’s most recent operations apply to, the initial
write value corresponding to this timestamp, and the number of rotate op-
erations this process has applied to this value. A write operation generates
a new timestamp, sets the written value to its input, and resets the rotate
count to 0. A rotate operation updates the timestamp and associated write
value to the most recent that the process sees, and adjusts the rotate count
as appropriate. A read operation combines all the rotate counts associated
with the most recent write to obtain the value of the simulated register.
Since each operation requires one snapshot and at most one update, the
cost is O(n) using the linear-time snapshot algorithm of Inoue et al. [IMCT94].
Linearizability is easily verified by observing that ordering all operations by
the maximum timestamp/process tuple that they compute and then by the
total number of rotations that they observe produces an ordering consis-
tent with the concurrent execution for which all return values of reads are
correct.

B.6.2 A randomized two-process test-and-set

Algorithm B.4 gives pseudocode for a protocol for two processes p0 and p1 .
It uses two shared unbounded single-writer atomic registers r0 and r1 , both
initially 0. Each process also has a local variable s.

1. Show that any return values of the protocol are consistent with a
linearizable, single-use test-and-set.

2. Will this protocol always terminate with probability 1 assuming an

oblivious adversary?

3. Will this protocol always terminate with probability 1 assuming an

adaptive adversary?
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 305

1 procedure write(A, v)
2 s ← snapshot(A)
3 A[id] ← hmaxi s[i].timestamp + 1, id, v, 0i
4 procedure RotateLeft(A)
5 s ← snapshot(A)
6 Let i maximize hs[i].timestamp, s[i].processi
7 if s[i].timestamp = A[id].timestamp and
s[i].process = A[id].process then
// Increment my rotation count
8 A[id].rotations ← A[id].rotations + 1
9 else
// Reset and increment my rotation count
10 A[id] ← hs[i].timestamp, s[i].process, s[i].value, 1i

11 procedure read(A)
12 s ← snapshot(A)
13 Let i maximize hs[i].timestamp, s[i].processi
14 Let
P
r = j,s[j].timestamp=s[i].timestamp∧s[j].process=s[i].process s[j].rotations
15 return s[i].value rotated r times.
Algorithm B.3: Implementation of a rotate register

1 procedure TASi ()
2 while true do
3 with probability 1/2 do
4 ri ← ri + 1
5 else
6 ri ← ri
7 s ← r¬i
8 if s > ri then
9 return 1
10 else if s < ri − 1 do
11 return 0

Algorithm B.4: Randomized two-process test-and-set for B.6.2

APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 306

Solution
1. To show that this implements a linearizable test-and-set, we need to
show that exactly one process returns 0 and the other 1, and that
if one process finishes before the other starts, the first process to go
returns 1.
Suppose that pi finishes before p¬i starts. Then pi reads only 0 from
r¬i , and cannot observe ri < r¬i : pi returns 0 in this case.
We now show that the two processes cannot return the same value.
Suppose that both processes terminate. Let i be such that pi reads r¬i
for the last time before p¬i reads ri for the last time. If pi returns 0,
then it observes ri ≥ r¬i + 2 at the time of its read; p¬i can increment
r¬i at most once before reading ri again, and so observed r¬i < ri and
returns 1.
Alternatively, if pi returns 1, it observed ri < r¬i . Since it performs
no more increments on ri , pi also observes ri < r¬i in all subsequent
reads, and so cannot also return 1.

2. Let’s run the protocol with an oblivious adversary, and track the value
of r0t − r1t over time, where rit is the value of ri after t writes (to either
register). Each write to r0 increases this value by 1/2 on average, with
a change of 0 or 1 equally likely, and each write to r1 decreases it by
1/2 on average.
To make things look symmetric, let ∆t be the change caused by the
t-th write and write ∆t as ct + X t where ct = ±1/2 is a constant
determined by whether p0 or p1 does the t-th write and X t = ±1/2 is
a random variable with expectation 0. Observe that the X t variables
are independent of each other and the constants ct (which depend only
on the schedule).
For the protocol to run forever, at every time t it must hold that
r0t − r1t ≤ 3; otherwise, even after one or both processes does its
0 0
next write, we will have r0t − r1t and the next process to read will
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 307

terminate. But
t
X
r0t − r1t = ∆s
s=1
Xt
= (cs + Xs )
s=1
Xt t
X
= cs + Xs .
s=1 s=1

The left-hand sum is a constant, while the right-hand sum has a bi-
nomial distribution. For any fixed constant, the probability that a
binomial distribution lands within ±2 of the constant goes to zero in
the limit as t → ∞, so with probability 1 there is some t for which
this event does not occur.

3. For an adaptive adversary, the following strategy prevents agreement:

(a) Run p0 until it is about to increment r0 .

(b) Run p1 until it is about to increment r1 .
(c) Allow both increments to proceed and repeat.

The effect is that both processes always observe r0 = r1 whenever

they do a read, and so never finish. This works because the adaptive
adversary can see the coin-flips done by the processes before they act
on them; it would not work with an oblivious adversary or in a model
that supported probabilistic writes.

B.7 CS465/CS565 Final Exam, May 2nd, 2014

Write your answers in the blue book(s). Justify your answers. Work alone.
Do not use any notes or books.
There are four problems on this exam, each worth 20 points, for a total
of 80 points. You have approximately three hours to complete this exam.

B.7.1 Maxima (20 points)

Some deterministic processes organized in an anonymous, synchronous ring
are each given an integer input (which may or may not be distinct from other
processes’ inputs), but otherwise run the same code and do not know the
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 308

size of the ring. We would like the processes to each compute the maximum
input. As usual, each process may only return an output once, and must do
so after a finite number of rounds, although it may continue to participate
in the protocol (say, by relaying messages) even after it returns an output.
Prove or disprove: It is possible to solve this problem in this model.

Solution
It’s not possible.
Consider an execution with n = 3 processes, each with input 0. If the
protocol is correct, then after some finite number of rounds t, each process
returns 0. By symmetry, the processes all have the same states and send
the same messages throughout this execution.
Now consider a ring of size 2(t + 1) where every process has input 0,
except for one process p that has input 1. Let q be the process at maximum
distance from p. By induction on r, we can show that after r rounds of
communication, every process that is more than r + 1 hops away from p has
the same state as all of the processes in the 3-process execution above. So
in particular, after t rounds, process q (at distance t + 1) is in the same state
as it would be in the 3-process execution, and thus it returns 0. But—as it
learns to its horror, one round too late—the correct maximum is 1.

B.7.2 Historyless objects (20 points)

Recall that a shared-memory object is historyless if any operation on the
object either (a) always leaves the object in the same state as before the
operation, or (b) always leaves the object in a new state that doesn’t depend
on the state before the operation.
What is the maximum possible consensus number for a historyless ob-
ject? That is, for what value n is it possible to solve wait-free consensus
for n processes using some particular historyless object but not possible to
solve wait-free consensus for n + 1 processes using any historyless object?

Solution
Test-and-sets are (a) historyless, and (b) have consensus number 2, so n is
at least 2.
To show that no historyless object can solve wait-free 3-process consen-
sus, consider an execution that starts in a bivalent configuration and runs
to a configuration C with two pending operations x and y such that Cx is
0-valent and Cy is 1-valent. By the usual arguments x and y must both be
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 309

operations on the same object. If either of x and y is a read operation, then

(0-valent) Cxy and (1-valent) Cyx are indistinguishable to a third process
pz if run alone, because the object is left in the same state in both configura-
tions; whichever way pz decides, it will give a contradiction in an execution
starting with one of these configurations. If neither of x and y is a read,
then x overwrites y, and Cx is indistinguishable from Cyxto pz if pz runs
alone; again we get a contradiction.

B.7.3 Hams (20 points)

Hamazon, LLC, claims to be the world’s biggest delivery service for canned
hams, with guaranteed delivery of a canned ham to your home anywhere
on Earth via suborbital trajectory from secret launch facilities at the North
and South Poles. Unfortunately, these launch facilities may be subject to
crash failures due to inclement weather, trademark infringement actions, or
military retaliation for misdirected hams.
For this problem, you are to evaluate Hamazon’s business model from
the perspective of distributed algorithms. Consider a system consisting of
a client process and two server processes (corresponding to the North and
South Pole facilities) that communicate by means of asynchronous message
passing. In addition to the usual message-passing actions, each server also
has an irrevocable launch action that launches a ham at the client. As with
messages, hams are delivered asynchronously: it is impossible for the client
to tell if a ham has been launched until it arrives.
A ham protocol is correct provided (a) a client that orders no ham re-
ceives no ham; and (b) a client that orders a ham receives exactly one ham.
Show that there can be no correct deterministic protocol for this problem if
one of the servers can crash.

Solution
Consider an execution in which the client orders ham. Run the northern
server together with the client until the server is about to issue a launch
action (if it never does so, the client receives no ham when the southern
server is faulty).
Now run the client together with the southern server. There are two
cases:

1. If the southern server ever issues launch, execute both this and the
northern server’s launch actions: the client gets two hams.
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 310

2. If the southern server never issues launch, never run the northern
server again: the client gets no hams.

In either case, the one-ham rule is violated, and the protocol is not
correct.5

B.7.4 Mutexes (20 points)

A swap register s has an operation swap(s, v) that returns the argument to
the previous call to swap, or ⊥ if it is the first such operation applied to the
register. It’s easy to build a mutex from a swap register by treating it as a
test-and-set: to grab the mutex, I swap in 1, and if I get back ⊥ I win (and
otherwise try again); and to release the mutex, I put back ⊥.
Unfortunately, this implementation is not starvation-free: some other
process acquiring the mutex repeatedly might always snatch the ⊥ away
just before I try to swap it out. Algorithm B.5 uses a swap object s along
with an atomic register r to try to fix this.

1 procedure mutex()
2 predecessor ← swap(s, myId)
3 while r 6= predecessor do
4 try again
// Start of critical section
5 ...
// End of critical section
6 r ← myId
Algorithm B.5: Mutex using a swap object and register

Prove that Algorithm B.5 gives a starvation-free mutex, or give an ex-

ample of an execution where it fails. You should assume that s and r are
both initialized to ⊥.
5
It’s tempting to try to solve this problem by reduction from a known impossibility
result, like Two Generals or FLP. For these specific problems, direct reductions don’t
appear to work. Two Generals assumes message loss, but in this model, messages are not
lost. FLP needs any process to be able to fail, but in this model, the client never fails.
Indeed, we can solve consensus in the Hamazon model by just having the client transmit
its input to both servers.
APPENDIX B. SAMPLE ASSIGNMENTS FROM SPRING 2014 311

Solution
Because processes use the same id if they try to access the mutex twice, the
algorithm doesn’t work.
Here’s an example of a bad execution:

1. Process 1 swaps 1 into s and gets ⊥, reads ⊥ from r, performs its

critical section, and writes 1 to r.

2. Process 2 swaps 2 into s and gets 1, reads 1 from r, and enters the
critical section.

3. Process 1 swaps 1 into s and gets 2, and spins waiting to see 2 in r.

4. Process 3 swaps 3 into s and gets 1. Because r is still 1, process 3 reads

this 1 and enters the critical section. We now have two processes in
the critical section, violating mutual exclusion.

I believe this works if each process adopts a new id every time it calls
mutex, but the proof is a little tricky.6

6
The simplest proof I can come up with is to apply an invariant that says that (a)
the processes that have executed swap(s, myId) but have not yet left the while loop have
predecessor values that form a linked list, with the last pointer either equal to ⊥ (if no
process has yet entered the critical section) or the last process to enter the critical section;
(b) r is ⊥ if no process has yet left the critical section, or the last process to leave the
critical section otherwise; and (c) if there is a process that is in the critical section, its
predecessor field points to the last process to leave the critical section. Checking the effects
of each operation shows that this invariant is preserved through the execution, and (a)
combined with (c) show that we can’t have two processes in the critical section at the
same time. Additional work is still needed to show starvation-freedom. It’s a good thing
this algorithm doesn’t work as written.
Appendix C

Sample assignments from

Fall 2011

C.1 Assignment 1: due Wednesday, 2011-09-28, at

17:00
Bureaucratic part
Send me email! My address is aspnes@cs.yale.edu.
In your message, include:

1. Your name.

2. Your status: whether you are an undergraduate, grad student, auditor,

etc.

3. Anything else you’d like to say.

(You will not be graded on the bureaucratic part, but you should do it
anyway.)

C.1.1 Anonymous algorithms on a torus

An n × m torus is a two-dimensional version of a ring, where a node at
position (i, j) has a neighbor to the north at (i, j − 1), the east at (i + 1, j),
the south at (i, j + 1), and the west at (i − 1, j). These values wrap around
modulo n for the first coordinate and modulo m for the second; so (0, 0) has
neighbors (0, m − 1), (1, 0), (0, 1), and (n − 1, 0).

312
APPENDIX C. SAMPLE ASSIGNMENTS FROM FALL 2011 313

Suppose that we have a synchronous message-passing network in the

form of an n × m torus, consisting of anonymous, identical processes that do
not know n, m, or their own coordinates, but do have a sense of direction
(meaning they can tell which of their neighbors is north, east, etc.).
Prove or disprove: Under these conditions, there is a deterministic1 al-
gorithm that computes whether n > m.

Solution
Disproof: Consider two executions, one in an n × m torus and one in an
m × n torus where n > m and both n and m are at least 2.2 Using the same
argument as in Lemma 6.1.1, show by induction on the round number that,
for each round r, all processes in both executions have the same state. It
follows that if the processes correctly detect n > m in the n × m execution,
then they incorrectly report m > n in the m × n execution.

C.1.2 Clustering
Suppose that k of the nodes in an asynchronous message-passing network
are designated as cluster heads, and we want to have each node learn the
identity of the nearest head. Given the most efficient algorithm you can for
this problem, and compute its worst-case time and message complexities.
You may assume that processes have unique identifiers and that all pro-
cesses know how many neighbors they have.3

Solution
The simplest approach would be to run either of the efficient distributed
breadth-first search algorithms from Chapter 5 simultaneously starting at
all cluster heads, and have each process learn the distance to all cluster heads
at once and pick the nearest one. This gives O(D2 ) time and O(k(E + V D))
messages if we use layering and O(D) time and O(kDE) messages using
local synchronization.
We can get rid of the dependence on k in the local-synchronization algo-
rithm by running it almost unmodified, with the only difference being the
attachment of a cluster head id to the exactly messages. The simplest way to
show that the resulting algorithm works is to imagine coalescing all cluster
1
Clarification added 2011-09-28.
2
This last assumption is not strictly necessary, but it avoids having to worry about
what it means when a process sends a message to itself.
3
Clarification added 2011-09-26.
APPENDIX C. SAMPLE ASSIGNMENTS FROM FALL 2011 314

heads into a single initiator; the clustering algorithm effectively simulates

the original algorithm running in this modified graph, and the same proof
goes through. The running time is still O(D) and the message complexity
O(DE).

C.1.3 Negotiation
Two merchants A and B are colluding to fix the price of some valuable
commodity, by sending messages to each other for r rounds in a synchronous
message-passing system. To avoid the attention of antitrust regulators, the
merchants are transmitting their messages via carrier pigeons, which are
unreliable and may become lost. Each merchant has an initial price pA or
pB , which are integer values satisfying 0 ≤ p ≤ m for some known value
m, and their goal is to choose new prices p0A and p0B , where |p0A − p0B | ≤ 1.
If pA = pB and no messages are lost, they want the stronger goal that
p0A = p0B = pA = pB .
Prove the best lower bound you can on r, as a function of m, for all
protocols that achieve these goals.

Solution
This is a thinly-disguised version of the Two Generals Problem from Chap-
ter 3, with the agreement condition p0A = p0B replaced by an approximate
agreement condition |p0A − p0B | ≤ 1. We can use a proof based on the
indistinguishability argument in §3.2 to show that r ≥ m/2.
Fix r, and suppose that in a failure-free execution both processes send
messages in all rounds (we can easily modify an algorithm that does not
have this property to have it, without increasing r). We will start with a
sequence of executions with pA = pB = 0. Let X0 be the execution in which
no messages are lost, X1 the execution in which A’s last message is lost,
X2 the execution in which both A and B’s last messages are lost, and so
on, with Xk for 0 ≤ k ≤ 2r losing k messages split evenly between the two
processes, breaking ties in favor of losing messages from A.
When i is even, Xi is indistinguishable from Xi+1 by A; it follows that
p0A is the same in both executions. Because we no longer have agreement,
it may be that p0B (Xi ) and p0B (Xi+1 ) are not the same as p0A in either ex-
ecution; but since both are within 1 of p0A , the difference between them is
at most 2. Next, because Xi+1 to Xi+2 are indistinguishable to B, we have
p0B (Xi+1 ) = p0B (Xi+2 ), which we can combine with the previous claim to get
|p0B (Xi ) − p0B (Xi+2 )|. A simple induction then gives p0B (X2r ) ≤ 2r, where
APPENDIX C. SAMPLE ASSIGNMENTS FROM FALL 2011 315

X2r is an execution in which all messages are lost.

Now construct executions X2r+1 and X2r+2 by changing pA and pB to
m one at a time. Using essentially the same argument as before, we get
|p0B (X2r ) − p0B (X2r+2 )| ≤ 2 and thus p0B (X2r+2 ) ≤ 2r + 2.
Repeat the initial 2r steps backward to get to an execution X4r+2 with
pA = pB = m and no messages lost. Applying the same reasoning as above
shows m = p0B (X4r+2 ) ≤ 4r + 2 or r ≥ m−2 4 = Ω(m).
Though it is not needed for the solution, it is not too hard to unwind
the lower bound argument to extract an algorithm that matches the lower
bound up to a small constant factor. For simplicity, let’s assume m is even.
The protocol is to send my input in the first message and then use m/2−1
subsequent acknowledgments, stopping immediately if I ever fail to receive
a message in some round; the total number of rounds r is exactly m/2. If
I receive s messages in the first s rounds, I decide on min(pA , pB ) if that
value lies in [m/2 − s, m/2 + s] and the nearest endpoint otherwise. (Note
that if s = 0, I don’t need to compute min(pA , pB ), and if s > 0, I can do
so because I know both inputs.)
This satisfies the approximate agreement condition because if I see only
s messages, you see at most s + 1, because I stop sending once I miss a
message. So either we both decide min(pA , pB ) or we choose endpoints
m/2 ± sA and m/2 ± sB that are within 1 of each other. It also satisfies the
validity condition p0A = p0B = pA = pB when both inputs are equal and no
messages are lost (and even the stronger requirement that p0A = p0B when no
messages are lost), because in this case [m/2 − s, m/2 + s] is exactly [0, m]
and both processes decide min(pA , pB ).
There is still a factor-of-2 gap between the upper and lower bounds. My
guess would be that the correct bound is very close to m/2 on both sides,
and that my lower bound proof is not quite clever enough.

C.2 Assignment 2: due Wednesday, 2011-11-02, at

17:00
C.2.1 Consensus with delivery notifications
The FLP bound (Chapter 11) shows that we can’t solve consensus in an
asynchronous system with one crash failure. Part of the reason for this is
that only the recipient can detect when a message is delivered, so the other
processes can’t distinguish between a configuration in which a message has
or has not been delivered to a faulty process.
APPENDIX C. SAMPLE ASSIGNMENTS FROM FALL 2011 316

Suppose that we augment the system so that senders are notified imme-
diately when their messages are delivered. We can model this by making the
delivery of a single message an event that updates the state of both sender
and recipient, both of which may send additional messages in response. Let
us suppose that this includes attempted deliveries to faulty processes, so
that any non-faulty process that sends a message m is eventually notified
that m has been delivered (although it might not have any effect on the
recipient if the recipient has already crashed).

1. Show that this system can solve consensus with one faulty process
when n = 2.
2. Show that this system cannot solve consensus with two faulty processes
when n = 3.

Solution
1. To solve consensus, each process sends its input to the other. Whichever
input is delivered first becomes the output value for both processes.
2. To show impossibility with n = 3 and two faults, run the usual FLP
proof until we get to a configuration C with events e0 and e such that
Ce is 0-valent and Ce0 e is 1-valent (or vice versa). Observe that e
and e0 may involve two processes each (sender and receiver), for up
to four processes total, but only a process that is involved in both e
and e0 can tell which happened first. There can be at most two such
processes. Kill both, and get that Ce0 e is indistinguishable from Cee0
for the remaining process, giving the usual contradiction.

C.2.2 A circular failure detector

Suppose we equip processes 0 . . . n − 1 in an asynchronous message-passing
system with n processes subject to crash failures with a failure detector that
is strongly accurate (no non-faulty process is ever suspected) and causes
process i + 1 (mod n) to eventually permanently suspect process i if process
i crashes. Note that this failure detector is not even weakly complete (if
both i and i + 1 crash, no non-faulty process suspects i). Note also that the
ring structure of the failure detector doesn’t affect the actual network: even
though only process i + 1 (mod n) may suspect process i, any process can
send messages to any other process.
Prove the best upper and lower bounds you can on the largest number
of failures f that allows solving consensus in this system.
APPENDIX C. SAMPLE ASSIGNMENTS FROM FALL 2011 317

Solution
There is an easy reduction to FLP that shows f ≤ n/2 is necessary (when n
√
is even), and a harder reduction that shows f < 2 n − 1 is necessary. The
easy reduction is based on crashing every other process; now no surviving
process can suspect any other survivor, and we are back in an asynchronous
message-passing system with no failure detector and 1 remaining failure (if
f is at least n/2 + 1).
√
The harder reduction is to crash every ( n)-th process. This partitions
√ √
the ring into n segments of length n − 1 each, where there is no failure
detector in any segment that suspects any process in another segment. If an
algorithm exists that solves consensus in this situation, then it does so even
if (a) all processes in each segment have the same input, (b) if any process
√
in one segment crashes, all n − 1 process in the segment crash, and (c) if
any process in a segment takes a step, all take a step, in some fixed order.
Under this additional conditions, each segment can be simulated by a single
process in an asynchronous system with no failure detectors, and the extra
√ √
n − 1 failures in 2 n − 1 correspond to one failure in the simulation. But
we can’t solve consensus in the simulating system (by FLP), so we can’t
solve it in the original system either.
On the other side, let’s first boost completeness of the failure detector,
by having any process that suspects another transmit this submission by
reliable broadcast. So now if any non-faulty process i suspects i + 1, all the
non-faulty processes will suspect i + 1. Now with up to t failures, whenever
I learn that process i is faulty (through a broadcast message passing on the
suspicion of the underlying failure detector, I will suspect processes i + 1
through i + t − f as well, where f is the number of failures I have heard
about directly. I don’t need to suspect process i + t − f + 1 (unless there is
some intermediate process that has also failed), because the only way that
this process will not be suspected eventually is if every process in the range
i to i + t − f is faulty, which can’t happen given the bound t.
Now if t is small enough that I can’t cover the entire ring with these
segments, then there is some non-faulty processes that is far enough away
from the nearest preceding faulty process that it is never suspected: this
gives us an eventually strong failure detector, and we can solve consensus
using the standard Chandra-Toueg ♦S algorithm from §13.4 or [CT96]. The
inequality I am looking for is f (t − f ) < n, where the √ left-hand side is
maximized by setting 2
f = t/2, which gives t /4 < n or t < 2n. This leaves
√
a gap of about 2 between the upper and lower bounds; I don’t know which
one can be improved.
APPENDIX C. SAMPLE ASSIGNMENTS FROM FALL 2011 318

√
I am indebted to Hao Pan for suggesting the Θ( n) upper and lower
bounds, which corrected an error in my original draft solution to this prob-
lem.

C.2.3 An odd problem

Suppose that each of n processes in a message-passing system with a com-
plete network is attached to a sensor. Each sensor has two states, active
and inactive; initially, all sensors are off. When the sensor changes state,
the corresponding process is notified immediately, and can update its state
and send messages to other processes in response to this event. It is also
guaranteed that if a sensor changes state, it does not change state again for
at least two time units. We would like to detect when an odd number of
sensors are active, by having at least one process update its state to set off
an alarm at a time when this condition holds.
A correct protocol for this problem should satisfy two conditions:

No false positives If a process sets of an alarm, then an odd number of

sensors are active.

Termination If at some time an odd number of sensors are active, and from
that point on no sensor changes its state, then some process eventually
sets off an alarm.

For what values of n is it possible to construct such a protocol?

Solution
It is feasible to solve the problem for n < 3.
For n = 1, the unique process sets off its alarm as soon as its sensor
becomes active.
For n = 2, have each process send a message to the other containing
its sensor state whenever the sensor state changes. Let s1 and s2 be the
state of the two process’s sensors, with 0 representing inactive and 1 active,
and let pi set off its alarm if it receives a message s such that s ⊕ si = 1.
This satisfies termination, because if we reach a configuration with an odd
number of active sensors, the last sensor to change causes a message to be
sent to the other process that will cause it to set off its alarm. It satisfies
no-false-positives, because if pi sets off its alarm, then s¬i = s because at
most one time unit has elapsed since p¬i sent s; it follows that s¬i ⊕ si = 1
and an odd number of sensors are active.
APPENDIX C. SAMPLE ASSIGNMENTS FROM FALL 2011 319

No such protocol is possible for n ≥ 3. Make p1 ’s sensor active. Run the

protocol until some process pi is about to enter an alarm state (this occurs
eventually because otherwise we violate termination). Let pj be one of p2 or
p3 with j 6= i, activate pj ’s sensor (we can do this without violating the once-
per-time-unit restriction because it has never previously been activated) and
then let pi set off its alarm. We have now violated no-false-positives.

C.3 Assignment 3: due Friday, 2011-12-02, at 17:00

C.3.1 A restricted queue
Suppose you have an atomic queue Q that supports operations enq and deq,
restricted so that:

• enq(Q) always pushes the identity of the current process onto the tail
of the queue.

• deq(Q) tests if the queue is nonempty and its head is equal to the
identity of the current process. If so, it pops the head and returns
true. If not, it does nothing and returns false.

The rationale for these restrictions is that this is the minimal version of
a queue needed to implement a starvation-free mutex using Algorithm 17.2.
What is the consensus number of this object?

Solution
The restricted queue has consensus number 1.
Suppose we have 2 processes, and consider all pairs of operations on Q
that might get us out of a bivalent configuration C. Let x be an operation
carried out by p that leads to a b-valent state, and y an operation by q that
leads to a (¬b)-valent state. There are three cases:

• Two deq operations. If Q is empty, the operations commute. If the

head of the Q is p, then y is a no-op and p can’t distinguish between
Cx and Cyx. Similarly for q if the head is q.

• One enq and one deq operation. Suppose x is an enq and y a deq. If
Q is empty or the head is not q, then y is a no-op: p can’t distinguish
Cx from Cyx. If the head is q, then x and y commute. The same
holds in reverse if x is a deq and y an enq.
APPENDIX C. SAMPLE ASSIGNMENTS FROM FALL 2011 320

• Two enq operations. This is a little tricky, because Cxy and Cyx
are different states. However, if Q is nonempty in C, whichever pro-
cess isn’t at the head of Q can’t distinguish them, because any deq
operation returns false and never reaches the newly-enqueued values.
This leaves the case where Q is empty in C. Run p until it is poised
to do x0 = deq(Q) (if this never happens, p can’t distinguish Cxy
from Cyx); then run q until it is poised to do y 0 = deq(Q) as well
(same argument as for p). Now allow both deq operations to proceed
in whichever order causes them both to succeed. Since the processes
can’t tell which deq happened first, they can’t tell which enq hap-
pened first either. Slightly more formally, if we let α be the sequence
of operations leading up to the two deq operations, we’ve just shown
Cxyαx0 y 0 is indistinguishable from Cyxαy 0 x0 to both processes.

In all cases, we find that we can’t escape bivalence. It follows that Q can’t
solve 2-process consensus.

C.3.2 Writable fetch-and-increment

Suppose you are given an unlimited supply of atomic registers and fetch-
and-increment objects, where the fetch-and-increment objects are all initial-
ized to 0 and supply only a fetch-and-increment operation that increments
the object and returns the old value. Show how to use these objects to
construct a wait-free, linearizable implementation of an augmented fetch-
and-increment that also supports a write operation that sets the value of
the fetch-and-increment and returns nothing.

Solution
We’ll use a snapshot object a to control access to an infinite array f of fetch-
and-increments, where each time somebody writes to the implemented ob-
ject, we switch to a new fetch-and-increment. Each cell in a holds (timestamp, base),
where base is the starting value of the simulated fetch-and-increment. We’ll
also use an extra fetch-and-increment T to hand out timestamps.
Code is in Algorithm C.1.
Since this is all straight-line code, it’s trivially wait-free.
Proof of linearizability is by grouping all operations by timestamp, us-
ing s[i].timestamp for FetchAndIncrement operations and t for write opera-
tions, then putting write before FetchAndIncrement, then ordering FetchAndIncrement
by return value. Each group will consist of a write(v) for some v followed by
zero or more FetchAndIncrement operations, which will return increasing
APPENDIX C. SAMPLE ASSIGNMENTS FROM FALL 2011 321

1 procedure FetchAndIncrement()
2 s ← snapshot(a)
3 i ← argmaxi (s[i].timestamp)
4 return f [s[i].timestamp] + s[i].base

5 procedure write(v)
6 t ← FetchAndIncrement(T )
7 a[myId] ← (t, v)
Algorithm C.1: Resettable fetch-and-increment

values starting at v since they are just returning values from the underlying
FetchAndIncrement object; the implementation thus meets the specifica-
tion.
To show consistency with the actual execution order, observe that time-
stamps only increase over time and that the use of snapshot means that
any process that observes or writes a timestamp t does so at a time later
than any process that observes or writes any t0 < t; this shows the group
order is consistent. Within each group, the write writes a[myId] before
any FetchAndIncrement reads it, so again we have consistency between the
write and any FetchAndIncrement operations. The FetchAndIncrement
operations are linearized in the order in which they access the underlying
f [. . . ] object, so we win here too.

C.3.3 A box object

Suppose you want to implement an object representing a w × h box whose
width (w) and height (h) can be increased if needed. Initially, the box is
1 × 1, and the coordinates can be increased by 1 each using IncWidth and
IncHeight operations. There is also a GetArea operation that returns the
area w · h of the box.
Give an obstruction-free deterministic implementation of this object from
atomic registers that optimizes the worst-case individual step complexity of
GetArea, and show that your implementation is optimal by this measure up
to constant factors.

Solution
Let b be the box object. Represent b by a snapshot object a, where a[i]
holds a pair (∆wi , ∆hi ) representing the number of times process i has
APPENDIX C. SAMPLE ASSIGNMENTS FROM FALL 2011 322

executed IncWidth and IncHeight; these operations simply increment the

appropriate value and update the snapshot object. Let GetArea take a
P P
snapshot and return ( i ∆wi ) ( i ∆hi ); the cost of the snapshot is O(n).
To see that this is optimal, observe that we can use IncWidth and
GetArea to represent inc and read for a standard counter. The Jayanti-
Tan-Toueg bound applies to counters, giving a worst-case cost of Ω(n) for
GetArea.

C.4 CS465/CS565 Final Exam, December 12th,

2011
Write your answers in the blue book(s). Justify your answers. Work alone.
Do not use any notes or books.
There are four problems on this exam, each worth 20 points, for a total
of 80 points. You have approximately three hours to complete this exam.

General clarifications added during exam Assume all processes have

unique ids and know n. Assume that the network is complete in the message-
passing model.

C.4.1 Lockable registers (20 points)

Most memory-management units provide the ability to control access to
specific memory pages, allowing a page to be marked (for example) read-
only. Suppose that we model this by a lockable register that has the usual
register operations read(r) and write(r, v) plus an additional operation
lock(r). The behavior of the register is just like a normal atomic register
until somebody calls lock(r); after this, any call to write(r) has no effect.
What is the consensus number of this object?

Solution
The consensus number is ∞; a single lockable register solves consensus for
any number of processes. Code is in Algorithm C.2.

1 write(r, input)
2 lock(r)
3 return read(r)
Algorithm C.2: Consensus using a lockable register
APPENDIX C. SAMPLE ASSIGNMENTS FROM FALL 2011 323

Termination and validity are trivial. Agreement follows from the fact
that whatever value is in r when lock(r) is first called will never change,
and thus will be read and returned by all processes.

C.4.2 Byzantine timestamps (20 points)

Suppose you have an asynchronous message passing system with exactly one
Byzantine process.
You would like the non-faulty processes to be able to acquire an increas-
ing sequence of timestamps. A process should be able to execute the time-
stamp protocol as often as it likes, and it should be guaranteed that when a
process is non-faulty, it eventually obtains a timestamp that is larger than
any timestamp returned in any execution of the protocol by a non-faulty
process that finishes before the current process’s execution started.
Note that there is no bound on the size of a timestamp, so having the
Byzantine process run up the timestamp values is not a problem, as long as
it can’t cause the timestamps to go down.
For what values of n is it possible to solve this problem?

Solution
It is possible to solve the problem for all n except n = 3. For n = 1, there are
no non-faulty processes, so the specification is satisfied trivially. For n = 2,
there is only one non-faulty process: it can just keep its own counter and
return an increasing sequence of timestamps without talking to the other
process at all.
For n = 3, it is not possible. Consider an execution in which messages
between non-faulty processes p and q are delayed indefinitely. If the Byzan-
tine process r acts to each of p and q as it would if the other had crashed,
this execution is indistinguishable to p and q from an execution in which r
is correct and the other is faulty. Since there is no communication between
p and q, it is easy to construct and execution in which the specification is
violated.
For n ≥ 4, the protocol given in Algorithm C.3 works.
The idea is similar to the Attiya, Bar-Noy, Dolev distributed shared
memory algorithm [ABND95]. A process that needs a timestamp polls n − 1
other processes for the maximum values they’ve seen and adds 1 to it; before
returning, it sends the new timestamp to all other processes and waits to
receive n − 1 acknowledgments. The Byzantine process may choose not to
answer, but this is not enough to block completion of the protocol.
APPENDIX C. SAMPLE ASSIGNMENTS FROM FALL 2011 324

1 procedure getTimestamp()
2 ci ← ci + 1
3 send probe(ci ) to all processes
4 wait to receive response(ci , vj ) from n − 1 processes
5 vi ← (maxj vj ) + 1
6 send newTimestamp(ci , vi ) to all processes
7 wait to receive ack(ci ) from n − 1 processes
8 return vi

9 upon receiving probe(cj ) from j do

10 send response(cj , vi ) to j

11 upon receiving newTimestamp(cj , vj ) from j do

12 vi ← max(vi , vj )
13 send ack(cj ) to j
Algorithm C.3: Timestamps with n ≥ 3 and one Byzantine process

To show the timestamps are increasing, observe that after the completion
of any call by i to getTimestamp, at least n − 2 non-faulty processes j have
a value vj ≥ vi . Any call to getTimestamp that starts later sees at least
n − 3 > 0 of these values, and so computes a max that is at least as big as
vi and then adds 1 to it, giving a larger value.

C.4.3 Failure detectors and k-set agreement (20 points)

Recall that in the k-set agreement problem we want each of n processes
to choose a decision value, with the property that the set of decision values
has at most k distinct elements. It is known that k-set agreement cannot
be solved deterministically in an asynchronous message-passing or shared-
memory system with k or more crash failures.
Suppose that you are working in an asynchronous message-passing sys-
tem with an eventually strong (♦S) failure detector. Is it possible to solve
k-set agreement deterministically with f crash failures, when k ≤ f < n/2?

Solution
Yes. With f < n/2 and ♦S, we can solve consensus using Chandra-
Toueg [CT96]. Since this gives a unique decision value, it solves k-set agree-
APPENDIX C. SAMPLE ASSIGNMENTS FROM FALL 2011 325

ment for any k ≥ 1.

C.4.4 A set data structure (20 points)

Consider a data structure that represents a set S, with an operation add(S, x)
that adds x to S by setting S ← S ∪ {x}), and an operation size(S) that
returns the number of distinct4 elements |S| of S. There are no restrictions
on the types or sizes of elements that can be added to the set.
Show that any deterministic wait-free implementation of this object from
atomic registers has individual step complexity Ω(n) for some operation in
the worst case.

Solution
Algorithm C.4 implements a counter from a set object, where the counter
read consists of a single call to size(S). The idea is that each increment
is implemented by inserting a new element into S, so |S| is always equal to
the number of increments.

1 procedure inc(S)
2 nonce ← nonce + 1
3 add(S, hmyId, noncei).

4 procedure read(S)
5 return size(S)
Algorithm C.4: Counter from set object

Since the Jayanti-Tan-Toueg lower bound [JTT00] gives a lower bound

of Ω(n) on the worst-case cost of a counter read, there exists an execution
in which size(S) takes Ω(n) steps.
(We could also apply JTT directly by showing that the set object is
perturbable; this follows because adding an element not added by anybody
else is always visible to the reader.)

4
Clarification added during exam.
Appendix D

Additional sample final

exams

This appendix contains final exams from previous times the course was of-
fered, and is intended to give a rough guide to the typical format and content
of a final exam. Note that the topics covered in past years were not neces-
sarily the same as those covered this year.

D.1 CS425/CS525 Final Exam, December 15th,

2005
Write your answers in the blue book(s). Justify your answers. Work alone.
Do not use any notes or books.
There are three problems on this exam, each worth 20 points, for a total
of 60 points. You have approximately three hours to complete this exam.

D.1.1 Consensus by attrition (20 points)

Suppose you are given a bounded fetch-and-subtract register that holds
a non-negative integer value and supports an operation fetch-and-subtract(k)
for each k > 0 that (a) sets the value of the register to the previous value
minus k, or zero if this result would be negative, and (b) returns the previous
value of the register.
Determine the consensus number of bounded fetch-and-subtract under
the assumptions that you can use arbitrarily many such objects, that you can
supplement them with arbitrarily many multiwriter/multireader read/write
registers, that you can initialize all registers of both types to initial values

326
APPENDIX D. ADDITIONAL SAMPLE FINAL EXAMS 327

of your choosing, and that the design of the consensus protocol can depend
on the number of processes N .

Solution
The consensus number is 2.
To implement 2-process wait-free consensus, use a single fetch-and-subtract
register initialized to 1 plus two auxiliary read/write registers to hold the
input values of the processes. Each process writes its input to its own regis-
ter, then performs a fetch-and-subtract(1) on the fetch-and-subtract register.
Whichever process gets 1 from the fetch-and-subtract returns its own input;
the other process (which gets 0) returns the winning process’s input (which
it can read from the winning process’s read/write register.)
To show that the consensus number is at most 2, observe that any
two fetch-and-subtract operations commute: starting from state x, after
fetch-and-subtract(k1 ) and fetch-and-subtract(k2 ) the value in the fetch-
and-subtract register is max(0, x − k1 − k2 ) regardless of the order of the
operations.

D.1.2 Long-distance agreement (20 points)

Consider an asynchronous message-passing model consisting of N processes
p1 . . . pN arranged in a line, so that each process i can send messages only
to processes i − 1 and i + 1 (if they exist). Assume that there are no failures,
that local computation takes zero time, and that every message is delivered
at most 1 time unit after it is sent no matter how many messages are sent
on the same edge.
Now suppose that we wish to solve agreement in this model, where the
agreement protocol is triggered by a local input event at one or more pro-
cesses and it terminates when every process executes a local decide event.
As with all agreement problems, we want Agreement (all processes decide
the same value), Termination (all processes eventually decide), and Validity
(the common decision value previously appeared in some input). We also
want no false starts: the first action of any process should either be an input
action or the receipt of a message.
Define the time cost of a protocol for this problem as the worst-case
time between the first input event and the last decide event. Give the best
upper and lower bounds you can on this time as function of N . Your upper
and lower bounds should be exact: using no asymptotic notation or hidden
constant factors. Ideally, they should also be equal.
APPENDIX D. ADDITIONAL SAMPLE FINAL EXAMS 328

Solution
Upper bound
Because there are no failures, we can appoint a leader and have it decide.
The natural choice is some process near the middle, say pb(N +1)/2c . Upon
receiving an input, either directly through an input event or indirectly from
another process, the process sends the input value along the line toward the
leader. The leader takes the first input it receives and broadcasts it back
out in both directions as the decision value. The worst case is when the
protocol is initiated at pN ; then we pay 2(N − b(N + 1)/2c) time to send all
messages out and back, which is N time units when N is even and N − 1
time units when N is odd.

Lower bound
Proving an almost-matching lower bound of N − 1 time units is trivial: if
p1 is the only initiator and it starts at time t0 , then by an easy induction
argument,in the worst case pi doesn’t learn of any input until time t0 +(i−1),
and in particular pN doesn’t find out until after N − 1 time units. If pN
nonetheless decides early, its decision value will violate validity in some
executions.
But we can actually prove something stronger than this: that N time
units are indeed required when N is odd. Consider two slow executions Ξ0
and Ξ1 , where (a) all messages are delivered after exactly one time unit in
each execution; (b) in Ξ0 only p1 receives an input and the input is 0; and
(c) in Ξ1 only pN receives an input and the input is 1. For each of the
executions, construct a causal ordering on events in the usual fashion: a
send is ordered before a receive, two events of the same process are ordered
by time, and other events are partially ordered by the transitive closure of
this relation.
Now consider for Ξ0 the set of all events that precede the decide(0)
event of p1 and for Ξ1 the set of all events that precede the decide(1) event
of pN . Consider further the sets of processes S0 and S1 at which these events
occur; if these two sets of processes do not overlap, then we can construct
an execution in which both sets of events occur, violating Agreement.
Because S0 and S1 overlap, we must have |S0 | + |S1 | ≥ N + 1, and so at
least one of the two sets has size at least d(N + 1)/2e, which is N/2 + 1 when
N is even. Suppose that it is S0 . Then in order for any event to occur at
pN/2+1 at all some sequence of messages must travel from the initial input
to p1 to process pN/2+1 (taking N/2 time units), and the causal ordering
APPENDIX D. ADDITIONAL SAMPLE FINAL EXAMS 329

implies that an additional sequence of messages travels back from pN/2+1 to

p1 before p1 decides (taking and additional N/2 time units). The total time
is thus N .

D.1.3 Mutex appendages (20 points)

An append register supports standard read operations plus an append op-
eration that appends its argument to the list of values already in the register.
An append-and-fetch register is similar to an append register, except that
it returns the value in the register after performing the append operation.
Suppose that you have an failure-free asynchronous system with anonymous
deterministic processes (i.e., deterministic processes that all run exactly the
same code). Prove or disprove each of the following statements:

1. It is possible to solve mutual exclusion using only append registers.

2. It is possible to solve mutual exclusion using only append-and-fetch

registers.

In either case, the solution should work for arbitrarily many processes—solving
mutual exclusion when N = 1 is not interesting. You are also not required
in either case to guarantee lockout-freedom.

Clarification given during exam

1. If it helps, you may assume that the processes know N . (It probably
doesn’t help.)

Solution
1. Disproof: With append registers only, it is not possible to solve mutual
exclusion. To prove this, construct a failure-free execution in which
the processes never break symmetry. In the initial configuration, all
processes have the same state and thus execute either the same read
operation or the same append operation; in either case we let all N
operations occur in some arbitrary order. If the operations are all
reads, all processes read the same value and move to the same new
state. If the operations are all appends, then no values are returned
and again all processes enter the same new state. (It’s also the case
that the processes can’t tell from the register’s state which of the
identical append operations went first, but we don’t actually need to
use this fact.)
APPENDIX D. ADDITIONAL SAMPLE FINAL EXAMS 330

Since we get a fair failure-free execution where all processes move

through the same sequence of states, if any process decides it’s in its
critical section, all do. We thus can’t solve mutual exclusion in this
model.

2. Since the processes are anonymous, any solution that depends on them
having identifiers isn’t going to work. But there is a simple solution
that requires only appending single bits to the register.
Each process trying to enter a critical section repeatedly executes an
append-and-fetch operation with argument 0; if the append-and-fetch
operation returns either a list consisting only of a single 0 or a list
whose second-to-last element is 1, the process enters its critical section.
To leave the critical section, the process does append-and-fetch(1).

D.2 CS425/CS525 Final Exam, May 8th, 2008

Write your answers in the blue book(s). Justify your answers. Work alone.
Do not use any notes or books.
There are four problems on this exam, each worth 20 points, for a total
of 80 points. You have approximately three hours to complete this exam.

D.2.1 Message passing without failures (20 points)

Suppose you have an asynchronous message-passing system with a com-
plete communication graph, unique node identities, and no failures. Show
that any deterministic atomic shared-memory object can be simulated in
this model, or give an example of a shared-memory object that can’t be
simulated.

Solution
Pick some leader node to implement the object. To execute an operation,
send the operation to the leader node, then have the leader carry out the
operation (sequentially) on its copy of the object and send the results back.

D.2.2 A ring buffer (20 points)

Suppose you are given a ring buffer object that consists of k ≥ 1 memory
locations a[0] . . . a[k − 1] with an atomic shift-and-fetch operation that takes
an argument v and (a) shifts v into the buffer, so that a[i] ← a[i + 1] for
APPENDIX D. ADDITIONAL SAMPLE FINAL EXAMS 331

each i less than k − 1 and a[k − 1] ← v; and (b) returns a snapshot of the
new contents of the array (after the shift).
What is the consensus number of this object as a function of k?

Solution
We can clearly solve consensus for at least k processes: each process calls
shift-and-fetch on its input, and returns the first non-null value in the buffer.
So now we want to show that we can’t solve consensus for k+1 processes.
Apply the usual FLP-style argument to get to a bivalent configuration C
where each of the k + 1 processes has a pending operation that leads to
a univalent configuration. Let e0 and e1 be particular operations leading
to 0-valent and 1-valent configurations, respectively, and let e2 . . . ek be the
remaining k − 1 pending operations.
We need to argue first that no two distinct operations ei and ej are
operations of different objects. Suppose that Cei is 0-valent and Cej is
1-valent; then if ei and ej are on different objects, Cei ej (still 0-valent) is
indistinguishable by all processes from Cej ei (still 1-valent), a contradiction.
Alternatively, if ei and ej are both b-valent, there exists some (1−b)-valent ek
such that ei and ej both operate on the same object as ek , by the preceding
argument. So all of e0 . . . ek are operations on the same object.
By the usual argument we know that this object can’t be a register. Let’s
show it can’t be a ring buffer either. Consider the configurations Ce0 e1 . . . ek
and Ce1 . . . ek . These are indistinguishable to the process carrying out ek
(because its sees only the inputs to e1 through ek in its snapshot). So they
must have the same valence, a contradiction.
It follows that the consensus number of a k-element ring buffer is exactly
k.

D.2.3 Leader election on a torus (20 points)

An n × n torus is a graph consisting of n2 nodes, where each node (i, j),
0 ≤ i, j ≤ n − 1, is connected to nodes (i − 1, j), (i + 1, j), (i, j − 1), and
(i, j + 1), where all computation is done mod n.
Suppose you have an asynchronous message-passing system with a com-
munication graph in the form of an n × n torus. Suppose further that each
node has a unique identifier (some large natural number) but doesn’t know
the value of n. Give an algorithm for leader election in this model with the
best message complexity you can come up with.
APPENDIX D. ADDITIONAL SAMPLE FINAL EXAMS 332

Solution
First observe that each row and column of the torus is a bidirectional ring,
so we can run e.g. Hirschbirg and Sinclair’s O(n log n)-message protocol
within each of these rings to find the smallest identifier in the ring. We’ll
use this to construct the following algorithm:
1. Run Hirschbirg-Sinclair in each row to get a local leader for each row;
this takes n × O(n log n) = O(n2 log n) messages. Use an additional n
messages per row to distribute the identifier for the row leader to all
nodes and initiate the next stage of the protocol.
2. Run Hirschbirg-Sinclair in each column with each node adopting the
row leader identifier as its own. This costs another O(n2 log n) mes-
sages; at the end, every node knows the minimum identifier of all nodes
in the torus.
The total message complexity is O(n2 log n). (I suspect this is optimal,
but I don’t have a proof.)

D.2.4 An overlay network (20 points)

A collection of n nodes—in an asynchronous message-passing system with a
connected, bidirectional communications graph with O(1) links per node—wish
to engage in some strictly legitimate file-sharing. Each node starts with some
input pair (k, v), where k is a key and v is a value, and the search problem
is to find the value v corresponding to a particular key k.
1. Suppose that we can’t do any preparation ahead of time. Give an
algorithm for searching with the smallest asymptotic worst-case mes-
sage complexity you can find as a function of n. You may assume that
there are no limits on time complexity, message size, or storage space
at each node.
2. Suppose now that some designated leader node can initiate a protocol
ahead of time to pre-process the data in the nodes before any query is
initiated. Give a pre-processing algorithm (that does not depend on
which key is eventually searched for) and associated search algorithm
such that the search algorithm minimizes the asymptotic worst-case
message complexity. Here you may assume that there are no limits on
time complexity, message size, or storage space for either algorithm,
and that you don’t care about the message complexity of the pre-
processing algorithm.
APPENDIX D. ADDITIONAL SAMPLE FINAL EXAMS 333

3. Give the best lower bound you can on the total message complexity of
the pre-processing and search algorithms in the case above.

Solution
1. Run depth-first search to find the matching key and return the corre-
sponding value back up the tree. Message complexity is O(|E|) = O(n)
(since each node has only O(1) links).

2. Basic idea: give each node a copy of all key-value pairs, then searches
take zero messages. To give each node a copy of all key-value pairs we
could do convergecast followed by broadcast (O(n) message complex-
ity) or just flood each pair O(n2 ). Either is fine since we don’t care
about the message complexity of the pre-processing stage.

3. Suppose the total message complexity of both the pre-processing stage

and the search protocol is less than n − 1. Then there is some node
other than the initiator of the search that sends no messages at any
time during the protocol. If this is the node with the matching key-
value pair, we don’t find it. It follows that any solution to the search
problem. requires a total of Ω(n) messages in the pre-processing and
search protocols.

D.3 CS425/CS525 Final Exam, May 10th, 2010

Write your answers in the blue book(s). Justify your answers. Work alone.
Do not use any notes or books.
There are four problems on this exam, each worth 20 points, for a total
of 80 points. You have approximately three hours to complete this exam.

D.3.1 Anti-consensus (20 points)

A wait-free anti-consensus protocol satisfies the conditions:

Wait-free termination Every process decides in a bounded number of its

own steps.

Non-triviality There is at least one process that decides different values

in different executions.

Disagreement If at least two processes decide, then some processes decide

on different values.
APPENDIX D. ADDITIONAL SAMPLE FINAL EXAMS 334

Show that there is no deterministic wait-free anti-consensus protocol

using only atomic registers for two processes and two possible output values,
but there is one for three processes and three possible output values.

Clarification: You should assume processes have distinct identities.

Solution
No protocol for two: turn an anti-consensus protocol with outputs in {0, 1}
into a consensus protocol by having one of the processes always negate its
output.
A protocol for three: Use a splitter.

D.3.2 Odd or even (20 points)

Suppose you have a protocol for a synchronous message-passing ring that is
anonymous (all processes run the same code) and uniform (this code is the
same for rings of different sizes). Suppose also that the processes are given
inputs marking some, but not all, of them as leaders. Give an algorithm
for determining if the size of the ring is odd or even, or show that no such
algorithm is possible.

Clarification: Assume a bidirectional, oriented ring and a deterministic

algorithm.

Solution
Here is an impossibility proof. Suppose there is such an algorithm, and let
it correctly decide “odd” on a ring of size 2k + 1 for some k and some set
of leader inputs. Now construct a ring of size 4k + 2 by pasting two such
rings together (assigning the same values to the leader bits in each copy)
and run the algorithm on this ring. By the usual symmetry argument,
every corresponding process sends the same messages and makes the same
decisions in both rings, implying that the processes incorrectly decide the
ring of size 4k + 2 is odd.

D.3.3 Atomic snapshot arrays using message-passing (20 points)

Consider the following variant of Attiya-Bar-Noy-Dolev for obtaining snap-
shots of an array instead of individual register values, in an asynchronous
message-passing system with t < n/4 crash failures. The data structure we
APPENDIX D. ADDITIONAL SAMPLE FINAL EXAMS 335

are simulating is an array a consisting of an atomic register a[i] for each

process i, with the ability to perform atomic snapshots.
Values are written by sending a set of hi, v, ti i values to all processes,
where i specifies the segment a[i] of the array to write, v gives a value for
this segment, and ti is an increasing timestamp used to indicate more recent
values. We use a set of values because (as in ABD) some values may be
obtained indirectly.
To update segment a[i] with value v, process i generates a new timestamp
ti , sends {hi, v, ti i} to all processes, and waits for acknowledgments from at
least 3n/4 processes.
Upon receiving a message containing one or more hi, v, ti i triples, a pro-
cess updates its copy of a[i] for any i with a higher timestamp than previously
seen, and responds with an acknowledgment (we’ll assume use of nonces so
that it’s unambiguous which message is being acknowledged).
To perform a snapshot, a process sends snapshot to all processes, and
waits to receive responses from at least 3n/4 processes, which will consist of
the most recent values of each a[i] known by each of these processes together
with their timestamps (it’s a set of triples as above). The snapshot process
then takes the most recent versions of a[i] for each of these responses and
updates its own copy, then sends its entire snapshot vector to all processes
and waits to receive at least 3n/4 acknowledgments. When it has received
these acknowledgments, it returns its own copy of a[i] for all i.
Prove or disprove: The above procedure implements an atomic snap-
shot array in an asynchronous message-passing system with t < n/4 crash
failures.

Solution
Disproof: Let s1 and s2 be processes carrying out snapshots and let w1
and w2 be processes carrying out writes. Suppose that each wi initiates a
write of 1 to a[wi ], but all of its messages to other processes are delayed
after it updates its own copy awi [wi ]. Now let each si receive responses
from 3n/4 − 1 processes not otherwise mentioned plus wi . Then s1 will
return a vector with a[w1 ] = 1 and a[w2 ] = 0 while s2 will return a vector
with a[w1 ] = 0 and a[w2 ] = 1, which is inconsistent. The fact that these
vectors are also disseminated throughout at least 3n/4 other processes is a
red herring.
APPENDIX D. ADDITIONAL SAMPLE FINAL EXAMS 336

D.3.4 Priority queues (20 points)

Let Q be a priority queue whose states are multisets of natural numbers and
that has operations enq(v) and deq(), where enq(p) adds a new value v to
the queue, and deq() removes and returns the smallest value in the queue,
or returns null if the queue is empty. (If there is more than one copy of the
smallest value, only one copy is removed.)
What is the consensus number of this object?

Solution
The consensus number is 2. The proof is similar to that for a queue.
To show we can do consensus for n = 2, start with a priority queue with
a single value in it, and have each process attempt to dequeue this value. If
a process gets the value, it decides on its own input; if it gets null, it decides
on the other process’s input.
To show we can’t do consensus for n = 3, observe first that starting from
any states C of the queue, given any two operations x and y that are both
enqueues or both dequeues, the states Cxy and Cyx are identical. This
means that a third process can’t tell which operation went first, meaning
that a pair of enqueues or a pair of dequeues can’t get us out of a bivalent
configuration in the FLP argument. We can also exclude any split involving
two operations on different queues (or other objects) But we still need to
consider the case of a dequeue operation d and an enqueue operation e on
the same queue Q. This splits into several subcases, depending on the state
C of the queue in some bivalent configuration:
1. C = {}. Then Ced = Cd = {}, and a third process can’t tell which of
d or e went first.

2. C is nonempty and e = enq(v), where v is greater than or equal to the

smallest value in C. Then Cde and Ced are identical, and no third
process can tell which of d or e went first.

3. C is nonempty and e = enq(v), where v is less than any value in C.

Consider the configurations Ced and Cde. Here the process pd that
performs d can tell which operation went first, because it either obtains
v or some other value v 0 6= v. Kill this process. No other process in Ced
or Cde can distinguish the two states without dequeuing whichever of
v or v 0 was not dequeued by pd . So consider two parallel executions
Cedσ and Cdeσ where σ consists of an arbitrary sequence of operations
ending with a deq on Q by some process p (if no process ever attempts
APPENDIX D. ADDITIONAL SAMPLE FINAL EXAMS 337

to dequeue from Q, then we have already won, since the survivors can’t
distinguish Ced from Cde). Now the state of all objects is the same
after Cedσ and Cdeσ, and only pd and p have different states in these
two configurations. So any third process is out of luck.
Appendix E

I/O automata

E.1 Low-level view: I/O automata

An I/O automaton A is an automaton where transitions are labeled by
actions, which come in three classes: input actions, triggered by the
outside world; output actions triggered by the automaton and visible to
the outside world; and internal actions, triggered by the automaton but
not visible to the outside world. These classes correspond to inputs, outputs,
and internal computation steps of the automaton; the latter are provided
mostly to give merged input/output actions a place to go when automata are
composed together. A transition relation trans(A) relates states(A) ×
acts(A) × states(A); if (s, a, s0 ) is in trans(A), it means that A can move
from state s to state s0 by executing action a.
There is also an equivalence relation task(A) on the output and internal
actions, which is used for enforcing fairness conditions—the basic idea is that
in a fair execution some action in each equivalence class must be executed
eventually (a more accurate definition will be given below).
The I/O automaton model carries with it a lot of specialized jargon.
We’ll try to avoid it as much as possible. One thing that will be difficult to
avoid in reading [Lyn96] is the notion of a signature, which is just the tuple
sig(A) = (in(A), out(A), int(A)) describing the actions of an automaton
A.

E.1.1 Enabled actions

An action a is enabled in some state s if trans(A) contains at least one
transition (s, a, s0 ). Input actions are always enabled—this is a requirement

338
APPENDIX E. I/O AUTOMATA 339

of the model. Output and internal actions—the “locally controlled” ac-

tions—are not subject to this restriction. A state s is quiescent if only
input actions are enabled in s.

E.1.2 Executions, fairness, and traces

An execution of A is a sequence s0 a0 s1 a1 . . . where each triple (si , ai si+1 )
is in trans(A). Executions may be finite or infinite; if finite, they must end
in a state.
A trace of A is a subsequence of some execution consisting precisely
of the external (i.e., input and output) actions, with states and internal
actions omitted. If we don’t want to get into the guts of a particular I/O
automaton—and we usually don’t, unless we can’t help it because we have to
think explicitly about states for some reason—we can describe its externally-
visible behavior by just giving its set of traces.

E.1.3 Composition of automata

Composing a set of I/O automata yields a new super-automaton whose state
set is the Cartesian product of the state sets of its components and whose
action set is the union of the action sets of its components. A transition
with a given action a updates the states of all components that have a as an
action and has no effect on the states of other components. The classification
of actions into the three classes is used to enforce some simple compatibility
rules on the component automata; in particular:

1. An internal action of a component is never an action of another com-

ponent—internal actions are completely invisible.

2. No output action of a component can be an output action of another

component.

3. No action is shared by infinitely many components.1 In practice this

means that no action can be an input action of infinitely many com-
ponents, since the preceding rules mean that any action is an output
or internal action of at most one component.

All output actions of the components are also output actions of the
composition. An input action of a component is an input of the composition
only if some other component doesn’t supply it as an output; in this case
1
Note that infinite (but countable) compositions are permitted.
APPENDIX E. I/O AUTOMATA 340

it becomes an output action of the composition. Internal actions remain

internal (and largely useless, except for bookkeeping purposes).
The task equivalence relation is the union of the task relations for the
components: this turns out to give a genuine equivalence relation on output
and internal actions precisely because the first two compatibility rules hold.
Given an execution or trace X of a composite automaton that includes
A, we can construct the corresponding execution or trace X|A of A which
just includes the states of A and the actions visible to A (events that don’t
change the state of A drop out). The definition of composition is chosen so
that X|A is in fact an execution/trace of A whenever X is.

E.1.4 Hiding actions

Composing A and B continues to expose the outputs of A even if they line
up with inputs of B. While this may sometimes be desirable, often we want
to shove such internal communication under the rug. The model lets us do
this by redefining the signature of an automaton to make some or all of the
output actions into internal actions.

E.1.5 Fairness
I/O automata come with a built-in definition of fair executions, where an
execution of A is fair if, for each equivalence class C of actions in task(A),

1. the execution is finite and no action in C is enabled in the final state,

or

2. the execution is infinite and there are infinitely many occurrences of

actions in C, or

3. the execution is infinite and there are infinitely many states in which
no action in C is enabled.

If we think of C as corresponding to some thread or process, this says

that C gets infinitely many chances to do something in an infinite execution,
but may not actually do them if it gives ups and stops waiting (the third
case). The finite case essentially says that a finite execution isn’t fair unless
nobody is waiting at the end. The motivation for this particular definition
is that it guarantees (a) that any finite execution can be extended to a fair
execution and (b) that the restriction X|A of a fair execution or trace X is
also fair.
APPENDIX E. I/O AUTOMATA 341

Fairness is useful e.g. for guaranteeing message delivery in a message-

passing system: make each message-delivery action its own task class and
each message will eventually be delivered; similarly make each message-
sending action its own task class and a process will eventually send every
message it intends to send. Tweaking the task classes can allow for possi-
bilities of starvation, e.g. if all message-delivery actions are equivalent then
a spammer can shut down the system in a “fair” execution where only his
(infinitely many) messages are delivered.

E.1.6 Specifying an automaton

The typical approach is to write down preconditions and effects for each
action (for input actions, the preconditions are empty). An example would
be the spambot in Algorithm E.1.

1 input action setMessage(m)

2 effects
3 state ← m

4 output action spam(m)

5 precondition
6 spam = m
7 effects
8 none (keep spamming)

Algorithm E.1: Spambot as an I/O automaton

(Plus an initial state, e.g. state = ⊥, where ⊥ is not a possible message,

and a task partition, of which we will speak more below when we talk about
liveness properties.)

E.2 High-level view: traces

When studying the behavior of a system, traces are what we really care
about, and we want to avoid talking about states as much as possible. So
what we’ll aim to do is to get rid of the states early by computing the set of
traces (or fair traces) of each automaton in our system, then compose traces
to get traces for the system as a whole. Our typical goal will be to show
that the resulting set of traces has some desirable properties, usually of the
form (1) nothing bad happens (a safety property); (2) something good
APPENDIX E. I/O AUTOMATA 342

eventually happens (a liveness property); or (3) the horribly complex

composite automaton representing this concrete system acts just like that
nice clean automaton representing a specification (a simulation).
Very formally, a trace property specifies both the signature of the
automaton and a set of traces, such that all traces (or perhaps fair traces)
of the automata appear in the set. We’ll usually forget about the first part.
Tricky detail: It’s OK if not all traces in P are generated by A (we want
trace(A) ⊆ P , but not necessarily trace(A) = P ). But trace(A) will be
pretty big (it includes, for example, all finite sequences of input actions)
so hopefully the fact that A has to do something with inputs will tell us
something useful.

E.2.1 Example
A property we might demand of the spambot above (or some other ab-
straction of a message channel) is that it only delivers messages that have
previously been given to it. As a trace property this says that in any trace
t, if tk = spam(m), then tj = setMessage(m) for some j < k. (As a set, this
is just the set of all sequences of external spambot-actions that have this
property.) Call this property P .
To prove that the spambot automaton given above satisfies P , we might
argue that for any execution s0 a0 s1 a1 . . . , that si = m in the last setMessage
action preceding si , or ⊥ if there is no such action. This is easily proved
by induction on i. It then follows that since spam(m) can only transmit the
current state, that if spam(m) follows si = m that it follows some earlier
setMessage(m) as claimed.
However, there are traces that satisfy P that don’t correspond to execu-
tions of the spambot; for example, consider the trace setMessage(0)setMessage(1)spam(0).
This satisfies P (0 was previously given to the automaton spam(0)), but the
automaton won’t generate it because the 0 was overwritten by the later
setMessage(1) action. Whether this is indicates a problem with our automa-
ton not being nondeterministic enough or our trace property being too weak
is a question about what we really want the automaton to do.

E.2.2 Types of trace properties

E.2.2.1 Safety properties
P is a safety property if

1. P is nonempty.
APPENDIX E. I/O AUTOMATA 343

2. P is prefix-closed, i.e. if xy is in P then x is in P .

3. P is limit-closed, i.e. if x1 , x1 x2 , x1 x2 x3 , . . . are all in P , then so is

the infinite sequence obtained by taking their limit.

Because of the last restrictions, it’s enough to prove that P holds for all
finite traces of A to show that it holds for all traces (and thus for all fair
traces), since any trace is a limit of finite traces. Conversely, if there is some
trace or fair trace for which P fails, the second restriction says that P fails
on any finite prefix of P , so again looking at only finite prefixes is enough.
The spambot property mentioned above is a safety property.
Safety properties are typically proved using invariants, properties that
are shown by induction to hold in all reachable states.

E.2.2.2 Liveness properties

P is a liveness property of A if any finite sequence of actions in acts(A)
has an extension in P . Note that liveness properties will in general include
many sequences of actions that aren’t traces of A, since they are extensions
of finite sequences that A can’t do (e.g. starting the execution with an
action not enabled in the initial state). If you want to restrict yourself only
to proper executions of A, use a safety property. (It’s worth noting that the
same property P can’t do both: any P that is both a liveness and a safety
property includes all sequences of actions because of the closure rules.)
Liveness properties are those that are always eventually satisfiable; as-
serting one says that the property is eventually satisfied. The typical way
to prove a liveness property is with a progress function, a function f on
states that (a) drops by at least 1 every time something that happens in-
finitely often happens (like an action from an always-enabled task class) and
(b) guarantees P once it reaches 0.
An example would be the following property we might demand of our
spambot: any trace with at least one setMessage(. . . ) action contains in-
finitely many spam(. . . ) actions. Whether the spambot automaton will
satisfy this property (in fair traces) depends on its task partition. If all
spam(. . . ) actions are in the same equivalence class, then any execution with
at least one setMessage will have some spam (. . . ) action enabled at all times
thereafter, so a fair trace containing a setMessage can’t be finite (since spam
is enabled in the last state) and if infinite contains infinitely many spam mes-
sages (since spam messages of some sort are enabled in all but an initial finite
prefix). On the other hand, if spam(m1 ) and spam(m2 ) are not equivalent in
APPENDIX E. I/O AUTOMATA 344

task(A), then the spambot doesn’t satisfy the liveness property: in an exe-
cution that alternates setMessage(m1 )setMessage(m2 )setMessage(m1 )setMessage(m2 ) . . .
there are infinitely many states in which spam(m1 ) is not enabled, so fairness
doesn’t require doing it even once, and similarly for spam(m2 ).

E.2.2.3 Other properties

Any other property P can be expressed as the intersection of a safety prop-
erty (the closure of P ) and a liveness property (the union of P and the set of
all finite sequences that aren’t prefixes of traces in P ). The intuition is that
the safety property prunes out the excess junk we threw into the liveness
property to make it a liveness property, since any sequence that isn’t a prefix
of a trace in P won’t go into the safety property. This leaves only the traces
in P .
Example: Let P = {0n 1∞ } be the set of traces where we eventually give
up on our pointless 0-action and start doing only 1-actions forever. Then P
is the intersection of the safety property S = {0n 1m } ∪ P (the extra junk is
from prefix-closure) and the liveness property L = {0n 11m 0x|xin{0, 1}∗ }∪P .
Property S says that once we do a 1 we never do a 0, but allows finite
executions of the form 0n where we never do a 1. Property L says that we
eventually do a 1-action, but that we can’t stop unless we later do at least
one 0-action.

E.2.3 Compositional arguments

The product of trace properties P1 , P2 . . . is the trace property P where
T is in P if and only if T |sig(Pi ) is in Pi for each i. If the {Ai } satisfy
corresponding propertties {Pi } individually, then their composition satisfies
the product property. (For safety properties, often we prove something
weaker about the Ai , which is that each Ai individually is not the first to
violate P —i.e., it can’t leave P by executing an internal or output action.
In an execution where inputs by themselves can’t violate P , P then holds.)
Product properties let us prove trace properties by smashing together
properties of the component automata, possibly with some restrictions on
the signatures to get rid of unwanted actions. The product operation itself
is in a sense a combination of a Cartesian product (pick traces ti and smash
them together) filtered by a consistency rule (the smashed trace must be
consistent); it acts much like intersection (and indeed can be made identical
to intersection if we treat a trace property with a given signature as a way
of describing the set of all T such that T |sig(Pi ) is in Pi ).
APPENDIX E. I/O AUTOMATA 345

E.2.3.1 Example
Consider two spambots A1 and A2 where we identify the spam(m) operation
of A1 with the setMessage(m) operation of A2 ; we’ll call this combined action
spam1 (m) to distinguish it from the output actions of A2 . We’d like to
argue that the composite automaton A1 + A2 satisfies the safety property
(call it Pm ) that any occurrence of spam(m) is preceded by an occurrence
of setMessage(m), where the signature of Pm includes setMessage(m) and
spam(m) for some specific m but no other operations. (This is an example
of where trace property signatures can be useful without being limited to
actions of any specific component automaton.)
To do so, we’ll prove a stronger property Pm 0 , which is P
m modified
to include the spam1 (m) action in its signature. Observe that Pm 0 is the

product of the safety properties for A1 and A2 restricted to sig(Pm 0 ), since

the later says that any trace that includes spam(m) has a previous spam1 (m)
and the former says that any trace that includes spam1 (m) has a previous
setMessage(m). Since these properties hold for the individual A1 and A2 ,
their product, and thus the restriction Pm0 , holds for A + A , and so P (as
1 2 m
a further restriction) holds for A1 + A2 as well.
Now let’s prove the liveness property for A1 + A2 , that at least one
occurrence of setMessage yields infinitely many spam actions. Here we
let L1 = {at least one setMessage action ⇒ infinitely many spam1 actions}
and L2 = {at least one spam1 action ⇒ infinitely many spam actions}. The
product of these properties is all sequences with (a) no setMessage actions
or (b) infinitely many spam actions, which is what we want. This product
holds if the individual properties L1 and L2 hold for A1 + A2 , which will be
the case if we set task(A1 ) and task(A2 ) correctly.

E.2.4 Simulation arguments

Show that traces(A) is a subset of traces(B) (possibly after hiding some
actions of A) by showing a simulation relation f : states(A) → states(B)
between states of A and states of B. Requirements on f are

1. If s is in start(A), then f (s) includes some element of start(B).

2. If (s, a, s0 ) is in trans(A) and s is reachable, then for any reachable u
in f (s), there is a sequence of actions x that takes u to some v in f (s0 )
with trace(x) = trace(a).

Using these we construct an execution of B matching (in trace) an execu-

tion of A by starting in f (s0 ) and applying the second part of the definition
APPENDIX E. I/O AUTOMATA 346

to each action in the A execution (including the hidden ones!)

E.2.4.1 Example
A single spambot A can simulate the conjoined spambots A1 +A2 . Proof: Let
f (s) = (s, s). Then f (⊥) = (⊥, ⊥) is a start state of A1 + A2 . Now consider
a transition (s, a, s0 ) of A; the action a is either (a) setMessage(m), giving
s0 = m; here we let x = setMessage(m)spam1 (m) with trace(x) = trace(a)
since spam1 (m) is internal and f (s0 ) = (m, m) the result of applying x; or (b)
a = spam(m), which does not change s or f (s); the matching x is spam(m),
which also does not change f (s) and has the same trace.
A different proof could take advantage of f being a relation by defining
f (s) = {(s, s0 )|s0 ∈ states(A2 )}. Now we don’t care about the state of
A2 , and treat a setMessage(m) action of A as the sequence setMessage(m)
in A1 + A2 (which updates the first component of the state correctly) and
treat a spam(m) action as spam1 (m)spam(m) (which updates the second
component—which we don’t care about—and has the correct trace.) In
some cases an approach of this sort is necessary because we don’t know
which simulated state we are heading for until we get an action from A.
Note that the converse doesn’t work: A1 + A2 don’t simulate A, since
there are traces of A1 +A2 (e.g. setMessage(0)spam1 (0)setMessage(1)spam(0))
that don’t restrict to traces of A. See [Lyn96, §8.5.5] for a more complicated
example of how one FIFO queue can simulate two FIFO queues and vice
versa (a situation called bisimulation).
Since we are looking at traces rather than fair traces, this kind of simula-
tion doesn’t help much with liveness properties, but sometimes the connec-
tion between states plus a liveness proof for B can be used to get a liveness
proof for A (essentially we have to argue that A can’t do infinitely many
action without triggering a B-action in an appropriate task class). Again
see [Lyn96, §8.5.5].
Bibliography

[AA11] Dan Alistarh and James Aspnes. Sub-logarithmic test-and-set

against a weak adversary. In Distributed Computing: 25th In-
ternational Symposium, DISC 2011, volume 6950 of Lecture
Notes in Computer Science, pages 97–109. Springer-Verlag,
September 2011.

[AAC09] James Aspnes, Hagit Attiya, and Keren Censor. Max registers,
counters, and monotone circuits. In Proceedings of the 28th An-
nual ACM Symposium on Principles of Distributed Computing,
PODC 2009, Calgary, Alberta, Canada, August 10-12, 2009,
pages 36–45, August 2009.

[AACH+ 11] Dan Alistarh, James Aspnes, Keren Censor-Hillel, Seth

Gilbert, and Morteza Zadimoghaddam. Optimal-time adaptive
tight renaming, with applications to counting. In Proceedings
of the Thirtieth Annual ACM SIGACT-SIGOPS Symposium
on Principles of Distributed Computing, pages 239–248, June
2011.

[AACHE12] James Aspnes, Hagit Attiya, Keren Censor-Hillel, and Faith

Ellen. Faster than optimal snapshots (for a while). In
2012 ACM Symposium on Principles of Distributed Comput-
ing, pages 375–384, July 2012.

[AACHH12] James Aspnes, Hagit Attiya, Keren Censor-Hillel, and Danny

Hendler. Lower bounds for restricted-use objects. In Twenty-
Fourth ACM Symposium on Parallel Algorithms and Architec-
tures, pages 172–181, June 2012.

[AAD+ 93] Yehuda Afek, Hagit Attiya, Danny Dolev, Eli Gafni, Michael
Merritt, and Nir Shavit. Atomic snapshots of shared memory.
J. ACM, 40(4):873–890, 1993.

347
BIBLIOGRAPHY 348

[AAG+ 10] Dan Alistarh, Hagit Attiya, Seth Gilbert, Andrei Giurgiu, and
Rachid Guerraoui. Fast randomized test-and-set and renam-
ing. In Nancy A. Lynch and Alexander A. Shvartsman, editors,
Distributed Computing, 24th International Symposium, DISC
2010, Cambridge, MA, USA, September 13-15, 2010. Proceed-
ings, volume 6343 of Lecture Notes in Computer Science, pages
94–108. Springer, 2010.

[AAGG11] Dan Alistarh, James Aspnes, Seth Gilbert, and Rachid Guer-
raoui. The complexity of renaming. In Fifty-Second Annual
IEEE Symposium on Foundations of Computer Science, pages
718–727, October 2011.

[AAGW13] Dan Alistarh, James Aspnes, George Giakkoupis, and Philipp

Woelfel. Randomized loose renmaing in o(log log n) time. In
2013 ACM Symposium on Principles of Distributed Computing,
pages 200–209, July 2013.

[ABND+ 90] Hagit Attiya, Amotz Bar-Noy, Danny Dolev, David Peleg, and
Rüdiger Reischuk. Renaming in an asynchronous environment.
J. ACM, 37(3):524–548, 1990.

[ABND95] Hagit Attiya, Amotz Bar-Noy, and Danny Dolev. Sharing mem-
ory robustly in message-passing systems. Journal of the ACM,
42(1):124–142, 1995.

[Abr88] Karl Abrahamson. On achieving consensus using a shared

memory. In Proceedings of the 7th Annual ACM Symposium
on Principles of Distributed Computing (PODC), pages 291–
302, 1988.

[AC08] Hagit Attiya and Keren Censor. Tight bounds for asynchronous
randomized consensus. Journal of the ACM, 55(5):20, October
2008.

[AC09] James Aspnes and Keren Censor. Approximate shared-memory

counting despite a strong adversary. In SODA ’09: Proceedings
of the Nineteenth Annual ACM -SIAM Symposium on Discrete
Algorithms, pages 441–450, Philadelphia, PA, USA, 2009. So-
ciety for Industrial and Applied Mathematics.
BIBLIOGRAPHY 349

[ACH10] Hagit Attiya and Keren Censor-Hillel. Lower bounds for ran-
domized consensus under a weak adversary. SIAM J. Comput.,
39(8):3885–3904, 2010.

[ACH13] James Aspnes and Keren Censor-Hillel. Atomic snapshots in

o(log 3 n) steps using randomized helping. In Yehuda Afek, ed-
itor, Distributed Computing: 27th International Symposium,
DISC 2013, Jerusalem, Israel, October 14–18, 2013. Proceed-
ings, volume 8205 of Lecture Notes in Computer Science, pages
254–268. Springer Berlin Heidelberg, 2013.

[ACHS13] Dan Alistarh, Keren Censor-Hillel, and Nir Shavit. Are lock-
free concurrent algorithms practically wait-free? arXiv preprint
arXiv:1311.3200, 2013.

[AE11] James Aspnes and Faith Ellen. Tight bounds for anonymous
adopt-commit objects. In 23rd Annual ACM Symposium on
Parallelism in Algorithms and Architectures, pages 317–324,
June 2011.

[AEH75] E. A. Akkoyunlu, K. Ekanadham, and R. V. Huber. Some con-

straints and tradeoffs in the design of network communications.
SIGOPS Oper. Syst. Rev., 9:67–74, November 1975.

[AF01] Hagit Attiya and Arie Fouren. Adaptive and efficient algo-
rithms for lattice agreement and renaming. SIAM Journal on
Computing, 31(2):642–664, 2001.

[AFL83] Eshrat Arjomandi, Michael J. Fischer, and Nancy A. Lynch.

Efficiency of synchronous versus asynchronous distributed sys-
tems. J. ACM, 30(3):449–456, 1983.

[AG91] Yehuda Afek and Eli Gafni. Time and message bounds for
election in synchronous and asynchronous complete networks.
SIAM Journal on Computing, 20(2):376–394, 1991.

[AGGT09] Dan Alistarh, Seth Gilbert, Rachid Guerraoui, and Corentin

Travers. Of choices, failures and asynchrony: The many faces
of set agreement. In Yingfei Dong, Ding-Zhu Du, and Oscar H.
Ibarra, editors, ISAAC, volume 5878 of Lecture Notes in Com-
puter Science, pages 943–953. Springer, 2009.
BIBLIOGRAPHY 350

[AGHK06] Hagit Attiya, Rachid Guerraoui, Danny Hendler, and Petr

Kouznetsov. Synchronizing without locks is inherently expen-
sive. In PODC ’06: Proceedings of the twenty-fifth annual ACM
symposium on Principles of distributed computing, pages 300–
307, New York, NY, USA, 2006. ACM.

[AGTV92] Yehuda Afek, Eli Gafni, John Tromp, and Paul M. B. Vitányi.
Wait-free test-and-set (extended abstract). In Adrian Segall
and Shmuel Zaks, editors, Distributed Algorithms, 6th Inter-
national Workshop, WDAG ’92, Haifa, Israel, November 2-4,
1992, Proceedings, volume 647 of Lecture Notes in Computer
Science, pages 85–94. Springer, 1992.

[AH90a] James Aspnes and Maurice Herlihy. Fast randomized consensus

using shared memory. Journal of Algorithms, 11(3):441–461,
September 1990.

[AH90b] James Aspnes and Maurice Herlihy. Wait-free data structures

in the asynchronous PRAM model. In Second Annual ACM
Symposium on Parallel Algorithms and Architectures, pages
340–349, July 1990.

[AHM09] Hagit Attiya, Eshcar Hillel, and Alessia Milani. Inherent lim-
itations on disjoint-access parallel implementations of trans-
actional memory. In Friedhelm Meyer auf der Heide and
Michael A. Bender, editors, SPAA 2009: Proceedings of the
21st Annual ACM Symposium on Parallelism in Algorithms
and Architectures, Calgary, Alberta, Canada, August 11-13,
2009, pages 69–78. ACM, 2009.

[AHR95] Hagit Attiya, Maurice Herlihy, and Ophir Rachman. Atomic

snapshots using lattice agreement. Distributed Computing,
8(3):121–132, 1995.

[AHS94] James Aspnes, Maurice Herlihy, and Nir Shavit. Counting net-
works. Journal of the ACM, 41(5):1020–1048, September 1994.

[AHW08] Hagit Attiya, Danny Hendler, and Philipp Woelfel. Tight rmr
lower bounds for mutual exclusion and other problems. In Pro-
ceedings of the 40th annual ACM symposium on Theory of com-
puting, STOC ’08, pages 217–226, New York, NY, USA, 2008.
ACM.
BIBLIOGRAPHY 351

[AKP+ 06] Hagit Attiya, Fabian Kuhn, C. Greg Plaxton, Mirjam Watten-
hofer, and Roger Wattenhofer. Efficient adaptive collect using
randomization. Distributed Computing, 18(3):179–188, 2006.

[AKS83] M. Ajtai, J. Komlós, and E. Szemerédi. An o(n log n) sorting

network. In Proceedings of the fifteenth annual ACM sympo-
sium on Theory of computing, pages 1–9, New York, NY, USA,
1983. ACM.

[AM93] James H. Anderson and Mark Moir. Towards a necessary and

sufficient condition for wait-free synchronization (extended ab-
stract). In André Schiper, editor, Distributed Algorithms, 7th
International Workshop, WDAG ’93, Lausanne, Switzerland,
September 27-29, 1993, Proceedings, volume 725 of Lecture
Notes in Computer Science, pages 39–53. Springer, 1993.

[AM94] Hagit Attiya and Marios Mavronicolas. Efficiency of semisyn-

chronous versus asynchronous networks. Mathematical Systems
Theory, 27(6):547–571, November 1994.

[AM99] Yehuda Afek and Michael Merritt. Fast, wait-free (2k − 1)-
renaming. In PODC, pages 105–112, 1999.

[And90] Thomas E. Anderson. The performance of spin lock alterna-

tives for shared-money multiprocessors. IEEE Trans. Parallel
Distrib. Syst., 1(1):6–16, 1990.

[And94] James H. Anderson. Multi-writer composite registers. Dis-

tributed Computing, 7(4):175–195, 1994.

[Ang80] Dana Angluin. Local and global properties in networks of

processors (extended abstract). In Proceedings of the twelfth
annual ACM symposium on Theory of computing, STOC ’80,
pages 82–93, New York, NY, USA, 1980. ACM.

[Asp10] James Aspnes. Slightly smaller splitter networks. Technical

Report YALEU/DCS/TR-1438, Yale University Department
of Computer Science, November 2010.

[Asp11] James Aspnes. Notes on randomized algorithms. http://www.

cs.yale.edu/homes/aspnes/classes/469/notes.pdf, July
2011.
BIBLIOGRAPHY 352

[Asp12a] James Aspnes. Faster randomized consensus with an oblivi-

ous adversary. In 2012 ACM Symposium on Principles of Dis-
tributed Computing, pages 1–8, July 2012.

[Asp12b] James Aspnes. A modular approach to shared-memory con-

sensus, with applications to the probabilistic-write model. Dis-
tributed Computing, 25(2):179–188, May 2012.

[ASW88] Hagit Attiya, Marc Snir, and Manfred K. Warmuth. Com-

puting on an anonymous ring. J. ACM, 35:845–875, October
1988.

[Att14] Hagit Attiya. Lower bounds and impossibility results for trans-
actional memory computing. Bulletin of the European Associ-
ation for Computer Science, 112:38–52, February 2014.

[Aum97] Yonatan Aumann. Efficient asynchronous consensus with the

weak adversary scheduler. In PODC ’97: Proceedings of the
Sixteenth Annual ACM Symposium on Principles of Distributed
Computing, pages 209–218, New York, NY, USA, 1997. ACM.

[AW99] Yehuda Afek and Eytan Weisberger. The instancy of snapshots

and commuting objects. J. Algorithms, 30(1):68–105, 1999.

[AW04] Hagit Attiya and Jennifer Welch. Distributed Computing: Fun-

damentals, Simulations, and Advanced Topics. Wiley, second
edition, 2004. On-line version: http://dx.doi.org/10.1002/
0471478210. (This may not work outside Yale.).

[Awe85] Baruch Awerbuch. Complexity of network synchronization. J.

ACM, 32:804–823, October 1985.

[AWW93] Yehuda Afek, Eytan Weisberger, and Hanan Weisman. A com-

pleteness theorem for a class of synchronization objects (ex-
tended abstract). In Proceedings of the Twelfth Annual ACM
Symposium on Principles of Distributed Computing, pages 159–
170, 1993.

[Bat68] K. E. Batcher. Sorting networks and their applications. In

Proceedings of the AFIPS Spring Joint Computer Conference
32, pages 307–314, 1968.
BIBLIOGRAPHY 353

[BDLP08] Christian Boulinier, Ajoy K. Datta, Lawrence L. Larmore, and

Franck Petit. Space efficient and time optimal distributed BFS
tree construction. Information Processing Letters, 108(5):273–
278, November 2008. http://dx.doi.org/10.1016/j.ipl.
2008.05.016.

[Bel03] S. Bellovin. The Security Flag in the IPv4 Header. RFC 3514
(Informational), April 2003.

[BEW11] Alex Brodsky, Faith Ellen, and Philipp Woelfel. Fully-adaptive

algorithms for long-lived renaming. Distributed Computing,
24(2):119–134, 2011.

[BG93] Elizabeth Borowsky and Eli Gafni. Generalized flp impossibil-

ity result for t-resilient asynchronous computations. In STOC,
pages 91–100, 1993.

[BG97] Elizabeth Borowsky and Eli Gafni. A simple algorithmically

reasoned characterization of wait-free computations (extended
abstract). In PODC, pages 189–198, 1997.

[BG11] Michael A. Bender and Seth Gilbert. Mutual exclusion

with o(log2 log n) amortized work. Unpublished manuscript,
available at http://www.cs.sunysb.edu/~bender/newpub/
2011-focs-BenderGi-mutex.pdf as of 2011-12-02, 2011.

[BGA94] Elizabeth Borowsky, Eli Gafni, and Yehuda Afek. Consensus

power makes (some) sense! (extended abstract). In PODC,
pages 363–372, 1994.

[BGLR01] E. Borowsky, E. Gafni, N. Lynch, and S. Rajsbaum. The bg

distributed simulation algorithm. Distrib. Comput., 14(3):127–
146, October 2001.

[BGP89] Piotr Berman, Juan A. Garay, and Kenneth J. Perry. To-

wards optimal distributed consensus (extended abstract). In
30th Annual Symposium on Foundations of Computer Science,
30 October-1 November 1989, Research Triangle Park, North
Carolina, USA, pages 410–415, 1989.

[BL93] James E. Burns and Nancy A. Lynch. Bounds on shared mem-

ory for mutual exclusion. Inf. Comput., 107(2):171–184, 1993.
BIBLIOGRAPHY 354

[BND89] A. Bas-Noy and D. Dolev. Shared-memory vs. message-passing

in an asynchronous distributed environment. In Proceedings of
the eighth annual ACM Symposium on Principles of distributed
computing, PODC ’89, pages 307–318, New York, NY, USA,
1989. ACM.

[Bor95] Elizabeth Borowsky. Capturing the Power of Resiliency and

Set Consensus in Distributed Systems. PhD thesis, University
of California, Los Angeles, 1995.

[BPSV06] Harry Buhrman, Alessandro Panconesi, Riccardo Silvestri, and

Paul Vitányi. On the importance of having an identity or,
is consensus really universal? Distrib. Comput., 18:167–176,
February 2006.

[BR91] Gabriel Bracha and Ophir Rachman. Randomized consensus in

expected O(n2 log n) operations. In Sam Toueg, Paul G. Spi-
rakis, and Lefteris M. Kirousis, editors, Distributed Algorithms,
5th International Workshop, volume 579 of Lecture Notes in
Computer Science, pages 143–150, Delphi, Greece, 7–9 Octo-
ber 1991. Springer, 1992.

[Bur80] James E. Burns. A formal model for message passing systems.

Technical Report 91, Computer Science Department, Indiana
University, September 1980. http://www.cs.indiana.edu/
pub/techreports/TR91.pdf.

[Cha93] Soma Chaudhuri. More choices allow more faults: Set consen-
sus problems in totally asynchronous systems. Inf. Comput.,
105(1):132–158, 1993.

[Cha96] Tushar Deepak Chandra. Polylog randomized wait-free con-

sensus. In Proceedings of the Fifteenth Annual ACM Sympo-
sium on Principles of Distributed Computing, pages 166–175,
Philadelphia, Pennsylvania, USA, 23–26 May 1996.

[CHT96] Tushar Deepak Chandra, Vassos Hadzilacos, and Sam Toueg.

The weakest failure detector for solving consensus. J. ACM,
43:685–722, July 1996.

[CIL94] Benny Chor, Amos Israeli, and Ming Li. Wait-free consensus
using asynchronous hardware. SIAM J. Comput., 23(4):701–
712, 1994.
BIBLIOGRAPHY 355

[CL85] K. Mani Chandy and Leslie Lamport. Distributed snapshots:

Determining global states of distributed systems. ACM Trans.
Comput. Syst., 3(1):63–75, 1985.
[CR79] Ernest Chang and Rosemary Roberts. An improved algorithm
for decentralized extrema-finding in circular configurations of
processes. Commun. ACM, 22:281–283, May 1979.
[CR08] Armando Castañeda and Sergio Rajsbaum. New combina-
torial topology upper and lower bounds for renaming. In
Rida A. Bazzi and Boaz Patt-Shamir, editors, Proceedings of
the Twenty-Seventh Annual ACM Symposium on Principles of
Distributed Computing, PODC 2008, Toronto, Canada, August
18-21, 2008, pages 295–304. ACM, 2008.
[CT96] Tushar Deepak Chandra and Sam Toueg. Unreliable failure
detectors for reliable distributed systems. J. ACM, 43:225–267,
March 1996.
[DH04] Robert Danek and Vassos Hadzilacos. Local-spin group mutual
exclusion algorithms. In Rachid Guerraoui, editor, Distributed
Computing, 18th International Conference, DISC 2004, Ams-
terdam, The Netherlands, October 4-7, 2004, Proceedings, vol-
ume 3274 of Lecture Notes in Computer Science, pages 71–85.
Springer, 2004.
[DHW97] Cynthia Dwork, Maurice Herlihy, and Orli Waarts. Contention
in shared memory algorithms. J. ACM, 44(6):779–805, 1997.
[DLP+ 86] Danny Dolev, Nancy A. Lynch, Shlomit S. Pinter, Eugene W.
Stark, and William E. Weihl. Reaching approximate agreement
in the presence of faults. J. ACM, 33(3):499–516, 1986.
[DLS88] Cynthia Dwork, Nancy A. Lynch, and Larry J. Stockmeyer.
Consensus in the presence of partial synchrony. J. ACM,
35(2):288–323, 1988.
[DS83] Danny Dolev and H. Raymond Strong. Authenticated algo-
rithms for byzantine agreement. SIAM J. Comput., 12(4):656–
666, 1983.
[EHS12] Faith Ellen, Danny Hendler, and Nir Shavit. On the inherent
sequentiality of concurrent objects. SIAM Journal on Comput-
ing, 41(3):519–536, 2012.
BIBLIOGRAPHY 356

[FH07] Keir Fraser and Timothy L. Harris. Concurrent programming

without locks. ACM Trans. Comput. Syst., 25(2), 2007.

[FHS98] Faith Ellen Fich, Maurice Herlihy, and Nir Shavit. On the space
complexity of randomized synchronization. J. ACM, 45(5):843–
862, 1998.

[FHS05] Faith Ellen Fich, Danny Hendler, and Nir Shavit. Linear lower
bounds on real-world implementations of concurrent objects.
In Foundations of Computer Science, Annual IEEE Sympo-
sium on, pages 165–173, Los Alamitos, CA, USA, 2005. IEEE
Computer Society.

[Fic05] Faith Fich. How hard is it to take a snapshot? In Peter Vojtáš,

Mária Bieliková, Bernadette Charron-Bost, and Ondrej Sýkora,
editors, SOFSEM 2005: Theory and Practice of Computer Sci-
ence, volume 3381 of Lecture Notes in Computer Science, pages
28–37. Springer Berlin / Heidelberg, 2005.

[Fid91] Colin J. Fidge. Logical time in distributed computing systems.

IEEE Computer, 24(8):28–33, 1991.

[FK07] Panagiota Fatourou and Nikolaos D. Kallimanis. Time-

optimal, space-efficient single-scanner snapshots & multi-
scanner snapshots using CAS. In Indranil Gupta and Roger
Wattenhofer, editors, Proceedings of the Twenty-Sixth An-
nual ACM Symposium on Principles of Distributed Comput-
ing, PODC 2007, Portland, Oregon, USA, August 12-15, 2007,
pages 33–42. ACM, 2007.

[FL82] Michael J. Fischer and Nancy A. Lynch. A lower bound for

the time to assure interactive consistency. Inf. Process. Lett.,
14(4):183–186, 1982.

[FL87] Greg N. Frederickson and Nancy A. Lynch. Electing a leader

in a synchronous ring. J. ACM, 34(1):98–115, 1987.

[FL06] Rui Fan and Nancy A. Lynch. An ω(n log n) lower bound on the
cost of mutual exclusion. In Eric Ruppert and Dahlia Malkhi,
editors, Proceedings of the Twenty-Fifth Annual ACM Sym-
posium on Principles of Distributed Computing, PODC 2006,
Denver, CO, USA, July 23-26, 2006, pages 275–284. ACM,
2006.
BIBLIOGRAPHY 357

[FLM86] Michael J. Fischer, Nancy A. Lynch, and Michael Merritt. Easy

impossibility proofs for distributed consensus problems. Dis-
tributed Computing, 1(1):26–39, 1986.

[FLMS05] Faith Ellen Fich, Victor Luchangco, Mark Moir, and Nir
Shavit. Obstruction-free algorithms can be practically wait-
free. In Pierre Fraigniaud, editor, Distributed Computing,
19th International Conference, DISC 2005, Cracow, Poland,
September 26-29, 2005, Proceedings, volume 3724 of Lecture
Notes in Computer Science, pages 78–92. Springer, 2005.

[FLP85] Michael J. Fischer, Nancy A. Lynch, and Michael S. Paterson.

Impossibility of distributed consensus with one faulty process.
Journal of the ACM, 32(2):374–382, April 1985.

[Gaf98] Eli Gafni. Round-by-round fault detectors: Unifying synchrony

and asynchrony (extended abstract). In Proceedings of the Sev-
enteenth Annual ACM Symposium on Principles of Distributed
Computing, pages 143–152, 1998.

[Gaf09] Eli Gafni. The extended BG-simulation and the characteri-

zation of t-resiliency. In Proceedings of the 41st annual ACM
symposium on Theory of computing, pages 85–92. ACM, 2009.

[Gal82] Robert G. Gallager. Distributed minimum hop algorithms.

Technical Report LIDS-P-1175, M.I.T. Laboratory for Infor-
mation and Decision Systems, January 1982.

[GHHW13] George Giakkoupis, Maryam Helmi, Lisa Higham, and Philipp

√
Woelfel. An o( n) space bound for obstruction-free leader
election. In Proceedings of the 27th International Symposium
on Distributed Computing (DISC), pages 46–60, October 14–18
2013.

[GM98] Juan A. Garay and Yoram Moses. Fully polynomial byzantine

agreement for n > 3t processors in t + 1 rounds. SIAM J.
Comput., 27(1):247–290, 1998.

[Gol11] Wojciech M. Golab. A complexity separation between the

cache-coherent and distributed shared memory models. In Cyril
Gavoille and Pierre Fraigniaud, editors, Proceedings of the 30th
BIBLIOGRAPHY 358

Annual ACM Symposium on Principles of Distributed Comput-

ing, PODC 2011, San Jose, CA, USA, June 6-8, 2011, pages
109–118. ACM, 2011.

[Gra78] Jim Gray. Notes on data base operating systems. In Operating

Systems, An Advanced Course, pages 393–481. Springer-Verlag,
London, UK, 1978.

[GRS90] Ronald L. Graham, Bruce L. Rothschild, and Joel H. Spencer.

Ramsey Theory. Wiley-Interscience, 2nd edition, 1990.

[GW12a] George Giakkoupis and Philipp Woelfel. On the time and space
complexity of randomized test-and-set. In Darek Kowalski and
Alessandro Panconesi, editors, ACM Symposium on Principles
of Distributed Computing, PODC ’12, Funchal, Madeira, Por-
tugal, July 16-18, 2012, pages 19–28. ACM, 2012.

[GW12b] George Giakkoupis and Philipp Woelfel. A tight rmr lower

bound for randomized mutual exclusion. In Proceedings of
the 44th symposium on Theory of Computing, pages 983–1002.
ACM, 2012.

[Her91a] Maurice Herlihy. Impossibility results for asynchronous PRAM

(extended abstract). In Proceedings of the third annual ACM
symposium on Parallel algorithms and architectures, SPAA ’91,
pages 327–336, New York, NY, USA, 1991. ACM.

[Her91b] Maurice Herlihy. Wait-free synchronization. ACM Transac-

tions on Programming Languages and Systems, 13(1):124–149,
January 1991.

[Her93] Maurice Herlihy. A methodology for implementing highly con-

current objects. ACM Trans. Program. Lang. Syst., 15(5):745–
770, 1993.

[HFP02] Timothy L. Harris, Keir Fraser, and Ian A. Pratt. A practi-

cal multi-word compare-and-swap operation. In Dahlia Malkhi,
editor, Distributed Computing, 16th International Conference,
DISC 2002, Toulouse, France, October 28-30, 2002 Proceed-
ings, volume 2508 of Lecture Notes in Computer Science, pages
265–279. Springer, 2002.
BIBLIOGRAPHY 359

[HLM03] Maurice Herlihy, Victor Luchangco, and Mark Moir.

Obstruction-free synchronization: Double-ended queues as an
example. In 23rd International Conference on Distributed
Computing Systems (ICDCS 2003), 19-22 May 2003, Provi-
dence, RI, USA, pages 522–529. IEEE Computer Society, 2003.

[HM93] Maurice Herlihy and J. Eliot B. Moss. Transactional memory:

Architectural support for lock-free data structures. In ISCA,
pages 289–300, 1993.

[HS80] Daniel S. Hirschberg and J. B. Sinclair. Decentralized extrema-

finding in circular configurations of processors. Commun.
ACM, 23(11):627–628, 1980.

[HS99] Maurice Herlihy and Nir Shavit. The topological structure of

asynchronous computability. J. ACM, 46(6):858–923, 1999.

[HW90] Maurice Herlihy and Jeannette M. Wing. Linearizability: A

correctness condition for concurrent objects. ACM Trans. Pro-
gram. Lang. Syst., 12(3):463–492, 1990.

[HW11] Danny Hendler and Philipp Woelfel. Randomized mutual ex-

clusion with sub-logarithmic rmr-complexity. Distributed Com-
puting, 24(1):3–19, 2011.

[IMCT94] Michiko Inoue, Toshimitsu Masuzawa, Wei Chen, and Nobuki

Tokura. Linear-time snapshot using multi-writer multi-reader
registers. In Gerard Tel and Paul Vitányi, editors, Distributed
Algorithms, volume 857 of Lecture Notes in Computer Science,
pages 130–140. Springer Berlin / Heidelberg, 1994.

[IR09] Damien Imbs and Michel Raynal. Visiting Gafni’s reduction

land: From the BG simulation to the extended BG simulation.
In Stabilization, Safety, and Security of Distributed Systems,
pages 369–383. Springer, 2009.

[Jay97] Prasad Jayanti. Robust wait-free hierarchies. J. ACM,

44(4):592–614, 1997.

[Jay02] Prasad Jayanti. f -arrays: implementation and applications. In

Proceedings of the twenty-first annual symposium on Principles
of distributed computing, PODC ’02, pages 270–279, New York,
NY, USA, 2002. ACM.
BIBLIOGRAPHY 360

[Jay11] Prasad Jayanti. personal communication, 19 October 2011.

[JTT00] Prasad Jayanti, King Tan, and Sam Toueg. Time and space
lower bounds for nonblocking implementations. SIAM J. Com-
put., 30(2):438–456, 2000.

[Kaw00] Jawal Y. Kawash. Limitations and Capabilities of Weak Mem-

ory Consistency Systems. PhD thesis, University of Calgary,
January 2000.

[LAA87] Michael C. Loui and Hosame H. Abu-Amara. Memory require-

ments for agreement among unreliable asynchronous processes.
In Franco P. Preparata, editor, Parallel and Distributed Com-
puting, volume 4 of Advances in Computing Research, pages
163–183. JAI Press, 1987.

[Lam77] Leslie Lamport. Concurrent reading and writing. Communica-

tions of the ACM, 20(11):806–811, November 1977.

[Lam78] Leslie Lamport. Time, clocks, and the ordering of events in a

distributed system. Commun. ACM, 21(7):558–565, 1978.

[Lam79] L. Lamport. How to make a multiprocessor computer that

correctly executes multiprocess programs. Computers, IEEE
Transactions on, C-28(9):690–691, Sept 1979.

[Lam83] Leslie Lamport. The weak byzantine generals problem. J.

ACM, 30(3):668–676, 1983.

[Lam87] Leslie Lamport. A fast mutual exclusion algorithm. ACM

Trans. Comput. Syst., 5(1):1–11, 1987.

[Lam98] Leslie Lamport. The part-time parliament. ACM Trans. Com-

put. Syst., 16(2):133–169, 1998.

[Lam01] Leslie Lamport. Paxos made simple. SIGACT News, 32(4):18–

25, 2001.

[LL77] Gérard Le Lann. Distributed systems—towards a formal ap-

proach. In B. Gilchrist, editor, Information Processing 77,
pages 155–160. North-Holland, 1977.

[Lyn96] Nancy A. Lynch. Distributed Algorithms. Morgan Kaufmann,

1996.
BIBLIOGRAPHY 361

[MA95] Mark Moir and James H. Anderson. Wait-free algorithms for

fast, long-lived renaming. Sci. Comput. Program., 25(1):1–39,
1995.

[Mat93] Friedemann Mattern. Efficient algorithms for distributed snap-

shots and global virtual time approximation. J. Parallel Dis-
trib. Comput., 18(4):423–434, 1993.

[MCS91] John M. Mellor-Crummey and Michael L. Scott. Algorithms

for scalable synchronization on shared-memory multiproces-
sors. ACM Trans. Comput. Syst., 9(1):21–65, 1991.

[Mor95] Shlomo Moran. Using approximate agreement to obtain com-

plete disagreement: the output structure of input-free asyn-
chronous computations. In Third Israel Symposium on the
Theory of Computing and Systems, pages 251–257, January
1995.

[MR98] Dahlia Malkhi and Michael K. Reiter. Byzantine quorum sys-

tems. Distributed Computing, 11(4):203–213, 1998.

[MR10] Michael Merideth and Michael Reiter. Selected results from

the latest decade of quorum systems research. In Bernadette
Charron-Bost, Fernando Pedone, and André Schiper, editors,
Replication, volume 5959 of Lecture Notes in Computer Sci-
ence, pages 185–206. Springer, 2010.

[MRRT08] Achour Mostefaoui, Sergio Rajsbaum, Michel Raynal, and

Corentin Travers. The combined power of conditions and infor-
mation on failures to solve asynchronous set agreement. SIAM
Journal on Computing, 38(4):1574–1601, 2008.

[MRWW01] Dahlia Malkhi, Michael K. Reiter, Avishai Wool, and Re-

becca N. Wright. Probabilistic quorum systems. Inf. Comput.,
170(2):184–206, 2001.

[NT87] Gil Neiger and Sam Toueg. Substituting for real time and
common knowledge in asynchronous distributed systems. In
Proceedings of the sixth annual ACM Symposium on Principles
of distributed computing, PODC ’87, pages 281–293, New York,
NY, USA, 1987. ACM.
BIBLIOGRAPHY 362

[NW98] Moni Naor and Avishai Wool. The load, capacity, and avail-
ability of quorum systems. SIAM J. Comput., 27(2):423–447,
1998.
[Oka99] Chris Okasaki. Purely Functional Data Structures. Cambridge
University Press, 1999.
[Pet81] Gary L. Peterson. Myths about the mutual exclusion problem.
Inf. Process. Lett., 12(3):115–116, 1981.
[Pet82] Gary L. Peterson. An O(n log n) unidirectional algorithm for
the circular extrema problem. ACM Trans. Program. Lang.
Syst., 4(4):758–762, 1982.
[PF77] Gary L. Peterson and Michael J. Fischer. Economical solu-
tions for the critical section problem in a distributed system
(extended abstract). In John E. Hopcroft, Emily P. Fried-
man, and Michael A. Harrison, editors, Proceedings of the 9th
Annual ACM Symposium on Theory of Computing, May 4-6,
1977, Boulder, Colorado, USA, pages 91–97. ACM, 1977.
[Plo89] S. A. Plotkin. Sticky bits and universality of consensus. In
Proceedings of the eighth annual ACM Symposium on Princi-
ples of distributed computing, PODC ’89, pages 159–175, New
York, NY, USA, 1989. ACM.
[PSL80] M. Pease, R. Shostak, and L. Lamport. Reaching agreements
in the presence of faults. Journal of the ACM, 27(2):228–234,
April 1980.
[PW95] David Peleg and Avishai Wool. The availability of quorum
systems. Inf. Comput., 123(2):210–223, 1995.
[PW97a] David Peleg and Avishai Wool. The availability of crumbling
wall quorum systems. Discrete Applied Mathematics, 74(1):69–
83, 1997.
[PW97b] David Peleg and Avishai Wool. Crumbling walls: A class of
practical and efficient quorum systems. Distributed Computing,
10(2):87–97, 1997.
[RST01] Yaron Riany, Nir Shavit, and Dan Touitou. Towards a practical
snapshot algorithm. Theor. Comput. Sci., 269(1-2):163–201,
2001.
BIBLIOGRAPHY 363

[Rup00] Eric Ruppert. Determining consensus numbers. SIAM J. Com-

put., 30(4):1156–1168, 2000.
[Sch95] Eric Schenk. Faster approximate agreement with multi-writer
registers. In 36th Annual Symposium on Foundations of Com-
puter Science, Milwaukee, Wisconsin, 23-25 October 1995,
pages 714–723. IEEE Computer Society, 1995.
[Spe28] E. Sperner. Neuer Beweis für die Invarianz der Dimensions-
zahl und des Gebietes. Abhandlungen aus dem Mathema-
tischen Seminar der Universität Hamburg, 6:265–272, 1928.
10.1007/BF02940617.
[SSW91] M. Saks, N. Shavit, and H. Woll. Optimal time randomized
consensus - making resilient algorithms fast in practice. In Proc.
of the 2nd ACM Symposium on Discrete Algorithms (SODA),
pages 351–362, 1991.
[ST97] Nir Shavit and Dan Touitou. Software transactional memory.
Distributed Computing, 10(2):99–116, 1997.
[SZ00] Michael E. Saks and Fotios Zaharoglou. Wait-free k-set agree-
ment is impossible: The topology of public knowledge. SIAM
J. Comput., 29(5):1449–1483, 2000.
[TV02] John Tromp and Paul M. B. Vitányi. Randomized two-process
wait-free test-and-set. Distributed Computing, 15(3):127–135,
2002.
[VL92] George Varghese and Nancy A. Lynch. A tradeoff between
safety and liveness for randomized coordinated attack proto-
cols. In Proceedings of the Eleventh Annual ACM Symposium
on Principles of Distributed Computing, PODC ’92, pages 241–
250, New York, NY, USA, 1992. ACM.
[Wel87] Jennifer L. Welch. Simulating synchronous processors. Inf.
Comput., 74(2):159–170, 1987.
[YA95] Jae-Heon Yang and James H. Anderson. A fast, scalable mutual
exclusion algorithm. Distributed Computing, 9(1):51–60, 1995.
[Yu06] Haifeng Yu. Signed quorum systems. Distributed Computing,
18(4):307–323, 2006.
Index

δ-high-quality quorum, 109 strong, 197

-agreement, 278 value-oblivious, 197
-intersecting quorum system, 108 weak, 197
b-disseminating quorum system, 107 agreement, 17, 62, 81
b-masking quorum system, 108 -, 278
k-connectivity, 275 k-set, 259, 260, 324
k-neighborhood, 45 approximate, 278, 314
k-set agreement, 259, 260, 324 Byzantine, 69
0-valent, 82 probabilistic, 200
1-valent, 82 randomized, 19
safe, 253
abstract simplicial complex, 262 simplex, 273
accepter, 86 synchronous, 62
accessible state, 9 alpha synchronizer, 32, 56
accuracy, 93 anonymous, 36
action, 338 anti-consensus, 333
input, 338 append, 329
internal, 338 append-and-fetch, 329
output, 338 approximate agreement, 278, 314
active round, 45 array
adaptive, 212, 220 max, 185
adaptive adversary, 197 asynchronous, 10
adaptive collect, 221 Asynchronous Computability Theo-
admissible, 3, 10 rem, 272
adopt-commit object, 199 asynchronous message-passing, 3, 10
adopt-commit protocol, 199 atomic, 119, 226
adversary atomic queue, 4
adaptive, 197 atomic register, 112
content-oblivious, 19, 197 atomic registers, 4
intermediate, 197 atomic snapshot object, 157
location-oblivious, 197 average-case complexity, 42
oblivious, 19, 197

364
INDEX 365

barrier client-server, 11
memory, 4 clock
beta synchronizer, 32 logical, 47
BFS, 29 Lamport, 49
BG simulation, 253 Neiger-Toueg-Welch, 50
extended, 257 coherence, 199
big-step, 116 collect, 124, 157
binary consensus, 68 adaptive, 221
biologically inspired systems, 4 coordinated, 171
birthday paradox, 220 colorless task, 257
bisimulation, 346 common node, 75
bit complexity, 117 common2, 190
bivalence, 82 communication pattern, 64
bivalent, 82 commuting object, 190
Borowsky-Gafni simulation, 253 commuting operations, 147
bounded, 17 comparability, 163
bounded bypass, 126 comparators, 222
bounded fetch-and-subtract, 326 compare-and-swap, 4, 117, 149
bounded wait-free, 278 comparison-based algorithm, 44
breadth-first search, 29 complement, 107
broadcast completeness, 93, 190
reliable, 99 complex
terminating reliable, 102 input, 263
busy-waiting, 116 output, 263
Byzantine agreement, 69 protocol, 272
weak, 72 simplicial, 260
Byzantine failure, 3, 69 complexity
bit, 117
cache-coherent, 139 message, 15
capacity, 105 obstruction-free step, 236
CAS, 149 space, 117
causal ordering, 47 step, 14
causal shuffle, 48 individual, 14, 116
Chandra-Toueg consensus protocol, 98 per-process, 116
channel total, 14, 116
FIFO, 12 time, 14, 116
chemical reaction networks, 4 composite register, 158
chromatic subdivision, 272 computation event, 9
class G, 245 conciliator, 200
client, 11 concurrency detector, 300
INDEX 366

concurrent execution, 114 decision bit, 199

configuration, 2, 8 delivery event, 8
initial, 9 depth
connected, 263 sorting network, 223
simply, 275 deque, 237
consensus, 16, 196 detector
binary, 68 failure, 92
Chandra-Toeug, 98 deterministic, 9, 36
randomized, 196 deterministic renaming, 213
synchronous, 62 diameter, 30
universality of, 154 direct scan, 158, 162
consensus number, 144 disk, 275
consensus object, 153 distributed breadth-first search, 25
consistency property, 115 distributed shared memory, 119, 140
consistent cut, 52 distributed shared-memory, 4
consistent scan, 158 distributed systems, 1
consistent snapshot, 52 downward validity, 163
content-oblivious adversary, 19, 197 dual graph, 107
contention, 117, 244 dynamic transaction, 226
contention management, 244
contention manager, 234 Elias gamma code, 183
continuous function, 271 enabled, 3, 338
convergecast, 26 equivalence relation, 36
Convergence., 199 event, 2, 8
coordinated attack, 16 computation, 9
randomized, 18 delivery, 8
coordinated collect, 171 receive, 47
coordinator, 98 send, 47
copy eventually perfect failure detector, 93,
memory-to-memory, 150 95
counter, 299 eventually strong failure detector, 95,
counting network, 223 324
course staff, xix execution, 2, 9, 339
cover, 137 concurrent, 114
crash failure, 3, 63 fair, 340
crashes fully, 64 execution segment, 9
critical, 125 exiting, 125
critical section, 125 exponential information gathering, 66
extended BG simulation, 257
deadlock, 126
INDEX 367

failure wait-free, 144

Byzantine, 3, 69 high quality quorum, 109
crash, 3, 63 historyless, 308
omission, 3 historyless object, 177, 190
failure detector, 3, 91, 92 homeomorphism, 261
eventually perfect, 93, 95 homotopy, 275
eventually strong, 95, 324
perfect, 95 I/O automaton, 338
strong, 95 identity, 37
failure proability IIS, 265
quorum system, 105 immediacy, 265
fair, 340 impossibility, 5, 6
fair execution, 340 indirect scan, 158, 162
fairness, 5, 10 indistinguishability, 6, 63
fast path, 135 indistinguishability proof, 17
fault-tolerance indistinguishability proofs, 10
quorum system, 105 indistinguishable, 17
faulty, 63 individual step complexity, 14, 116
fence, 4 individual work, 116
memory, 4 initial configuration, 9
fetch-and-add, 117, 148 initiator, 29
fetch-and-cons, 118, 149, 156 input action, 338
fetch-and-increment, 320 input complex, 263
fetch-and-subtract instructor, xix
bounded, 326 interfering operations, 147
FIFO channel, 12 intermediate adversary, 197
flooding, 22, 29 internal action, 338
Frankenexecution, 70 interval, 114
full-information algorithm, 67 invariant, 5, 6
full-information protocol, 45 invariants, 343
function invocation, 114, 119
continuous, 271 iterated immediate snapshot, 265

global synchronizer, 55 join, 163

Gnutella, 23
König’s lemma, 17
handshake, 161
Lamport clock, 49
happens-before, 47, 56
lattice, 163
hierarchy
lattice agreement, 163
robust, 144
leader election, 15
INDEX 368

learner, 86 memory-to-memory copy, 150

left null, 237 memory-to-memory swap, 149
limit-closed, 343 message complexity, 15
line message-passing, 3
world, 12 asynchronous, 3, 10
linearizability, 115, 122 semi-synchronous, 3
linearizable, 115, 119 synchronous, 3, 12
linearization, 122 multi-writer multi-reader register, 113
linearization point, 115, 160 multi-writer register, 113
liveness, 5, 6 mutual exclusion, 117, 125
liveness property, 342, 343 mutual exclusion protocol, 125
LL/SC, 149, 228
load, 105 Neiger-Toueg-Welch clock, 50
load balancing, 220 network, 10
load-linked, 149, 228 counting, 223
load-linked/store-conditional, 4, 171, overlay, 10
228 renaming, 222
load-linked/stored-conditional, 149 sorting, 222
local coin, 196 node
local synchronizer, 55 common, 75
location-oblivious adversary, 197 non-blocking, 226, 233
lock-free, 233 non-triviality, 63
lockable register, 322 nonce, 121
lockout, 126 nondeterminism, 1
lockout-freedom, 126 nondeterministic solo termination, 209
logical clock, 47, 49 null
Lamport, 49 left, 237
Neiger-Toueg-Welch, 50 right, 237
long-lived renaming, 215, 220 null path, 275
long-lived strong renaming, 220 null-homotopic, 275
lower bound, 6 number
sequence, 13
map
simplicial, 271 object, 112
max array, 185 commuting, 190
max register, 180 historyless, 177, 190
meet, 163 resilient, 175
memory barrier, 4 ring buffer, 330
memory fence, 4 snapshot, 157, 175
memory stall, 245 swap, 177
INDEX 369

oblivious adversary, 19, 197 progress function, 343

obstruction-free, 233 progress measure, 6
obstruction-free step complexity, 236 proof
omission failure, 3 impossibility, 6
one-time building blocks, 134 invariant, 6
operation, 85 liveness, 6
operations lower bound, 6
commuting, 147 safety, 6
interfering, 147 termination, 6
overwriting, 147 property
oracle, 237 stable, 53
order-equivalent, 44 proposer, 86
order-preserving renaming, 213 protocol complex, 272
ordering
causal, 47 queue
output action, 338 atomic, 4
output complex, 263 wait-free, 148
overlay network, 10 with peek, 149
overwriting operations, 147 quiesce, 24
quiescent, 44, 339
participating set, 273 quorum
path δ-high-quality, 109
null, 275 high quality, 109
path-connected, 271 quorum size, 105
Paxos, 85 quorum system, 104
per-process step complexity, 116 -intersecting, 108
per-process work, 116 b-disseminating, 107
perfect failure detector, 95 b-masking, 108
persona, 205 probabilistic, 108
perturbable, 177, 178 signed, 110
phase king, 76 strict, 108
preference, 281
prefix code, 182 racing counters, 235
prefix-closed, 343 Ramsey theory, 45
probabilistic agreement, 200 Ramsey’s Theorem, 45
probabilistic quorum system, 108 randomization, 3, 37
probabilistic termination, 196 randomized agreement, 19
process, 8 randomized consensus, 196
product, 344 randomized coordinated attack, 18
progress, 126 randomized splitter, 221
INDEX 370

RatRace, 221 robust hierarchy, 144

read, 119 round, 12, 88, 116
read-modify-write, 117, 126
receive event, 47 safe agreement, 253
register, 112, 119 safety, 4
atomic, 4, 112 safety property, 6, 341, 342
composite, 158 scan
lockable, 322 direct, 158, 162
max, 180 indirect, 158, 162
multi-writer, 113 schedule, 3, 9
single-reader, 113 admissible, 10
single-writer, 113 self-stabilization, 4
relation semi-lattice, 164
simulation, 345 semi-synchronous message-passing, 3
reliable broadcast, 99 semisynchrony
terminating, 102 unknown-bound, 239
remainder, 125 send event, 47
remote memory reference, 116, 139 sense of direction, 36
renaming, 134, 211 sequence number, 13
deterministic, 213 sequential consistency, 115
long-lived, 215 sequential execution, 119
long-lived strong, 220 server, 11
order-preserving, 213 session, 60
strong, 212, 220 session problem, 60
tight, 212 shared memory, 3
renaming network, 222 distributed, 119
replicated state machine, 54, 85 sifter, 203
representative, 206 signature, 338
request, 11 signed quorum system, 110
reset, 126 similar, 45, 47
ReShuffle, 222 simplex, 261
resilience, 175 simplex agreement, 273
resilient object, 175 simplicial complex, 260, 261
response, 11, 85, 114, 119 abstract, 262
restriction, 9 simplicial map, 271
right null, 237 simply connected, 275
ring, 36 simulation, 5, 342
ring buffer object, 330 simulation relation, 345
RMR, 116, 139 single-reader single-writer register, 113
RMW, 117 single-use swap object, 192
INDEX 371

single-writer multi-reader register, 113 suspect, 92

single-writer register, 113 swap, 148
slow path, 135 memory-to-memory, 149
snapshot, 157 swap object, 177
snapshot object, 175 single-use, 192
software transactional memory, 226 symmetry, 36
solo termination, 177 symmetry breaking, 36
solo-terminating, 177, 233 synchronizer, 12, 29, 49, 55
sorting network, 222 alpha, 32, 56
space complexity, 117 beta, 32, 56
special action, 60 gamma, 57
Sperner’s Lemma, 260 global, 55
sphere, 275 local, 55
splitter, 133, 134 synchronizers, 5
randomized, 221 synchronous agreement, 62
splitters, 216 synchronous message-passing, 3, 12
spread, 279
stable property, 53 task
staff, xix colorless, 257
stall, 245 terminating reliable broadcast, 102
starvation, 5 termination, 6, 17, 62, 81, 196
state, 2, 8, 85 solo, 177
accessible, 9 test-and-set, 117, 126, 148, 192
static transaction, 226 tight renaming, 212
step complexity, 14 time complexity, 14, 116
individual, 116 time-to-live, 23
obstruction-free, 236 timestamp, 49
per-process, 116 torus, 312
total, 116 total step complexity, 14, 116
sticky bit, 118, 149 total work, 116
two-writer, 302 trace, 339
sticky register, 118 trace property, 342
STM, 226 transaction, 226
store-conditional, 149, 228 dynamic, 226
strict quorum system, 108 static, 226
strong adversary, 197 transactional memory
strong failure detector, 95 software, 226
strong renaming, 212, 220 transition function, 9
subdivision, 265 transition relation, 338
chromatic, 272 transitive closure, 48
INDEX 372

triangulation, 266
trying, 125
Two Generals, 5, 16
two-writer sticky bit, 302

unidirectional ring, 37
uniform, 43
univalent, 82
universality of consensus, 154
unknown-bound semisynchrony, 239
unsafe, 254
upward validity, 163

validity, 17, 62, 81

downward, 163
upward, 163
value-oblivious adversary, 197
values, 199
vector clock, 49, 51

wait-free, 114, 144, 233

bounded, 278
wait-free hierarchy, 144
wait-free queue, 148
wait-freedom, 164, 233
weak adversary, 197
weak Byzantine agreement, 72
weird condition, 248
width, 117
sorting network, 223
wire, 222
work
individual, 116
per-process, 116
total, 116
world line, 12
write, 119